[BUGFIX] clear request (#4320)

Co-authored-by: ltd0924 <luotingdan@baidu.com>

.github/workflows/Codestyle-Check.yml

@@ -2,7 +2,9 @@ name: Codestyle-Check
 
 on:
   pull_request:
-    branches: ["develop"]
+    branches:
+      - develop
+      - 'release/*'
 
 jobs:
   pre-commit:
@@ -11,7 +13,7 @@ jobs:
     runs-on: ubuntu-latest
     env:
       PR_ID: ${{ github.event.pull_request.number }}
-      BRANCH: develop
+      BRANCH: ${{ github.event.pull_request.base.ref }}
 
     steps:
       - name: Cleanup

.github/workflows/_accuracy_test.yml

@@ -0,0 +1,186 @@
+name: Accuracy Test
+description: "Run Accuracy Tests"
+
+on:
+  workflow_call:
+    inputs:
+      DOCKER_IMAGE:
+        description: "Build Images"
+        required: true
+        type: string
+        default: "ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleqa:cuda126-py310"
+      FASTDEPLOY_ARCHIVE_URL:
+        description: "URL of the compressed FastDeploy code archive."
+        required: true
+        type: string
+      FASTDEPLOY_WHEEL_URL:
+        description: "URL of the FastDeploy Wheel."
+        required: true
+        type: string
+      CACHE_DIR:
+        description: "Cache Dir Use"
+        required: false
+        type: string
+        default: ""
+      MODEL_CACHE_DIR:
+        description: "Cache Dir Use"
+        required: false
+        type: string
+        default: ""
+
+jobs:
+  accuracy_tests:
+    runs-on: [self-hosted, GPU-h20-1Cards]
+    timeout-minutes: 60
+    steps:
+      - name: Code Prepare
+        shell: bash
+        env:
+          docker_image: ${{ inputs.DOCKER_IMAGE }}
+          fd_archive_url: ${{ inputs.FASTDEPLOY_ARCHIVE_URL }}
+        run: |
+            set -x
+            REPO="https://github.com/${{ github.repository }}.git"
+            FULL_REPO="${{ github.repository }}"
+            REPO_NAME="${FULL_REPO##*/}"
+            BASE_BRANCH="${{ github.base_ref }}"
+
+            # Clean the repository directory before starting
+            docker run --rm --net=host -v $(pwd):/workspace -w /workspace \
+            -e "REPO_NAME=${REPO_NAME}" \
+            ${docker_image} /bin/bash -c '
+              if [ -d ${REPO_NAME} ]; then
+                echo "Directory ${REPO_NAME} exists, removing it..."
+                rm -rf ${REPO_NAME}*
+              fi
+            '
+
+            wget -q ${fd_archive_url}
+            tar -xf FastDeploy.tar.gz
+            rm -rf FastDeploy.tar.gz
+            cd FastDeploy
+            git config --global user.name "FastDeployCI"
+            git config --global user.email "fastdeploy_ci@example.com"
+            git log -n 3 --oneline
+
+      - name: Run FastDeploy Base Tests
+        shell: bash
+        env:
+          docker_image: ${{ inputs.DOCKER_IMAGE }}
+          fastdeploy_wheel_url: ${{ inputs.FASTDEPLOY_WHEEL_URL }}
+          CACHE_DIR: ${{ inputs.CACHE_DIR }}
+          MODEL_CACHE_DIR: ${{ inputs.MODEL_CACHE_DIR }}
+        run: |
+          runner_name="${{ runner.name }}"
+          CARD_ID=$(echo "${runner_name}" | awk -F'-' '{print $NF}')
+          DEVICES=$(echo "$CARD_ID" | fold -w1 | paste -sd,)
+          DEVICE_PORT=$(echo "$DEVICES" | cut -d',' -f1)
+
+          FLASK_PORT=$((42068 + DEVICE_PORT * 100))
+          FD_API_PORT=$((42088 + DEVICE_PORT * 100))
+          FD_ENGINE_QUEUE_PORT=$((42058 + DEVICE_PORT * 100))
+          FD_METRICS_PORT=$((42078 + DEVICE_PORT * 100))
+          FD_CACHE_QUEUE_PORT=$((42098 + DEVICE_PORT * 100))
+          echo "Test ENV Parameter:"
+          echo "========================================================="
+          echo "FLASK_PORT=${FLASK_PORT}"
+          echo "FD_API_PORT=${FD_API_PORT}"
+          echo "FD_ENGINE_QUEUE_PORT=${FD_ENGINE_QUEUE_PORT}"
+          echo "FD_METRICS_PORT=${FD_METRICS_PORT}"
+          echo "FD_CACHE_QUEUE_PORT=${FD_CACHE_QUEUE_PORT}"
+          echo "DEVICES=${DEVICES}"
+          echo "========================================================="
+
+          CACHE_DIR="${CACHE_DIR:-$(dirname "$(dirname "${{ github.workspace }}")")}"
+          echo "CACHE_DIR is set to ${CACHE_DIR}"
+          if [ ! -f "${CACHE_DIR}/gitconfig" ]; then
+            touch "${CACHE_DIR}/gitconfig"
+          fi
+          if [ ! -d "${MODEL_CACHE_DIR}" ]; then
+            echo "Error: MODEL_CACHE_DIR '${MODEL_CACHE_DIR}' does not exist."
+            exit 1
+          fi
+
+          PORTS=($FLASK_PORT $FD_API_PORT $FD_ENGINE_QUEUE_PORT $FD_METRICS_PORT $FD_CACHE_QUEUE_PORT)
+          LOG_FILE="./port_cleanup_$(date +%Y%m%d_%H%M%S).log"
+          echo "==== LOG_FILE is ${LOG_FILE} ===="
+
+          echo "==== PORT CLEAN BEFORE TASK RUN ====" | tee -a $LOG_FILE
+
+          for port in "${PORTS[@]}"; do
+              PIDS=$(lsof -t -i :$port || true)
+              if [ -n "$PIDS" ]; then
+                  echo "Port $port is occupied by PID(s): $PIDS" | tee -a $LOG_FILE
+                  echo "$PIDS" | xargs -r kill -9
+                  echo "Port $port cleared" | tee -a $LOG_FILE
+              else
+                  echo "Port $port is free" | tee -a $LOG_FILE
+              fi
+          done
+
+          echo "==== PORT CLEAN COMPLETE ====" | tee -a $LOG_FILE
+
+          echo "========================================================="
+          echo "Ensuring no stale container named ${runner_name} ..."
+          if [ "$(docker ps -a -q -f name=${runner_name})" ]; then
+            echo "Removing stale container: ${runner_name}"
+            docker rm -f ${runner_name} || true
+          fi
+
+          docker run --rm --ipc=host --pid=host --net=host \
+          --name ${runner_name} \
+          -v $(pwd):/workspace \
+          -w /workspace \
+          -e fastdeploy_wheel_url=${fastdeploy_wheel_url} \
+          -e "FD_API_PORT=${FD_API_PORT}" \
+          -e "FD_ENGINE_QUEUE_PORT=${FD_ENGINE_QUEUE_PORT}" \
+          -e "FD_METRICS_PORT=${FD_METRICS_PORT}" \
+          -e "FD_CACHE_QUEUE_PORT=${FD_CACHE_QUEUE_PORT}" \
+          -e "FLASK_PORT=${FLASK_PORT}" \
+          -v "${MODEL_CACHE_DIR}:/MODELDATA" \
+          -v "${CACHE_DIR}/gitconfig:/etc/gitconfig:ro" \
+          -v "${CACHE_DIR}/.cache:/root/.cache" \
+          -v "${CACHE_DIR}/ConfigDir:/root/.config" \
+          -e TZ="Asia/Shanghai" \
+          --gpus '"device='"${DEVICES}"'"' ${docker_image} /bin/bash -xc '
+          python -m pip install paddlepaddle-gpu==3.2.0 -i https://www.paddlepaddle.org.cn/packages/stable/cu126/
+
+          pip config set global.index-url https://mirrors.tuna.tsinghua.edu.cn/pypi/web/simple
+
+          python -m pip install ${fastdeploy_wheel_url}
+          python -m pip install pytest
+
+          wget https://paddle-qa.bj.bcebos.com/zhengtianyu/tools/llm-deploy-linux-amd64
+          chmod +x ./llm-deploy-linux-amd64
+          ./llm-deploy-linux-amd64 -python python3.10 \
+          -model_name ERNIE-4.5-0.3B-Paddle \
+          -model_path /MODELDATA \
+          --skip install
+
+          git config --global --add safe.directory /workspace/FastDeploy
+          cd FastDeploy
+          pushd tests/ce/deploy
+          python3.10 deploy.py > dd.log 2>&1 &
+          sleep 3
+          curl -X POST http://0.0.0.0:${FLASK_PORT}/start \
+            -H "Content-Type: application/json" \
+            -d "{\"--model\": \"/MODELDATA/ERNIE-4.5-0.3B-Paddle\"}"
+
+          curl -X POST http://localhost:${FLASK_PORT}/wait_for_infer?timeout=90
+          popd
+
+          pushd tests/ce/accuracy_cases
+          export URL=http://localhost:${FD_API_PORT}/v1/chat/completions
+          export TEMPLATE=TOKEN_LOGPROB
+          export MODEL_SIZE=0.3B
+          TEST_EXIT_CODE=0
+          python gsm8k.py || TEST_EXIT_CODE=1
+          popd
+          echo "TEST_EXIT_CODE=${TEST_EXIT_CODE}" >> /workspace/FastDeploy/exit_code.env
+          '
+          if [ -f ./FastDeploy/exit_code.env ]; then
+            source ./FastDeploy/exit_code.env
+            cat ./FastDeploy/exit_code.env >> $GITHUB_ENV
+          fi
+          echo "TEST_EXIT_CODE=${TEST_EXIT_CODE}"
+          exit ${TEST_EXIT_CODE}

.github/workflows/_base_test.yml

@@ -0,0 +1,229 @@
+name: Base Test
+description: "Run Base Tests"
+
+on:
+  workflow_call:
+    inputs:
+      DOCKER_IMAGE:
+        description: "Build Images"
+        required: true
+        type: string
+        default: "ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleqa:cuda126-py310"
+      FASTDEPLOY_ARCHIVE_URL:
+        description: "URL of the compressed FastDeploy code archive."
+        required: true
+        type: string
+      FASTDEPLOY_WHEEL_URL:
+        description: "URL of the FastDeploy Wheel."
+        required: true
+        type: string
+      CACHE_DIR:
+        description: "Cache Dir Use"
+        required: false
+        type: string
+        default: ""
+      MODEL_CACHE_DIR:
+        description: "Cache Dir Use"
+        required: false
+        type: string
+        default: ""
+
+jobs:
+  base_tests:
+    runs-on: [self-hosted, GPU-h20-1Cards]
+    timeout-minutes: 60
+    steps:
+      - name: Code Prepare
+        shell: bash
+        env:
+          docker_image: ${{ inputs.DOCKER_IMAGE }}
+          fd_archive_url: ${{ inputs.FASTDEPLOY_ARCHIVE_URL }}
+        run: |
+            set -x
+            REPO="https://github.com/${{ github.repository }}.git"
+            FULL_REPO="${{ github.repository }}"
+            REPO_NAME="${FULL_REPO##*/}"
+            BASE_BRANCH="${{ github.base_ref }}"
+
+            # Clean the repository directory before starting
+            docker run --rm --net=host -v $(pwd):/workspace -w /workspace \
+            -e "REPO_NAME=${REPO_NAME}" \
+            ${docker_image} /bin/bash -c '
+              if [ -d ${REPO_NAME} ]; then
+                echo "Directory ${REPO_NAME} exists, removing it..."
+                rm -rf ${REPO_NAME}*
+              fi
+            '
+
+            wget -q ${fd_archive_url}
+            tar -xf FastDeploy.tar.gz
+            rm -rf FastDeploy.tar.gz
+            cd FastDeploy
+            git config --global user.name "FastDeployCI"
+            git config --global user.email "fastdeploy_ci@example.com"
+            git log -n 3 --oneline
+
+      - name: Run FastDeploy Base Tests
+        shell: bash
+        env:
+          docker_image: ${{ inputs.DOCKER_IMAGE }}
+          fastdeploy_wheel_url: ${{ inputs.FASTDEPLOY_WHEEL_URL }}
+          CACHE_DIR: ${{ inputs.CACHE_DIR }}
+          MODEL_CACHE_DIR: ${{ inputs.MODEL_CACHE_DIR }}
+        run: |
+          runner_name="${{ runner.name }}"
+          CARD_ID=$(echo "${runner_name}" | awk -F'-' '{print $NF}')
+          DEVICES=$(echo "$CARD_ID" | fold -w1 | paste -sd,)
+          DEVICE_PORT=$(echo "$DEVICES" | cut -d',' -f1)
+
+          FLASK_PORT=$((42068 + DEVICE_PORT * 100))
+          FD_API_PORT=$((42088 + DEVICE_PORT * 100))
+          FD_ENGINE_QUEUE_PORT=$((42058 + DEVICE_PORT * 100))
+          FD_METRICS_PORT=$((42078 + DEVICE_PORT * 100))
+          FD_CACHE_QUEUE_PORT=$((42098 + DEVICE_PORT * 100))
+          echo "Test ENV Parameter:"
+          echo "========================================================="
+          echo "FLASK_PORT=${FLASK_PORT}"
+          echo "FD_API_PORT=${FD_API_PORT}"
+          echo "FD_ENGINE_QUEUE_PORT=${FD_ENGINE_QUEUE_PORT}"
+          echo "FD_METRICS_PORT=${FD_METRICS_PORT}"
+          echo "FD_CACHE_QUEUE_PORT=${FD_CACHE_QUEUE_PORT}"
+          echo "DEVICES=${DEVICES}"
+          echo "========================================================="
+
+          CACHE_DIR="${CACHE_DIR:-$(dirname "$(dirname "${{ github.workspace }}")")}"
+          echo "CACHE_DIR is set to ${CACHE_DIR}"
+          if [ ! -f "${CACHE_DIR}/gitconfig" ]; then
+            touch "${CACHE_DIR}/gitconfig"
+          fi
+          if [ ! -d "${MODEL_CACHE_DIR}" ]; then
+            echo "Error: MODEL_CACHE_DIR '${MODEL_CACHE_DIR}' does not exist."
+            exit 1
+          fi
+
+          PORTS=($FLASK_PORT $FD_API_PORT $FD_ENGINE_QUEUE_PORT $FD_METRICS_PORT $FD_CACHE_QUEUE_PORT)
+          LOG_FILE="./port_cleanup_$(date +%Y%m%d_%H%M%S).log"
+          echo "==== LOG_FILE is ${LOG_FILE} ===="
+
+          echo "==== PORT CLEAN BEFORE TASK RUN ====" | tee -a $LOG_FILE
+
+          for port in "${PORTS[@]}"; do
+              PIDS=$(lsof -t -i :$port || true)
+              if [ -n "$PIDS" ]; then
+                  echo "Port $port is occupied by PID(s): $PIDS" | tee -a $LOG_FILE
+                  echo "$PIDS" | xargs -r kill -9
+                  echo "Port $port cleared" | tee -a $LOG_FILE
+              else
+                  echo "Port $port is free" | tee -a $LOG_FILE
+              fi
+          done
+
+          echo "==== PORT CLEAN COMPLETE ====" | tee -a $LOG_FILE
+
+          echo "========================================================="
+          echo "Ensuring no stale container named ${runner_name} ..."
+          if [ "$(docker ps -a -q -f name=${runner_name})" ]; then
+            echo "Removing stale container: ${runner_name}"
+            docker rm -f ${runner_name} || true
+          fi
+
+          docker run --rm --ipc=host --pid=host --net=host \
+          --name ${runner_name} \
+          -v $(pwd):/workspace \
+          -w /workspace \
+          -e fastdeploy_wheel_url=${fastdeploy_wheel_url} \
+          -e "FD_API_PORT=${FD_API_PORT}" \
+          -e "FD_ENGINE_QUEUE_PORT=${FD_ENGINE_QUEUE_PORT}" \
+          -e "FD_METRICS_PORT=${FD_METRICS_PORT}" \
+          -e "FLASK_PORT=${FLASK_PORT}" \
+          -e "FD_CACHE_QUEUE_PORT=${FD_CACHE_QUEUE_PORT}" \
+          -v "${MODEL_CACHE_DIR}:/MODELDATA" \
+          -v "${CACHE_DIR}/gitconfig:/etc/gitconfig:ro" \
+          -v "${CACHE_DIR}/.cache:/root/.cache" \
+          -v "${CACHE_DIR}/ConfigDir:/root/.config" \
+          -e TZ="Asia/Shanghai" \
+          --gpus '"device='"${DEVICES}"'"' ${docker_image} /bin/bash -xc '
+          python -m pip install paddlepaddle-gpu==3.2.0 -i https://www.paddlepaddle.org.cn/packages/stable/cu126/
+
+          pip config set global.index-url https://mirrors.tuna.tsinghua.edu.cn/pypi/web/simple
+
+          python -m pip install ${fastdeploy_wheel_url}
+          python -m pip install pytest
+
+          wget https://paddle-qa.bj.bcebos.com/zhengtianyu/tools/llm-deploy-linux-amd64
+          chmod +x ./llm-deploy-linux-amd64
+          ./llm-deploy-linux-amd64 -python python3.10 \
+          -model_name ERNIE-4.5-0.3B-Paddle \
+          -model_path /MODELDATA \
+          --skip install
+
+          git config --global --add safe.directory /workspace/FastDeploy
+          cd FastDeploy
+          pushd tests/ce/deploy
+          python3.10 deploy.py > dd.log 2>&1 &
+          sleep 3
+          curl -X POST http://0.0.0.0:${FLASK_PORT}/start \
+            -H "Content-Type: application/json" \
+            -d "{\"--model\": \"/MODELDATA/ERNIE-4.5-0.3B-Paddle\"}"
+
+          check_service() {
+            local timeout=${1:-90}
+            local url="http://localhost:${FLASK_PORT}/wait_for_infer?timeout=${timeout}"
+            local resp
+
+            resp=$(curl -s -X POST "$url")
+
+            if echo "$resp" | grep -q "服务启动超时"; then
+              exit 8
+            fi
+          }
+
+          check_service 90
+          popd
+
+          pushd tests/ce/server
+          export URL=http://localhost:${FD_API_PORT}/v1/chat/completions
+          export TEMPLATE=TOKEN_LOGPROB
+          TEST_EXIT_CODE=0
+          python -m pytest -sv test_base_chat.py test_compare_top_logprobs.py test_logprobs.py test_params_boundary.py test_seed_usage.py test_stream.py test_evil_cases.py test_completions.py test_return_token_ids.py || TEST_EXIT_CODE=1
+          curl -X POST http://0.0.0.0:${FLASK_PORT}/switch \
+            -H "Content-Type: application/json" \
+            -d "{\"--model\": \"/MODELDATA/ERNIE-4.5-0.3B-Paddle\", \"--early-stop-config\": \"{\\\"enable_early_stop\\\":true, \\\"window_size\\\":6, \\\"threshold\\\":0.93}\"}"
+          check_service 90
+          python -m pytest -sv test_repetition_early_stop.py || TEST_EXIT_CODE=1
+
+          curl -X POST http://0.0.0.0:${FLASK_PORT}/switch \
+            -H "Content-Type: application/json" \
+            -d "{ \"--model\": \"/MODELDATA/ERNIE-4.5-0.3B-Paddle\", \"--max-concurrency\": 5, \"--max-waiting-time\": 1 }"
+          check_service 90
+          python -m pytest -sv test_max_concurrency.py || TEST_EXIT_CODE=1
+
+          curl -X POST http://0.0.0.0:${FLASK_PORT}/switch \
+            -H "Content-Type: application/json" \
+            -d "{ \"--model\": \"/MODELDATA/ERNIE-4.5-0.3B-Paddle\", \"--max-concurrency\": 5000, \"--max-waiting-time\": 1 }"
+          check_service 90
+          python -m pytest -sv test_max_waiting_time.py || TEST_EXIT_CODE=1
+
+          curl -X POST http://0.0.0.0:${FLASK_PORT}/switch \
+            -H "Content-Type: application/json" \
+            -d "{\"--model\": \"/MODELDATA/ernie-4_5-21b-a3b-bf16-paddle\", \"--config\": \"21b_mtp.yaml\", \"--enable-logprob\": \"False\"}"
+          check_service 180
+          export TEMPLATE=TOKEN_NORMAL
+          python -m pytest -sv test_seed_usage.py -k "not test_seed_stream" || TEST_EXIT_CODE=1
+
+          curl -X POST http://0.0.0.0:${FLASK_PORT}/switch \
+            -H "Content-Type: application/json" \
+            -d "{\"--model\": \"/MODELDATA/ernie-4_5-21b-a3b-bf16-paddle\", \"--config\": \"21b_sot.yaml\", \"--enable-logprob\": \"False\"}"
+          check_service 360
+          export TEMPLATE=TOKEN_NORMAL
+          python -m pytest -sv test_seed_usage.py -k "not test_seed_stream" || TEST_EXIT_CODE=1
+
+          popd
+          echo "TEST_EXIT_CODE=${TEST_EXIT_CODE}" >> /workspace/FastDeploy/exit_code.env
+          '
+          if [ -f ./FastDeploy/exit_code.env ]; then
+            source ./FastDeploy/exit_code.env
+            cat ./FastDeploy/exit_code.env >> $GITHUB_ENV
+          fi
+          echo "TEST_EXIT_CODE=${TEST_EXIT_CODE}"
+          exit ${TEST_EXIT_CODE}

.github/workflows/_build_linux.yml

@@ -22,12 +22,22 @@ on:
         description: "Enable nightly build mode (e.g. add date suffix to version)"
         required: false
         type: string
-        default: "ON"
+        default: "OFF"
       FD_VERSION:
         description: "FastDeploy Package Version"
         required: false
         type: string
         default: ""
+      PADDLEVERSION:
+        description: "Paddle Version Build Use"
+        required: false
+        type: string
+        default: ""
+      PADDLE_WHL_URL:
+        description: "Paddle Wheel Package URL"
+        required: false
+        type: string
+        default: ""
       UPLOAD:
         description: "Upload Package"
         required: false
@@ -44,7 +54,8 @@ on:
         value: ${{ jobs.fd-build.outputs.wheel_path }}
 jobs:
   fd-build:
-    runs-on: [self-hosted, GPU-h1z1-4Cards]
+    runs-on: [self-hosted, GPU-Build]
+    timeout-minutes: 240
     outputs:
       wheel_path: ${{ steps.set_output.outputs.wheel_path }}
     steps:
@@ -85,13 +96,17 @@ jobs:
           compile_arch: ${{ inputs.COMPILE_ARCH }}
           fd_version: ${{ inputs.FD_VERSION }}
           CACHE_DIR: ${{ inputs.CACHE_DIR }}
+          BRANCH_REF: ${{ github.ref_name }}
+          PADDLEVERSION: ${{ inputs.PADDLEVERSION }}
+          PADDLE_WHL_URL: ${{ inputs.PADDLE_WHL_URL }}
+          WITH_NIGHTLY_BUILD: ${{ inputs.WITH_NIGHTLY_BUILD }}
         run: |
             set -x
             runner_name="${{ runner.name }}"
-            CARD_ID=$(echo "${runner_name}" | cut -d'-' -f2)
+            CARD_ID=$(echo "${runner_name}" | awk -F'-' '{print $NF}')
             gpu_id=$(echo "$CARD_ID" | fold -w1 | paste -sd,)
 
-            CACHE_DIR=${CACHE_DIR:-${{ github.workspace }}}
+            CACHE_DIR="${CACHE_DIR:-$(dirname "$(dirname "${{ github.workspace }}")")}"
             echo "CACHE_DIR is set to ${CACHE_DIR}"
             if [ ! -f "${CACHE_DIR}/gitconfig" ]; then
               touch "${CACHE_DIR}/gitconfig"
@@ -103,11 +118,15 @@ jobs:
             -v $(pwd):/workspace -w /workspace \
             -v "${CACHE_DIR}/gitconfig:/etc/gitconfig:ro" \
             -v "${CACHE_DIR}/.cache:/root/.cache" \
+            -v "${CACHE_DIR}/.ccache:/root/.ccache" \
             -v "${CACHE_DIR}/ConfigDir:/root/.config" \
             -e TZ="Asia/Shanghai" \
             -e "COMPILE_ARCH=${compile_arch}" \
             -e "FD_VERSION=${fd_version}" \
             -e "WITH_NIGHTLY_BUILD=${WITH_NIGHTLY_BUILD}" \
+            -e "PADDLEVERSION=${PADDLEVERSION}" \
+            -e "PADDLE_WHL_URL=${PADDLE_WHL_URL}" \
+            -e "BRANCH_REF=${BRANCH_REF}" \
             --gpus "\"device=${gpu_id}\"" ${docker_image} /bin/bash -c '
             if [[ -n "${FD_VERSION}" ]]; then
               export FASTDEPLOY_VERSION=${FD_VERSION}
@@ -115,6 +134,7 @@ jobs:
             fi
 
             git config --global --add safe.directory /workspace/FastDeploy
+            chown -R $(whoami) /workspace/FastDeploy
             cd FastDeploy
             if [[ "${WITH_NIGHTLY_BUILD}" == "ON" ]];then
               GIT_COMMIT_TIME=$(git --no-pager show -s --format=%ci HEAD)
@@ -123,14 +143,20 @@ jobs:
               echo "Date Only: $DATE_ONLY"
               export FASTDEPLOY_VERSION="${FASTDEPLOY_VERSION}.dev${DATE_ONLY}"
             fi
-            pip config set global.index-url http://pip.baidu.com/root/baidu/+simple/
-            pip config set install.trusted-host  pip.baidu.com
-            pip config set global.extra-index-url https://mirrors.tuna.tsinghua.edu.cn/pypi/web/simple
+            # 针对不同分支和tag使用不同的PaddlePaddle安装包
+            if [[ "${PADDLE_WHL_URL}" != "" ]];then
+              python -m pip install ${PADDLE_WHL_URL}
+            elif [[ "${PADDLEVERSION}" != "" ]];then
+              python -m pip install paddlepaddle-gpu==${PADDLEVERSION} -i https://www.paddlepaddle.org.cn/packages/stable/cu126/
+            else
+              python -m pip install paddlepaddle-gpu==3.2.0 -i https://www.paddlepaddle.org.cn/packages/stable/cu126/
+            fi
+
+            pip config set global.index-url https://mirrors.tuna.tsinghua.edu.cn/pypi/web/simple
 
             python -m pip install --upgrade pip
             python -m pip install -r requirements.txt
             python -m pip install wheel
-            python -m pip install --pre paddlepaddle-gpu -i https://www.paddlepaddle.org.cn/packages/nightly/cu126/
             # 编译RDMA
             export ENABLE_FD_RDMA=1
             bash build.sh 1 python false [${COMPILE_ARCH}]

.github/workflows/_clone_linux.yml

@@ -68,7 +68,7 @@ jobs:
             branch_name=${{ github.ref_name }}
             target_path=paddle-github-action/BRANCH/FastDeploy/${branch_name}/${commit_id}
           fi
-          wget  -q --no-proxy --no-check-certificate https://paddle-qa.bj.bcebos.com/CodeSync/develop/PaddlePaddle/PaddleTest/tools/bos_tools.py
+          wget -O bos_tools.py -q --no-proxy --no-check-certificate https://paddle-qa.bj.bcebos.com/CodeSync/develop/PaddlePaddle/PaddleTest/tools/bos_tools.py
           push_file=$(realpath bos_tools.py)
           python -m pip install bce-python-sdk==0.9.29
           ls

.github/workflows/_logprob_test_linux.yml

@@ -0,0 +1,184 @@
+name: Run FastDeploy LogProb Tests
+description: "Run FastDeploy LogProb Tests"
+
+on:
+  workflow_call:
+    inputs:
+      DOCKER_IMAGE:
+        description: "Build Images"
+        required: true
+        type: string
+        default: "ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleqa:cuda126-py310"
+      PADDLETEST_ARCHIVE_URL:
+        description: "URL of the compressed FastDeploy code archive."
+        required: true
+        type: string
+        default: "https://xly-devops.bj.bcebos.com/PaddleTest/PaddleTest.tar.gz"
+      FASTDEPLOY_WHEEL_URL:
+        description: "URL of the FastDeploy Wheel."
+        required: true
+        type: string
+      CACHE_DIR:
+        description: "Cache Dir Use"
+        required: false
+        type: string
+        default: ""
+      MODEL_CACHE_DIR:
+        description: "Cache Dir Use"
+        required: false
+        type: string
+        default: ""
+
+jobs:
+  run_tests_logprob:
+    runs-on: [self-hosted, GPU-h20-1Cards]
+    steps:
+      - name: Code Prepare
+        shell: bash
+        env:
+          docker_image: ${{ inputs.DOCKER_IMAGE }}
+          paddletest_archive_url: ${{ inputs.PADDLETEST_ARCHIVE_URL }}
+        run: |
+            # Clean the repository directory before starting
+            docker run --rm --net=host -v $(pwd):/workspace -w /workspace \
+            -e "REPO_NAME=${REPO_NAME}" \
+            -e "BASE_BRANCH=${BASE_BRANCH}" \
+            ${docker_image} /bin/bash -c '
+            rm -rf /workspace/*
+            '
+            wget -q ${paddletest_archive_url}
+            tar -xf PaddleTest.tar.gz
+            rm -rf PaddleTest.tar.gz
+            cd PaddleTest
+            git config --global user.name "FastDeployCI"
+            git config --global user.email "fastdeploy_ci@example.com"
+            git log -n 3 --oneline
+      - name: logprob test
+        shell: bash
+        env:
+          docker_image: ${{ inputs.DOCKER_IMAGE }}
+          fastdeploy_wheel_url: ${{ inputs.FASTDEPLOY_WHEEL_URL }}
+          CACHE_DIR: ${{ inputs.CACHE_DIR }}
+          MODEL_CACHE_DIR: ${{ inputs.MODEL_CACHE_DIR }}
+        run: |
+          runner_name="${{ runner.name }}"
+          CARD_ID=$(echo "${runner_name}" | awk -F'-' '{print $NF}')
+          DEVICES=$(echo "$CARD_ID" | fold -w1 | paste -sd,)
+          DEVICE_PORT=$(echo "$DEVICES" | cut -d',' -f1)
+
+          FLASK_PORT=$((42068 + DEVICE_PORT * 100))
+          FD_API_PORT=$((42088 + DEVICE_PORT * 100))
+          FD_ENGINE_QUEUE_PORT=$((42058 + DEVICE_PORT * 100))
+          FD_METRICS_PORT=$((42078 + DEVICE_PORT * 100))
+          FD_CACHE_QUEUE_PORT=$((42098 + DEVICE_PORT * 100))
+          echo "Test ENV Parameter:"
+          echo "========================================================="
+          echo "FLASK_PORT=${FLASK_PORT}"
+          echo "FD_API_PORT=${FD_API_PORT}"
+          echo "FD_ENGINE_QUEUE_PORT=${FD_ENGINE_QUEUE_PORT}"
+          echo "FD_METRICS_PORT=${FD_METRICS_PORT}"
+          echo "FD_CACHE_QUEUE_PORT=${FD_CACHE_QUEUE_PORT}"
+          echo "DEVICES=${DEVICES}"
+          echo "========================================================="
+
+          CACHE_DIR="${CACHE_DIR:-$(dirname "$(dirname "${{ github.workspace }}")")}"
+          echo "CACHE_DIR is set to ${CACHE_DIR}"
+          if [ ! -f "${CACHE_DIR}/gitconfig" ]; then
+            touch "${CACHE_DIR}/gitconfig"
+          fi
+          if [ ! -d "${MODEL_CACHE_DIR}" ]; then
+            echo "Error: MODEL_CACHE_DIR '${MODEL_CACHE_DIR}' does not exist."
+            exit 1
+          fi
+
+          PORTS=($FLASK_PORT $FD_API_PORT $FD_ENGINE_QUEUE_PORT $FD_METRICS_PORT $FD_CACHE_QUEUE_PORT)
+          LOG_FILE="./port_cleanup_$(date +%Y%m%d_%H%M%S).log"
+          echo "==== LOG_FILE is ${LOG_FILE} ===="
+
+          echo "==== PORT CLEAN BEFORE TASK RUN ====" | tee -a $LOG_FILE
+
+          for port in "${PORTS[@]}"; do
+              PIDS=$(lsof -t -i :$port || true)
+              if [ -n "$PIDS" ]; then
+                  echo "Port $port is occupied by PID(s): $PIDS" | tee -a $LOG_FILE
+                  echo "$PIDS" | xargs -r kill -9
+                  echo "Port $port cleared" | tee -a $LOG_FILE
+              else
+                  echo "Port $port is free" | tee -a $LOG_FILE
+              fi
+          done
+
+          echo "==== PORT CLEAN COMPLETE ====" | tee -a $LOG_FILE
+
+          echo "========================================================="
+          echo "Ensuring no stale container named ${runner_name} ..."
+          if [ "$(docker ps -a -q -f name=${runner_name})" ]; then
+            echo "Removing stale container: ${runner_name}"
+            docker rm -f ${runner_name} || true
+          fi
+
+          docker run --rm --ipc=host --pid=host --net=host \
+          --name ${runner_name} \
+          -v $(pwd):/workspace \
+          -w /workspace \
+          -e fastdeploy_wheel_url=${fastdeploy_wheel_url} \
+          -e "FD_API_PORT=${FD_API_PORT}" \
+          -e "FD_ENGINE_QUEUE_PORT=${FD_ENGINE_QUEUE_PORT}" \
+          -e "FD_METRICS_PORT=${FD_METRICS_PORT}" \
+          -e "FD_CACHE_QUEUE_PORT=${FD_CACHE_QUEUE_PORT}" \
+          -e "FLASK_PORT=${FLASK_PORT}" \
+          -v "${MODEL_CACHE_DIR}:/MODELDATA" \
+          -v "${CACHE_DIR}/gitconfig:/etc/gitconfig:ro" \
+          -v "${CACHE_DIR}/.cache:/root/.cache" \
+          -v "${CACHE_DIR}/ConfigDir:/root/.config" \
+          -e TZ="Asia/Shanghai" \
+          --gpus '"device='"${DEVICES}"'"' ${docker_image} /bin/bash -xc '
+          python -m pip install paddlepaddle-gpu==3.2.0 -i https://www.paddlepaddle.org.cn/packages/stable/cu126/
+
+          pip config set global.index-url https://mirrors.tuna.tsinghua.edu.cn/pypi/web/simple
+
+          python -m pip install ${fastdeploy_wheel_url}
+
+          wget https://paddle-qa.bj.bcebos.com/zhengtianyu/tools/llm-deploy-linux-amd64
+          chmod +x ./llm-deploy-linux-amd64
+          ./llm-deploy-linux-amd64 -python python3.10 \
+          -model_name ERNIE-4.5-0.3B-Paddle \
+          -model_path /MODELDATA \
+          --skip install
+
+          cd PaddleTest/framework/ServeTest
+          python3.10 deploy.py > dd.log 2>&1 &
+          sleep 3
+          curl -X POST http://0.0.0.0:${FLASK_PORT}/start \
+              -H "Content-Type: application/json" \
+              -d "{\"--model\": \"/MODELDATA/ERNIE-4.5-0.3B-Paddle\"}"
+
+          curl -X POST http://localhost:${FLASK_PORT}/wait_for_infer?timeout=90
+          curl -s -o /dev/null -w "%{http_code}" -m 2 "http://0.0.0.0:${FD_API_PORT}/health"
+          curl -X POST "http://0.0.0.0:${FD_API_PORT}/v1/chat/completions" \
+            -H "Content-Type: application/json" \
+            -d "{\"messages\": [{\"role\": \"user\", \"content\": \"1+1=?\"}], \"logprobs\": true}"
+          set +e
+          rm -rf ./baseline_output
+          cp -r baseline/ERNIE-4.5-0.3B-Paddle ./baseline_output
+          LOGPROB_EXIT_CODE=0
+          python3.10 lanucher.py --request_template TOKEN_LOGPROB --url http://localhost:${FD_API_PORT}/v1/chat/completions  --case ./cases/demo.yaml  --concurrency 1 --name demo --exe logprob || LOGPROB_EXIT_CODE=$?
+          echo "LOGPROB_EXIT_CODE=${LOGPROB_EXIT_CODE}" > /workspace/exit_code.env
+          curl -X POST http://localhost:${FLASK_PORT}/stop
+          sleep 10s
+          cat *result.log
+          exit 0
+          '
+          if [ $? -ne 0 ];then
+            exit 1
+          fi
+
+          if [ -f exit_code.env ]; then
+            cat exit_code.env >> $GITHUB_ENV
+          fi
+      - name: logprob test result
+        if: ${{ env.LOGPROB_EXIT_CODE != 0 }}
+        shell: bash
+        run: |
+          echo "logprob test failed with exit code ${{ env.LOGPROB_EXIT_CODE }}"
+          exit 8

.github/workflows/_pre_ce_test.yml

@@ -0,0 +1,148 @@
+name: Pre-CE-Test
+
+on:
+  workflow_call:
+    inputs:
+      DOCKER_IMAGE:
+        description: "Build Images"
+        required: true
+        type: string
+        default: "ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddle:fastdeploy-ciuse-cuda126"
+      FASTDEPLOY_ARCHIVE_URL:
+        description: "URL of the compressed FastDeploy code archive."
+        required: true
+        type: string
+      FASTDEPLOY_WHEEL_URL:
+        description: "URL of the FastDeploy Wheel."
+        required: true
+        type: string
+      CACHE_DIR:
+        description: "Cache Dir Use"
+        required: false
+        type: string
+        default: ""
+      MODEL_CACHE_DIR:
+        description: "Cache Dir Use"
+        required: false
+        type: string
+        default: ""
+
+jobs:
+  run_ce_cases:
+    runs-on: [self-hosted, PRE_CE_RUN_2Card]
+    timeout-minutes: 60
+    steps:
+      - name: Print current runner name
+        run: |
+          echo "Current runner name: ${{ runner.name }}"
+      - name: Code Prepare
+        shell: bash
+        env:
+          docker_image: ${{ inputs.DOCKER_IMAGE }}
+          fd_archive_url: ${{ inputs.FASTDEPLOY_ARCHIVE_URL }}
+        run: |
+            set -x
+            REPO="https://github.com/${{ github.repository }}.git"
+            FULL_REPO="${{ github.repository }}"
+            REPO_NAME="${FULL_REPO##*/}"
+            BASE_BRANCH="${{ github.base_ref }}"
+
+            # Clean the repository directory before starting
+            docker run --rm --net=host -v $(pwd):/workspace -w /workspace \
+            -e "REPO_NAME=${REPO_NAME}" \
+            ${docker_image} /bin/bash -c '
+              if [ -d ${REPO_NAME} ]; then
+                echo "Directory ${REPO_NAME} exists, removing it..."
+                rm -rf ${REPO_NAME}*
+              fi
+            '
+
+            wget -q ${fd_archive_url}
+            tar -xf FastDeploy.tar.gz
+            rm -rf FastDeploy.tar.gz
+            cd FastDeploy
+            git config --global user.name "FastDeployCI"
+            git config --global user.email "fastdeploy_ci@example.com"
+            git log -n 3 --oneline
+
+      - name: Run CI unittest
+        env:
+          docker_image: ${{ inputs.DOCKER_IMAGE }}
+          fd_wheel_url: ${{ inputs.FASTDEPLOY_WHEEL_URL }}
+          CACHE_DIR: ${{ inputs.CACHE_DIR }}
+          MODEL_CACHE_DIR: ${{ inputs.MODEL_CACHE_DIR }}
+        run: |
+          runner_name="${{ runner.name }}"
+          CARD_ID=$(echo "${runner_name}" | awk -F'-' '{print $NF}')
+          DEVICES=$(echo "$CARD_ID" | fold -w1 | paste -sd,)
+          DEVICE_PORT=$(echo "$DEVICES" | cut -d',' -f1)
+
+          FLASK_PORT=$((42068 + DEVICE_PORT * 100))
+          FD_API_PORT=$((42088 + DEVICE_PORT * 100))
+          FD_ENGINE_QUEUE_PORT=$((42058 + DEVICE_PORT * 100))
+          FD_METRICS_PORT=$((42078 + DEVICE_PORT * 100))
+          FD_CACHE_QUEUE_PORT=$((42098 + DEVICE_PORT * 100))
+          echo "Test ENV Parameter:"
+          echo "========================================================="
+          echo "FLASK_PORT=${FLASK_PORT}"
+          echo "FD_API_PORT=${FD_API_PORT}"
+          echo "FD_ENGINE_QUEUE_PORT=${FD_ENGINE_QUEUE_PORT}"
+          echo "FD_METRICS_PORT=${FD_METRICS_PORT}"
+          echo "FD_CACHE_QUEUE_PORT=${FD_CACHE_QUEUE_PORT}"
+          echo "DEVICES=${DEVICES}"
+          echo "========================================================="
+
+          CACHE_DIR="${CACHE_DIR:-$(dirname "$(dirname "${{ github.workspace }}")")}"
+          echo "CACHE_DIR is set to ${CACHE_DIR}"
+          if [ ! -f "${CACHE_DIR}/gitconfig" ]; then
+            touch "${CACHE_DIR}/gitconfig"
+          fi
+
+          PORTS=($FLASK_PORT $FD_API_PORT $FD_ENGINE_QUEUE_PORT $FD_METRICS_PORT $FD_CACHE_QUEUE_PORT)
+          LOG_FILE="./port_cleanup_$(date +%Y%m%d_%H%M%S).log"
+          echo "==== LOG_FILE is ${LOG_FILE} ===="
+
+          echo "==== PORT CLEAN BEFORE TASK RUN ====" | tee -a $LOG_FILE
+
+          for port in "${PORTS[@]}"; do
+              PIDS=$(lsof -t -i :$port || true)
+              if [ -n "$PIDS" ]; then
+                  echo "Port $port is occupied by PID(s): $PIDS" | tee -a $LOG_FILE
+                  echo "$PIDS" | xargs -r kill -9
+                  echo "Port $port cleared" | tee -a $LOG_FILE
+              else
+                  echo "Port $port is free" | tee -a $LOG_FILE
+              fi
+          done
+
+          echo "==== PORT CLEAN COMPLETE ====" | tee -a $LOG_FILE
+
+          echo "========================================================="
+          echo "Ensuring no stale container named ${runner_name} ..."
+          if [ "$(docker ps -a -q -f name=${runner_name})" ]; then
+            echo "Removing stale container: ${runner_name}"
+            docker rm -f ${runner_name} || true
+          fi
+
+          docker run --rm --net=host \
+          --name ${runner_name} \
+          -v $(pwd):/workspace \
+          -w /workspace \
+          -v "${CACHE_DIR}/gitconfig:/etc/gitconfig:ro" \
+          -v "${CACHE_DIR}/.cache:/root/.cache" \
+          -v "${CACHE_DIR}/ConfigDir:/root/.config" \
+          -v "${MODEL_CACHE_DIR}:/ModelData:ro" \
+          -e "MODEL_PATH=/ModelData" \
+          -e "FD_API_PORT=${FD_API_PORT}" \
+          -e "FD_ENGINE_QUEUE_PORT=${FD_ENGINE_QUEUE_PORT}" \
+          -e "FD_METRICS_PORT=${FD_METRICS_PORT}" \
+          -e "FD_CACHE_QUEUE_PORT=${FD_CACHE_QUEUE_PORT}" \
+          -e "FLASK_PORT=${FLASK_PORT}" \
+          -e "fd_wheel_url=${fd_wheel_url}" \
+          --gpus "\"device=${DEVICES}\"" ${docker_image} /bin/bash -c '
+          git config --global --add safe.directory /workspace/FastDeploy
+          cd FastDeploy
+          python -m pip install paddlepaddle-gpu==3.2.0 -i https://www.paddlepaddle.org.cn/packages/stable/cu126/
+          python -m pip install ${fd_wheel_url}
+          bash scripts/run_pre_ce.sh
+          '

.github/workflows/_stable_test.yml

@@ -0,0 +1,170 @@
+name: Stable Test
+description: "Run Stable Tests"
+
+on:
+  workflow_call:
+    inputs:
+      DOCKER_IMAGE:
+        description: "Build Images"
+        required: true
+        type: string
+        default: "ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleqa:cuda126-py310"
+      FASTDEPLOY_ARCHIVE_URL:
+        description: "URL of the compressed FastDeploy code archive."
+        required: true
+        type: string
+      FASTDEPLOY_WHEEL_URL:
+        description: "URL of the FastDeploy Wheel."
+        required: true
+        type: string
+      CACHE_DIR:
+        description: "Cache Dir Use"
+        required: false
+        type: string
+        default: ""
+      MODEL_CACHE_DIR:
+        description: "Cache Dir Use"
+        required: false
+        type: string
+        default: ""
+
+jobs:
+  stable_tests:
+    runs-on: [self-hosted, GPU-h1z1-2Cards]
+    timeout-minutes: 60
+    steps:
+      - name: Code Prepare
+        shell: bash
+        env:
+          docker_image: ${{ inputs.DOCKER_IMAGE }}
+          fd_archive_url: ${{ inputs.FASTDEPLOY_ARCHIVE_URL }}
+        run: |
+            set -x
+            REPO="https://github.com/${{ github.repository }}.git"
+            FULL_REPO="${{ github.repository }}"
+            REPO_NAME="${FULL_REPO##*/}"
+            BASE_BRANCH="${{ github.base_ref }}"
+
+            # Clean the repository directory before starting
+            docker run --rm --net=host -v $(pwd):/workspace -w /workspace \
+            -e "REPO_NAME=${REPO_NAME}" \
+            ${docker_image} /bin/bash -c '
+              if [ -d ${REPO_NAME} ]; then
+                echo "Directory ${REPO_NAME} exists, removing it..."
+                rm -rf ${REPO_NAME}*
+              fi
+            '
+
+            wget -q ${fd_archive_url}
+            tar -xf FastDeploy.tar.gz
+            rm -rf FastDeploy.tar.gz
+            cd FastDeploy
+            git config --global user.name "FastDeployCI"
+            git config --global user.email "fastdeploy_ci@example.com"
+            git log -n 3 --oneline
+
+      - name: Run FastDeploy Stable Tests
+        shell: bash
+        env:
+          docker_image: ${{ inputs.DOCKER_IMAGE }}
+          fastdeploy_wheel_url: ${{ inputs.FASTDEPLOY_WHEEL_URL }}
+          CACHE_DIR: ${{ inputs.CACHE_DIR }}
+          MODEL_CACHE_DIR: ${{ inputs.MODEL_CACHE_DIR }}
+        run: |
+          runner_name="${{ runner.name }}"
+          CARD_ID=$(echo "${runner_name}" | awk -F'-' '{print $NF}')
+          DEVICES=$(echo "$CARD_ID" | fold -w1 | paste -sd,)
+          DEVICE_PORT=$(echo "$DEVICES" | cut -d',' -f1)
+
+          FLASK_PORT=$((42068 + DEVICE_PORT * 100))
+          FD_API_PORT=$((42088 + DEVICE_PORT * 100))
+          FD_ENGINE_QUEUE_PORT=$((42058 + DEVICE_PORT * 100))
+          FD_METRICS_PORT=$((42078 + DEVICE_PORT * 100))
+          FD_CACHE_QUEUE_PORT=$((42038 + DEVICE_PORT * 100))
+          FD_INFERENCE_MSG_QUEUE_ID=$(( 42048 + DEVICE_PORT * 100))
+          echo "Test ENV Parameter:"
+          echo "========================================================="
+          echo "FLASK_PORT=${FLASK_PORT}"
+          echo "FD_API_PORT=${FD_API_PORT}"
+          echo "FD_ENGINE_QUEUE_PORT=${FD_ENGINE_QUEUE_PORT}"
+          echo "FD_METRICS_PORT=${FD_METRICS_PORT}"
+          echo "FD_INFERENCE_MSG_QUEUE_ID=${FD_INFERENCE_MSG_QUEUE_ID}"
+          echo "FD_CACHE_QUEUE_PORT=${FD_CACHE_QUEUE_PORT}"
+          echo "DEVICES=${DEVICES}"
+          echo "========================================================="
+
+          CACHE_DIR="${CACHE_DIR:-$(dirname "$(dirname "${{ github.workspace }}")")}"
+          echo "CACHE_DIR is set to ${CACHE_DIR}"
+          if [ ! -f "${CACHE_DIR}/gitconfig" ]; then
+            touch "${CACHE_DIR}/gitconfig"
+          fi
+          if [ ! -d "${MODEL_CACHE_DIR}" ]; then
+            echo "Error: MODEL_CACHE_DIR '${MODEL_CACHE_DIR}' does not exist."
+            exit 1
+          fi
+
+          PORTS=($FLASK_PORT $FD_API_PORT $FD_ENGINE_QUEUE_PORT $FD_METRICS_PORT)
+          LOG_FILE="./port_cleanup_$(date +%Y%m%d_%H%M%S).log"
+          echo "==== LOG_FILE is ${LOG_FILE} ===="
+
+          echo "==== PORT CLEAN BEFORE TASK RUN ====" | tee -a $LOG_FILE
+
+          for port in "${PORTS[@]}"; do
+              PIDS=$(lsof -t -i :$port || true)
+              if [ -n "$PIDS" ]; then
+                  echo "Port $port is occupied by PID(s): $PIDS" | tee -a $LOG_FILE
+                  echo "$PIDS" | xargs -r kill -9
+                  echo "Port $port cleared" | tee -a $LOG_FILE
+              else
+                  echo "Port $port is free" | tee -a $LOG_FILE
+              fi
+          done
+
+          echo "==== PORT CLEAN COMPLETE ====" | tee -a $LOG_FILE
+
+          echo "========================================================="
+          echo "Ensuring no stale container named ${runner_name} ..."
+          if [ "$(docker ps -a -q -f name=${runner_name})" ]; then
+            echo "Removing stale container: ${runner_name}"
+            docker rm -f ${runner_name} || true
+          fi
+
+          docker run --rm --ipc=host --pid=host --net=host \
+          --name ${runner_name} \
+          -v $(pwd):/workspace \
+          -w /workspace \
+          -e fastdeploy_wheel_url=${fastdeploy_wheel_url} \
+          -e "FD_API_PORT=${FD_API_PORT}" \
+          -e "FD_ENGINE_QUEUE_PORT=${FD_ENGINE_QUEUE_PORT}" \
+          -e "FD_METRICS_PORT=${FD_METRICS_PORT}" \
+          -e "FLASK_PORT=${FLASK_PORT}" \
+          -e "FD_INFERENCE_MSG_QUEUE_ID=${FD_INFERENCE_MSG_QUEUE_ID}" \
+          -e "FD_CACHE_QUEUE_PORT=${FD_CACHE_QUEUE_PORT}" \
+          -v "${MODEL_CACHE_DIR}:/MODELDATA" \
+          -v "${CACHE_DIR}/gitconfig:/etc/gitconfig:ro" \
+          -v "${CACHE_DIR}/.cache:/root/.cache" \
+          -v "${CACHE_DIR}/ConfigDir:/root/.config" \
+          -e TZ="Asia/Shanghai" \
+          --gpus '"device='"${DEVICES}"'"' ${docker_image} /bin/bash -xc '
+          python -m pip install paddlepaddle-gpu==3.2.0 -i https://www.paddlepaddle.org.cn/packages/stable/cu126/
+
+          pip config set global.index-url https://mirrors.tuna.tsinghua.edu.cn/pypi/web/simple
+
+          python -m pip install ${fastdeploy_wheel_url}
+          python -m pip install pytest
+
+          git config --global --add safe.directory /workspace/FastDeploy
+          cd FastDeploy
+          TEST_EXIT_CODE=0
+          pushd tests/ce/stable_cases
+          bash launch_model.sh /MODELDATA
+          bash run.sh || TEST_EXIT_CODE=1
+          popd
+          echo "TEST_EXIT_CODE=${TEST_EXIT_CODE}" >> /workspace/FastDeploy/exit_code.env
+          '
+          if [ -f ./FastDeploy/exit_code.env ]; then
+            source ./FastDeploy/exit_code.env
+            cat ./FastDeploy/exit_code.env >> $GITHUB_ENV
+          fi
+          echo "TEST_EXIT_CODE=${TEST_EXIT_CODE}"
+          exit ${TEST_EXIT_CODE}

.github/workflows/_unit_test_coverage.yml

@@ -0,0 +1,322 @@
+name: Coverage Check
+description: "Run FastDeploy Unit Tests and Coverage"
+
+on:
+  workflow_call:
+    inputs:
+      DOCKER_IMAGE:
+        description: "Build Images"
+        required: true
+        type: string
+        default: "ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleqa:cuda126-py310"
+      FASTDEPLOY_ARCHIVE_URL:
+        description: "URL of the compressed FastDeploy code archive."
+        required: true
+        type: string
+      FASTDEPLOY_WHEEL_URL:
+        description: "URL of the FastDeploy Wheel."
+        required: true
+        type: string
+      CACHE_DIR:
+        description: "Cache Dir Use"
+        required: false
+        type: string
+        default: ""
+      MODEL_CACHE_DIR:
+        description: "Cache Dir Use"
+        required: false
+        type: string
+        default: ""
+    secrets:
+      github-token:
+        required: true
+
+jobs:
+  check_cov_skip:
+    uses: ./.github/workflows/check-bypass.yml
+    secrets:
+      github-token: ${{ secrets.github-token }}
+    with:
+      workflow-name: coverage
+
+  run_tests_with_coverage:
+    runs-on: [self-hosted, GPU-h1z1-2Cards]
+    timeout-minutes: 60
+    needs: check_cov_skip
+    if: needs.check_cov_skip.outputs.can-skip != 'true'
+    outputs:
+      diff_cov_file_url: ${{ steps.cov_upload.outputs.diff_cov_file_url }}
+      unittest_failed_url: ${{ steps.cov_upload.outputs.unittest_failed_url }}
+      diff_cov_result_json_url: ${{ steps.cov_upload.outputs.diff_cov_result_json_url }}
+    steps:
+      - name: Code Prepare
+        shell: bash
+        env:
+          docker_image: ${{ inputs.DOCKER_IMAGE }}
+          fd_archive_url: ${{ inputs.FASTDEPLOY_ARCHIVE_URL }}
+        run: |
+            set -x
+            REPO="https://github.com/${{ github.repository }}.git"
+            FULL_REPO="${{ github.repository }}"
+            REPO_NAME="${FULL_REPO##*/}"
+            BASE_BRANCH="${{ github.base_ref }}"
+
+            # Clean the repository directory before starting
+            docker run --rm --net=host -v $(pwd):/workspace -w /workspace \
+            -e "REPO_NAME=${REPO_NAME}" \
+            ${docker_image} /bin/bash -c '
+              if [ -d ${REPO_NAME} ]; then
+                echo "Directory ${REPO_NAME} exists, removing it..."
+                rm -rf ${REPO_NAME}*
+              fi
+            '
+
+            wget -q ${fd_archive_url}
+            tar -xf FastDeploy.tar.gz
+            rm -rf FastDeploy.tar.gz
+            cd FastDeploy
+            git config --global user.name "FastDeployCI"
+            git config --global user.email "fastdeploy_ci@example.com"
+            git log -n 3 --oneline
+      - name: Run FastDeploy Unit Tests and Coverage
+        shell: bash
+        env:
+          docker_image: ${{ inputs.DOCKER_IMAGE }}
+          fd_wheel_url: ${{ inputs.FASTDEPLOY_WHEEL_URL }}
+          CACHE_DIR: ${{ inputs.CACHE_DIR }}
+          BASE_REF: ${{ github.event.pull_request.base.ref }}
+          MODEL_CACHE_DIR: ${{ inputs.MODEL_CACHE_DIR }}
+          IS_PR: ${{ github.event_name == 'pull_request' }}
+        run: |
+          if [[ "$IS_PR" == "true" ]]; then
+            echo "Running on PR"
+          else
+            echo "Not a PR"
+          fi
+          runner_name="${{ runner.name }}"
+          CARD_ID=$(echo "${runner_name}" | awk -F'-' '{print $NF}')
+          DEVICES=$(echo "$CARD_ID" | fold -w1 | paste -sd,)
+          DEVICE_PORT=$(echo "$DEVICES" | cut -d',' -f1)
+
+          FLASK_PORT=$((42068 + DEVICE_PORT * 100))
+          FD_API_PORT=$((42088 + DEVICE_PORT * 100))
+          FD_ENGINE_QUEUE_PORT=$((42058 + DEVICE_PORT * 100))
+          FD_METRICS_PORT=$((42078 + DEVICE_PORT * 100))
+          FD_CACHE_QUEUE_PORT=$((42098 + DEVICE_PORT * 100))
+          echo "Test ENV Parameter:"
+          echo "========================================================="
+          echo "FLASK_PORT=${FLASK_PORT}"
+          echo "FD_API_PORT=${FD_API_PORT}"
+          echo "FD_ENGINE_QUEUE_PORT=${FD_ENGINE_QUEUE_PORT}"
+          echo "FD_METRICS_PORT=${FD_METRICS_PORT}"
+          echo "FD_CACHE_QUEUE_PORT=${FD_CACHE_QUEUE_PORT}"
+          echo "DEVICES=${DEVICES}"
+          echo "========================================================="
+
+          CACHE_DIR="${CACHE_DIR:-$(dirname "$(dirname "${{ github.workspace }}")")}"
+          echo "CACHE_DIR is set to ${CACHE_DIR}"
+          if [ ! -f "${CACHE_DIR}/gitconfig" ]; then
+            touch "${CACHE_DIR}/gitconfig"
+          fi
+
+          PORTS=($FLASK_PORT $FD_API_PORT $FD_ENGINE_QUEUE_PORT $FD_METRICS_PORT $FD_CACHE_QUEUE_PORT)
+          LOG_FILE="./port_cleanup_$(date +%Y%m%d_%H%M%S).log"
+          echo "==== LOG_FILE is ${LOG_FILE} ===="
+
+          echo "==== PORT CLEAN BEFORE TASK RUN ====" | tee -a $LOG_FILE
+
+          for port in "${PORTS[@]}"; do
+              PIDS=$(lsof -t -i :$port || true)
+              if [ -n "$PIDS" ]; then
+                  echo "Port $port is occupied by PID(s): $PIDS" | tee -a $LOG_FILE
+                  echo "$PIDS" | xargs -r kill -9
+                  echo "Port $port cleared" | tee -a $LOG_FILE
+              else
+                  echo "Port $port is free" | tee -a $LOG_FILE
+              fi
+          done
+
+          echo "==== PORT CLEAN COMPLETE ====" | tee -a $LOG_FILE
+
+          echo "========================================================="
+          echo "Ensuring no stale container named ${runner_name} ..."
+          if [ "$(docker ps -a -q -f name=${runner_name})" ]; then
+            echo "Removing stale container: ${runner_name}"
+            docker rm -f ${runner_name} || true
+          fi
+
+          docker run --rm --net=host \
+          --name ${runner_name} \
+          --cap-add=SYS_PTRACE --shm-size=64G \
+          -v $(pwd):/workspace -w /workspace \
+          -v "${CACHE_DIR}/gitconfig:/etc/gitconfig:ro" \
+          -v "${CACHE_DIR}/.cache:/root/.cache" \
+          -v "${CACHE_DIR}/ConfigDir:/root/.config" \
+          -v "${MODEL_CACHE_DIR}:/ModelData:ro" \
+          -e "MODEL_PATH=/ModelData" \
+          -e "FD_API_PORT=${FD_API_PORT}" \
+          -e "FD_ENGINE_QUEUE_PORT=${FD_ENGINE_QUEUE_PORT}" \
+          -e "FD_METRICS_PORT=${FD_METRICS_PORT}" \
+          -e "FLASK_PORT=${FLASK_PORT}" \
+          -e "FD_CACHE_QUEUE_PORT=${FD_CACHE_QUEUE_PORT}" \
+          -e TZ="Asia/Shanghai" \
+          -e "fd_wheel_url=${fd_wheel_url}" \
+          -e "BASE_REF=${BASE_REF}" \
+          -e "IS_PR=${IS_PR}" \
+          --gpus "\"device=${DEVICES}\"" ${docker_image} /bin/bash -c '
+
+          git config --global --add safe.directory /workspace/FastDeploy
+          cd FastDeploy
+          git diff origin/${BASE_REF}..HEAD --unified=0 > diff.txt
+          python -m pip install paddlepaddle-gpu==3.2.0 -i https://www.paddlepaddle.org.cn/packages/stable/cu126/
+          pip config set global.extra-index-url https://mirrors.tuna.tsinghua.edu.cn/pypi/web/simple
+
+          python -m pip install coverage
+          python -m pip install diff-cover
+          python -m pip install pytest-cov
+          python -m pip install jsonschema aistudio_sdk==0.3.5
+          python -m pip install ${fd_wheel_url}
+          rm -rf fastdeploy
+          # coverage subprocess use
+          python -m pip install ${fd_wheel_url} --no-deps --target=/workspace/FastDeploy
+          export PYTHONPATH=/workspace/FastDeploy/
+          if [ -d "tests/plugins" ]; then
+              cd tests/plugins
+              python setup.py install
+              cd ../..
+          else
+              echo "Warning: tests/plugins directory not found, skipping setup.py install"
+          fi
+          export COVERAGE_FILE=/workspace/FastDeploy/coveragedata/.coverage
+          export COVERAGE_RCFILE=/workspace/FastDeploy/scripts/.coveragerc
+          TEST_EXIT_CODE=0
+          bash scripts/coverage_run.sh || TEST_EXIT_CODE=8
+          echo "TEST_EXIT_CODE=${TEST_EXIT_CODE}" >> exit_code.env
+          coverage combine coveragedata/ || echo "No data to combine"
+          coverage report
+          coverage xml -o python_coverage_all.xml
+          COVERAGE_EXIT_CODE=0
+          if [[ "$IS_PR" == "true" ]]; then
+            diff-cover python_coverage_all.xml --diff-file=diff.txt --fail-under=80 --json-report diff_coverage.json || COVERAGE_EXIT_CODE=9
+            python scripts/generate_diff_coverage_xml.py diff.txt python_coverage_all.xml
+          else
+            echo "Not a PR, skipping diff-cover"
+          fi
+          echo "COVERAGE_EXIT_CODE=${COVERAGE_EXIT_CODE}" >> exit_code.env
+          '
+          if [ -f FastDeploy/exit_code.env ]; then
+            cat FastDeploy/exit_code.env >> $GITHUB_ENV
+          fi
+      - name: Upload unit resule and diff coverage to bos
+        id: cov_upload
+        shell: bash
+        run: |
+          cd FastDeploy
+          commit_id=${{ github.event.pull_request.head.sha }}
+          pr_num=${{ github.event.pull_request.number }}
+          target_path=paddle-github-action/PR/FastDeploy/${pr_num}/${commit_id}/SM${compile_arch//,/_}
+          wget  -q --no-proxy --no-check-certificate https://paddle-qa.bj.bcebos.com/CodeSync/develop/PaddlePaddle/PaddleTest/tools/bos_tools.py -O bos_tools.py
+          push_file=$(realpath bos_tools.py)
+          python -m pip install bce-python-sdk==0.9.29
+          diff_cov_file="diff_coverage.xml"
+          if [ -f ${diff_cov_file} ];then
+            python ${push_file} ${diff_cov_file} ${target_path}/CoverageData
+            target_path_stripped="${target_path#paddle-github-action/}"
+            DIFF_COV_FILE_URL=https://paddle-github-action.bj.bcebos.com/${target_path_stripped}/CoverageData/${diff_cov_file}
+            echo "diff_cov_file_url=${DIFF_COV_FILE_URL}" >> $GITHUB_OUTPUT
+            echo "diff_cov_file_url=${DIFF_COV_FILE_URL}" >> $GITHUB_ENV
+          fi
+          diff_cov_result_json="diff_coverage.json"
+          if [ -f ${diff_cov_result_json} ];then
+            python ${push_file} ${diff_cov_result_json} ${target_path}/CoverageData
+            target_path_stripped="${target_path#paddle-github-action/}"
+            DIFF_COV_JSON_URL=https://paddle-github-action.bj.bcebos.com/${target_path_stripped}/CoverageData/${diff_cov_result_json}
+            echo "diff_cov_result_json_url=${DIFF_COV_JSON_URL}" >> $GITHUB_OUTPUT
+            echo "diff_cov_result_json_url=${DIFF_COV_JSON_URL}" >> $GITHUB_ENV
+          fi
+          unittest_result="failed_tests.log"
+          if [ -s ${unittest_result} ];then
+            python ${push_file} ${unittest_result} ${target_path}/UnitTestResult
+            target_path_stripped="${target_path#paddle-github-action/}"
+            UNIT_TEST_RESULT_URL=https://paddle-github-action.bj.bcebos.com/${target_path_stripped}/UnitTestResult/${unittest_result}
+            echo "unittest_failed_url=${UNIT_TEST_RESULT_URL}" >> $GITHUB_OUTPUT
+            echo "unittest_failed_url=${UNIT_TEST_RESULT_URL}" >> $GITHUB_ENV
+          fi
+      - name: Check Unit Test Success
+        shell: bash
+        run: |
+          cd FastDeploy
+          if [ "$TEST_EXIT_CODE" -eq 8 ]; then
+            filename=$(basename "$unittest_failed_url")
+            if [ -z "${unittest_failed_url}" ]; then
+              echo "No diff unit failed file URL provided."
+            else
+              rm -rf "${filename}"
+              wget -O ${filename} ${unittest_failed_url} || echo "Download unittest file failed, but continuing..."
+            fi
+            echo "Unit tests failed (exit code 8)"
+            if [ -f "${filename}" ];then
+              echo "Failed test cases:"
+              cat "${filename}"
+            fi
+            exit "$TEST_EXIT_CODE"
+          fi
+          echo "All tests passed"
+
+      - name: Verify Code Coverage Threshold (80%)
+        if: ${{ github.event_name == 'pull_request' }}
+        shell: bash
+        run: |
+          cd FastDeploy
+          if [ "$COVERAGE_EXIT_CODE" -eq 9 ]; then
+            echo "Coverage generation failed (exit code 9)"
+            filename=$(basename "$diff_cov_result_json_url")
+            if [ -z "${diff_cov_result_json_url}" ]; then
+              echo "No diff cov result file URL provided."
+            else
+              rm -rf "${filename}"
+              wget -O ${filename} ${diff_cov_result_json_url} || echo "Download cov json file failed, but continuing..."
+            fi
+            if [ -f "${filename}" ];then
+              echo "Failed test cases:"
+              if command -v jq >/dev/null 2>&1; then
+                  jq . "${filename}"
+              else
+                  cat "${filename}"
+              fi
+            fi
+            exit "$COVERAGE_EXIT_CODE"
+          fi
+          echo "coverage passed"
+          exit 0
+
+  diff_coverage_report:
+    needs: run_tests_with_coverage
+    if: always()
+    runs-on: ubuntu-latest
+    env:
+      fd_archive_url: ${{ inputs.FASTDEPLOY_ARCHIVE_URL }}
+    steps:
+      - name: coverage diff file download
+        shell: bash
+        env:
+          diff_cov_file_url: ${{ needs.run_tests_with_coverage.outputs.diff_cov_file_url }}
+        run: |
+          wget ${fd_archive_url}
+          tar -xf FastDeploy.tar.gz
+          cd FastDeploy
+          if [ -z "${diff_cov_file_url}" ]; then
+            echo "No diff coverage file URL provided."
+            exit 0
+          fi
+          wget "${diff_cov_file_url}" -O ./diff_coverage.xml || echo "Download cov file failed, but continuing..."
+      - name: Upload diff coverage report
+        if: ${{ needs.run_tests_with_coverage.outputs.diff_cov_file_url != null && needs.run_tests_with_coverage.outputs.diff_cov_file_url != '' }}
+        uses: codecov/codecov-action@v5
+        with:
+          files: ./FastDeploy/diff_coverage.xml
+          name: python diff coverage
+          verbose: true
+          disable_search: true
+          commit_parent: false
+          flags: diff

.github/workflows/approve.yml

@@ -0,0 +1,42 @@
+name: Approval
+
+on:
+  pull_request:
+    branches:
+      - develop
+      - 'release/*'
+
+env:
+  GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+
+jobs:
+  Approval:
+    name: Approval
+    if: ${{ github.repository_owner == 'PaddlePaddle' }}
+    runs-on: ubuntu-latest
+    env:
+      PR_ID: ${{ github.event.pull_request.number }}
+      BRANCH: ${{ github.event.pull_request.base.ref }}
+    steps:
+      - name: Checkout base repo
+        uses: actions/checkout@v4
+        with:
+          ref: ${{ github.event.pull_request.base.ref }}
+          fetch-depth: 1000
+
+      - name: Merge PR to test branch
+        run: |
+          git fetch origin pull/${PR_ID}/merge
+          git checkout -b test FETCH_HEAD
+          git log -n 3 --oneline
+          git remote add upstream https://github.com/PaddlePaddle/FastDeploy.git
+          git fetch upstream $BRANCH
+
+      - name: Setup python3.10
+        uses: actions/setup-python@v5
+        with:
+          python-version: '3.10'
+
+      - name: Run approval check script
+        run: |
+          bash scripts/check_approval.sh

.github/workflows/ce_job.yml

@@ -0,0 +1,248 @@
+name: CE Compile Job
+
+on:
+  workflow_dispatch:
+  push:
+    branches:
+      - develop
+      - 'release/*'
+permissions: read-all
+
+concurrency:
+  group: ${{ github.ref }}-${{ github.sha }}
+  cancel-in-progress: true
+
+jobs:
+  ce_job_pre_check:
+    runs-on: ubuntu-latest
+    env:
+      COMPILE_BRANCH: ${{ vars.COMPILE_BRANCH }}
+      CE_COMPILE_SELECTION: ${{ vars.CE_COMPILE_SELECTION }}
+      COMPILE_USE_PADDLE_WHL_URL_MAPPINGS: ${{ vars.COMPILE_USE_PADDLE_WHL_URL_MAPPINGS }}
+    outputs:
+      branch_match: ${{ steps.set_output.outputs.branch_match }}
+      compile_use_paddle_whl_url: ${{ steps.set_output.outputs.compile_use_paddle_whl_url }}
+      sm8689_match: ${{ steps.set_output.outputs.sm8689_match }}
+      sm8090_match: ${{ steps.set_output.outputs.sm8090_match }}
+
+    steps:
+      - name: Set Version
+        id: set_output
+        env:
+          COMPILE_BRANCH: ${{ env.COMPILE_BRANCH }}
+          CE_COMPILE_SELECTION: ${{ env.CE_COMPILE_SELECTION }}
+          COMPILE_USE_PADDLE_WHL_URL_MAPPINGS: ${{ env.COMPILE_USE_PADDLE_WHL_URL_MAPPINGS }}
+          GITHUB_REF_NAME: ${{ github.ref_name }}
+        run: |
+          # 选择要触发编译任务的分支 done
+          # 选择指定分支要编译的任务 8090或者8689
+          # 指定分支编译要使用的Paddle的安装包,默认使用nightly最新的
+
+          IFS=',' read -ra BRANCHES <<< "$COMPILE_BRANCH"
+          MATCH=false
+          for b in "${BRANCHES[@]}"; do
+            if [[ "$b" == "${GITHUB_REF_NAME}" ]]; then
+              MATCH=true
+              break
+            fi
+          done
+          echo "branch_match=$MATCH" >> $GITHUB_OUTPUT
+
+          # 通过变量CE_COMPILE_SELECTION中的映射关系,决定分支是编译sm8090还是sm8689
+          for pair in $(echo "$CE_COMPILE_SELECTION" | tr ';' ' '); do
+            branch=$(echo "$pair" | cut -d',' -f1)
+            compile_task_list=$(echo "$pair" | cut -d',' -f2)
+
+            if [[ "$branch" == "$GITHUB_REF_NAME" ]]; then
+
+              # 判断里面是否包含 sm8090 或 sm8689
+              if [[ "$compile_task_list" == *"sm8090"* ]]; then
+                echo "sm8090_match=true" >> $GITHUB_OUTPUT
+              fi
+              if [[ "$compile_task_list" == *"sm8689"* ]]; then
+                echo "sm8689_match=true" >> $GITHUB_OUTPUT
+              fi
+              break
+            fi
+          done
+
+          # 通过变量COMPILE_USE_PADDLE_WHL_URL_MAPPINGS中的映射关系,决定是否是安装指定版本的Paddle还是直接安装URL
+          for pair in $(echo $COMPILE_USE_PADDLE_WHL_URL_MAPPINGS | tr ';' ' '); do
+            branch=$(echo "$pair" | cut -d',' -f1)
+            paddle_whl_url=$(echo "$pair" | cut -d',' -f2)
+            if [[ "$branch" == "${{ github.ref_name }}" ]]; then
+              FOUND_PADDLE_URL="$paddle_whl_url"
+              echo "compile_use_paddle_whl_url=${FOUND_PADDLE_URL}" >> $GITHUB_OUTPUT
+              break
+            fi
+          done
+
+  print_ce_job_pre_check_outputs:
+    runs-on: ubuntu-latest
+    needs: ce_job_pre_check
+    steps:
+      - name: Print outputs as JSON
+        run: |
+          echo '${{ toJSON(needs.ce_job_pre_check.outputs) }}'
+
+
+  clone:
+    environment: CodeSync
+    name: FD-Clone-Linux
+    runs-on: ubuntu-latest
+    needs: ce_job_pre_check
+    if: ${{ needs.ce_job_pre_check.outputs.branch_match == 'true' }}
+    outputs:
+      repo_archive_url: ${{ steps.set_output.outputs.repo_archive_url }}
+    steps:
+      - name: Clone FastDeploy
+        uses: actions/checkout@v4
+        with:
+          ref: ${{ github.event_name == 'pull_request'
+                && github.event.pull_request.base.ref
+                || github.ref_name }}
+          submodules: 'recursive'
+          fetch-depth: 1000
+
+      - name: Python Setup
+        uses: actions/setup-python@v5
+        with:
+          python-version: '3.10'
+      - name: Code Info Show and Upload
+        id: set_output
+        env:
+          AK: ${{ secrets.BOS_AK }}
+          SK: ${{ secrets.BOS_SK }}
+        run: |
+          git config --unset http.https://github.com/.extraheader
+          git submodule foreach --recursive sh -c "git config --local --unset-all 'http.https://github.com/.extraheader'"
+          git submodule foreach --recursive sh -c "git config remote.origin.fetch '+refs/heads/*:refs/remotes/origin/*'"
+          echo "Current HEAD Log:"
+          git log --oneline -n 5
+          ls
+          cd ..
+          tar -zcf FastDeploy.tar.gz FastDeploy
+          commit_id=${{ github.sha }}
+          branch_name=${{ github.ref_name }}
+          target_path=paddle-qa/BRANCH/FastDeploy/${branch_name}/${commit_id}
+          wget  -q --no-proxy --no-check-certificate https://paddle-qa.bj.bcebos.com/CodeSync/develop/PaddlePaddle/PaddleTest/tools/bos_tools.py
+          push_file=$(realpath bos_tools.py)
+          python -m pip install bce-python-sdk==0.9.29
+          ls
+          python ${push_file} FastDeploy.tar.gz ${target_path}
+          target_path_stripped="${target_path#paddle-qa/}"
+          REPO_ARCHIVE_URL=https://paddle-qa.bj.bcebos.com/${target_path_stripped}/FastDeploy.tar.gz
+          echo "repo_archive_url=${REPO_ARCHIVE_URL}" >> $GITHUB_OUTPUT
+
+  resultshow:
+    name: Show Code Archive Output
+    needs: clone
+    runs-on: ubuntu-latest
+    steps:
+      - name: Print wheel path
+        run: |
+          echo "The code archive is located at: ${{ needs.clone.outputs.repo_archive_url }}"
+
+  build_sm8090:
+    name: BUILD_SM8090
+    needs: [clone, ce_job_pre_check]
+    if: ${{ needs.ce_job_pre_check.outputs.sm8090_match == 'true' }}
+    uses: ./.github/workflows/_build_linux.yml
+    with:
+      DOCKER_IMAGE: ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleqa:fastdeploy-ciuse-cuda126-dailyupdate
+      FASTDEPLOY_ARCHIVE_URL: ${{ needs.clone.outputs.repo_archive_url }}
+      COMPILE_ARCH: "80,90"
+      WITH_NIGHTLY_BUILD: OFF
+      FD_VERSION: 0.0.0
+      PADDLE_WHL_URL: ${{ needs.ce_job_pre_check.outputs.compile_use_paddle_whl_url }}
+
+  build_sm8689:
+    name: BUILD_SM8689
+    needs: [clone, ce_job_pre_check]
+    if: ${{ needs.ce_job_pre_check.outputs.sm8689_match == 'true' }}
+    uses: ./.github/workflows/_build_linux.yml
+    with:
+      DOCKER_IMAGE: ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleqa:fastdeploy-ciuse-cuda126-dailyupdate
+      FASTDEPLOY_ARCHIVE_URL: ${{ needs.clone.outputs.repo_archive_url }}
+      COMPILE_ARCH: "86,89"
+      WITH_NIGHTLY_BUILD: OFF
+      FD_VERSION: 0.0.0
+      PADDLE_WHL_URL: ${{ needs.ce_job_pre_check.outputs.compile_use_paddle_whl_url }}
+
+  ce_upload_sm8090:
+    environment: CodeSync
+    name: CE_UPLOAD
+    needs: build_sm8090
+    runs-on: ubuntu-latest
+    env:
+      AK: ${{ secrets.BOS_AK }}
+      SK: ${{ secrets.BOS_SK }}
+      FASTDEPLOY_WHEEL_URL: ${{ needs.build_sm8090.outputs.wheel_path }}
+      COMPILE_ARCH: "80,90"
+    steps:
+      - uses: actions/setup-python@v5
+        with:
+          python-version: '3.10'
+      - name: Wheel Info Show and Upload
+        run: |
+          echo "The wheel is located at: ${{ needs.build_sm8090.outputs.wheel_path }}"
+          wget -q --no-check-certificate ${{ needs.build_sm8090.outputs.wheel_path }}
+          filename=$(basename ${{ needs.build_sm8090.outputs.wheel_path }})
+
+          commit_id=${{ github.sha }}
+          branch_name=${{ github.ref_name }}
+          target_path=paddle-qa/paddle-pipeline/FastDeploy_ActionCE/SM${COMPILE_ARCH//,/_}/${branch_name}/${commit_id}
+
+          wget  -q --no-proxy --no-check-certificate https://paddle-qa.bj.bcebos.com/CodeSync/develop/PaddlePaddle/PaddleTest/tools/bos_tools.py
+          push_file=$(realpath bos_tools.py)
+          python -m pip install bce-python-sdk==0.9.29
+          ls
+          python ${push_file} ${filename} ${target_path}
+          target_path_stripped="${target_path#paddle-qa/}"
+          WHEEL_PATH=https://paddle-qa.bj.bcebos.com/${target_path_stripped}/${fd_wheel_name}
+          echo "commit wheel url is ${WHEEL_PATH}"
+
+          target_path_latest=paddle-qa/paddle-pipeline/FastDeploy_ActionCE/SM${COMPILE_ARCH//,/_}/${branch_name}/latest
+          python ${push_file} ${filename} ${target_path_latest}
+          target_path_stripped_latest="${target_path_latest#paddle-qa/}"
+          WHEEL_PATH_LATEST=https://paddle-qa.bj.bcebos.com/${target_path_stripped_latest}/${fd_wheel_name}
+          echo "latest wheel url is ${WHEEL_PATH_LATEST}"
+
+  ce_upload_sm8689:
+    environment: CodeSync
+    name: CE_UPLOAD
+    needs: build_sm8689
+    runs-on: ubuntu-latest
+    env:
+      AK: ${{ secrets.BOS_AK }}
+      SK: ${{ secrets.BOS_SK }}
+      FASTDEPLOY_WHEEL_URL: ${{ needs.build_sm8689.outputs.wheel_path }}
+      COMPILE_ARCH: "86,89"
+    steps:
+      - uses: actions/setup-python@v5
+        with:
+          python-version: '3.10'
+      - name: Wheel Info Show and Upload
+        run: |
+          echo "The wheel is located at: ${{ needs.build_sm8090.outputs.wheel_path }}"
+          wget -q --no-check-certificate ${{ needs.build_sm8090.outputs.wheel_path }}
+          filename=$(basename ${{ needs.build_sm8090.outputs.wheel_path }})
+
+          commit_id=${{ github.sha }}
+          branch_name=${{ github.ref_name }}
+          target_path=paddle-qa/paddle-pipeline/FastDeploy_ActionCE/SM${COMPILE_ARCH//,/_}/${branch_name}/${commit_id}
+
+          wget  -q --no-proxy --no-check-certificate https://paddle-qa.bj.bcebos.com/CodeSync/develop/PaddlePaddle/PaddleTest/tools/bos_tools.py
+          push_file=$(realpath bos_tools.py)
+          python -m pip install bce-python-sdk==0.9.29
+          ls
+          python ${push_file} ${filename} ${target_path}
+          target_path_stripped="${target_path#paddle-qa/}"
+          WHEEL_PATH=https://paddle-qa.bj.bcebos.com/${target_path_stripped}/${fd_wheel_name}
+          echo "commit wheel url is ${WHEEL_PATH}"
+
+          target_path_latest=paddle-qa/paddle-pipeline/FastDeploy_ActionCE/SM${COMPILE_ARCH//,/_}/${branch_name}/latest
+          python ${push_file} ${filename} ${target_path_latest}
+          target_path_stripped_latest="${target_path_latest#paddle-qa/}"
+          WHEEL_PATH_LATEST=https://paddle-qa.bj.bcebos.com/${target_path_stripped_latest}/${fd_wheel_name}
+          echo "latest wheel url is ${WHEEL_PATH_LATEST}"

.github/workflows/check-bypass.yml

@@ -0,0 +1,51 @@
+on:
+  workflow_call:
+    inputs:
+      workflow-name:
+        required: true
+        type: string
+    secrets:
+      github-token:
+        required: true
+    outputs:
+      can-skip:
+        description: "Whether the workflow can be skipped."
+        value: ${{ jobs.check-bypass.outputs.can-skip }}
+
+jobs:
+  check-bypass:
+    name: Check bypass
+    runs-on: ubuntu-latest
+    permissions:
+      contents: read
+    env:
+      CI_TEAM_MEMBERS: '["yuanlehome","YuanRisheng","Jiang-Jia-Jun","DDDivano","XieYunshen"]'
+    outputs:
+      can-skip: ${{ steps.check-bypass.outputs.can-skip }}
+    steps:
+      - name: Cleanup
+        run: |
+          rm -rf * .[^.]*
+
+      - id: check-bypass
+        name: Check Bypass
+        uses: PFCCLab/ci-bypass@v1
+        with:
+          github-token: ${{ secrets.github-token }}
+          non-pull-request-event-strategy: 'never-skipped'
+          type: 'composite'
+          composite-rule: |
+            {
+              "any": [
+                {
+                  "type": "labeled",
+                  "label": ["skip-ci: ${{ inputs.workflow-name }}", "skip-ci: all"],
+                  "username": ${{ env.CI_TEAM_MEMBERS }}
+                },
+                {
+                  "type": "commented",
+                  "comment-pattern": [".*/skip-ci ${{ inputs.workflow-name }}.*", ".*/skip-ci all.*"],
+                  "username": ${{ env.CI_TEAM_MEMBERS }}
+                }
+              ]
+            }

.github/workflows/ci.yml

@@ -1,109 +0,0 @@
-name: CI
-
-on:
-  pull_request:
-    branches:
-      - develop
-      - 'release/*'
-  workflow_dispatch:
-
-concurrency:
-  group: ${{ github.event.pull_request.number }}
-  cancel-in-progress: true
-
-jobs:
-  build:
-    runs-on: [self-hosted, GPU-L20-4Card]
-    steps:
-      - name: Print current runner name
-        run: |
-          echo "Current runner name: ${{ runner.name }}"
-      # Because the system version is lower than 2.23, the checkout cannot be used.
-      # - name: Checkout code
-      #   uses: actions/checkout@v4
-
-      - name: Code Checkout
-        env:
-          docker_image: ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddle:fastdeploy-ciuse-cuda126
-        run: |
-          REPO="https://github.com/${{ github.repository }}.git"
-          FULL_REPO="${{ github.repository }}"
-          REPO_NAME="${FULL_REPO##*/}"
-          BASE_BRANCH="${{ github.base_ref }}"
-          # Clean the repository directory before starting
-          docker run --rm --net=host -v $(pwd):/workspace -w /workspace \
-          -e "REPO_NAME=${REPO_NAME}" \
-          -e "BASE_BRANCH=${BASE_BRANCH}" \
-          ${docker_image} /bin/bash -c '
-            if [ -d ${REPO_NAME} ]; then
-              echo "Directory ${REPO_NAME} exists, removing it..."
-              rm -rf ${REPO_NAME}
-            fi
-          '
-          git config --global user.name "FastDeployCI"
-          git config --global user.email "fastdeploy_ci@example.com"
-          git clone ${REPO} ${REPO_NAME} -b ${BASE_BRANCH}
-          cd FastDeploy
-          if [ "${{ github.event_name }}" = "pull_request" ]; then
-            git fetch origin pull/${{ github.event.pull_request.number }}/head:pr/${{ github.event.pull_request.number }}
-            git merge pr/${{ github.event.pull_request.number }}
-            git log -n 3 --oneline
-          else
-            git checkout ${{ github.sha }}
-            git log -n 3 --oneline
-          fi
-
-      - name: Run CI unittest
-        env:
-          docker_image: ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddle:fastdeploy-ciuse-cuda126
-        run: |
-          runner_name="${{ runner.name }}"
-          last_char="${runner_name: -1}"
-
-          if [ "${last_char}" = "1" ]; then
-            gpu_id=2
-            DEVICES="2,3"
-          else
-            gpu_id=0
-            DEVICES="0,1"
-          fi
-          
-          FLASK_PORT=$((41068 + gpu_id * 100))
-          FD_API_PORT=$((41088 + gpu_id * 100))
-          FD_ENGINE_QUEUE_PORT=$((41058 + gpu_id * 100))
-          FD_METRICS_PORT=$((41078 + gpu_id * 100))
-          
-          PORTS=($FLASK_PORT $FD_API_PORT $FD_ENGINE_QUEUE_PORT $FD_METRICS_PORT)
-          LOG_FILE="./port_cleanup_$(date +%Y%m%d_%H%M%S).log"
-          echo "==== LOG_FILE is ${LOG_FILE} ===="
-
-          echo "==== PORT CLEAN BEFORE TASK RUN ====" | tee -a $LOG_FILE
-
-          for port in "${PORTS[@]}"; do
-              PIDS=$(lsof -t -i :$port || true)
-              if [ -n "$PIDS" ]; then
-                  echo "Port $port is occupied by PID(s): $PIDS" | tee -a $LOG_FILE
-                  echo "$PIDS" | xargs -r kill -9
-                  echo "Port $port cleared" | tee -a $LOG_FILE
-              else
-                  echo "Port $port is free" | tee -a $LOG_FILE
-              fi
-          done
-          echo "==== PORT CLEAN COMPLETE ====" | tee -a $LOG_FILE
-
-          PARENT_DIR=$(dirname "$WORKSPACE")
-          echo "PARENT_DIR:$PARENT_DIR"
-          docker run --rm --net=host -v $(pwd):/workspace -w /workspace \
-          -v "/ssd4/GithubActions/gitconfig:/etc/gitconfig:ro" \
-          -v "/ssd4/GithubActions/ModelData:/ModelData:ro" \
-          -v "/ssd4/GithubActions/CacheDir:/root/.cache" \
-          -v "/ssd4/GithubActions/ConfigDir:/root/.config" \
-          -e "MODEL_PATH=/ModelData" \
-          -e "FD_API_PORT=${FD_API_PORT}" \
-          -e "FD_ENGINE_QUEUE_PORT=${FD_ENGINE_QUEUE_PORT}" \
-          -e "FD_METRICS_PORT=${FD_METRICS_PORT}" \
-          --gpus '"device='"${DEVICES}"'"' ${docker_image} /bin/bash -c "
-          git config --global --add safe.directory /workspace/FastDeploy
-          cd FastDeploy
-          bash scripts/run_ci.sh
-          "

.github/workflows/ci_gcu.yml

@@ -0,0 +1,98 @@
+name: CI_GCU
+
+on:
+  pull_request:
+    branches:
+      - develop
+      - 'release/*'
+  workflow_dispatch:
+
+concurrency:
+  group: ${{ github.event.pull_request.number }}-gcu-ci
+  cancel-in-progress: true
+
+jobs:
+  CI_GCU:
+    runs-on:
+      group: GCU
+    steps:
+      - name: Print current runner name
+        run: |
+          echo "Current runner name: ${{ runner.name }}"
+
+      - name: Code Checkout
+        env:
+          docker_image: ccr-2vdh3abv-pub.cnc.bj.baidubce.com/device/paddle-gcu:topsrider3.5.102-ubuntu20-x86_64-gcc84
+        run: |
+          REPO="https://github.com/${{ github.repository }}.git"
+          FULL_REPO="${{ github.repository }}"
+          REPO_NAME="${FULL_REPO##*/}"
+          BASE_BRANCH="${{ github.base_ref }}"
+          # Clean the repository directory before starting
+          docker run --rm --net=host -v $(pwd):/workspace \
+          -v ${{ github.workspace }}/../../..:${{ github.workspace }}/../../.. \
+          -w /workspace \
+          -e "REPO_NAME=${REPO_NAME}" \
+          -e "BASE_BRANCH=${BASE_BRANCH}" \
+          ${docker_image} /bin/bash -c '
+            if [ -d ${REPO_NAME} ]; then
+              echo "Directory ${REPO_NAME} exists, removing it..."
+              rm -rf ${REPO_NAME}
+            fi
+          '
+          git config --global user.name "FastDeployCI"
+          git config --global user.email "fastdeploy_ci@example.com"
+          source ${{ github.workspace }}/../../../proxy
+          git clone ${REPO} ${REPO_NAME} -b ${BASE_BRANCH}
+          cd FastDeploy
+          if [ "${{ github.event_name }}" = "pull_request" ]; then
+            git fetch origin pull/${{ github.event.pull_request.number }}/head:pr/${{ github.event.pull_request.number }}
+            git merge pr/${{ github.event.pull_request.number }}
+            git log -n 3 --oneline
+          else
+            git checkout ${{ github.sha }}
+            git log -n 3 --oneline
+          fi
+          echo "Copy models..."
+          sudo mkdir -p ci_models && sudo cp -r /work/deps/ERNIE-4.5-21B-A3B-Paddle ci_models
+          echo "Copy models done."
+
+      - name: Run CI unittest
+        env:
+          docker_image: ccr-2vdh3abv-pub.cnc.bj.baidubce.com/device/paddle-gcu:topsrider3.5.102-ubuntu20-x86_64-gcc84
+        run: |
+          runner_name="${{ runner.name }}"
+          last_char="${runner_name: -1}"
+
+          if [[ "$last_char" =~ [0-3] ]]; then
+            gcu_id="$last_char"
+          else
+            gcu_id="0"
+          fi
+          FD_API_PORT=$((9180 + gcu_id * 100))
+          FD_ENGINE_QUEUE_PORT=$((9150 + gcu_id * 100))
+          FD_METRICS_PORT=$((9170 + gcu_id * 100))
+
+          PARENT_DIR=$(dirname "$WORKSPACE")
+          echo "PARENT_DIR:$PARENT_DIR"
+          echo "Install drivers..."
+          cd /work/deps
+          sudo bash TopsRider_i3x_*_deb_amd64.run --driver --no-auto-load -y
+          cd -
+          echo "Create docker..."
+          docker run --rm --network=host --ipc=host --privileged \
+          -v $(pwd):/workspace \
+          -v /home:/home \
+          -v /work:/work \
+          -w /workspace \
+          -e "MODEL_PATH=./ci_models" \
+          -e "http_proxy=$(git config --global --get http.proxy)" \
+          -e "https_proxy=$(git config --global --get https.proxy)" \
+          -e "FD_API_PORT=${FD_API_PORT}" \
+          -e "FD_ENGINE_QUEUE_PORT=${FD_ENGINE_QUEUE_PORT}" \
+          -e "FD_METRICS_PORT=${FD_METRICS_PORT}" \
+          ${docker_image} /bin/bash -c "
+          git config --global --add safe.directory /workspace/FastDeploy
+          cd FastDeploy
+          bash scripts/run_ci_gcu.sh
+          "

.github/workflows/ci_iluvatar.yml

@@ -11,7 +11,8 @@ concurrency:
 
 jobs:
   CI_ILUVATAR:
-    runs-on: [self-hosted, IXUCA]
+    runs-on:
+      group: IXUCA
     steps:
       - name: Print current runner name
         run: |

.github/workflows/ci_xpu.yml

@@ -24,7 +24,7 @@ jobs:
 
       - name: Code Checkout
         env:
-          docker_image: ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/fastdeploy-xpu:2.0.0
+          docker_image: ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/fastdeploy-xpu:2.1.0
         run: |
           REPO="https://github.com/${{ github.repository }}.git"
           FULL_REPO="${{ github.repository }}"
@@ -55,7 +55,7 @@ jobs:
 
       - name: Run CI unittest
         env:
-          docker_image: ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/fastdeploy-xpu:2.0.0
+          docker_image: ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/fastdeploy-xpu:2.1.0
         run: |
           runner_name="${{ runner.name }}"
           last_char="${runner_name: -1}"
@@ -77,6 +77,7 @@ jobs:
           -e "MODEL_PATH=/ssd3/model" \
           -e "http_proxy=$(git config --global --get http.proxy)" \
           -e "https_proxy=$(git config --global --get https.proxy)" \
+          -e "no_proxy=bcebos.com,mirrors.tuna.tsinghua.edu.cn,127.0.0.1,localhost" \
           -e "FD_API_PORT=${FD_API_PORT}" \
           -e "FD_ENGINE_QUEUE_PORT=${FD_ENGINE_QUEUE_PORT}" \
           -e "FD_METRICS_PORT=${FD_METRICS_PORT}" \

.github/workflows/gh-pages.yml

@@ -15,7 +15,7 @@ jobs:
       - uses: actions/setup-python@v5
         with:
           python-version: 3.x
-      - run: pip install mkdocs-material mkdocs-get-deps mkdocs-material-extensions mkdocs-multilang
+      - run: pip install mkdocs-material mkdocs-get-deps mkdocs-material-extensions mkdocs-multilang mkdocs-static-i18n
       - name: Deploy to GitHub Pages
         env:
           GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}

.github/workflows/pr_build_and_test.yml

@@ -19,9 +19,9 @@ jobs:
     needs: clone
     uses: ./.github/workflows/_build_linux.yml
     with:
-      DOCKER_IMAGE: ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleqa:cuda126-py310
+      DOCKER_IMAGE: ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleqa:fastdeploy-ciuse-cuda126-dailyupdate
       FASTDEPLOY_ARCHIVE_URL: ${{ needs.clone.outputs.repo_archive_url }}
-      COMPILE_ARCH: "90"
+      COMPILE_ARCH: "89,90"
       WITH_NIGHTLY_BUILD: "OFF"
       FD_VERSION: "0.0.0"
 
@@ -33,3 +33,65 @@ jobs:
       - name: Print wheel path
         run: |
           echo "The built wheel is located at: ${{ needs.build.outputs.wheel_path }}"
+
+  unittest_coverage:
+    name: Run FastDeploy Unit Tests and Coverage
+    needs: [clone,build]
+    uses: ./.github/workflows/_unit_test_coverage.yml
+    with:
+      DOCKER_IMAGE: ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleqa:fastdeploy-ciuse-cuda126-dailyupdate
+      FASTDEPLOY_ARCHIVE_URL: ${{ needs.clone.outputs.repo_archive_url }}
+      FASTDEPLOY_WHEEL_URL: ${{ needs.build.outputs.wheel_path }}
+      MODEL_CACHE_DIR: "/ssd2/actions-runner/ModelData"
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+
+  logprob_test:
+    name: Run FastDeploy LogProb Tests
+    needs: [build]
+    uses: ./.github/workflows/_logprob_test_linux.yml
+    with:
+      DOCKER_IMAGE: ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleqa:fastdeploy-ciuse-cuda126-dailyupdate
+      PADDLETEST_ARCHIVE_URL: "https://xly-devops.bj.bcebos.com/PaddleTest/PaddleTest.tar.gz"
+      FASTDEPLOY_WHEEL_URL: ${{ needs.build.outputs.wheel_path }}
+      MODEL_CACHE_DIR: "/ssd2/actions-runner/ModelData"
+
+  pre_ce_test:
+    name: Extracted partial CE model tasks to run in CI.
+    needs: [clone,build]
+    uses: ./.github/workflows/_pre_ce_test.yml
+    with:
+      DOCKER_IMAGE: ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleqa:fastdeploy-ciuse-cuda126-dailyupdate
+      FASTDEPLOY_ARCHIVE_URL: ${{ needs.clone.outputs.repo_archive_url }}
+      FASTDEPLOY_WHEEL_URL: ${{ needs.build.outputs.wheel_path }}
+      MODEL_CACHE_DIR: "/ssd2/actions-runner/ModelData"
+
+  base_test:
+    name: Run Base Tests
+    needs: [clone,build]
+    uses: ./.github/workflows/_base_test.yml
+    with:
+      DOCKER_IMAGE: ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleqa:fastdeploy-ciuse-cuda126-dailyupdate
+      FASTDEPLOY_ARCHIVE_URL: ${{ needs.clone.outputs.repo_archive_url }}
+      FASTDEPLOY_WHEEL_URL: ${{ needs.build.outputs.wheel_path }}
+      MODEL_CACHE_DIR: "/ssd2/actions-runner/ModelData"
+
+  accuracy_test:
+    name: Run Accuracy Tests
+    needs: [clone,build]
+    uses: ./.github/workflows/_accuracy_test.yml
+    with:
+      DOCKER_IMAGE: ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleqa:fastdeploy-ciuse-cuda126-dailyupdate
+      FASTDEPLOY_ARCHIVE_URL: ${{ needs.clone.outputs.repo_archive_url }}
+      FASTDEPLOY_WHEEL_URL: ${{ needs.build.outputs.wheel_path }}
+      MODEL_CACHE_DIR: "/ssd2/actions-runner/ModelData"
+
+  stable_test:
+    name: Run Stable Tests
+    needs: [clone,build]
+    uses: ./.github/workflows/_stable_test.yml
+    with:
+      DOCKER_IMAGE: ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleqa:fastdeploy-ciuse-cuda126-dailyupdate
+      FASTDEPLOY_ARCHIVE_URL: ${{ needs.clone.outputs.repo_archive_url }}
+      FASTDEPLOY_WHEEL_URL: ${{ needs.build.outputs.wheel_path }}
+      MODEL_CACHE_DIR: "/ssd2/actions-runner/ModelData"

.github/workflows/publish_job.yml

@@ -0,0 +1,321 @@
+name: Publish Job
+
+on:
+  workflow_dispatch:
+  schedule:
+    - cron: '0 18 * * *'   # 2:00 AM China Standard Time (UTC+8)
+  push:
+    # branches:
+    #   - develop
+    tags:
+      - '*'
+
+permissions: read-all
+
+concurrency:
+  group: ${{ github.ref }}-${{ github.sha }}
+  cancel-in-progress: true
+
+
+jobs:
+  publish_pre_check:
+    runs-on: ubuntu-latest
+    if: |
+      github.event.repository.fork == false &&
+      (
+        (github.event_name == 'schedule' && github.ref_name == 'develop') ||
+        (github.event_name == 'push' && github.ref_type == 'tag') ||
+        ((github.event_name == 'workflow_dispatch') &&
+          (github.ref_name == 'develop' || github.ref_type == 'tag'))
+      )
+    env:
+      TAG_VERSION_MAPPINGS: ${{ vars.TAG_VERSION_MAPPINGS }}
+      FD_VERSION_DEV: ${{ vars.FD_VERSION_DEV }}
+      COMPILE_USE_PADDLE_WHL_URL_MAPPINGS: ${{ vars.COMPILE_USE_PADDLE_WHL_URL_MAPPINGS }}
+    outputs:
+      compile_use_paddle_version: ${{ steps.set_output.outputs.compile_use_paddle_version }}
+      compile_continue: ${{ steps.set_output.outputs.compile_continue }}
+      fd_version: ${{ steps.set_output.outputs.fd_version }}
+      with_nightly_build: ${{ steps.set_output.outputs.with_nightly_build }}
+      compile_use_paddle_whl_url: ${{ steps.set_output.outputs.compile_use_paddle_whl_url }}
+
+    steps:
+      - name: Get tag version
+        if: github.ref_type == 'tag'
+        run: |
+          TAG_NAME="${GITHUB_REF##*/}"       # 提取 tag 名称，比如 v2.1.0
+          TAG_VERSION="${TAG_NAME#v}"            # 去掉前缀 v
+          echo "FD_VERSION=$TAG_VERSION" >> $GITHUB_ENV
+
+      - name: Check FD version to Paddle version mapping
+        if: github.ref_type == 'tag'
+        env:
+          TARGET_FD: ${{ env.FD_VERSION }}
+        run: |
+          FOUND_PADDLE=""
+          # 遍历映射
+          for pair in $(echo $TAG_VERSION_MAPPINGS | tr ';' ' '); do
+            fd=$(echo "$pair" | cut -d',' -f1)
+            paddle=$(echo "$pair" | cut -d',' -f2)
+            if [[ "$fd" == "$TARGET_FD" ]]; then
+              FOUND_PADDLE="$paddle"
+              break
+            fi
+          done
+
+          if [[ -z "$FOUND_PADDLE" ]]; then
+            echo "No Paddle version found for FD $TARGET_FD"
+          else
+            echo "FD $TARGET_FD maps to Paddle $FOUND_PADDLE"
+            echo "PADDLE_VERSION=$FOUND_PADDLE" >> $GITHUB_ENV
+          fi
+      - name: Set Version
+        id: set_output
+        env:
+          PADDLE_VERSION: ${{ env.PADDLE_VERSION }}
+          FD_VERSION: ${{ env.FD_VERSION }}
+        run: |
+          if [[ "${{ github.ref_type }}" == "tag" ]]; then
+            if [[ -z "$PADDLE_VERSION" ]]; then
+              compile_continue=false
+            else
+              compile_use_paddle_version=$PADDLE_VERSION
+              compile_continue=true
+            fi
+            fd_version=$FD_VERSION
+          fi
+          if [[ "${{ github.ref_name }}" == "develop" ]];then
+            compile_continue=true
+            compile_use_paddle_version=""
+            fd_version=${FD_VERSION_DEV}
+            with_nightly_build=ON
+          fi
+          # Todo
+          # 通过变量COMPILE_USE_PADDLE_WHL_URL_MAPPINGS中的映射关系,决定是否是安装指定版本的Paddle还是直接安装URL
+          for pair in $(echo $COMPILE_USE_PADDLE_WHL_URL_MAPPINGS | tr ';' ' '); do
+            branch=$(echo "$pair" | cut -d',' -f1)
+            paddle_whl_url=$(echo "$pair" | cut -d',' -f2)
+            if [[ "$branch" == "${{ github.ref_name }}" ]]; then
+              FOUND_PADDLE_URL="$paddle_whl_url"
+              echo "compile_use_paddle_whl_url=${FOUND_PADDLE_URL}" >> $GITHUB_OUTPUT
+              compile_continue=true
+              break
+            fi
+          done
+          echo "compile_continue=${compile_continue}" >> $GITHUB_OUTPUT
+          echo "compile_use_paddle_version=${compile_use_paddle_version}" >> $GITHUB_OUTPUT
+          echo "fd_version=${fd_version}" >> $GITHUB_OUTPUT
+          echo "with_nightly_build=${with_nightly_build:-OFF}" >> $GITHUB_OUTPUT
+
+  print_publish_pre_check_outputs:
+    runs-on: ubuntu-latest
+    needs: publish_pre_check
+    steps:
+      - name: Print outputs as JSON
+        run: |
+          echo '${{ toJSON(needs.publish_pre_check.outputs) }}'
+
+  clone:
+    environment: CodeSync
+    name: FD-Clone-Linux
+    runs-on: ubuntu-latest
+    needs: publish_pre_check
+    if: ${{ needs.publish_pre_check.outputs.compile_continue == 'true' }}
+    outputs:
+      repo_archive_url: ${{ steps.set_output.outputs.repo_archive_url }}
+    steps:
+      - name: Clone FastDeploy
+        uses: actions/checkout@v4
+        with:
+          ref: ${{ github.ref_name }}
+          submodules: 'recursive'
+          fetch-depth: 1000
+
+      - name: Python Setup
+        uses: actions/setup-python@v5
+        with:
+          python-version: '3.10'
+      - name: Code Info Show and Upload
+        id: set_output
+        env:
+          AK: ${{ secrets.BOS_AK }}
+          SK: ${{ secrets.BOS_SK }}
+        run: |
+          git config --unset http.https://github.com/.extraheader
+          git submodule foreach --recursive sh -c "git config --local --unset-all 'http.https://github.com/.extraheader'"
+          git submodule foreach --recursive sh -c "git config remote.origin.fetch '+refs/heads/*:refs/remotes/origin/*'"
+          echo "Current HEAD Log:"
+          git log --oneline -n 5
+          ls
+          cd ..
+          tar -zcf FastDeploy.tar.gz FastDeploy
+          if [[ "${{ github.ref_type }}" == "tag" ]]; then
+            commit_id=${{ github.sha }}
+            tag_name=${{ github.ref_name }}
+            target_path=paddle-qa/TAG/FastDeploy/${tag_name}/${commit_id}
+          else
+            commit_id=${{ github.sha }}
+            branch_name=${{ github.ref_name }}
+            target_path=paddle-qa/BRANCH/FastDeploy/${branch_name}/${commit_id}
+          fi
+          wget  -q --no-proxy --no-check-certificate https://paddle-qa.bj.bcebos.com/CodeSync/develop/PaddlePaddle/PaddleTest/tools/bos_tools.py
+          push_file=$(realpath bos_tools.py)
+          python -m pip install bce-python-sdk==0.9.29
+          ls
+          python ${push_file} FastDeploy.tar.gz ${target_path}
+          target_path_stripped="${target_path#paddle-qa/}"
+          REPO_ARCHIVE_URL=https://paddle-qa.bj.bcebos.com/${target_path_stripped}/FastDeploy.tar.gz
+          echo "repo_archive_url=${REPO_ARCHIVE_URL}" >> $GITHUB_OUTPUT
+
+  resultshow:
+    name: Show Code Archive Output
+    needs: clone
+    runs-on: ubuntu-latest
+    steps:
+      - name: Print wheel path
+        run: |
+          echo "The code archive is located at: ${{ needs.clone.outputs.repo_archive_url }}"
+
+  build_sm8090:
+    name: BUILD_SM8090
+    needs: [clone, publish_pre_check]
+    uses: ./.github/workflows/_build_linux.yml
+    with:
+      DOCKER_IMAGE: ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleqa:fastdeploy-ciuse-cuda126-dailyupdate
+      FASTDEPLOY_ARCHIVE_URL: ${{ needs.clone.outputs.repo_archive_url }}
+      COMPILE_ARCH: "80,90"
+      WITH_NIGHTLY_BUILD: ${{ needs.publish_pre_check.outputs.with_nightly_build }}
+      FD_VERSION: ${{ needs.publish_pre_check.outputs.fd_version }}
+      PADDLEVERSION: ${{ needs.publish_pre_check.outputs.compile_use_paddle_version }}
+      PADDLE_WHL_URL: ${{ needs.publish_pre_check.outputs.compile_use_paddle_whl_url }}
+
+  build_sm8689:
+    name: BUILD_SM8689
+    needs: [clone, publish_pre_check]
+    uses: ./.github/workflows/_build_linux.yml
+    with:
+      DOCKER_IMAGE: ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleqa:fastdeploy-ciuse-cuda126-dailyupdate
+      FASTDEPLOY_ARCHIVE_URL: ${{ needs.clone.outputs.repo_archive_url }}
+      COMPILE_ARCH: "86,89"
+      WITH_NIGHTLY_BUILD: ${{ needs.publish_pre_check.outputs.with_nightly_build }}
+      FD_VERSION: ${{ needs.publish_pre_check.outputs.fd_version }}
+      PADDLEVERSION: ${{ needs.publish_pre_check.outputs.compile_use_paddle_version }}
+      PADDLE_WHL_URL: ${{ needs.publish_pre_check.outputs.compile_use_paddle_whl_url }}
+
+  paddle_pypi_upload_sm8090:
+    environment: PaddleSourceUpload
+    name: PADDLE_PYPI_UPLOAD_8090
+    needs: build_sm8090
+    runs-on: ubuntu-latest
+    env:
+      AK: ${{ secrets.BOS_AK }}
+      SK: ${{ secrets.BOS_SK }}
+      FASTDEPLOY_WHEEL_URL: ${{ needs.build_sm8090.outputs.wheel_path }}
+      COMPILE_ARCH: "80,90"
+    steps:
+      - uses: actions/setup-python@v5
+        with:
+          python-version: '3.10'
+      - name: Wheel Info Show and Upload
+        if: github.ref_name == 'develop' || github.ref_type == 'tag'
+        run: |
+          echo "The wheel is located at: ${FASTDEPLOY_WHEEL_URL}"
+          wget -q --no-check-certificate ${FASTDEPLOY_WHEEL_URL}
+          filename=$(basename ${FASTDEPLOY_WHEEL_URL})
+          if [[ "${{ github.ref_name }}" == "develop" ]];then
+            target_path=paddle-whl/nightly/fastdeploy-gpu-${COMPILE_ARCH//,/_}/fastdeploy-gpu
+          elif [[ "${{ github.ref_type }}" == "tag" ]]; then
+            target_path=paddle-whl/stable/fastdeploy-gpu-${COMPILE_ARCH//,/_}/fastdeploy-gpu
+          else
+            echo "Not develop or tag, do nothing"
+          fi
+          wget -q --no-proxy --no-check-certificate https://paddle-qa.bj.bcebos.com/CodeSync/develop/PaddlePaddle/PaddleTest/tools/bos_tools.py
+          push_file=$(realpath bos_tools.py)
+          python -m pip install bce-python-sdk==0.9.29
+          ls
+          python ${push_file} ${filename} ${target_path}
+
+  paddle_pypi_upload_sm8689:
+    environment: PaddleSourceUpload
+    name: PADDLE_PYPI_UPLOAD_8689
+    needs: build_sm8689
+    runs-on: ubuntu-latest
+    env:
+      AK: ${{ secrets.BOS_AK }}
+      SK: ${{ secrets.BOS_SK }}
+      FASTDEPLOY_WHEEL_URL: ${{ needs.build_sm8689.outputs.wheel_path }}
+      COMPILE_ARCH: "86,89"
+    steps:
+      - uses: actions/setup-python@v5
+        with:
+          python-version: '3.10'
+      - name: Wheel Info Show and Upload
+        if: github.ref_name == 'develop' || github.ref_type == 'tag'
+        run: |
+          echo "The wheel is located at: ${FASTDEPLOY_WHEEL_URL}"
+          wget -q --no-check-certificate ${FASTDEPLOY_WHEEL_URL}
+          filename=$(basename ${FASTDEPLOY_WHEEL_URL})
+          if [[ "${{ github.ref_name }}" == "develop" ]];then
+            target_path=paddle-whl/nightly/fastdeploy-gpu-${COMPILE_ARCH//,/_}/fastdeploy-gpu
+          elif [[ "${{ github.ref_type }}" == "tag" ]]; then
+            target_path=paddle-whl/stable/fastdeploy-gpu-${COMPILE_ARCH//,/_}/fastdeploy-gpu
+          else
+            echo "Not develop or tag, do nothing"
+          fi
+          wget -q --no-proxy --no-check-certificate https://paddle-qa.bj.bcebos.com/CodeSync/develop/PaddlePaddle/PaddleTest/tools/bos_tools.py
+          push_file=$(realpath bos_tools.py)
+          python -m pip install bce-python-sdk==0.9.29
+          ls
+          python ${push_file} ${filename} ${target_path}
+
+  unittest_coverage:
+    name: Run FastDeploy Unit Tests and Coverage
+    needs: [clone,build_sm8090]
+    uses: ./.github/workflows/_unit_test_coverage.yml
+    with:
+      DOCKER_IMAGE: ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleqa:fastdeploy-ciuse-cuda126-dailyupdate
+      FASTDEPLOY_ARCHIVE_URL: ${{ needs.clone.outputs.repo_archive_url }}
+      FASTDEPLOY_WHEEL_URL: ${{ needs.build_sm8090.outputs.wheel_path }}
+      MODEL_CACHE_DIR: "/ssd2/actions-runner/ModelData"
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+
+  logprob_test:
+    name: Run FastDeploy LogProb Tests
+    needs: [build_sm8090]
+    uses: ./.github/workflows/_logprob_test_linux.yml
+    with:
+      DOCKER_IMAGE: ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleqa:fastdeploy-ciuse-cuda126-dailyupdate
+      PADDLETEST_ARCHIVE_URL: "https://xly-devops.bj.bcebos.com/PaddleTest/PaddleTest.tar.gz"
+      FASTDEPLOY_WHEEL_URL: ${{ needs.build_sm8090.outputs.wheel_path }}
+      MODEL_CACHE_DIR: "/ssd2/actions-runner/ModelData"
+
+  pre_ce_test:
+    name: Extracted partial CE model tasks to run in CI.
+    needs: [clone,build_sm8090]
+    uses: ./.github/workflows/_pre_ce_test.yml
+    with:
+      DOCKER_IMAGE: ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleqa:fastdeploy-ciuse-cuda126-dailyupdate
+      FASTDEPLOY_ARCHIVE_URL: ${{ needs.clone.outputs.repo_archive_url }}
+      FASTDEPLOY_WHEEL_URL: ${{ needs.build_sm8090.outputs.wheel_path }}
+      MODEL_CACHE_DIR: "/ssd2/actions-runner/ModelData"
+
+  base_test:
+    name: Run Base Tests
+    needs: [clone,build_sm8090]
+    uses: ./.github/workflows/_base_test.yml
+    with:
+      DOCKER_IMAGE: ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleqa:fastdeploy-ciuse-cuda126-dailyupdate
+      FASTDEPLOY_ARCHIVE_URL: ${{ needs.clone.outputs.repo_archive_url }}
+      FASTDEPLOY_WHEEL_URL: ${{ needs.build_sm8090.outputs.wheel_path }}
+      MODEL_CACHE_DIR: "/ssd2/actions-runner/ModelData"
+
+  accuracy_test:
+    name: Run Accuracy Tests
+    needs: [clone,build_sm8090]
+    uses: ./.github/workflows/_accuracy_test.yml
+    with:
+      DOCKER_IMAGE: ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleqa:fastdeploy-ciuse-cuda126-dailyupdate
+      FASTDEPLOY_ARCHIVE_URL: ${{ needs.clone.outputs.repo_archive_url }}
+      FASTDEPLOY_WHEEL_URL: ${{ needs.build_sm8090.outputs.wheel_path }}
+      MODEL_CACHE_DIR: "/ssd2/actions-runner/ModelData"

.gitignore

@@ -121,7 +121,7 @@ dmypy.json
 FETCH_HEAD
 
 #log
-log*/
+log/
 
 checkpoints/
 checkpoints_origin/
@@ -156,6 +156,12 @@ nohup.out
 custom_ops/gpu_ops/fp8_deep_gemm/deep_gemm/include/cutlass
 custom_ops/gpu_ops/fp8_deep_gemm/deep_gemm/include/cute
 
+#marlin_kernel
+custom_ops/gpu_ops/moe/moe_wna16_marlin_utils/kernel_*.cu
+
+#machete_kernel
+custom_ops/gpu_ops/machete/generated
+
 # buff
 custom_ops/tmp*
 
@@ -164,3 +170,9 @@ build
 .ccls-cache
 
 third_party
+
+custom_ops/gpu_ops/w4afp8_gemm/w4afp8_gemm_*.cu
+custom_ops/gpu_ops/w4afp8_gemm/w4afp8_gemm_template.h
+
+custom_ops/gpu_ops/wfp8afp8_sparse_gemm/wfp8Afp8_sparse_gemm_*.cu
+custom_ops/gpu_ops/wfp8afp8_sparse_gemm/wfp8Afp8_sparse_gemm_template.h

.gitmodules

@@ -0,0 +1,9 @@
+[submodule "custom_ops/third_party/DeepGEMM"]
+	path = custom_ops/third_party/DeepGEMM
+	url = https://github.com/deepseek-ai/DeepGEMM.git
+[submodule "custom_ops/third_party/cutlass"]
+	path = custom_ops/third_party/cutlass
+	url = https://github.com/NVIDIA/cutlass.git
+[submodule "custom_ops/third_party/nlohmann_json"]
+	path = custom_ops/third_party/nlohmann_json
+	url = https://github.com/nlohmann/json.git

README.md

@@ -1,3 +1,4 @@
+English | [简体中文](README_CN.md)
 <p align="center">
   <a href="https://github.com/PaddlePaddle/FastDeploy/releases"><img src="https://github.com/user-attachments/assets/42b0039f-39e3-4279-afda-6d1865dfbffb" width="500"></a>
 </p>
@@ -22,11 +23,12 @@
 </p>
 
 --------------------------------------------------------------------------------
-# FastDeploy 2.0: Inference and Deployment Toolkit for LLMs and VLMs based on PaddlePaddle
+# FastDeploy : Inference and Deployment Toolkit for LLMs and VLMs based on PaddlePaddle
 
 ## News
+**[2025-09] 🔥 FastDeploy v2.2 is newly released!** It now offers compatibility with models in the HuggingFace ecosystem, has further optimized performance, and newly adds support for [baidu/ERNIE-21B-A3B-Thinking](https://huggingface.co/baidu/ERNIE-4.5-21B-A3B-Thinking)!
 
-**[2025-07] 《FastDeploy2.0推理部署实测》专题活动已上线!** 完成文心4.5系列开源模型的推理部署等任务，即可获得骨瓷马克杯等FastDeploy2.0官方周边及丰富奖金！🎁 欢迎大家体验反馈～ 📌[报名地址](https://www.wjx.top/vm/meSsp3L.aspx#)   📌[活动详情](https://github.com/PaddlePaddle/FastDeploy/discussions/2728)
+**[2025-08] 🔥 Released FastDeploy v2.1:** A brand-new KV Cache scheduling strategy has been introduced, and expanded support for PD separation and CUDA Graph across more models. Enhanced hardware support has been added for platforms like Kunlun and Hygon, along with comprehensive optimizations to improve the performance of both the service and inference engine.
 
 **[2025-07] The FastDeploy 2.0 Inference Deployment Challenge is now live!** Complete the inference deployment task for the ERNIE 4.5 series open-source models to win official FastDeploy 2.0 merch and generous prizes! 🎁 You're welcome to try it out and share your feedback! 📌[Sign up here](https://www.wjx.top/vm/meSsp3L.aspx#) 📌[Event details](https://github.com/PaddlePaddle/FastDeploy/discussions/2728)
 
@@ -50,14 +52,16 @@
 
 ## Installation
 
-FastDeploy supports inference deployment on **NVIDIA GPUs**, **Kunlunxin XPUs**, **Iluvatar GPUs**, **Enflame GCUs**, and other hardware. For detailed installation instructions:
+FastDeploy supports inference deployment on **NVIDIA GPUs**, **Kunlunxin XPUs**, **Iluvatar GPUs**, **Enflame GCUs**, **Hygon DCUs** and other hardware. For detailed installation instructions:
 
 - [NVIDIA GPU](./docs/get_started/installation/nvidia_gpu.md)
 - [Kunlunxin XPU](./docs/get_started/installation/kunlunxin_xpu.md)
 - [Iluvatar GPU](./docs/get_started/installation/iluvatar_gpu.md)
 - [Enflame GCU](./docs/get_started/installation/Enflame_gcu.md)
+- [Hygon DCU](./docs/get_started/installation/hygon_dcu.md)
+- [MetaX GPU](./docs/get_started/installation/metax_gpu.md.md)
 
-**Note:** We are actively working on expanding hardware support. Additional hardware platforms including Ascend NPU, Hygon DCU, and MetaX GPU are currently under development and testing. Stay tuned for updates!
+**Note:** We are actively working on expanding hardware support. Additional hardware platforms including Ascend NPU are currently under development and testing. Stay tuned for updates!
 
 ## Get Started
 
@@ -67,19 +71,12 @@ Learn how to use FastDeploy through our documentation:
 - [ERNIE-4.5-VL Multimodal Model Deployment](./docs/get_started/ernie-4.5-vl.md)
 - [Offline Inference Development](./docs/offline_inference.md)
 - [Online Service Deployment](./docs/online_serving/README.md)
-- [Full Supported Models List](./docs/supported_models.md)
+- [Best Practices](./docs/best_practices/README.md)
 
 ## Supported Models
 
-| Model | Data Type | PD Disaggregation | Chunked Prefill | Prefix Caching |  MTP | CUDA Graph | Maximum Context Length |
-|:--- | :------- | :---------- | :-------- | :-------- | :----- | :----- | :----- |
-|ERNIE-4.5-300B-A47B | BF16/WINT4/WINT8/W4A8C8/WINT2/FP8 | ✅| ✅ | ✅|✅(WINT4)| WIP |128K |
-|ERNIE-4.5-300B-A47B-Base| BF16/WINT4/WINT8 | ✅| ✅ | ✅|✅(WINT4)| WIP | 128K |
-|ERNIE-4.5-VL-424B-A47B | BF16/WINT4/WINT8 | WIP | ✅ | WIP | ❌ | WIP |128K |
-|ERNIE-4.5-VL-28B-A3B | BF16/WINT4/WINT8 | ❌ | ✅ | WIP | ❌ | WIP |128K |
-|ERNIE-4.5-21B-A3B | BF16/WINT4/WINT8/FP8  |  ❌ |  ✅ |  ✅ | WIP | ✅|128K |
-|ERNIE-4.5-21B-A3B-Base | BF16/WINT4/WINT8/FP8  |  ❌ |  ✅ |  ✅ | WIP | ✅|128K |
-|ERNIE-4.5-0.3B | BF16/WINT8/FP8  |  ❌ |  ✅ |  ✅ | ❌ | ✅| 128K |
+Learn how to download models, enable using the torch format, and more:
+- [Full Supported Models List](./docs/supported_models.md)
 
 ## Advanced Usage
 

README_CN.md

@@ -0,0 +1,89 @@
+[English](README.md) | 简体中文
+<p align="center">
+  <a href="https://github.com/PaddlePaddle/FastDeploy/releases"><img src="https://github.com/user-attachments/assets/42b0039f-39e3-4279-afda-6d1865dfbffb" width="500"></a>
+</p>
+<p align="center">
+    <a href=""><img src="https://img.shields.io/badge/python-3.10-aff.svg"></a>
+    <a href=""><img src="https://img.shields.io/badge/os-linux-pink.svg"></a>
+    <a href="https://github.com/PaddlePaddle/FastDeploy/graphs/contributors"><img src="https://img.shields.io/github/contributors/PaddlePaddle/FastDeploy?color=9ea"></a>
+    <a href="https://github.com/PaddlePaddle/FastDeploy/commits"><img src="https://img.shields.io/github/commit-activity/m/PaddlePaddle/FastDeploy?color=3af"></a>
+    <a href="https://github.com/PaddlePaddle/FastDeploy/issues"><img src="https://img.shields.io/github/issues/PaddlePaddle/FastDeploy?color=9cc"></a>
+    <a href="https://github.com/PaddlePaddle/FastDeploy/stargazers"><img src="https://img.shields.io/github/stars/PaddlePaddle/FastDeploy?color=ccf"></a>
+
+</p>
+
+<p align="center">
+     <a href="https://trendshift.io/repositories/4046" target="_blank"><img src="https://trendshift.io/api/badge/repositories/4046" alt="PaddlePaddle%2FFastDeploy | Trendshift" style="width: 250px; height: 55px;" width="250" height="55"/></a></br>
+    <a href="https://paddlepaddle.github.io/FastDeploy/zh/get_started/installation/nvidia_gpu/"><b> 安装指导 </b></a>
+    |
+    <a href="https://paddlepaddle.github.io/FastDeploy/zh/get_started/quick_start"><b> 快速入门 </b></a>
+    |
+    <a href="https://paddlepaddle.github.io/FastDeploy/zh/supported_models/"><b> 支持模型列表 </b></a>
+
+</p>
+
+--------------------------------------------------------------------------------
+# FastDeploy ：基于飞桨的大语言模型与视觉语言模型推理部署工具包
+
+## 最新活动
+**[2025-09] 🔥 FastDeploy v2.2 全新发布**: HuggingFace生态模型兼容，性能进一步优化，更新增对[baidu/ERNIE-21B-A3B-Thinking](https://huggingface.co/baidu/ERNIE-4.5-21B-A3B-Thinking)支持!
+
+**[2025-08] FastDeploy v2.1 发布**:全新的KV Cache调度策略，更多模型支持PD分离和CUDA Graph，昆仑、海光等更多硬件支持增强，全方面优化服务和推理引擎的性能。
+
+**[2025-07] 《FastDeploy2.0推理部署实测》专题活动已上线!** 完成文心4.5系列开源模型的推理部署等任务，即可获得骨瓷马克杯等FastDeploy2.0官方周边及丰富奖金！🎁 欢迎大家体验反馈～ 📌[报名地址](https://www.wjx.top/vm/meSsp3L.aspx#)   📌[活动详情](https://github.com/PaddlePaddle/FastDeploy/discussions/2728)
+
+## 关于
+
+**FastDeploy** 是基于飞桨（PaddlePaddle）的大语言模型（LLM）与视觉语言模型（VLM）推理部署工具包，提供**开箱即用的生产级部署方案**，核心技术特性包括：
+
+- 🚀 **负载均衡式PD分解**：工业级解决方案，支持上下文缓存与动态实例角色切换，在保障SLO达标和吞吐量的同时优化资源利用率
+- 🔄 **统一KV缓存传输**：轻量级高性能传输库，支持智能NVLink/RDMA选择
+- 🤝 **OpenAI API服务与vLLM兼容**：单命令部署，兼容[vLLM](https://github.com/vllm-project/vllm/)接口
+- 🧮 **全量化格式支持**：W8A16、W8A8、W4A16、W4A8、W2A16、FP8等
+- ⏩ **高级加速技术**：推测解码、多令牌预测（MTP）及分块预填充
+- 🖥️ **多硬件支持**：NVIDIA GPU、昆仑芯XPU、海光DCU、昇腾NPU、天数智芯GPU、燧原GCU、沐曦GPU等
+
+## 要求
+
+- 操作系统: Linux
+- Python: 3.10 ~ 3.12
+
+## 安装
+
+FastDeploy 支持在**英伟达（NVIDIA）GPU**、**昆仑芯（Kunlunxin）XPU**、**天数（Iluvatar）GPU**、**燧原（Enflame）GCU**、**海光（Hygon）DCU** 以及其他硬件上进行推理部署。详细安装说明如下：
+
+- [英伟达 GPU](./docs/zh/get_started/installation/nvidia_gpu.md)
+- [昆仑芯 XPU](./docs/zh/get_started/installation/kunlunxin_xpu.md)
+- [天数 CoreX](./docs/zh/get_started/installation/iluvatar_gpu.md)
+- [燧原 S60](./docs/zh/get_started/installation/Enflame_gcu.md)
+- [海光 DCU](./docs/zh/get_started/installation/hygon_dcu.md)
+- [沐曦 GPU](./docs/zh/get_started/installation/metax_gpu.md.md)
+
+**注意:** 我们正在积极拓展硬件支持范围。目前，包括昇腾（Ascend）NPU 等其他硬件平台正在开发测试中。敬请关注更新！
+
+## 入门指南
+
+通过我们的文档了解如何使用 FastDeploy：
+- [10分钟快速部署](./docs/zh/get_started/quick_start.md)
+- [ERNIE-4.5 部署](./docs/zh/get_started/ernie-4.5.md)
+- [ERNIE-4.5-VL 部署](./docs/zh/get_started/ernie-4.5-vl.md)
+- [离线推理](./docs/zh/offline_inference.md)
+- [在线服务](./docs/zh/online_serving/README.md)
+- [最佳实践](./docs/zh/best_practices/README.md)
+
+## 支持模型列表
+
+通过我们的文档了解如何下载模型，如何支持torch格式等：
+- [模型支持列表](./docs/zh/supported_models.md)
+
+## 进阶用法
+
+- [量化](./docs/zh/quantization/README.md)
+- [分离式部署](./docs/zh/features/disaggregated.md)
+- [投机解码](./docs/zh/features/speculative_decoding.md)
+- [前缀缓存](./docs/zh/features/prefix_caching.md)
+- [分块预填充](./docs/zh/features/chunked_prefill.md)
+
+## 致谢
+
+FastDeploy 依据 [Apache-2.0 开源许可证](./LICENSE). 进行授权。在开发过程中，我们参考并借鉴了 [vLLM](https://github.com/vllm-project/vllm) 的部分代码，以保持接口兼容性，在此表示衷心感谢。

benchmarks/benchmark_serving.py

@@ -361,8 +361,7 @@ async def benchmark(
 
     if not test_output.success:
         raise ValueError(
-            "Initial test run failed - Please make sure benchmark arguments "
-            f"are correctly specified. Error: {test_output.error}"
+            f"Initial test run failed - Please make sure that 1. benchmark arguments are correctly specified and 2. the http_proxy and https_proxy are turned off. Error: {test_output.error}"
         )
     else:
         print("Initial test run completed. Starting main benchmark run...")

benchmarks/yaml/eb45-8k-fp8-tp1-dp8_ep.yaml

@@ -0,0 +1,6 @@
+num_gpu_blocks_override: 1024
+max_model_len: 8192
+max_num_seqs: 64
+data_parallel_size: 8
+tensor_parallel_size: 1
+enable_expert_parallel: True

benchmarks/yaml/request_yaml/x1.yaml

@@ -0,0 +1,8 @@
+top_p: 0.95
+temperature: 0.6
+metadata:
+  min_tokens: 1
+max_tokens: 65535
+repetition_penalty: 1.0
+frequency_penalty: 0
+presence_penalty: 0

benchmarks/yaml/x1-64k-w4a8c8-tp4.yaml

@@ -0,0 +1,10 @@
+reasoning-parser: ernie_x1
+tool_call_parser: ernie_x1
+tensor_parallel_size: 4
+max_model_len: 65536
+max_num_seqs: 128
+enable_prefix_caching: True
+enable_chunked_prefill: True
+gpu_memory_utilization: 0.85
+use_cudagraph: True
+enable_custom_all_reduce: True

build.sh

@@ -34,7 +34,6 @@ EGG_DIR="fastdeploy.egg-info"
 
 # custom_ops directory config
 OPS_SRC_DIR="custom_ops"
-OPS_TMP_DIR_BASE="tmp_base"
 OPS_TMP_DIR="tmp"
 
 # command line log config
@@ -71,25 +70,20 @@ function copy_ops(){
     PY_VERSION="py${PY_MAIN_VERSION}.${PY_SUB_VERSION}"
     SYSTEM_VERSION=`${python} -c "import platform; print(platform.system().lower())"`
     PROCESSOR_VERSION=`${python} -c "import platform; print(platform.processor())"`
-    WHEEL_BASE_NAME="fastdeploy_base_ops-${OPS_VERSION}-${PY_VERSION}-${SYSTEM_VERSION}-${PROCESSOR_VERSION}.egg"
     WHEEL_NAME="fastdeploy_ops-${OPS_VERSION}-${PY_VERSION}-${SYSTEM_VERSION}-${PROCESSOR_VERSION}.egg"
     WHEEL_CPU_NAME="fastdeploy_cpu_ops-${OPS_VERSION}-${PY_VERSION}-${SYSTEM_VERSION}-${PROCESSOR_VERSION}.egg"
     is_rocm=`$python -c "import paddle; print(paddle.is_compiled_with_rocm())"`
     if [ "$is_rocm" = "True" ]; then
       DEVICE_TYPE="rocm"
-      mkdir -p ../fastdeploy/model_executor/ops/base
-      cp -r ./${OPS_TMP_DIR_BASE}/${WHEEL_BASE_NAME}/* ../fastdeploy/model_executor/ops/base
       cp -r ./${OPS_TMP_DIR}/${WHEEL_NAME}/* ../fastdeploy/model_executor/ops/gpu
-      echo -e "BASE and ROCM ops have been copy to fastdeploy"
+      echo -e "ROCM ops have been copy to fastdeploy"
       return
     fi
-    mkdir -p ../fastdeploy/model_executor/ops/base
     is_cuda=`$python -c "import paddle; print(paddle.is_compiled_with_cuda())"`
     if [ "$is_cuda" = "True" ]; then
       DEVICE_TYPE="gpu"
-      cp -r ./${OPS_TMP_DIR_BASE}/${WHEEL_BASE_NAME}/* ../fastdeploy/model_executor/ops/base
       cp -r ./${OPS_TMP_DIR}/${WHEEL_NAME}/* ../fastdeploy/model_executor/ops/gpu
-      echo -e "BASE and CUDA ops have been copy to fastdeploy"
+      echo -e "CUDA ops have been copy to fastdeploy"
       return
     fi
 
@@ -112,9 +106,8 @@ function copy_ops(){
     if_corex=`$python -c "import paddle; print(paddle.is_compiled_with_custom_device(\"iluvatar_gpu\"))"`
     if [ "$if_corex" = "True" ]; then
       DEVICE_TYPE="iluvatar-gpu"
-      cp -r ./${OPS_TMP_DIR_BASE}/${WHEEL_BASE_NAME}/* ../fastdeploy/model_executor/ops/base
       cp -r ./${OPS_TMP_DIR}/${WHEEL_NAME}/* ../fastdeploy/model_executor/ops/iluvatar
-      echo -e "BASE and Iluvatar ops have been copy to fastdeploy"
+      echo -e "Iluvatar ops have been copy to fastdeploy"
       return
     fi
 
@@ -126,20 +119,26 @@ function copy_ops(){
       return
     fi
 
+    is_maca=`$python -c "import paddle; print(paddle.device.is_compiled_with_custom_device('metax_gpu'))"`
+    if [ "$is_maca" = "True" ]; then
+      DEVICE_TYPE="metax_gpu"
+      mkdir -p ../fastdeploy/model_executor/ops/base
+      cp -r ./${OPS_TMP_DIR_BASE}/${WHEEL_BASE_NAME}/* ../fastdeploy/model_executor/ops/base
+      cp -r ./${OPS_TMP_DIR}/${WHEEL_NAME}/* ../fastdeploy/model_executor/ops/gpu
+      echo -e "MACA ops have been copy to fastdeploy"
+      return
+    fi
+
     DEVICE_TYPE="cpu"
-    cp -r ./${OPS_TMP_DIR_BASE}/${WHEEL_BASE_NAME}/* ../fastdeploy/model_executor/ops/base
     cd ../../../../
     cp -r ${OPS_TMP_DIR}/${WHEEL_CPU_NAME}/* ../fastdeploy/model_executor/ops/cpu
-    echo -e "BASE and CPU ops have been copy to fastdeploy"
+    echo -e "CPU ops have been copy to fastdeploy"
     return
 }
 
 function build_and_install_ops() {
   cd $OPS_SRC_DIR
   export no_proxy=bcebos.com,paddlepaddle.org.cn,${no_proxy}
-  echo -e "${BLUE}[build]${NONE} build and install fastdeploy_base_ops..."
-  ${python} setup_ops_base.py install --install-lib ${OPS_TMP_DIR_BASE}
-  find ${OPS_TMP_DIR_BASE} -type f -name "*.o" -exec rm -f {} \;
   echo -e "${BLUE}[build]${NONE} build and install fastdeploy_ops..."
   TMP_DIR_REAL_PATH=`readlink -f ${OPS_TMP_DIR}`
   is_xpu=`$python -c "import paddle; print(paddle.is_compiled_with_xpu())"`
@@ -213,7 +212,6 @@ function cleanup() {
   fi
 
   rm -rf $OPS_SRC_DIR/$BUILD_DIR $OPS_SRC_DIR/$EGG_DIR
-  rm -rf $OPS_SRC_DIR/$OPS_TMP_DIR_BASE
   rm -rf $OPS_SRC_DIR/$OPS_TMP_DIR
 }
 

custom_ops/cpu_ops/get_padding_offset.cc

@@ -84,7 +84,6 @@ std::vector<paddle::Tensor> GetPaddingOffset(const paddle::Tensor &input_ids,
                    seq_length,
                    bsz);
     return {x_remove_padding,
-            cum_offsets_out,
             padding_offset,
             cu_seqlens_q,
             cu_seqlens_k};
@@ -97,7 +96,7 @@ std::vector<std::vector<int64_t>> GetPaddingOffsetInferShape(
     const std::vector<int64_t> &seq_len_shape) {
     int64_t bsz = seq_len_shape[0];
     int64_t seq_len = input_ids_shape[1];
-    return {{-1}, {bsz}, {-1}, {bsz + 1}, {bsz + 1}};
+    return {{-1}, {-1}, {bsz + 1}, {bsz + 1}};
 }
 
 std::vector<paddle::DataType> GetPaddingOffsetInferDtype(
@@ -106,7 +105,6 @@ std::vector<paddle::DataType> GetPaddingOffsetInferDtype(
     const paddle::DataType &token_num_dtype,
     const paddle::DataType &seq_len_dtype) {
     return {input_ids_dtype,
-            seq_len_dtype,
             seq_len_dtype,
             seq_len_dtype,
             seq_len_dtype};
@@ -115,7 +113,6 @@ std::vector<paddle::DataType> GetPaddingOffsetInferDtype(
 PD_BUILD_STATIC_OP(get_padding_offset_cpu)
     .Inputs({"input_ids", "cum_offsets", "token_num", "seq_len"})
     .Outputs({"x_remove_padding",
-              "cum_offsets_out",
               "padding_offset",
               "cu_seqlens_q",
               "cu_seqlens_k"})

custom_ops/cpu_ops/rebuild_padding.cc

@@ -1,4 +1,4 @@
-// Copyright (c) 2024 PaddlePaddle Authors. All Rights Reserved.
+// Copyright (c) 2025 PaddlePaddle Authors. All Rights Reserved.
 //
 // Licensed under the Apache License, Version 2.0 (the "License");
 // you may not use this file except in compliance with the License.
@@ -19,10 +19,11 @@
 #define PD_BUILD_STATIC_OP(name) PD_BUILD_OP(static_op_##name)
 #endif
 
+
 template <typename T>
 void RebuildPaddingCPUImpl(T *output_data,
                            const T *input_data,
-                           const int *cum_offsets_data,
+                           const int *cu_seqlens_q_data,
                            const int *seq_len_this_time_data,
                            const int *seq_lens_decoder_data,
                            const int *seq_lens_encoder_data,
@@ -40,11 +41,12 @@ void RebuildPaddingCPUImpl(T *output_data,
         if (seq_lens_decoder_data[bi] == 0 && seq_lens_encoder_data[bi] == 0) {
             continue;
         }
+
         if (seq_lens_encoder_data[bi] > 0) {
             seq_id = seq_lens_encoder_data[bi] - 1;
         }
-        const int ori_token_idx =
-            bi * max_input_length - cum_offsets_data[bi] + seq_id;
+
+        const int ori_token_idx = cu_seqlens_q_data[bi] + seq_id;
         const int src_offset = ori_token_idx * dim_embed + bias_idx;
 
         output_data[i] = input_data[src_offset];
@@ -54,7 +56,7 @@ void RebuildPaddingCPUImpl(T *output_data,
 template <typename T>
 void RebuildAppendPaddingCPUImpl(T *output_data,
                                  const T *input_data,
-                                 const int *cum_offsets_data,
+                                 const int *cu_seqlens_q_data,
                                  const int *seq_len_this_time_data,
                                  const int *seq_lens_decoder_data,
                                  const int *seq_lens_encoder_data,
@@ -69,30 +71,32 @@ void RebuildAppendPaddingCPUImpl(T *output_data,
         int bi = ori_token_id / max_input_length;
         if (seq_len_this_time_data[bi] == 0 ||
             (seq_lens_decoder_data[bi] == 0 &&
-             seq_lens_encoder_data[bi] == 0)) {
-            continue;
-        }
+            seq_lens_encoder_data[bi] == 0)) {
+                continue;
+            }
         int seq_id = 0;
+
         if (seq_lens_encoder_data[bi] > 0) {
             seq_id = seq_lens_encoder_data[bi] - 1;
         }
-        int input_token_id = ori_token_id - cum_offsets_data[bi] + seq_id;
+        int input_token_id = cu_seqlens_q_data[bi] + seq_id;
         int bias_idx = i % dim_embed;
         int src_offset = input_token_id * dim_embed + bias_idx;
+
         output_data[i] = input_data[src_offset];
     }
 }
 
 std::vector<paddle::Tensor> RebuildPaddingCPU(
     const paddle::Tensor &tmp_out,
-    const paddle::Tensor &cum_offsets,
+    const paddle::Tensor &cu_seqlens_q,
     const paddle::Tensor &seq_len_this_time,
     const paddle::Tensor &seq_lens_decoder,
     const paddle::Tensor &seq_lens_encoder,
     const paddle::optional<paddle::Tensor> &output_padding_offset,
     int max_input_length) {
     auto tmp_out_cpu = tmp_out.copy_to(paddle::CPUPlace(), true);
-    auto cum_offsets_cpu = cum_offsets.copy_to(paddle::CPUPlace(), true);
+    auto cu_seqlens_q_cpu = cu_seqlens_q.copy_to(paddle::CPUPlace(), true);
     auto seq_len_this_time_cpu =
         seq_len_this_time.copy_to(paddle::CPUPlace(), true);
     auto seq_lens_decoder_cpu =
@@ -107,7 +111,7 @@ std::vector<paddle::Tensor> RebuildPaddingCPU(
 
     int token_num = tmp_out_cpu.shape()[0];
     int dim_embed = tmp_out_cpu.shape()[1];
-    int bsz = cum_offsets_cpu.shape()[0];
+    int bsz = cu_seqlens_q_cpu.shape()[0] - 1;
 
     paddle::Tensor out;
     if (output_padding_offset_cpu) {
@@ -128,7 +132,7 @@ std::vector<paddle::Tensor> RebuildPaddingCPU(
             {bsz, dim_embed}, 0, tmp_out_cpu.dtype(), paddle::CPUPlace());
     }
 
-    const int *cum_offsets_data = cum_offsets_cpu.data<int>();
+    const int *cu_seqlens_q_data = cu_seqlens_q_cpu.data<int>();
     const int *seq_len_this_time_data = seq_len_this_time_cpu.data<int>();
     const int *seq_lens_decoder_data = seq_lens_decoder_cpu.data<int>();
     const int *seq_lens_encoder_data = seq_lens_encoder_cpu.data<int>();
@@ -141,7 +145,7 @@ std::vector<paddle::Tensor> RebuildPaddingCPU(
             case paddle::DataType::FLOAT32:
                 RebuildAppendPaddingCPUImpl<float>(out.data<float>(),
                                                    tmp_out_cpu.data<float>(),
-                                                   cum_offsets_data,
+                                                   cu_seqlens_q_data,
                                                    seq_len_this_time_data,
                                                    seq_lens_decoder_data,
                                                    seq_lens_encoder_data,
@@ -154,7 +158,7 @@ std::vector<paddle::Tensor> RebuildPaddingCPU(
                 RebuildAppendPaddingCPUImpl<paddle::float16>(
                     out.data<paddle::float16>(),
                     tmp_out_cpu.data<paddle::float16>(),
-                    cum_offsets_data,
+                    cu_seqlens_q_data,
                     seq_len_this_time_data,
                     seq_lens_decoder_data,
                     seq_lens_encoder_data,
@@ -167,7 +171,7 @@ std::vector<paddle::Tensor> RebuildPaddingCPU(
                 RebuildAppendPaddingCPUImpl<paddle::bfloat16>(
                     out.data<paddle::bfloat16>(),
                     tmp_out_cpu.data<paddle::bfloat16>(),
-                    cum_offsets_data,
+                    cu_seqlens_q_data,
                     seq_len_this_time_data,
                     seq_lens_decoder_data,
                     seq_lens_encoder_data,
@@ -186,7 +190,7 @@ std::vector<paddle::Tensor> RebuildPaddingCPU(
             case paddle::DataType::FLOAT32:
                 RebuildPaddingCPUImpl<float>(out.data<float>(),
                                              tmp_out_cpu.data<float>(),
-                                             cum_offsets_data,
+                                             cu_seqlens_q_data,
                                              seq_len_this_time_data,
                                              seq_lens_decoder_data,
                                              seq_lens_encoder_data,
@@ -198,7 +202,7 @@ std::vector<paddle::Tensor> RebuildPaddingCPU(
                 RebuildPaddingCPUImpl<paddle::float16>(
                     out.data<paddle::float16>(),
                     tmp_out_cpu.data<paddle::float16>(),
-                    cum_offsets_data,
+                    cu_seqlens_q_data,
                     seq_len_this_time_data,
                     seq_lens_decoder_data,
                     seq_lens_encoder_data,
@@ -207,11 +211,10 @@ std::vector<paddle::Tensor> RebuildPaddingCPU(
                     elem_nums);
                 break;
             case paddle::DataType::BFLOAT16:
-
                 RebuildPaddingCPUImpl<paddle::bfloat16>(
                     out.data<paddle::bfloat16>(),
                     tmp_out_cpu.data<paddle::bfloat16>(),
-                    cum_offsets_data,
+                    cu_seqlens_q_data,
                     seq_len_this_time_data,
                     seq_lens_decoder_data,
                     seq_lens_encoder_data,
@@ -230,7 +233,7 @@ std::vector<paddle::Tensor> RebuildPaddingCPU(
 
 std::vector<std::vector<int64_t>> RebuildPaddingInferShape(
     const std::vector<int64_t> &tmp_out_shape,
-    const std::vector<int64_t> &cum_offsets_shape,
+    const std::vector<int64_t> &cu_seqlens_q_shape,
     const std::vector<int64_t> &seq_len_this_time_shape,
     const std::vector<int64_t> &seq_lens_decoder_shape,
     const std::vector<int64_t> &seq_lens_encoder_shape,
@@ -239,14 +242,14 @@ std::vector<std::vector<int64_t>> RebuildPaddingInferShape(
     if (output_padding_offset_shape) {
         return {{-1, dim_embed}};
     } else {
-        int64_t bsz = cum_offsets_shape[0];
+        int64_t bsz = cu_seqlens_q_shape[0] - 1;
         return {{bsz, dim_embed}};
     }
 }
 
 std::vector<paddle::DataType> RebuildPaddingInferDtype(
     const paddle::DataType &tmp_out_dtype,
-    const paddle::DataType &cum_offsets_dtype,
+    const paddle::DataType &cu_seqlens_q_dtype,
     const paddle::DataType &seq_len_this_time_dtype,
     const paddle::DataType &seq_lens_decoder_dtype,
     const paddle::DataType &seq_lens_encoder_dtype,
@@ -256,7 +259,7 @@ std::vector<paddle::DataType> RebuildPaddingInferDtype(
 
 PD_BUILD_STATIC_OP(rebuild_padding_cpu)
     .Inputs({"tmp_out",
-             "cum_offsets",
+             "cu_seqlens_q",
              "seq_len_this_time",
              "seq_lens_decoder",
              "seq_lens_encoder",

custom_ops/gpu_ops/append_attention.cu

@@ -38,7 +38,7 @@ class type2value<phi::dtype::float16> {
 
 
 template <paddle::DataType D>
-std::vector<paddle::Tensor> AppendAttentionKernel(
+void AppendAttentionKernel(
     const AppendAttnMetaData& meta_data,
     const paddle::Tensor& qkv,
     const paddle::Tensor& key_cache,
@@ -60,6 +60,7 @@ std::vector<paddle::Tensor> AppendAttentionKernel(
     const paddle::Tensor& decoder_num_blocks,
     const paddle::Tensor& set_max_lengths,
     const paddle::Tensor& max_len_kv,
+    paddle::Tensor& fmha_out,
     const paddle::optional<paddle::Tensor>& rotary_embs,
     const paddle::optional<paddle::Tensor>& attn_mask,
     const paddle::optional<paddle::Tensor>& qkv_bias,
@@ -72,7 +73,11 @@ std::vector<paddle::Tensor> AppendAttentionKernel(
     const paddle::optional<paddle::Tensor>& cache_v_zp,
     const paddle::optional<paddle::Tensor>& out_linear_shifts,
     const paddle::optional<paddle::Tensor>& out_linear_smooths,
+    const paddle::optional<paddle::Tensor>& mask_offset,
     const paddle::optional<paddle::Tensor>& kv_signal_data,
+    const paddle::optional<paddle::Tensor>& q_norm_weight,
+    const paddle::optional<paddle::Tensor>& k_norm_weight,
+    const float rms_norm_eps,
     const std::string& cache_quant_type_str,
     const bool use_neox_rotary_style,
     const bool rope_3d,
@@ -118,27 +123,6 @@ std::vector<paddle::Tensor> AppendAttentionKernel(
   } else {
     qkv_out = qkv;
   }
-  paddle::Tensor fmha_out;
-  if (out_linear_in_scale > 0.0) {
-    if (fabs(quant_max_bound - 127.0f) < 0.000001) {
-      fmha_out = GetEmptyTensor(
-        {meta_data.token_nums, meta_data.q_num_heads * meta_data.head_dims},
-        paddle::DataType::INT8,
-        qkv.place());
-    } else if (fabs(quant_max_bound - 448.0f) < 0.000001) {
-      fmha_out = GetEmptyTensor(
-        {meta_data.token_nums, meta_data.q_num_heads * meta_data.head_dims},
-        paddle::DataType::FLOAT8_E4M3FN,
-        qkv.place());
-    }else{
-      PD_THROW("Only supported attr of quant_max_bound in ['127', '448'].");
-    }
-  } else {
-    fmha_out = GetEmptyTensor(
-        {meta_data.token_nums, meta_data.q_num_heads * meta_data.head_dims},
-        D,
-        qkv.place());
-  }
 
   auto dispatch_CascadeAppendAttentionKernel = [&](auto temp_args,
         const paddle::Tensor& lambda_batch_ids,
@@ -223,7 +207,10 @@ std::vector<paddle::Tensor> AppendAttentionKernel(
           main_stream,
           &qkv_out,
           const_cast<paddle::Tensor*>(&key_cache),
-          const_cast<paddle::Tensor*>(&value_cache));
+          const_cast<paddle::Tensor*>(&value_cache),
+          q_norm_weight,
+          k_norm_weight,
+          rms_norm_eps);
     };
 
     if (qkv_out_scales) {
@@ -339,7 +326,10 @@ std::vector<paddle::Tensor> AppendAttentionKernel(
             exec_stream,
             &qkv_out,
             const_cast<paddle::Tensor*>(&key_cache),
-            const_cast<paddle::Tensor*>(&value_cache));
+            const_cast<paddle::Tensor*>(&value_cache),
+            q_norm_weight,
+            k_norm_weight,
+            rms_norm_eps);
       } else {
         DecoderWriteCacheWithRoPEKernel<data_t, data_t>(
             meta_data,
@@ -363,7 +353,10 @@ std::vector<paddle::Tensor> AppendAttentionKernel(
             exec_stream,
             &qkv_out,
             const_cast<paddle::Tensor*>(&key_cache),
-            const_cast<paddle::Tensor*>(&value_cache));
+            const_cast<paddle::Tensor*>(&value_cache),
+            q_norm_weight,
+            k_norm_weight,
+            rms_norm_eps);
       }
     }
 
@@ -392,8 +385,6 @@ std::vector<paddle::Tensor> AppendAttentionKernel(
       cudaStreamWaitEvent(main_stream, decoder_event);
     }
   }
-
-  return {fmha_out, qkv_out};
 }
 
 std::vector<paddle::Tensor> AppendAttention(
@@ -429,7 +420,11 @@ std::vector<paddle::Tensor> AppendAttention(
     const paddle::optional<paddle::Tensor>& cache_v_zp,
     const paddle::optional<paddle::Tensor>& out_linear_shifts,
     const paddle::optional<paddle::Tensor>& out_linear_smooths,
+    const paddle::optional<paddle::Tensor>& mask_offset,
     const paddle::optional<paddle::Tensor>& kv_signal_data,
+    const paddle::optional<paddle::Tensor>& q_norm_weight,
+    const paddle::optional<paddle::Tensor>& k_norm_weight,
+    const float rms_norm_eps,
     const std::string& compute_dtype,
     const std::string& cache_quant_type_str,
     const bool use_neox_rotary_style,
@@ -464,8 +459,60 @@ std::vector<paddle::Tensor> AppendAttention(
   meta_data.block_size = key_cache.dims()[2];
   meta_data.batch_size = seq_lens_this_time.dims()[0];
 
-  auto dispatch_by_template = [&](auto temp_args) -> std::vector<paddle::Tensor> {
-      return AppendAttentionKernel<type2value<decltype(temp_args)>::value>(
+  // template dtype generation
+  phi::DataType dtype_id;
+  switch (qkv.dtype()) {
+    case paddle::DataType::FLOAT16:  {dtype_id = phi::DataType::FLOAT16;  break;}
+    case paddle::DataType::BFLOAT16: {dtype_id = phi::DataType::BFLOAT16; break;}
+    case paddle::DataType::INT32: {
+      if (compute_dtype == "bf16") {
+        dtype_id = phi::DataType::BFLOAT16;
+        break;
+      } else if (compute_dtype == "fp16") {
+        dtype_id = phi::DataType::FLOAT16;
+        break;
+      } else {
+        PD_THROW("Only supported attr of compute_dtype in ['fp16', 'bf16'].");
+        break;
+      }
+    }
+    default: {
+      PD_THROW(
+          "NOT supported data type. "
+          "Only float16 and bfloat16 are supported. ");
+      break;
+    }
+  }
+
+  // fmha_out generation, rewrite from AppendAttentionKernel
+  paddle::Tensor fmha_out;
+  if (out_linear_in_scale > 0.0) {
+    if (fabs(quant_max_bound - 127.0f) < 0.000001) {
+      fmha_out = GetEmptyTensor(
+        {meta_data.token_nums, meta_data.q_num_heads * meta_data.head_dims},
+        paddle::DataType::INT8,
+        qkv.place());
+    } else if (fabs(quant_max_bound - 448.0f) < 0.000001) {
+      fmha_out = GetEmptyTensor(
+        {meta_data.token_nums, meta_data.q_num_heads * meta_data.head_dims},
+        paddle::DataType::FLOAT8_E4M3FN,
+        qkv.place());
+    } else{
+      PD_THROW("Only supported attr of quant_max_bound in ['127', '448'].");
+    }
+  } else {
+    fmha_out = GetEmptyTensor(
+        {meta_data.token_nums, meta_data.q_num_heads * meta_data.head_dims},
+        dtype_id,
+        qkv.place());
+  }
+
+  if (mask_offset) {
+    meta_data.mask_offset = mask_offset.get().data<int>();
+  }
+
+  auto dispatch_by_template = [&](auto temp_args) -> void {
+      AppendAttentionKernel<type2value<decltype(temp_args)>::value>(
           meta_data,
           qkv,
           key_cache,
@@ -487,6 +534,7 @@ std::vector<paddle::Tensor> AppendAttention(
           decoder_num_blocks,
           set_max_lengths,
           max_len_kv,
+          fmha_out,
           rotary_embs,
           attn_mask,
           qkv_bias,
@@ -499,7 +547,11 @@ std::vector<paddle::Tensor> AppendAttention(
           cache_v_zp,
           out_linear_shifts,
           out_linear_smooths,
+          mask_offset,
           kv_signal_data,
+          q_norm_weight,
+          k_norm_weight,
+          rms_norm_eps,
           cache_quant_type_str,
           use_neox_rotary_style,
           rope_3d,
@@ -514,20 +566,183 @@ std::vector<paddle::Tensor> AppendAttention(
           speculate_max_draft_token_num,
           causal,
           speculate_decoder);
+  };
 
+
+  phi::dtype::float16 fp16_dtype;
+  phi::dtype::bfloat16 bp16_dtype;
+  switch (dtype_id){
+      case phi::DataType::FLOAT16: {
+        dispatch_by_template(fp16_dtype);
+        return {fmha_out};
+        }
+      case phi::DataType::BFLOAT16: {
+        dispatch_by_template(bp16_dtype);
+        return {fmha_out};
+        }
+      default:
+        PD_THROW(
+          "NOT supported data type. "
+          "Only float16 and bfloat16 are supported. ");
+        break;
+  }
+
+  return {paddle::Tensor{}};
+}
+
+void AppendAttentionWithOutput(
+    const paddle::Tensor& qkv,
+    const paddle::Tensor& key_cache,
+    const paddle::Tensor& value_cache,
+    const paddle::Tensor& seq_lens_encoder,
+    const paddle::Tensor& seq_lens_decoder,
+    const paddle::Tensor& seq_lens_this_time,
+    const paddle::Tensor& batch_id_per_token,
+    const paddle::Tensor& cu_seqlens_q,
+    const paddle::Tensor& block_tables,
+    const paddle::Tensor& encoder_batch_ids,
+    const paddle::Tensor& encoder_tile_ids_per_batch,
+    const paddle::Tensor& encoder_num_blocks,
+    const paddle::Tensor& kv_batch_ids,
+    const paddle::Tensor& kv_tile_ids_per_batch,
+    const paddle::Tensor& kv_num_blocks,
+    const paddle::Tensor& decoder_batch_ids,
+    const paddle::Tensor& decoder_tile_ids_per_batch,
+    const paddle::Tensor& decoder_num_blocks,
+    const paddle::Tensor& set_max_lengths,
+    const paddle::Tensor& max_len_kv,
+    paddle::Tensor& fmha_out,
+    const paddle::optional<paddle::Tensor>& rotary_embs,
+    const paddle::optional<paddle::Tensor>& attn_mask,
+    const paddle::optional<paddle::Tensor>& qkv_bias,
+    const paddle::optional<paddle::Tensor>& qkv_out_scales,
+    const paddle::optional<paddle::Tensor>& cache_k_quant_scales,
+    const paddle::optional<paddle::Tensor>& cache_v_quant_scales,
+    const paddle::optional<paddle::Tensor>& cache_k_dequant_scales,
+    const paddle::optional<paddle::Tensor>& cache_v_dequant_scales,
+    const paddle::optional<paddle::Tensor>& cache_k_zp,
+    const paddle::optional<paddle::Tensor>& cache_v_zp,
+    const paddle::optional<paddle::Tensor>& out_linear_shifts,
+    const paddle::optional<paddle::Tensor>& out_linear_smooths,
+    const paddle::optional<paddle::Tensor>& mask_offset,
+    const paddle::optional<paddle::Tensor>& kv_signal_data,
+    const paddle::optional<paddle::Tensor>& q_norm_weight,
+    const paddle::optional<paddle::Tensor>& k_norm_weight,
+    const float rms_norm_eps,
+    const std::string& compute_dtype,
+    const std::string& cache_quant_type_str,
+    const bool use_neox_rotary_style,
+    const bool rope_3d,
+    const int max_input_length,
+    const float quant_max_bound,
+    const float quant_min_bound,
+    const float out_linear_in_scale,
+    const int encoder_block_shape_q,
+    const int decoder_block_shape_q,
+    const int max_partition_size,
+    const int encoder_max_partition_size,
+    const int speculate_max_draft_token_num,
+    const bool causal,
+    const bool speculate_decoder) {
+  AppendAttnMetaData meta_data;
+
+  const auto& qkv_dims = qkv.dims();
+  const auto& key_cache_dims = key_cache.dims();
+  meta_data.token_nums = qkv_dims[0];
+  meta_data.kv_num_heads = key_cache_dims[1];
+  meta_data.head_dims = key_cache_dims[3];
+  // TODO: trick method support c4, add attr head_dims in the future
+  if (cache_quant_type_str == "cache_int4_zp") {
+    meta_data.head_dims *= 2;
+  }
+  const int total_num_head =
+      qkv_dims[qkv_dims.size() - 1] / meta_data.head_dims;
+  meta_data.q_num_heads = total_num_head - 2 * meta_data.kv_num_heads;
+
+  meta_data.max_blocks_per_seq = block_tables.dims()[1];
+  meta_data.block_size = key_cache.dims()[2];
+  meta_data.batch_size = seq_lens_this_time.dims()[0];
+
+  if (mask_offset) {
+    meta_data.mask_offset = mask_offset.get().data<int>();
+  }
+
+  auto dispatch_by_template = [&](auto temp_args) -> void {
+      AppendAttentionKernel<type2value<decltype(temp_args)>::value>(
+          meta_data,
+          qkv,
+          key_cache,
+          value_cache,
+          seq_lens_encoder,
+          seq_lens_decoder,
+          seq_lens_this_time,
+          batch_id_per_token,
+          cu_seqlens_q,
+          block_tables,
+          encoder_batch_ids,
+          encoder_tile_ids_per_batch,
+          encoder_num_blocks,
+          kv_batch_ids,
+          kv_tile_ids_per_batch,
+          kv_num_blocks,
+          decoder_batch_ids,
+          decoder_tile_ids_per_batch,
+          decoder_num_blocks,
+          set_max_lengths,
+          max_len_kv,
+          fmha_out,
+          rotary_embs,
+          attn_mask,
+          qkv_bias,
+          qkv_out_scales,
+          cache_k_quant_scales,
+          cache_v_quant_scales,
+          cache_k_dequant_scales,
+          cache_v_dequant_scales,
+          cache_k_zp,
+          cache_v_zp,
+          out_linear_shifts,
+          out_linear_smooths,
+          mask_offset,
+          kv_signal_data,
+          q_norm_weight,
+          k_norm_weight,
+          rms_norm_eps,
+          cache_quant_type_str,
+          use_neox_rotary_style,
+          rope_3d,
+          max_input_length,
+          quant_max_bound,
+          quant_min_bound,
+          out_linear_in_scale,
+          encoder_block_shape_q,
+          decoder_block_shape_q,
+          max_partition_size,
+          encoder_max_partition_size,
+          speculate_max_draft_token_num,
+          causal,
+          speculate_decoder);
   };
 
   phi::dtype::float16 fp16_dtype;
   phi::dtype::bfloat16 bp16_dtype;
 
   switch (qkv.dtype()) {
-    case paddle::DataType::FLOAT16: return dispatch_by_template(fp16_dtype);
-    case paddle::DataType::BFLOAT16: return dispatch_by_template(bp16_dtype);
+    case paddle::DataType::FLOAT16: {
+      dispatch_by_template(fp16_dtype);
+      break;
+    }
+    case paddle::DataType::BFLOAT16: {
+      dispatch_by_template(bp16_dtype);
+      break;
+    }
     case paddle::DataType::INT32: {
       if (compute_dtype == "bf16") {
-        return dispatch_by_template(bp16_dtype);
+        dispatch_by_template(bp16_dtype);
+        break;
       } else if (compute_dtype == "fp16") {
-        return dispatch_by_template(fp16_dtype);
+        dispatch_by_template(fp16_dtype);
+        break;
       } else {
         PD_THROW("Only supported attr of compute_dtype in ['fp16', 'bf16'].");
         break;
@@ -540,9 +755,9 @@ std::vector<paddle::Tensor> AppendAttention(
       break;
     }
   }
-  return {paddle::Tensor{}};
 }
 
+
 std::vector<std::vector<int64_t>> AppendAttentionInferShape(
     const std::vector<int64_t>& qkv_shape,
     const std::vector<int64_t>& key_cache_shape,
@@ -576,7 +791,11 @@ std::vector<std::vector<int64_t>> AppendAttentionInferShape(
     const paddle::optional<std::vector<int64_t>>& cache_v_zp_shape,
     const paddle::optional<std::vector<int64_t>>& out_linear_shifts_shape,
     const paddle::optional<std::vector<int64_t>>& out_linear_smooths_shape,
+    const paddle::optional<std::vector<int64_t>>& mask_offset_shape,
     const paddle::optional<std::vector<int64_t>>& kv_signal_data_shape,
+    const paddle::optional<std::vector<int64_t>>& q_norm_weight_shape,
+    const paddle::optional<std::vector<int64_t>>& k_norm_weight_shape,
+    const float rms_norm_eps,
     const std::string& compute_dtype,
     const std::string& cache_quant_type_str,
     const bool use_neox_rotary_style,
@@ -600,7 +819,7 @@ std::vector<std::vector<int64_t>> AppendAttentionInferShape(
   }
   const int total_num_head = qkv_shape[qkv_shape.size() - 1] / head_dim;
   const int num_heads = total_num_head - 2 * kv_num_heads;
-  return {{token_num, num_heads * head_dim}, qkv_shape};
+  return {{token_num, num_heads * head_dim}};
 }
 
 std::vector<paddle::DataType> AppendAttentionInferDtype(
@@ -636,7 +855,11 @@ std::vector<paddle::DataType> AppendAttentionInferDtype(
     const paddle::optional<paddle::DataType>& cache_v_zp_dtype,
     const paddle::optional<paddle::DataType>& out_linear_shifts_dtype,
     const paddle::optional<paddle::DataType>& out_linear_smooths_dtype,
+    const paddle::optional<paddle::DataType>& mask_offset_dtype,
     const paddle::optional<paddle::DataType>& kv_signal_data_dtype,
+    const paddle::optional<paddle::DataType>& q_norm_weight_dtype,
+    const paddle::optional<paddle::DataType>& k_norm_weight_dtype,
+    const float rms_norm_eps,
     const std::string& compute_dtype,
     const std::string& cache_quant_type_str,
     const bool use_neox_rotary_style,
@@ -655,32 +878,148 @@ std::vector<paddle::DataType> AppendAttentionInferDtype(
   if (compute_dtype == "bf16") {
     if (out_linear_in_scale > 0.0) {
       if (fabs(quant_max_bound - 127.0f) < 0.000001) {
-        return {paddle::DataType::INT8, paddle::DataType::BFLOAT16};
+        return {paddle::DataType::INT8};
       } else if (fabs(quant_max_bound - 448.0f) < 0.000001) {
-        return {paddle::DataType::FLOAT8_E4M3FN, paddle::DataType::BFLOAT16};
+        return {paddle::DataType::FLOAT8_E4M3FN};
       }else{
         PD_THROW("Only supported attr of quant_max_bound in ['127.0', '448.0'].");
       }
     } else {
-      return {paddle::DataType::BFLOAT16, paddle::DataType::BFLOAT16};
+      return {paddle::DataType::BFLOAT16};
     }
   } else if (compute_dtype == "fp16") {
     if (out_linear_in_scale > 0.0) {
       if (fabs(quant_max_bound - 127.0f) < 0.000001) {
-        return {paddle::DataType::INT8, paddle::DataType::FLOAT16};
+        return {paddle::DataType::INT8};
       } else if (fabs(quant_max_bound - 448.0f) < 0.000001) {
-        return {paddle::DataType::FLOAT8_E4M3FN, paddle::DataType::FLOAT16};
+        return {paddle::DataType::FLOAT8_E4M3FN};
       }else{
         PD_THROW("Only supported attr of quant_max_bound in ['127.0', '448.0'].");
       }
     } else {
-      return {paddle::DataType::FLOAT16, paddle::DataType::FLOAT16};
+      return {paddle::DataType::FLOAT16};
     }
   } else {
     PD_THROW("Only supported attr of compute_dtype in ['fp16', 'bf16'].");
   }
 }
 
+std::vector<std::vector<int64_t>> AppendAttentionWithOutputInferShape(
+    const std::vector<int64_t>& qkv_shape,
+    const std::vector<int64_t>& key_cache_shape,
+    const std::vector<int64_t>& value_cache_shape,
+    const std::vector<int64_t>& seq_lens_encoder_shape,
+    const std::vector<int64_t>& seq_lens_decoder_shape,
+    const std::vector<int64_t>& seq_lens_this_time_shape,
+    const std::vector<int64_t>& batch_id_per_token_shape,
+    const std::vector<int64_t>& cu_seqlens_q_shape,
+    const std::vector<int64_t>& block_tables_shape,
+    const std::vector<int64_t>& encoder_batch_ids_shape,
+    const std::vector<int64_t>& encoder_tile_ids_per_batch_shape,
+    const std::vector<int64_t>& encoder_num_blocks_shape,
+    const std::vector<int64_t>& kv_batch_ids_shape,
+    const std::vector<int64_t>& kv_tile_ids_per_batch_shape,
+    const std::vector<int64_t>& kv_num_blocks_shape,
+    const std::vector<int64_t>& decoder_batch_ids_shape,
+    const std::vector<int64_t>& decoder_tile_ids_per_batch_shape,
+    const std::vector<int64_t>& decoder_num_blocks_shape,
+    const std::vector<int64_t>& set_max_lengths_shape,
+    const std::vector<int64_t>& max_len_kv_shape,
+    const std::vector<int64_t>& fmha_out_shape,
+    const paddle::optional<std::vector<int64_t>>& rotary_embs_shape,
+    const paddle::optional<std::vector<int64_t>>& attn_mask_shape,
+    const paddle::optional<std::vector<int64_t>>& qkv_bias_shape,
+    const paddle::optional<std::vector<int64_t>>& qkv_out_scales_shape,
+    const paddle::optional<std::vector<int64_t>>& cache_k_quant_scales_shape,
+    const paddle::optional<std::vector<int64_t>>& cache_v_quant_scales_shape,
+    const paddle::optional<std::vector<int64_t>>& cache_k_dequant_scales_shape,
+    const paddle::optional<std::vector<int64_t>>& cache_v_dequant_scales_shape,
+    const paddle::optional<std::vector<int64_t>>& cache_k_zp_shape,
+    const paddle::optional<std::vector<int64_t>>& cache_v_zp_shape,
+    const paddle::optional<std::vector<int64_t>>& out_linear_shifts_shape,
+    const paddle::optional<std::vector<int64_t>>& out_linear_smooths_shape,
+    const paddle::optional<std::vector<int64_t>>& mask_offset_shape,
+    const paddle::optional<std::vector<int64_t>>& kv_signal_data_shape,
+    const paddle::optional<std::vector<int64_t>>& q_norm_weight_shape,
+    const paddle::optional<std::vector<int64_t>>& k_norm_weight_shape,
+    const float rms_norm_eps,
+    const std::string& compute_dtype,
+    const std::string& cache_quant_type_str,
+    const bool use_neox_rotary_style,
+    const bool rope_3d,
+    const int max_input_length,
+    const float quant_max_bound,
+    const float quant_min_bound,
+    const float out_linear_in_scale,
+    const int encoder_block_shape_q,
+    const int decoder_block_shape_q,
+    const int max_partition_size,
+    const int encoder_max_partition_size,
+    const int speculate_max_draft_token_num,
+    const bool causal,
+    const bool speculate_decoder) {
+  return {fmha_out_shape};
+}
+
+std::vector<paddle::DataType> AppendAttentionWithOutputInferDtype(
+    const paddle::DataType& qkv_dtype,
+    const paddle::DataType& key_cache_dtype,
+    const paddle::DataType& value_cache_dtype,
+    const paddle::DataType& seq_lens_encoder_dtype,
+    const paddle::DataType& seq_lens_decoder_dtype,
+    const paddle::DataType& seq_lens_this_time_dtype,
+    const paddle::DataType& batch_id_per_token_dtype,
+    const paddle::DataType& cu_seqlens_q_dtype,
+    const paddle::DataType& block_tables_dtype,
+    const paddle::DataType& encoder_batch_ids_dtype,
+    const paddle::DataType& encoder_tile_ids_per_batch_dtype,
+    const paddle::DataType& encoder_num_blocks_dtype,
+    const paddle::DataType& kv_batch_ids_dtype,
+    const paddle::DataType& kv_tile_ids_per_batch_dtype,
+    const paddle::DataType& kv_num_blocks_dtype,
+    const paddle::DataType& decoder_batch_ids_dtype,
+    const paddle::DataType& decoder_tile_ids_per_batch_dtype,
+    const paddle::DataType& decoder_num_blocks_dtype,
+    const paddle::DataType& set_max_lengths_dtype,
+    const paddle::DataType& max_len_kv_dtype,
+    const paddle::DataType& fmha_out_dtype,
+    const paddle::optional<paddle::DataType>& rotary_embs_dtype,
+    const paddle::optional<paddle::DataType>& attn_mask_dtype,
+    const paddle::optional<paddle::DataType>& qkv_bias_dtype,
+    const paddle::optional<paddle::DataType>& qkv_out_scales_dtype,
+    const paddle::optional<paddle::DataType>& cache_k_quant_scales_dtype,
+    const paddle::optional<paddle::DataType>& cache_v_quant_scales_dtype,
+    const paddle::optional<paddle::DataType>& cache_k_dequant_scales_dtype,
+    const paddle::optional<paddle::DataType>& cache_v_dequant_scales_dtype,
+    const paddle::optional<paddle::DataType>& cache_k_zp_dtype,
+    const paddle::optional<paddle::DataType>& cache_v_zp_dtype,
+    const paddle::optional<paddle::DataType>& out_linear_shifts_dtype,
+    const paddle::optional<paddle::DataType>& out_linear_smooths_dtype,
+    const paddle::optional<paddle::DataType>& mask_offset_dtype,
+    const paddle::optional<paddle::DataType>& kv_signal_data_dtype,
+    const paddle::optional<paddle::DataType>& q_norm_weight_dtype,
+    const paddle::optional<paddle::DataType>& k_norm_weight_dtype,
+    const float rms_norm_eps,
+    const std::string& compute_dtype,
+    const std::string& cache_quant_type_str,
+    const bool use_neox_rotary_style,
+    const bool rope_3d,
+    const int max_input_length,
+    const float quant_max_bound,
+    const float quant_min_bound,
+    const float out_linear_in_scale,
+    const int encoder_block_shape_q,
+    const int decoder_block_shape_q,
+    const int max_partition_size,
+    const int encoder_max_partition_size,
+    const int speculate_max_draft_token_num,
+    const bool causal,
+    const bool speculate_decoder) {
+  return {fmha_out_dtype};
+}
+
+
+
 PD_BUILD_STATIC_OP(append_attention)
     .Inputs({"qkv",
              "key_cache",
@@ -714,11 +1053,15 @@ PD_BUILD_STATIC_OP(append_attention)
              paddle::Optional("cache_v_zp"),
              paddle::Optional("out_linear_shifts"),
              paddle::Optional("out_linear_smooths"),
-             paddle::Optional("kv_signal_data")})
-    .Outputs({"fmha_out", "qkv_out", "key_cache_out", "value_cache_out"})
+             paddle::Optional("mask_offset"),
+             paddle::Optional("kv_signal_data"),
+             paddle::Optional("q_norm_weight"),
+             paddle::Optional("k_norm_weight")})
+    .Outputs({"fmha_out", "key_cache_out", "value_cache_out"})
     .SetInplaceMap({{"key_cache", "key_cache_out"},
                     {"value_cache", "value_cache_out"}})
-    .Attrs({"compute_type: std::string",
+    .Attrs({"rms_norm_eps: float",
+            "compute_type: std::string",
             "cache_quant_type: std::string",
             "use_neox_rotary_style: bool",
             "rope_3d: bool",
@@ -732,7 +1075,71 @@ PD_BUILD_STATIC_OP(append_attention)
             "encoder_max_partition_size: int",
             "speculate_max_draft_token_num: int",
             "causal: bool",
-            "speculate_decoder: bool"})
+            "speculate_decoder: bool",
+            })
     .SetKernelFn(PD_KERNEL(AppendAttention))
     .SetInferShapeFn(PD_INFER_SHAPE(AppendAttentionInferShape))
     .SetInferDtypeFn(PD_INFER_DTYPE(AppendAttentionInferDtype));
+
+PD_BUILD_STATIC_OP(append_attention_with_output)
+    .Inputs({"qkv",
+             "key_cache",
+             "value_cache",
+             "seq_lens_encoder",
+             "seq_lens_decoder",
+             "seq_lens_this_time",
+             "batch_id_per_token",
+             "cu_seqlens_q",
+             "block_tables",
+             "encoder_batch_ids",
+             "encoder_tile_ids_per_batch",
+             "encoder_num_blocks",
+             "kv_batch_ids",
+             "kv_tile_ids_per_batch",
+             "kv_num_blocks",
+             "decoder_batch_ids",
+             "decoder_tile_ids_per_batch",
+             "decoder_num_blocks",
+             "set_max_lengths",
+             "max_len_kv",
+             "fmha_out",
+             paddle::Optional("rotary_embs"),
+             paddle::Optional("attn_mask"),
+             paddle::Optional("qkv_bias"),
+             paddle::Optional("qkv_out_scales"),
+             paddle::Optional("cache_k_quant_scales"),
+             paddle::Optional("cache_v_quant_scales"),
+             paddle::Optional("cache_k_dequant_scales"),
+             paddle::Optional("cache_v_dequant_scales"),
+             paddle::Optional("cache_k_zp"),
+             paddle::Optional("cache_v_zp"),
+             paddle::Optional("out_linear_shifts"),
+             paddle::Optional("out_linear_smooths"),
+             paddle::Optional("mask_offset"),
+             paddle::Optional("kv_signal_data"),
+             paddle::Optional("q_norm_weight"),
+             paddle::Optional("k_norm_weight")})
+    .Outputs({"fmha_out_out", "qkv_out", "key_cache_out", "value_cache_out"})
+    .SetInplaceMap({{"fmha_out", "fmha_out_out"},
+                    {"key_cache", "key_cache_out"},
+                    {"value_cache", "value_cache_out"}})
+    .Attrs({"rms_norm_eps: float",
+            "compute_type: std::string",
+            "cache_quant_type: std::string",
+            "use_neox_rotary_style: bool",
+            "rope_3d: bool",
+            "max_input_length: int",
+            "quant_max_bound: float",
+            "quant_min_bound: float",
+            "out_linear_in_scale: float",
+            "encoder_block_shape_q: int",
+            "decoder_block_shape_q: int",
+            "max_partition_size: int",
+            "encoder_max_partition_size: int",
+            "speculate_max_draft_token_num: int",
+            "causal: bool",
+            "speculate_decoder: bool",
+            })
+    .SetKernelFn(PD_KERNEL(AppendAttentionWithOutput))
+    .SetInferShapeFn(PD_INFER_SHAPE(AppendAttentionWithOutputInferShape))
+    .SetInferDtypeFn(PD_INFER_DTYPE(AppendAttentionWithOutputInferDtype));

custom_ops/gpu_ops/append_attn/append_attention_c16_impl.cuh

@@ -43,6 +43,7 @@ __global__ void multi_query_append_attention_kernel(
     const int *__restrict__ tile_ids_per_batch,
     const int *__restrict__ cu_seqlens_q,
     const int *__restrict__ block_table,  // [bsz, block_num_per_seq]
+    const int *__restrict__ mask_offset,
     const int max_seq_len,
     const int max_dec_len,
     const int max_block_num_per_seq,
@@ -141,6 +142,7 @@ __global__ void multi_query_append_attention_kernel(
   } else {
     o_base_ptr_int8 = out + o_offset;
   }
+  const int *mask_offset_this_seq = mask_offset ? mask_offset + q_start_seq_id : nullptr;
   smem_t qo_smem(smem);
 
   uint32_t q_smem_offset_r = smem_t::get_permuted_offset<num_vecs_per_head>(
@@ -179,7 +181,7 @@ __global__ void multi_query_append_attention_kernel(
                          kv_len - q_len +
                              tile_id * num_rows_per_block / GROUP_SIZE,
                          chunk_start)))
-              : chunk_len) /
+              : mask_offset ? 0 : chunk_len) /
       (num_frags_z * 16);
   uint32_t k_smem_offset_r = smem_t::get_permuted_offset<num_vecs_per_head>(
       8 * (tid / 16) + tid % 8, (tid % 16) / 8);
@@ -245,12 +247,16 @@ __global__ void multi_query_append_attention_kernel(
              NUM_WARPS,
              num_frags_x,
              num_frags_y,
-             num_frags_z>(q_base_seq_id_this_block,
+             num_frags_z>(nullptr,
+                          q_base_seq_id_this_block,
                           kv_idx_base,
                           q_len,
                           kv_len,
                           chunk_end,
-                          s_frag);
+                          -1,
+                          s_frag,
+                          mask_offset_this_seq);
+
     }
 
     // update m,d
@@ -406,6 +412,8 @@ __global__ void multi_query_append_attention_warp1_4_kernel(
     const int *__restrict__ tile_ids_per_batch,
     const int *__restrict__ cu_seqlens_q,
     const int *__restrict__ block_table,  // [bsz, block_num_per_seq]
+    const int *__restrict__ mask_offset,
+    const bool *__restrict__ attn_mask,    // [bsz, max_q, max_q] for tree-mask
     const int max_seq_len,
     const int max_dec_len,
     const int max_block_num_per_seq,
@@ -419,7 +427,8 @@ __global__ void multi_query_append_attention_warp1_4_kernel(
     float *__restrict__ tmp_m,      // [token_num, num_chunks, num_heads]
     float *__restrict__ tmp_d,      // [token_num, num_chunks, num_heads]
     OutT *__restrict__ out,
-    const int speculate_max_draft_token_num = 5) {
+    const int speculate_max_draft_token_num = 5,
+    const uint32_t attn_mask_len = -1) {
   constexpr uint32_t num_vecs_per_head = HEAD_DIM / num_elems_per_128b<T>();
   static_assert(NUM_WARP_Q == 1, "NUM_WARP_Q must be 1");
   static_assert(NUM_WARP_KV == 4, "NUM_WARP_KV must be 4");
@@ -502,7 +511,7 @@ __global__ void multi_query_append_attention_warp1_4_kernel(
           tid % 8 * num_elems_per_128b<T>();
     }
   }
-
+  const int *mask_offset_this_seq = mask_offset ? mask_offset + q_start_seq_id : nullptr;
   smem_t qo_smem(smem);
 
   uint32_t q_smem_offset_r = smem_t::get_permuted_offset<num_vecs_per_head>(
@@ -540,10 +549,9 @@ __global__ void multi_query_append_attention_warp1_4_kernel(
   const uint32_t mask_check_iteration =
       (CAUSAL ? (min(chunk_len,
                      sub_if_greater_or_zero(
-                         kv_len - q_len +
-                             tile_id * num_rows_per_block / GROUP_SIZE,
+                         kv_len - q_len,
                          chunk_start)))
-              : chunk_len) /
+              : mask_offset ? 0 : chunk_len) /
       (NUM_WARP_KV * num_frags_z * 16);
 
   uint32_t k_smem_offset_r = smem_t::get_permuted_offset<num_vecs_per_head>(
@@ -611,12 +619,15 @@ __global__ void multi_query_append_attention_warp1_4_kernel(
              NUM_WARPS,
              num_frags_x,
              num_frags_y,
-             num_frags_z>(q_base_seq_id_this_block,
+             num_frags_z>(attn_mask ? attn_mask + batch_id * attn_mask_len *attn_mask_len : nullptr,
+                          q_base_seq_id_this_block,
                           kv_idx_base + wid * num_frags_z * 16,
                           q_len,
                           kv_len,
                           chunk_end,
-                          s_frag);
+                          attn_mask_len,
+                          s_frag,
+                          mask_offset_this_seq);
     }
 
     // update m,d
@@ -882,6 +893,7 @@ void MultiQueryAppendAttention(
           tile_ids_per_batch.data<int>(),
           cu_seqlens_q.data<int>(),
           block_table.data<int>(),
+          meta_data.mask_offset,
           max_seq_len,
           max_dec_len,
           max_block_num_per_seq,
@@ -939,6 +951,7 @@ void MultiQueryAppendAttention(
           tile_ids_per_batch.data<int>(),
           cu_seqlens_q.data<int>(),
           block_table.data<int>(),
+          meta_data.mask_offset,
           max_seq_len,
           max_dec_len,
           max_block_num_per_seq,
@@ -1061,12 +1074,18 @@ void MultiQueryAppendAttention(
     if (!is_decoder) {
       chunk_size = static_cast<uint32_t>(encoder_max_partition_size);
     }
-    const int num_chunks = div_up(max_dec_len, chunk_size);
 
+    uint32_t attn_mask_len;
+    if (attn_mask) {
+        attn_mask_len = attn_mask.get().shape()[1];
+    } else {
+        attn_mask_len = -1;
+    }
+
+    const int num_chunks = div_up(max_seq_len, chunk_size);
     dim3 grids(num_blocks_x_cpu, num_chunks, kv_num_heads);
     dim3 blocks(32, num_warps);
-
-    if (num_chunks <= 1) {
+    if (num_chunks <= 0) {
       auto nosplit_kv_kernel =
           multi_query_append_attention_warp1_4_kernel<NV_TYPE,
                                                       false,
@@ -1104,6 +1123,9 @@ void MultiQueryAppendAttention(
           tile_ids_per_batch.data<int>(),
           cu_seqlens_q.data<int>(),
           block_table.data<int>(),
+          meta_data.mask_offset,
+          attn_mask ? const_cast<bool *>(attn_mask.get().data<bool>())
+                        : nullptr,
           max_seq_len,
           max_dec_len,
           max_block_num_per_seq,
@@ -1116,7 +1138,8 @@ void MultiQueryAppendAttention(
           nullptr,
           nullptr,
           reinterpret_cast<OUT_NV_TYPE *>(out->data<OutT>()),
-          speculate_max_draft_token_num);
+          speculate_max_draft_token_num,
+          attn_mask_len);
     } else {
       phi::Allocator::AllocationPtr tmp_workspace, tmp_m, tmp_d;
       if (is_decoder) {
@@ -1161,8 +1184,8 @@ void MultiQueryAppendAttention(
           reinterpret_cast<NV_TYPE *>(const_cast<T *>(cache_k.data<T>())),
           reinterpret_cast<NV_TYPE *>(const_cast<T *>(cache_v.data<T>())),
           shift_bias ? reinterpret_cast<NV_TYPE *>(
-                           const_cast<T *>(shift_bias.get().data<T>()))
-                     : nullptr,
+                            const_cast<T *>(shift_bias.get().data<T>()))
+                      : nullptr,
           smooth_weight ? reinterpret_cast<NV_TYPE *>(
                               const_cast<T *>(smooth_weight.get().data<T>()))
                         : nullptr,
@@ -1172,6 +1195,9 @@ void MultiQueryAppendAttention(
           tile_ids_per_batch.data<int>(),
           cu_seqlens_q.data<int>(),
           block_table.data<int>(),
+          meta_data.mask_offset,
+          attn_mask ? const_cast<bool *>(attn_mask.get().data<bool>())
+                        : nullptr,
           max_seq_len,
           max_dec_len,
           max_block_num_per_seq,
@@ -1184,7 +1210,8 @@ void MultiQueryAppendAttention(
           static_cast<float *>(tmp_m->ptr()),
           static_cast<float *>(tmp_d->ptr()),
           reinterpret_cast<OUT_NV_TYPE *>(out->data<OutT>()),
-          speculate_max_draft_token_num);
+          speculate_max_draft_token_num,
+          attn_mask_len);
 
       // merge
       constexpr int vec_size = num_elems_per_128b<NV_TYPE>();
@@ -1208,8 +1235,8 @@ void MultiQueryAppendAttention(
                 seq_lens_encoder.data<int>(),
                 cu_seqlens_q.data<int>(),
                 shift_bias ? reinterpret_cast<NV_TYPE *>(
-                                 const_cast<T *>(shift_bias.get().data<T>()))
-                           : nullptr,
+                                  const_cast<T *>(shift_bias.get().data<T>()))
+                            : nullptr,
                 smooth_weight ? reinterpret_cast<NV_TYPE *>(const_cast<T *>(
                                     smooth_weight.get().data<T>()))
                               : nullptr,
@@ -1226,14 +1253,14 @@ void MultiQueryAppendAttention(
         constexpr int blockx = HEAD_DIM / vec_size;
         constexpr int blocky = (128 + blockx - 1) / blockx;
         dim3 grids_merge(min(sm_count * 4, token_num),
-                         num_heads);
+                          num_heads);
         dim3 blocks_merge(blockx, blocky);
         merge_multi_chunks_v2_kernel<NV_TYPE,
-                                     vec_size,
-                                     blocky,
-                                     HEAD_DIM,
-                                     OUT_NV_TYPE,
-                                     ENABLE_PREFILL>
+                                      vec_size,
+                                      blocky,
+                                      HEAD_DIM,
+                                      OUT_NV_TYPE,
+                                      ENABLE_PREFILL>
             <<<grids_merge, blocks_merge, 0, stream>>>(
                 reinterpret_cast<NV_TYPE *>(tmp_workspace->ptr()),
                 static_cast<float *>(tmp_m->ptr()),
@@ -1244,8 +1271,8 @@ void MultiQueryAppendAttention(
                 batch_id_per_token.data<int>(),
                 cu_seqlens_q.data<int>(),
                 shift_bias ? reinterpret_cast<NV_TYPE *>(
-                                 const_cast<T *>(shift_bias.get().data<T>()))
-                           : nullptr,
+                                  const_cast<T *>(shift_bias.get().data<T>()))
+                            : nullptr,
                 smooth_weight ? reinterpret_cast<NV_TYPE *>(const_cast<T *>(
                                     smooth_weight.get().data<T>()))
                               : nullptr,

custom_ops/gpu_ops/append_attn/append_attention_c4_impl.cuh

@@ -48,6 +48,7 @@ __global__ void multi_query_append_attention_c4_kernel(
     const int *__restrict__ tile_ids_per_batch,
     const int *__restrict__ cu_seqlens_q,
     const int *__restrict__ block_table,  // [bsz, block_num_per_seq]
+    const int *__restrict__ mask_offset,
     const int max_seq_len,
     const int max_dec_len,
     const int max_block_num_per_seq,
@@ -172,6 +173,7 @@ __global__ void multi_query_append_attention_c4_kernel(
   } else {
     o_base_ptr_int8 = out + o_offset;
   }
+  const int *mask_offset_this_seq = mask_offset ? mask_offset + q_start_seq_id : nullptr;
   smem_t qo_smem(smem);
 
   uint32_t q_smem_offset_r = smem_t::get_permuted_offset<num_vecs_per_head>(
@@ -248,7 +250,7 @@ __global__ void multi_query_append_attention_c4_kernel(
                          kv_len - q_len +
                              tile_id * num_rows_per_block / GROUP_SIZE,
                          chunk_start)))
-              : chunk_len) /
+              : mask_offset ? 0 : chunk_len) /
       (num_frags_z * 16);
 
   uint32_t k_smem_offset_r =
@@ -333,12 +335,15 @@ __global__ void multi_query_append_attention_c4_kernel(
              NUM_WARPS,
              num_frags_x,
              num_frags_y,
-             num_frags_z>(q_base_seq_id_this_block,
+             num_frags_z>(nullptr,
+                          q_base_seq_id_this_block,
                           kv_idx_base,
                           q_len,
                           kv_len,
                           chunk_end,
-                          s_frag);
+                          -1,
+                          s_frag,
+                          mask_offset_this_seq);
     }
 
     update_mdo_states<num_frags_x, num_frags_y, num_frags_z>(
@@ -505,6 +510,8 @@ __global__ void multi_query_append_attention_c4_warp1_4_kernel(
     const int *__restrict__ tile_ids_per_batch,
     const int *__restrict__ cu_seqlens_q,
     const int *__restrict__ block_table,  // [bsz, block_num_per_seq]
+    const int *__restrict__ mask_offset,
+    const bool *__restrict__ attn_mask,    // [bsz, max_q, max_q] for tree-mask
     const int max_seq_len,
     const int max_dec_len,
     const int max_block_num_per_seq,
@@ -518,7 +525,8 @@ __global__ void multi_query_append_attention_c4_warp1_4_kernel(
     float *__restrict__ tmp_m,      // [token_num, num_chunks, num_heads]
     float *__restrict__ tmp_d,      // [token_num, num_chunks, num_heads]
     OutT *__restrict__ out,
-    const int speculate_max_draft_token_num = 5) {
+    const int speculate_max_draft_token_num = 5,
+    const uint32_t attn_mask_len = -1) {
   constexpr uint32_t num_vecs_per_head = HEAD_DIM / num_elems_per_128b<T>();
   constexpr uint32_t num_vecs_per_head_k =
       HEAD_DIM / 2 / num_elems_per_128b<CacheT>();
@@ -627,7 +635,7 @@ __global__ void multi_query_append_attention_c4_warp1_4_kernel(
           tid % 8 * num_elems_per_128b<T>();
     }
   }
-
+  const int *mask_offset_this_seq = mask_offset ? mask_offset + q_start_seq_id : nullptr;
   smem_t qo_smem(smem);
 
   uint32_t q_smem_offset_r = smem_t::get_permuted_offset<num_vecs_per_head>(
@@ -703,10 +711,9 @@ __global__ void multi_query_append_attention_c4_warp1_4_kernel(
   const uint32_t mask_check_iteration =
       (CAUSAL ? (min(chunk_len,
                      sub_if_greater_or_zero(
-                         kv_len - q_len +
-                             tile_id * num_rows_per_block / GROUP_SIZE,
+                         kv_len - q_len,
                          chunk_start)))
-              : chunk_len) /
+              : mask_offset ? 0 : chunk_len) /
       (NUM_WARP_KV * num_frags_z * 16);
 
   uint32_t k_smem_offset_r =
@@ -788,12 +795,15 @@ __global__ void multi_query_append_attention_c4_warp1_4_kernel(
              NUM_WARPS,
              num_frags_x,
              num_frags_y,
-             num_frags_z>(q_base_seq_id_this_block,
+             num_frags_z>(attn_mask ? attn_mask + batch_id * attn_mask_len *attn_mask_len : nullptr,
+                          q_base_seq_id_this_block,
                           kv_idx_base + wid * num_frags_z * 16,
                           q_len,
                           kv_len,
                           chunk_end,
-                          s_frag);
+                          attn_mask_len,
+                          s_frag,
+                          mask_offset_this_seq);
     }
 
     update_mdo_states<num_frags_x, num_frags_y, num_frags_z>(
@@ -1088,6 +1098,7 @@ void MultiQueryAppendC4Attention(
           tile_ids_per_batch.data<int>(),
           cu_seqlens_q.data<int>(),
           block_table.data<int>(),
+          meta_data.mask_offset,
           max_seq_len,
           max_dec_len,
           max_block_num_per_seq,
@@ -1151,6 +1162,7 @@ void MultiQueryAppendC4Attention(
           tile_ids_per_batch.data<int>(),
           cu_seqlens_q.data<int>(),
           block_table.data<int>(),
+          meta_data.mask_offset,
           max_seq_len,
           max_dec_len,
           max_block_num_per_seq,
@@ -1285,10 +1297,18 @@ void MultiQueryAppendC4Attention(
     if (!is_decoder) {
       chunk_size = static_cast<uint32_t>(encoder_max_partition_size);
     }
-    const int num_chunks = div_up(max_dec_len, chunk_size);
+
+    const int num_chunks = div_up(max_seq_len, chunk_size);
+    uint32_t attn_mask_len;
+    if (attn_mask) {
+        attn_mask_len = attn_mask.get().shape()[1];
+    } else {
+        attn_mask_len = -1;
+    }
+
     dim3 grids(num_blocks_x_cpu, num_chunks, kv_num_heads);
     dim3 blocks(32, num_warps);
-    if (num_chunks <= 1) {
+    if (num_chunks <= 0) {
       auto nosplit_kv_kernel =
           multi_query_append_attention_c4_warp1_4_kernel<NV_TYPE,
                                                          uint8_t,
@@ -1334,6 +1354,9 @@ void MultiQueryAppendC4Attention(
           tile_ids_per_batch.data<int>(),
           cu_seqlens_q.data<int>(),
           block_table.data<int>(),
+          meta_data.mask_offset,
+          attn_mask ? const_cast<bool *>(attn_mask.get().data<bool>())
+                        : nullptr,
           max_seq_len,
           max_dec_len,
           max_block_num_per_seq,
@@ -1346,7 +1369,8 @@ void MultiQueryAppendC4Attention(
           nullptr,
           nullptr,
           reinterpret_cast<OUT_NV_TYPE *>(out->data<OutT>()),
-          speculate_max_draft_token_num);
+          speculate_max_draft_token_num,
+          attn_mask_len);
     } else {
       phi::Allocator::AllocationPtr tmp_workspace, tmp_m, tmp_d;
       if (is_decoder) {
@@ -1392,15 +1416,15 @@ void MultiQueryAppendC4Attention(
           const_cast<uint8_t *>(cache_v.data<uint8_t>()),
           reinterpret_cast<NV_TYPE *>(const_cast<T *>(cache_k_scale.data<T>())),
           cache_k_zp ? reinterpret_cast<NV_TYPE *>(
-                           const_cast<T *>(cache_k_zp.get().data<T>()))
-                     : nullptr,
+                            const_cast<T *>(cache_k_zp.get().data<T>()))
+                      : nullptr,
           reinterpret_cast<NV_TYPE *>(const_cast<T *>(cache_v_scale.data<T>())),
           cache_v_zp ? reinterpret_cast<NV_TYPE *>(
-                           const_cast<T *>(cache_v_zp.get().data<T>()))
-                     : nullptr,
+                            const_cast<T *>(cache_v_zp.get().data<T>()))
+                      : nullptr,
           shift_bias ? reinterpret_cast<NV_TYPE *>(
-                           const_cast<T *>(shift_bias.get().data<T>()))
-                     : nullptr,
+                            const_cast<T *>(shift_bias.get().data<T>()))
+                      : nullptr,
           smooth_weight ? reinterpret_cast<NV_TYPE *>(
                               const_cast<T *>(smooth_weight.get().data<T>()))
                         : nullptr,
@@ -1410,6 +1434,9 @@ void MultiQueryAppendC4Attention(
           tile_ids_per_batch.data<int>(),
           cu_seqlens_q.data<int>(),
           block_table.data<int>(),
+          meta_data.mask_offset,
+          attn_mask ? const_cast<bool *>(attn_mask.get().data<bool>())
+                        : nullptr,
           max_seq_len,
           max_dec_len,
           max_block_num_per_seq,
@@ -1422,7 +1449,8 @@ void MultiQueryAppendC4Attention(
           static_cast<float *>(tmp_m->ptr()),
           static_cast<float *>(tmp_d->ptr()),
           reinterpret_cast<OUT_NV_TYPE *>(out->data<OutT>()),
-          speculate_max_draft_token_num);
+          speculate_max_draft_token_num,
+          attn_mask_len);
       // merge
       constexpr int vec_size = num_elems_per_128b<NV_TYPE>();
       if (is_decoder) {
@@ -1445,8 +1473,8 @@ void MultiQueryAppendC4Attention(
                 seq_lens_encoder.data<int>(),
                 cu_seqlens_q.data<int>(),
                 shift_bias ? reinterpret_cast<NV_TYPE *>(
-                                 const_cast<T *>(shift_bias.get().data<T>()))
-                           : nullptr,
+                                  const_cast<T *>(shift_bias.get().data<T>()))
+                            : nullptr,
                 smooth_weight ? reinterpret_cast<NV_TYPE *>(const_cast<T *>(
                                     smooth_weight.get().data<T>()))
                               : nullptr,
@@ -1463,14 +1491,14 @@ void MultiQueryAppendC4Attention(
         constexpr int blockx = HEAD_DIM / vec_size;
         constexpr int blocky = (128 + blockx - 1) / blockx;
         dim3 grids_merge(min(sm_count * 4, token_num),
-                         num_heads);
+                          num_heads);
         dim3 blocks_merge(blockx, blocky);
         merge_multi_chunks_v2_kernel<NV_TYPE,
-                                     vec_size,
-                                     blocky,
-                                     HEAD_DIM,
-                                     OUT_NV_TYPE,
-                                     ENABLE_PREFILL>
+                                      vec_size,
+                                      blocky,
+                                      HEAD_DIM,
+                                      OUT_NV_TYPE,
+                                      ENABLE_PREFILL>
             <<<grids_merge, blocks_merge, 0, stream>>>(
                 reinterpret_cast<NV_TYPE *>(tmp_workspace->ptr()),
                 static_cast<float *>(tmp_m->ptr()),
@@ -1481,8 +1509,8 @@ void MultiQueryAppendC4Attention(
                 batch_id_per_token.data<int>(),
                 cu_seqlens_q.data<int>(),
                 shift_bias ? reinterpret_cast<NV_TYPE *>(
-                                 const_cast<T *>(shift_bias.get().data<T>()))
-                           : nullptr,
+                                  const_cast<T *>(shift_bias.get().data<T>()))
+                            : nullptr,
                 smooth_weight ? reinterpret_cast<NV_TYPE *>(const_cast<T *>(
                                     smooth_weight.get().data<T>()))
                               : nullptr,

custom_ops/gpu_ops/append_attn/append_attention_c8_impl.cuh

@@ -48,6 +48,7 @@ __global__ void multi_query_append_attention_c8_kernel(
     const int *__restrict__ tile_ids_per_batch,
     const int *__restrict__ cu_seqlens_q,
     const int *__restrict__ block_table,  // [bsz, block_num_per_seq]
+    const int *__restrict__ mask_offset,
     const int max_seq_len,
     const int max_dec_len,
     const int max_block_num_per_seq,
@@ -179,6 +180,7 @@ __global__ void multi_query_append_attention_c8_kernel(
   } else {
     o_base_ptr_int8 = out + o_offset;
   }
+  const int *mask_offset_this_seq = mask_offset ? mask_offset + q_start_seq_id : nullptr;
   smem_t qo_smem(smem);
 
   uint32_t q_smem_offset_r = smem_t::get_permuted_offset<num_vecs_per_head>(
@@ -216,7 +218,7 @@ __global__ void multi_query_append_attention_c8_kernel(
                          kv_len - q_len +
                              tile_id * num_rows_per_block / GROUP_SIZE,
                          chunk_start)))
-              : chunk_len) /
+              : mask_offset ? 0 : chunk_len) /
       (num_frags_z * 16);
 
   uint32_t k_smem_offset_r =
@@ -300,12 +302,15 @@ __global__ void multi_query_append_attention_c8_kernel(
              NUM_WARPS,
              num_frags_x,
              num_frags_y,
-             num_frags_z>(q_base_seq_id_this_block,
+             num_frags_z>(nullptr,
+                          q_base_seq_id_this_block,
                           kv_idx_base,
                           q_len,
                           kv_len,
                           chunk_end,
-                          s_frag);
+                          -1,
+                          s_frag,
+                          mask_offset_this_seq);
     }
 
     // update m,d
@@ -474,6 +479,8 @@ __global__ void multi_query_append_attention_c8_warp1_4_kernel(
     const int *__restrict__ tile_ids_per_batch,
     const int *__restrict__ cu_seqlens_q,
     const int *__restrict__ block_table,  // [bsz, block_num_per_seq]
+    const int *__restrict__ mask_offset,
+    const bool *__restrict__ attn_mask,   // [bsz, max_q, max_q] for tree-mask
     const int max_seq_len,
     const int max_dec_len,
     const int max_block_num_per_seq,
@@ -487,7 +494,8 @@ __global__ void multi_query_append_attention_c8_warp1_4_kernel(
     float *__restrict__ tmp_m,      // [token_num, num_chunks, num_heads]
     float *__restrict__ tmp_d,      // [token_num, num_chunks, num_heads]
     OutT *__restrict__ out,
-    const int speculate_max_draft_token_num = 5) {
+    const int speculate_max_draft_token_num = 5,
+    const uint32_t attn_mask_len = -1) {
   constexpr uint32_t num_vecs_per_head = HEAD_DIM / num_elems_per_128b<T>();
   constexpr uint32_t num_vecs_per_head_k =
       HEAD_DIM / num_elems_per_128b<CacheT>();
@@ -601,7 +609,7 @@ __global__ void multi_query_append_attention_c8_warp1_4_kernel(
           tid % 8 * num_elems_per_128b<T>();
     }
   }
-
+  const int *mask_offset_this_seq = mask_offset ? mask_offset + q_start_seq_id : nullptr;
   smem_t qo_smem(smem);
 
   uint32_t q_smem_offset_r = smem_t::get_permuted_offset<num_vecs_per_head>(
@@ -642,7 +650,7 @@ __global__ void multi_query_append_attention_c8_warp1_4_kernel(
                          kv_len - q_len +
                              tile_id * num_rows_per_block / GROUP_SIZE,
                          chunk_start)))
-              : chunk_len) /
+              : mask_offset ? 0 : chunk_len) /
       (NUM_WARP_KV * num_frags_z * 16);
 
   uint32_t k_smem_offset_r =
@@ -728,12 +736,16 @@ __global__ void multi_query_append_attention_c8_warp1_4_kernel(
              NUM_WARPS,
              num_frags_x,
              num_frags_y,
-             num_frags_z>(q_base_seq_id_this_block,
+             num_frags_z>(attn_mask ? attn_mask + batch_id * attn_mask_len *attn_mask_len : nullptr,
+                          q_base_seq_id_this_block,
                           kv_idx_base + wid * num_frags_z * 16,
                           q_len,
                           kv_len,
                           chunk_end,
-                          s_frag);
+                          attn_mask_len,
+                          s_frag,
+                          mask_offset_this_seq);
+
     }
 
     // update m,d
@@ -1054,6 +1066,7 @@ void MultiQueryAppendC8Attention(
           tile_ids_per_batch.data<int>(),
           cu_seqlens_q.data<int>(),
           block_table.data<int>(),
+          meta_data.mask_offset,
           max_seq_len,
           max_dec_len,
           max_block_num_per_seq,
@@ -1111,6 +1124,7 @@ void MultiQueryAppendC8Attention(
           tile_ids_per_batch.data<int>(),
           cu_seqlens_q.data<int>(),
           block_table.data<int>(),
+          meta_data.mask_offset,
           max_seq_len,
           max_dec_len,
           max_block_num_per_seq,
@@ -1254,10 +1268,17 @@ void MultiQueryAppendC8Attention(
       chunk_size = static_cast<uint32_t>(encoder_max_partition_size);
     }
 
-    const int num_chunks = div_up(max_dec_len, chunk_size);
+    const int num_chunks = div_up(max_seq_len, chunk_size);
+    uint32_t attn_mask_len;
+    if (attn_mask) {
+        attn_mask_len = attn_mask.get().shape()[1];
+    } else {
+        attn_mask_len = -1;
+    }
+
     dim3 grids(num_blocks_x_cpu, num_chunks, kv_num_heads);
     dim3 blocks(32, num_warps);
-    if (num_chunks <= 1) {
+    if (num_chunks <= 0) {
       auto nosplit_kv_kernel =
           multi_query_append_attention_c8_warp1_4_kernel<NV_TYPE,
                                                          uint8_t,
@@ -1318,6 +1339,9 @@ void MultiQueryAppendC8Attention(
           tile_ids_per_batch.data<int>(),
           cu_seqlens_q.data<int>(),
           block_table.data<int>(),
+          meta_data.mask_offset,
+          attn_mask ? const_cast<bool *>(attn_mask.get().data<bool>())
+                        : nullptr,
           max_seq_len,
           max_dec_len,
           max_block_num_per_seq,
@@ -1330,7 +1354,8 @@ void MultiQueryAppendC8Attention(
           nullptr,
           nullptr,
           reinterpret_cast<OUT_NV_TYPE *>(out->data<OutT>()),
-          speculate_max_draft_token_num);
+          speculate_max_draft_token_num,
+          attn_mask_len);
     } else {
       phi::Allocator::AllocationPtr tmp_workspace, tmp_m, tmp_d;
       if (is_decoder) {
@@ -1377,8 +1402,8 @@ void MultiQueryAppendC8Attention(
           reinterpret_cast<NV_TYPE *>(const_cast<T *>(cache_k_scale.data<T>())),
           reinterpret_cast<NV_TYPE *>(const_cast<T *>(cache_v_scale.data<T>())),
           shift_bias ? reinterpret_cast<NV_TYPE *>(
-                           const_cast<T *>(shift_bias.get().data<T>()))
-                     : nullptr,
+                            const_cast<T *>(shift_bias.get().data<T>()))
+                      : nullptr,
           smooth_weight ? reinterpret_cast<NV_TYPE *>(
                               const_cast<T *>(smooth_weight.get().data<T>()))
                         : nullptr,
@@ -1388,6 +1413,9 @@ void MultiQueryAppendC8Attention(
           tile_ids_per_batch.data<int>(),
           cu_seqlens_q.data<int>(),
           block_table.data<int>(),
+          meta_data.mask_offset,
+          attn_mask ? const_cast<bool *>(attn_mask.get().data<bool>())
+                        : nullptr,
           max_seq_len,
           max_dec_len,
           max_block_num_per_seq,
@@ -1400,7 +1428,8 @@ void MultiQueryAppendC8Attention(
           static_cast<float *>(tmp_m->ptr()),
           static_cast<float *>(tmp_d->ptr()),
           reinterpret_cast<OUT_NV_TYPE *>(out->data<OutT>()),
-          speculate_max_draft_token_num);
+          speculate_max_draft_token_num,
+          attn_mask_len);
       // merge
       constexpr int vec_size = num_elems_per_128b<NV_TYPE>();
       if (is_decoder) {
@@ -1418,8 +1447,8 @@ void MultiQueryAppendC8Attention(
                 seq_lens_encoder.data<int>(),
                 cu_seqlens_q.data<int>(),
                 shift_bias ? reinterpret_cast<NV_TYPE *>(
-                                 const_cast<T *>(shift_bias.get().data<T>()))
-                           : nullptr,
+                                  const_cast<T *>(shift_bias.get().data<T>()))
+                            : nullptr,
                 smooth_weight ? reinterpret_cast<NV_TYPE *>(const_cast<T *>(
                                     smooth_weight.get().data<T>()))
                               : nullptr,
@@ -1436,14 +1465,14 @@ void MultiQueryAppendC8Attention(
         constexpr int blockx = HEAD_DIM / vec_size;
         constexpr int blocky = (128 + blockx - 1) / blockx;
         dim3 grids_merge(min(sm_count * 4, token_num),
-                         num_heads);
+                          num_heads);
         dim3 blocks_merge(blockx, blocky);
         merge_multi_chunks_v2_kernel<NV_TYPE,
-                                     vec_size,
-                                     blocky,
-                                     HEAD_DIM,
-                                     OUT_NV_TYPE,
-                                     ENABLE_PREFILL>
+                                      vec_size,
+                                      blocky,
+                                      HEAD_DIM,
+                                      OUT_NV_TYPE,
+                                      ENABLE_PREFILL>
             <<<grids_merge, blocks_merge, 0, stream>>>(
                 reinterpret_cast<NV_TYPE *>(tmp_workspace->ptr()),
                 static_cast<float *>(tmp_m->ptr()),
@@ -1454,8 +1483,8 @@ void MultiQueryAppendC8Attention(
                 batch_id_per_token.data<int>(),
                 cu_seqlens_q.data<int>(),
                 shift_bias ? reinterpret_cast<NV_TYPE *>(
-                                 const_cast<T *>(shift_bias.get().data<T>()))
-                           : nullptr,
+                                  const_cast<T *>(shift_bias.get().data<T>()))
+                            : nullptr,
                 smooth_weight ? reinterpret_cast<NV_TYPE *>(const_cast<T *>(
                                     smooth_weight.get().data<T>()))
                               : nullptr,

custom_ops/gpu_ops/append_attn/append_attention_func.cuh

@@ -905,12 +905,15 @@ template <typename T,
           uint32_t num_frags_y,
           uint32_t num_frags_z,
           bool IS_SYSTEM = false>
-__device__ __forceinline__ void mask_s(const uint32_t qo_idx_base,
+__device__ __forceinline__ void mask_s(const bool* attn_mask,
+                                       const uint32_t qo_idx_base,
                                        const uint32_t kv_idx_base,
                                        const uint32_t qo_len,
                                        const uint32_t kv_len,
                                        const uint32_t chunk_end,
-                                       float (*s_frag)[num_frags_z][8]) {
+                                       const uint32_t attn_mask_len,
+                                       float (*s_frag)[num_frags_z][8],
+                                       const int *mask_offset = nullptr) {
   const uint32_t tx = threadIdx.x;
 #pragma unroll
   for (uint32_t fx = 0; fx < num_frags_x; ++fx) {
@@ -924,10 +927,21 @@ __device__ __forceinline__ void mask_s(const uint32_t qo_idx_base,
                                  group_size,
                          kv_idx = kv_idx_base + fz * 16 + 2 * (tx % 4) +
                                   8 * (reg_id / 4) + reg_id % 2;
-          const bool out_of_boundary =
-              (causal
-                   ? (kv_idx > kv_len + q_idx - qo_len || (kv_idx >= chunk_end))
-                   : kv_idx >= chunk_end);
+          bool out_of_boundary;
+          if (mask_offset) {
+            out_of_boundary = q_idx < qo_len ? (kv_idx > mask_offset[q_idx]) : true;
+          } else {
+            out_of_boundary =
+                (causal
+                    ? (kv_idx > kv_len + q_idx - qo_len || (kv_idx >= chunk_end))
+                    : kv_idx >= chunk_end);
+            if (attn_mask != nullptr && kv_idx > kv_len - qo_len && kv_idx < chunk_end && q_idx < attn_mask_len) {
+              const int32_t mask_idx = q_idx * attn_mask_len + kv_idx - kv_len + qo_len;
+              bool mask = attn_mask[mask_idx];
+              out_of_boundary |= mask;
+            }
+          }
+
           if constexpr (std::is_same<T, half>::value) {
             s_frag[fx][fz][reg_id] =
                 out_of_boundary ? -5e4f : s_frag[fx][fz][reg_id];
@@ -935,6 +949,7 @@ __device__ __forceinline__ void mask_s(const uint32_t qo_idx_base,
             s_frag[fx][fz][reg_id] =
                 out_of_boundary ? -3.0e+30f : s_frag[fx][fz][reg_id];
           }
+          // printf("tid: %d. qk[%u,%u] = %f, mask: %d \n ", threadIdx.x, kv_idx, q_idx, static_cast<float>(s_frag[fx][fz][reg_id]), int(out_of_boundary));
         } else {
           const uint32_t q_idx = qo_idx_base,
                          kv_idx = kv_idx_base + fz * 16 + 2 * (tx % 4) +

custom_ops/gpu_ops/append_attn/decoder_write_cache_with_rope_impl.cuh

@@ -18,6 +18,142 @@
 #include "mma_tensor_op.cuh"
 #include "utils.cuh"
 
+template <typename T, int VecSize = 1>
+__global__ void append_decode_cache_T_rope_qk_norm_kernel(
+    const T* __restrict__ quant_qkv,  // [bsz, num_heads + 2 * kv_num_heads,
+                                      // head_size]
+    T* __restrict__ key_cache,    // [num_blocks, kv_num_heads, block_size,
+                                  // head_size // 2]
+    T* __restrict__ value_cache,  // [num_blocks, kv_num_heads, block_size,
+                                  // head_size // 2]
+    T* __restrict__ qkv_out,
+    const int* __restrict__ block_tables,     // [bsz, max_blocks_per_seq]
+    const int* __restrict__ batch_id_per_token,  // [num_tokens]
+    const int* __restrict__ cu_seqlens_q,
+    const int* __restrict__ seq_lens,          // [bsz]
+    const int* __restrict__ seq_lens_encoder,  // [bsz]
+    const float* __restrict__ cos_emb,
+    const float* __restrict__ sin_emb,
+    const int max_seq_len,
+    const int max_blocks_per_seq,
+    const int num_heads,
+    const int head_size,
+    const int block_size,
+    const uint32_t elem_cnt,
+    const int kv_num_heads,
+    const bool rope_3d,
+    const float* q_norm_weight,
+    const float* k_norm_weight,
+    const float rms_norm_eps) {
+  using LoadT = AlignedVector<T, VecSize>;
+  using LoadBiasT = AlignedVector<T, VecSize>;
+  using LoadKVT = AlignedVector<T, VecSize>;
+  constexpr int HalfVecSize = VecSize / 2;
+  using LoadEmbT = AlignedVector<float, HalfVecSize>;
+  using LoadFloat = AlignedVector<float, VecSize>;
+  LoadT src_vec;
+  LoadBiasT out_vec;
+  LoadKVT cache_vec;
+  LoadEmbT cos_emb_vec;
+  LoadEmbT sin_emb_vec;
+  LoadFloat tmp_vec;
+  LoadFloat q_norm_vec, k_norm_vec;
+
+  int64_t global_warp_idx = blockDim.y * blockIdx.x + threadIdx.y;
+  int64_t all_warp_num = gridDim.x * blockDim.y;
+  int64_t all_head_dim = elem_cnt / head_size;
+
+  const int64_t hidden_size = (num_heads + 2 * kv_num_heads) * head_size;
+  const int half_head_size = head_size / 2;
+  for (int gloabl_hi = global_warp_idx; gloabl_hi < all_head_dim; gloabl_hi += all_warp_num) {
+    int64_t linear_index = gloabl_hi * head_size + threadIdx.x * VecSize;
+    const int ori_bi = linear_index / hidden_size;
+    const int bias = linear_index % hidden_size;
+    const int hi = bias / head_size;  // q + k + v
+    const int h_bias = bias % head_size;
+    const int start_token_idx = cu_seqlens_q[ori_bi];
+    if (seq_lens_encoder[ori_bi] > 0) return;
+    const int write_seq_id = seq_lens[ori_bi];
+    if (write_seq_id == 0) continue;
+
+    const int* block_table_now = nullptr;
+
+    block_table_now = block_tables + ori_bi * max_blocks_per_seq;
+    const int block_idx = block_table_now[write_seq_id / block_size];
+    const int block_offset = write_seq_id % block_size;
+    const uint32_t ori_idx =
+        start_token_idx * hidden_size + hi * head_size + h_bias;
+
+    const int bias_idx = hi * head_size + h_bias;
+    Load<T, VecSize>(&quant_qkv[ori_idx], &src_vec);
+    if (hi < num_heads + kv_num_heads) {
+      // q k rope
+      const uint32_t emb_idx = write_seq_id * half_head_size + h_bias / 2;
+      uint32_t new_emb_idx = rope_3d ? emb_idx + ori_bi * max_seq_len * head_size : emb_idx;
+      Load<float, HalfVecSize>(&cos_emb[new_emb_idx], &cos_emb_vec);
+      Load<float, HalfVecSize>(&sin_emb[new_emb_idx], &sin_emb_vec);
+    }
+    float thread_m2 = 0.0f;
+    float warp_m2 = 0.0f;
+
+#pragma unroll
+    for (int i = 0; i < HalfVecSize; i++) {
+      // dequant + add_bias + rope
+      float input_left = static_cast<float>(src_vec[2 * i]);
+      float input_right = static_cast<float>(src_vec[2 * i + 1]);
+
+      if (hi < num_heads + kv_num_heads) {
+        const float cos_tmp = cos_emb_vec[i];
+        const float sin_tmp = sin_emb_vec[i];
+        float tmp1 = input_left * cos_tmp - input_right * sin_tmp;
+        float tmp2 = input_right * cos_tmp + input_left * sin_tmp;
+        thread_m2 += tmp1 * tmp1 + tmp2 * tmp2;
+        tmp_vec[2 * i] = tmp1;
+        tmp_vec[2 * i + 1] = tmp2;
+      } else {
+        out_vec[2 * i] = src_vec[2 * i];
+        out_vec[2 * i + 1] = src_vec[2 * i + 1];
+      }
+    }
+    if (hi < (num_heads + kv_num_heads)) { // q k
+      WelfordWarpAllReduce<float, 32>(thread_m2, &warp_m2);
+      float row_variance =
+          max(warp_m2 / head_size, 0.0f);
+      float row_inv_var = Rsqrt(row_variance + rms_norm_eps);
+
+      if (hi < num_heads) { // q
+        Load<float, VecSize>(&q_norm_weight[threadIdx.x * VecSize], &q_norm_vec);
+        #pragma unroll
+        for (int i = 0; i < VecSize; i++) {
+          out_vec[i] = static_cast<T>(tmp_vec[i] * row_inv_var * q_norm_vec[i]);
+        }
+      } else { // k
+        Load<float, VecSize>(&k_norm_weight[threadIdx.x * VecSize], &k_norm_vec);
+        for (int i = 0; i < VecSize; i++) {
+          out_vec[i] = static_cast<T>(tmp_vec[i] * row_inv_var * k_norm_vec[i]);
+        }
+      }
+    }
+    if (hi < num_heads) {
+      // write q
+      Store<T, VecSize>(out_vec, &qkv_out[ori_idx]);
+    } else {
+      // quant + write k/v
+      const uint32_t kv_head_idx = (hi - num_heads) % kv_num_heads;
+      const uint32_t tgt_idx =
+          block_idx * kv_num_heads * block_size * head_size +
+          kv_head_idx * block_size * head_size + block_offset * head_size +
+          h_bias;
+      if (hi < num_heads + kv_num_heads) {
+        Store<T, VecSize>(out_vec, &key_cache[tgt_idx]);
+      } else {
+        Store<T, VecSize>(out_vec, &value_cache[tgt_idx]);
+      }
+    }
+
+  }
+}
+
 template <typename T, int VecSize = 1>
 __global__ void append_decode_cache_T_rope_kernel(
     const T* __restrict__ quant_qkv,  // [bsz, num_heads + 2 * kv_num_heads,
@@ -199,8 +335,9 @@ __global__ void append_decode_cache_T_rope_kernel(
     if (hi < num_heads + kv_num_heads) {
       // q k rope
       const uint32_t emb_idx = write_seq_id * half_head_size + h_bias / 2;
-      Load<float, HalfVecSize>(&cos_emb[emb_idx], &cos_emb_vec);
-      Load<float, HalfVecSize>(&sin_emb[emb_idx], &sin_emb_vec);
+      uint32_t new_emb_idx = rope_3d ? emb_idx + ori_bi * max_seq_len * head_size : emb_idx;
+      Load<float, HalfVecSize>(&cos_emb[new_emb_idx], &cos_emb_vec);
+      Load<float, HalfVecSize>(&sin_emb[new_emb_idx], &sin_emb_vec);
     }
 #pragma unroll
     for (int i = 0; i < HalfVecSize; i++) {
@@ -244,6 +381,142 @@ __global__ void append_decode_cache_T_rope_kernel(
   }
 }
 
+template <typename T, int VecSize = 1>
+__global__ void append_decode_cache_T_neox_partial_rope_kernel(
+    const T* __restrict__ qkv,    // [bsz, num_heads + 2 * kv_num_heads,
+                                  // head_size]
+    T* __restrict__ key_cache,    // [num_blocks, kv_num_heads, block_size,
+                                  // head_size // 2]
+    T* __restrict__ value_cache,  // [num_blocks, kv_num_heads, block_size,
+                                  // head_size // 2]
+    T* __restrict__ qkv_out,
+    const int* __restrict__ block_tables,     // [bsz, max_blocks_per_seq]
+    const int* __restrict__ cu_seqlens_q,
+    const int* __restrict__ seq_lens,          // [bsz]
+    const int* __restrict__ seq_lens_encoder,  // [bsz]
+    const float* __restrict__ cos_emb,         // [2, 1, max_model_len, 1, rotary_dim/2]
+    const float* __restrict__ sin_emb,         // [2, 1, max_model_len, 1, rotary_dim/2]
+    const int max_seq_len,
+    const int max_blocks_per_seq,
+    const int num_heads,
+    const int head_size,
+    const int rotary_dim,
+    const int block_size,
+    const uint32_t elem_cnt,
+    const int kv_num_heads,
+    const bool rope_3d) {
+  using LoadT = AlignedVector<T, VecSize>;
+  using LoadBiasT = AlignedVector<T, VecSize>;
+  using LoadKVT = AlignedVector<T, VecSize>;
+  constexpr int HalfVecSize = VecSize / 2;
+  using LoadEmbT = AlignedVector<float, VecSize>;
+
+  LoadT left_vec, right_vec;
+  LoadBiasT left_bias_vec, right_bias_vec;
+  LoadKVT left_cache_vec, right_cache_vec;
+  LoadEmbT cos_emb_vec;
+  LoadEmbT sin_emb_vec;
+
+  int64_t global_thread_idx = blockDim.x * blockIdx.x + threadIdx.x;
+  const int half_head_size = head_size / 2;
+  const int half_rotary_dim = rotary_dim / 2;
+  const int64_t hidden_size = (num_heads + 2 * kv_num_heads) * head_size;
+  const int64_t half_hidden_size = hidden_size / 2;
+  // const int64_t offset = 2 * hidden_size;
+
+  for (int32_t linear_index = global_thread_idx * VecSize,
+               step = gridDim.x * blockDim.x * VecSize;
+       linear_index < elem_cnt;
+       linear_index += step) {
+    const int ori_bi = linear_index / half_hidden_size;
+    const int bias = linear_index % half_hidden_size;
+    const int hi = bias / half_head_size;  // q + k + v
+    const int h_bias = bias % half_head_size;
+    if (hi < num_heads && h_bias >= half_rotary_dim){
+      continue;
+    }
+    if (seq_lens_encoder[ori_bi] > 0) continue;
+    const int write_seq_id = seq_lens[ori_bi];
+    if (write_seq_id == 0) continue;
+    const int start_token_idx = cu_seqlens_q[ori_bi];
+
+    const int* block_table_now = nullptr;
+
+    block_table_now = block_tables + ori_bi * max_blocks_per_seq;
+    const int block_idx = block_table_now[write_seq_id / block_size];
+    const int block_offset = write_seq_id % block_size;
+    uint32_t ori_idx_left =
+        start_token_idx * hidden_size + hi * head_size + h_bias;
+    uint32_t ori_idx_right = ori_idx_left + half_head_size;
+    if (hi < num_heads){
+      ori_idx_right = ori_idx_left + half_rotary_dim;
+    }else if (hi < num_heads + kv_num_heads){
+      if (h_bias < half_rotary_dim){
+        ori_idx_right = ori_idx_left + half_rotary_dim;
+      }else{
+        ori_idx_left = ori_idx_left + half_rotary_dim;
+        ori_idx_right = ori_idx_left + half_rotary_dim;
+      }
+    }
+
+    Load<T, VecSize>(&qkv[ori_idx_left], &left_vec);
+    Load<T, VecSize>(&qkv[ori_idx_right], &right_vec);
+
+    if (hi < num_heads + kv_num_heads) {
+      // q k rope
+      const uint32_t emb_idx = write_seq_id * half_rotary_dim + h_bias;
+      uint32_t new_emb_idx = rope_3d ? emb_idx + ori_bi * max_seq_len * head_size * 2 : emb_idx;
+      if (h_bias < half_rotary_dim){
+        Load<float, VecSize>(&cos_emb[new_emb_idx], &cos_emb_vec);
+        Load<float, VecSize>(&sin_emb[new_emb_idx], &sin_emb_vec);
+      }
+    }
+#pragma unroll
+    for (int i = 0; i < VecSize; i++) {
+      // rope
+      float input_left = static_cast<float>(left_vec[i]);
+      float input_right = static_cast<float>(right_vec[i]);
+      if (hi < num_heads + kv_num_heads && h_bias < half_rotary_dim) {
+        const float cos_tmp = cos_emb_vec[i];
+        const float sin_tmp = sin_emb_vec[i];
+        left_bias_vec[i] =
+            static_cast<T>(input_left * cos_tmp - input_right * sin_tmp);
+        right_bias_vec[i] =
+            static_cast<T>(input_right * cos_tmp + input_left * sin_tmp);
+      } else {
+        left_bias_vec[i] = static_cast<T>(input_left);
+        right_bias_vec[i] = static_cast<T>(input_right);
+      }
+    }
+    if (hi < num_heads) {
+      // write q
+      Store<T, VecSize>(left_bias_vec, &qkv_out[ori_idx_left]);
+      Store<T, VecSize>(right_bias_vec, &qkv_out[ori_idx_right]);
+    } else {
+      // write k/v
+      const uint32_t kv_head_idx = (hi - num_heads) % kv_num_heads;
+      uint32_t tgt_idx_left =
+          block_idx * kv_num_heads * block_size * head_size +
+          kv_head_idx * block_size * head_size + block_offset * head_size +
+          h_bias;
+      uint32_t tgt_idx_right = tgt_idx_left + half_head_size;
+      if (hi < num_heads + kv_num_heads) {
+        if (h_bias < half_rotary_dim) {
+          tgt_idx_right = tgt_idx_left + half_rotary_dim;
+        }else{
+          tgt_idx_left = tgt_idx_left + half_rotary_dim;
+          tgt_idx_right = tgt_idx_left + half_rotary_dim;
+        }
+        Store<T, VecSize>(left_bias_vec, &key_cache[tgt_idx_left]);
+        Store<T, VecSize>(right_bias_vec, &key_cache[tgt_idx_right]);
+      } else {
+        Store<T, VecSize>(left_bias_vec, &value_cache[tgt_idx_left]);
+        Store<T, VecSize>(right_bias_vec, &value_cache[tgt_idx_right]);
+      }
+    }
+  }
+}
+
 template <typename T, int VecSize = 1>
 __global__ void append_decode_cache_T_neox_rope_kernel(
     const T* __restrict__ qkv,    // [bsz, num_heads + 2 * kv_num_heads,
@@ -266,7 +539,8 @@ __global__ void append_decode_cache_T_neox_rope_kernel(
     const int head_size,
     const int block_size,
     const uint32_t elem_cnt,
-    const int kv_num_heads) {
+    const int kv_num_heads,
+    const bool rope_3d) {
   using LoadT = AlignedVector<T, VecSize>;
   using LoadBiasT = AlignedVector<T, VecSize>;
   using LoadKVT = AlignedVector<T, VecSize>;
@@ -313,8 +587,9 @@ __global__ void append_decode_cache_T_neox_rope_kernel(
     if (hi < num_heads + kv_num_heads) {
       // q k rope
       const uint32_t emb_idx = write_seq_id * head_size + h_bias;
-      Load<float, VecSize>(&cos_emb[emb_idx], &cos_emb_vec);
-      Load<float, VecSize>(&sin_emb[emb_idx], &sin_emb_vec);
+      uint32_t new_emb_idx = rope_3d ? emb_idx + ori_bi * max_seq_len * head_size * 2 : emb_idx;
+      Load<float, VecSize>(&cos_emb[new_emb_idx], &cos_emb_vec);
+      Load<float, VecSize>(&sin_emb[new_emb_idx], &sin_emb_vec);
     }
 #pragma unroll
     for (int i = 0; i < VecSize; i++) {
@@ -382,7 +657,8 @@ __global__ void append_decode_cache_T_neox_rope_kernel(
     const int head_size,
     const int block_size,
     const uint32_t elem_cnt,
-    const int kv_num_heads) {
+    const int kv_num_heads,
+    const bool rope_3d) {
   using LoadT = AlignedVector<int, VecSize>;
   using LoadBiasT = AlignedVector<T, VecSize>;
   using LoadOutScaleT = AlignedVector<float, VecSize>;
@@ -439,8 +715,9 @@ __global__ void append_decode_cache_T_neox_rope_kernel(
     if (hi < num_heads + kv_num_heads) {
       // q k rope
       const uint32_t emb_idx = write_seq_id * head_size + h_bias;
-      Load<float, VecSize>(&cos_emb[emb_idx], &cos_emb_vec);
-      Load<float, VecSize>(&sin_emb[emb_idx], &sin_emb_vec);
+      uint32_t new_emb_idx = rope_3d ? emb_idx + ori_bi * max_seq_len * head_size * 2 : emb_idx;
+      Load<float, VecSize>(&cos_emb[new_emb_idx], &cos_emb_vec);
+      Load<float, VecSize>(&sin_emb[new_emb_idx], &sin_emb_vec);
     }
 #pragma unroll
     for (int i = 0; i < VecSize; i++) {
@@ -512,7 +789,8 @@ __global__ void append_decode_cache_int8_rope_kernel(
     const int block_size,
     const float max_bound,
     const float min_bound,
-    const int kv_num_heads) {
+    const int kv_num_heads,
+    const bool rope_3d) {
   static_assert(HeadDim == 128, "just support HeadDim be 128 now!");
   static_assert(VecSize == 4, "just support VecSize be 4 now, 32 * 4!");
   constexpr int NUM_WARPS = 4;
@@ -555,8 +833,9 @@ __global__ void append_decode_cache_int8_rope_kernel(
 
       // q rope
       const uint32_t emb_idx = write_seq_id * half_head_size + head_bias / 2;
-      Load<float, HalfVecSize>(&cos_emb[emb_idx], &cos_emb_vec);
-      Load<float, HalfVecSize>(&sin_emb[emb_idx], &sin_emb_vec);
+      uint32_t new_emb_idx = rope_3d ? emb_idx + bid * max_seq_len * HeadDim : emb_idx;
+      Load<float, HalfVecSize>(&cos_emb[new_emb_idx], &cos_emb_vec);
+      Load<float, HalfVecSize>(&sin_emb[new_emb_idx], &sin_emb_vec);
 #pragma unroll
       for (int i = 0; i < HalfVecSize; i++) {
         // dequant + add_bias + rope
@@ -633,10 +912,11 @@ __global__ void append_decode_cache_int8_rope_kernel(
     const T *cache_v_scale_cur = cache_v_scale + v_head_idx * HeadDim + head_bias;
     if (head_idx < num_heads + kv_num_heads) {
       const uint32_t emb_idx = write_seq_id * half_head_size + head_bias / 2;
-      Load<float, 1>(&cos_emb[emb_idx], &cos_emb_vec1);
-      Load<float, 1>(&cos_emb[emb_idx + 4], &cos_emb_vec2);
-      Load<float, 1>(&sin_emb[emb_idx], &sin_emb_vec1);
-      Load<float, 1>(&sin_emb[emb_idx + 4], &sin_emb_vec2);
+      uint32_t new_emb_idx = rope_3d ? emb_idx + bid * max_seq_len * HeadDim : emb_idx;
+      Load<float, 1>(&cos_emb[new_emb_idx], &cos_emb_vec1);
+      Load<float, 1>(&cos_emb[new_emb_idx + 4], &cos_emb_vec2);
+      Load<float, 1>(&sin_emb[new_emb_idx], &sin_emb_vec1);
+      Load<float, 1>(&sin_emb[new_emb_idx + 4], &sin_emb_vec2);
       if constexpr (!is_scale_channel_wise) {
         scale = __ldg(&cache_k_scale[kv_head_idx]);
       }
@@ -763,7 +1043,8 @@ __global__ void append_decode_cache_int8_rope_kernel(
     const int block_size,
     const float max_bound,
     const float min_bound,
-    const int kv_num_heads) {
+    const int kv_num_heads,
+    const bool rope_3d) {
   static_assert(HeadDim == 128, "just support HeadDim be 128 now!");
   static_assert(VecSize == 4, "just support VecSize be 4 now, 32 * 4!");
   constexpr int NUM_WARPS = 4;
@@ -813,9 +1094,10 @@ __global__ void append_decode_cache_int8_rope_kernel(
 
       // q rope
       const uint32_t emb_idx = write_seq_id * half_head_size + head_bias / 2;
-      Load<float, HalfVecSize>(&cos_emb[emb_idx], &cos_emb_vec);
+      uint32_t new_emb_idx = rope_3d ? emb_idx + bid * max_seq_len * HeadDim : emb_idx;
+      Load<float, HalfVecSize>(&cos_emb[new_emb_idx], &cos_emb_vec);
 
-      Load<float, HalfVecSize>(&sin_emb[emb_idx], &sin_emb_vec);
+      Load<float, HalfVecSize>(&sin_emb[new_emb_idx], &sin_emb_vec);
 
 #pragma unroll
       for (int i = 0; i < HalfVecSize; i++) {
@@ -908,10 +1190,11 @@ __global__ void append_decode_cache_int8_rope_kernel(
     const T *cache_v_scale_cur = cache_v_scales + v_head_idx * HeadDim + head_bias;
     if (head_idx < num_heads + kv_num_heads) {
       const uint32_t emb_idx = write_seq_id * half_head_size + head_bias / 2;
-      Load<float, 1>(&cos_emb[emb_idx], &cos_emb_vec1);
-      Load<float, 1>(&cos_emb[emb_idx + 4], &cos_emb_vec2);
-      Load<float, 1>(&sin_emb[emb_idx], &sin_emb_vec1);
-      Load<float, 1>(&sin_emb[emb_idx + 4], &sin_emb_vec2);
+      uint32_t new_emb_idx = rope_3d ? emb_idx + bid * max_seq_len * HeadDim : emb_idx;
+      Load<float, 1>(&cos_emb[new_emb_idx], &cos_emb_vec1);
+      Load<float, 1>(&cos_emb[new_emb_idx + 4], &cos_emb_vec2);
+      Load<float, 1>(&sin_emb[new_emb_idx], &sin_emb_vec1);
+      Load<float, 1>(&sin_emb[new_emb_idx + 4], &sin_emb_vec2);
       if constexpr (!is_scale_channel_wise) {
         scale = __ldg(&cache_k_scales[kv_head_idx]);
       }
@@ -1061,7 +1344,8 @@ __global__ void append_decode_cache_int8_neox_rope_kernel(
     const int block_size,
     const float max_bound,
     const float min_bound,
-    const int kv_num_heads) {
+    const int kv_num_heads,
+    const bool rope_3d) {
   static_assert(HeadDim == 128, "just support HeadDim be 128 now!");
   static_assert(VecSize == 4, "just support VecSize be 4 now, 32 * 4!");
   constexpr int NUM_WARPS = 4;
@@ -1109,8 +1393,9 @@ __global__ void append_decode_cache_int8_neox_rope_kernel(
 
       // q rope
       const uint32_t emb_idx = write_seq_id * HeadDim + head_bias;
-      Load<float, VecSize>(&cos_emb[emb_idx], &cos_emb_vec);
-      Load<float, VecSize>(&sin_emb[emb_idx], &sin_emb_vec);
+      uint32_t new_emb_idx = rope_3d ? emb_idx + bid * max_seq_len * HeadDim * 2 : emb_idx;
+      Load<float, VecSize>(&cos_emb[new_emb_idx], &cos_emb_vec);
+      Load<float, VecSize>(&sin_emb[new_emb_idx], &sin_emb_vec);
 
 #pragma unroll
       for (int i = 0; i < VecSize; i++) {
@@ -1191,10 +1476,11 @@ __global__ void append_decode_cache_int8_neox_rope_kernel(
 
         T scale;
         const uint32_t emb_idx = write_seq_id * HeadDim + head_bias;
-        Load<float, HALF_K_VEC_SIZE>(&cos_emb[emb_idx], &cos_emb_vec1);
-        Load<float, HALF_K_VEC_SIZE>(&cos_emb[emb_idx + 8], &cos_emb_vec2);
-        Load<float, HALF_K_VEC_SIZE>(&sin_emb[emb_idx], &sin_emb_vec1);
-        Load<float, HALF_K_VEC_SIZE>(&sin_emb[emb_idx + 8], &sin_emb_vec2);
+        uint32_t new_emb_idx = rope_3d ? emb_idx + bid * max_seq_len * HeadDim * 2 : emb_idx;
+        Load<float, HALF_K_VEC_SIZE>(&cos_emb[new_emb_idx], &cos_emb_vec1);
+        Load<float, HALF_K_VEC_SIZE>(&cos_emb[new_emb_idx + 8], &cos_emb_vec2);
+        Load<float, HALF_K_VEC_SIZE>(&sin_emb[new_emb_idx], &sin_emb_vec1);
+        Load<float, HALF_K_VEC_SIZE>(&sin_emb[new_emb_idx + 8], &sin_emb_vec2);
         scale = __ldg(&cache_k_scales[kv_head_idx]);
 #pragma unroll
         for (int i = 0; i < HALF_K_VEC_SIZE; i++) {
@@ -1364,7 +1650,8 @@ __global__ void append_decode_cache_int8_neox_rope_kernel(
     const int block_size,
     const float max_bound,
     const float min_bound,
-    const int kv_num_heads) {
+    const int kv_num_heads,
+    const bool rope_3d) {
   static_assert(HeadDim == 128, "just support HeadDim be 128 now!");
   static_assert(VecSize == 4, "just support VecSize be 4 now, 32 * 4!");
   constexpr int NUM_WARPS = 4;
@@ -1424,8 +1711,10 @@ __global__ void append_decode_cache_int8_neox_rope_kernel(
 
       // q rope
       const uint32_t emb_idx = write_seq_id * HeadDim + head_bias;
-      Load<float, VecSize>(&cos_emb[emb_idx], &cos_emb_vec);
-      Load<float, VecSize>(&sin_emb[emb_idx], &sin_emb_vec);
+      uint32_t new_emb_idx = rope_3d ? emb_idx + bid * max_seq_len * HeadDim * 2 : emb_idx;
+      Load<float, VecSize>(&cos_emb[new_emb_idx], &cos_emb_vec);
+      Load<float, VecSize>(&sin_emb[new_emb_idx], &sin_emb_vec);
+
 
 #pragma unroll
       for (int i = 0; i < VecSize; i++) {
@@ -1533,10 +1822,11 @@ __global__ void append_decode_cache_int8_neox_rope_kernel(
 
         T scale;
         const uint32_t emb_idx = write_seq_id * HeadDim + head_bias;
-        Load<float, HALF_K_VEC_SIZE>(&cos_emb[emb_idx], &cos_emb_vec1);
-        Load<float, HALF_K_VEC_SIZE>(&cos_emb[emb_idx + 8], &cos_emb_vec2);
-        Load<float, HALF_K_VEC_SIZE>(&sin_emb[emb_idx], &sin_emb_vec1);
-        Load<float, HALF_K_VEC_SIZE>(&sin_emb[emb_idx + 8], &sin_emb_vec2);
+        uint32_t new_emb_idx = rope_3d ? emb_idx + bid * max_seq_len * HeadDim * 2 : emb_idx;
+        Load<float, HALF_K_VEC_SIZE>(&cos_emb[new_emb_idx], &cos_emb_vec1);
+        Load<float, HALF_K_VEC_SIZE>(&cos_emb[new_emb_idx + 8], &cos_emb_vec2);
+        Load<float, HALF_K_VEC_SIZE>(&sin_emb[new_emb_idx], &sin_emb_vec1);
+        Load<float, HALF_K_VEC_SIZE>(&sin_emb[new_emb_idx + 8], &sin_emb_vec2);
         scale = __ldg(&cache_k_scales[kv_head_idx]);
 #pragma unroll
         for (int i = 0; i < HALF_K_VEC_SIZE; i++) {
@@ -1755,7 +2045,8 @@ __global__ void append_decode_cache_int4_rope_kernel(
     const int block_size,
     const float max_bound,
     const float min_bound,
-    const int kv_num_heads) {
+    const int kv_num_heads,
+    const bool rope_3d) {
   static_assert(HeadDim == 128, "just support HeadDim be 128 now!");
   static_assert(VecSize == 4, "just support VecSize be 4 now, 32 * 4!");
   constexpr int NUM_WARPS = 4;
@@ -1799,8 +2090,9 @@ __global__ void append_decode_cache_int4_rope_kernel(
 
       // q rope
       const uint32_t emb_idx = write_seq_id * half_head_size + head_bias / 2;
-      Load<float, HalfVecSize>(&cos_emb[emb_idx], &cos_emb_vec);
-      Load<float, HalfVecSize>(&sin_emb[emb_idx], &sin_emb_vec);
+      uint32_t new_emb_idx = rope_3d ? emb_idx + bid * max_seq_len * HeadDim : emb_idx;
+      Load<float, HalfVecSize>(&cos_emb[new_emb_idx], &cos_emb_vec);
+      Load<float, HalfVecSize>(&sin_emb[new_emb_idx], &sin_emb_vec);
 #pragma unroll
       for (int i = 0; i < HalfVecSize; i++) {
         // dequant + add_bias + rope
@@ -1874,10 +2166,11 @@ __global__ void append_decode_cache_int4_rope_kernel(
     Load<T, HALF_K_VEC_SIZE>(&qkv_now[bias_idx + 8], &src_vec2);
     if (head_idx < num_heads + kv_num_heads) {
       const uint32_t emb_idx = write_seq_id * half_head_size + head_bias / 2;
-      Load<float, 1>(&cos_emb[emb_idx], &cos_emb_vec1);
-      Load<float, 1>(&cos_emb[emb_idx + 4], &cos_emb_vec2);
-      Load<float, 1>(&sin_emb[emb_idx], &sin_emb_vec1);
-      Load<float, 1>(&sin_emb[emb_idx + 4], &sin_emb_vec2);
+      uint32_t new_emb_idx = rope_3d ? emb_idx + bid * max_seq_len * HeadDim : emb_idx;
+      Load<float, 1>(&cos_emb[new_emb_idx], &cos_emb_vec1);
+      Load<float, 1>(&cos_emb[new_emb_idx + 4], &cos_emb_vec2);
+      Load<float, 1>(&sin_emb[new_emb_idx], &sin_emb_vec1);
+      Load<float, 1>(&sin_emb[new_emb_idx + 4], &sin_emb_vec2);
       Load<T, HALF_K_VEC_SIZE>(&cache_k_scale[cache_idx], &scale_vec1);
       Load<T, HALF_K_VEC_SIZE>(&cache_k_scale[cache_idx + 8], &scale_vec2);
       Load<T, HALF_K_VEC_SIZE>(&cache_k_zero_points[cache_idx], &zp_vec1);
@@ -2054,7 +2347,8 @@ __global__ void append_decode_cache_int4_rope_kernel(
     const int block_size,
     const float max_bound,
     const float min_bound,
-    const int kv_num_heads) {
+    const int kv_num_heads,
+    const bool rope_3d) {
   static_assert(HeadDim == 128, "just support HeadDim be 128 now!");
   static_assert(VecSize == 4, "just support VecSize be 4 now, 32 * 4!");
   constexpr int NUM_WARPS = 4;
@@ -2103,8 +2397,9 @@ __global__ void append_decode_cache_int4_rope_kernel(
       Load<float, VecSize>(&qkv_out_scales[bias_idx], &out_scale_vec);
       // q rope
       const uint32_t emb_idx = write_seq_id * half_head_size + head_bias / 2;
-      Load<float, HalfVecSize>(&cos_emb[emb_idx], &cos_emb_vec);
-      Load<float, HalfVecSize>(&sin_emb[emb_idx], &sin_emb_vec);
+      uint32_t new_emb_idx = rope_3d ? emb_idx + bid * max_seq_len * HeadDim : emb_idx;
+      Load<float, HalfVecSize>(&cos_emb[new_emb_idx], &cos_emb_vec);
+      Load<float, HalfVecSize>(&sin_emb[new_emb_idx], &sin_emb_vec);
 #pragma unroll
       for (int i = 0; i < HalfVecSize; i++) {
         // dequant + add_bias + rope
@@ -2191,10 +2486,11 @@ __global__ void append_decode_cache_int4_rope_kernel(
                                  &out_scale_vec2);
     if (head_idx < num_heads + kv_num_heads) {
       const uint32_t emb_idx = write_seq_id * half_head_size + head_bias / 2;
-      Load<float, 1>(&cos_emb[emb_idx], &cos_emb_vec1);
-      Load<float, 1>(&cos_emb[emb_idx + 4], &cos_emb_vec2);
-      Load<float, 1>(&sin_emb[emb_idx], &sin_emb_vec1);
-      Load<float, 1>(&sin_emb[emb_idx + 4], &sin_emb_vec2);
+      uint32_t new_emb_idx = rope_3d ? emb_idx + bid * max_seq_len * HeadDim : emb_idx;
+      Load<float, 1>(&cos_emb[new_emb_idx], &cos_emb_vec1);
+      Load<float, 1>(&cos_emb[new_emb_idx + 4], &cos_emb_vec2);
+      Load<float, 1>(&sin_emb[new_emb_idx], &sin_emb_vec1);
+      Load<float, 1>(&sin_emb[new_emb_idx + 4], &sin_emb_vec2);
       Load<T, HALF_K_VEC_SIZE>(&cache_k_scale[cache_idx], &scale_vec1);
       Load<T, HALF_K_VEC_SIZE>(&cache_k_scale[cache_idx + 8], &scale_vec2);
       Load<T, HALF_K_VEC_SIZE>(&cache_k_zero_points[cache_idx], &zp_vec1);
@@ -2378,7 +2674,8 @@ __global__ void append_decode_cache_int4_neox_rope_kernel(
     const int block_size,
     const float max_bound,
     const float min_bound,
-    const int kv_num_heads) {
+    const int kv_num_heads,
+    const bool rope_3d) {
   static_assert(HeadDim == 128, "just support HeadDim be 128 now!");
   static_assert(VecSize == 4, "just support VecSize be 4 now, 32 * 4!");
   constexpr int NUM_WARPS = 4;
@@ -2425,8 +2722,9 @@ __global__ void append_decode_cache_int4_neox_rope_kernel(
 
       // q rope
       const uint32_t emb_idx = write_seq_id * HeadDim + head_bias;
-      Load<float, VecSize>(&cos_emb[emb_idx], &cos_emb_vec);
-      Load<float, VecSize>(&sin_emb[emb_idx], &sin_emb_vec);
+      uint32_t new_emb_idx = rope_3d ? emb_idx + bid * max_seq_len * HeadDim * 2 : emb_idx;
+      Load<float, VecSize>(&cos_emb[new_emb_idx], &cos_emb_vec);
+      Load<float, VecSize>(&sin_emb[new_emb_idx], &sin_emb_vec);
 #pragma unroll
       for (int i = 0; i < VecSize; i++) {
         // dequant + add_bias + rope
@@ -2507,10 +2805,11 @@ __global__ void append_decode_cache_int4_neox_rope_kernel(
         Load<T, HALF_K_VEC_SIZE>(&qkv_now[right_bias_idx], &right_src_vec1);
         Load<T, HALF_K_VEC_SIZE>(&qkv_now[right_bias_idx + 8], &right_src_vec2);
         const uint32_t emb_idx = write_seq_id * HeadDim + head_bias;
-        Load<float, HALF_K_VEC_SIZE>(&cos_emb[emb_idx], &cos_emb_vec1);
-        Load<float, HALF_K_VEC_SIZE>(&cos_emb[emb_idx + 8], &cos_emb_vec2);
-        Load<float, HALF_K_VEC_SIZE>(&sin_emb[emb_idx], &sin_emb_vec1);
-        Load<float, HALF_K_VEC_SIZE>(&sin_emb[emb_idx + 8], &sin_emb_vec2);
+        uint32_t new_emb_idx = rope_3d ? emb_idx + bid * max_seq_len * HeadDim * 2 : emb_idx;
+        Load<float, HALF_K_VEC_SIZE>(&cos_emb[new_emb_idx], &cos_emb_vec1);
+        Load<float, HALF_K_VEC_SIZE>(&cos_emb[new_emb_idx + 8], &cos_emb_vec2);
+        Load<float, HALF_K_VEC_SIZE>(&sin_emb[new_emb_idx], &sin_emb_vec1);
+        Load<float, HALF_K_VEC_SIZE>(&sin_emb[new_emb_idx + 8], &sin_emb_vec2);
         Load<T, HALF_K_VEC_SIZE>(&cache_k_scale[left_cache_idx],
                                  &left_scale_vec1);
         Load<T, HALF_K_VEC_SIZE>(&cache_k_scale[left_cache_idx + 8],
@@ -2752,7 +3051,8 @@ __global__ void append_decode_cache_int4_neox_rope_kernel(
     const int block_size,
     const float max_bound,
     const float min_bound,
-    const int kv_num_heads) {
+    const int kv_num_heads,
+    const bool rope_3d) {
   static_assert(HeadDim == 128, "just support HeadDim be 128 now!");
   static_assert(VecSize == 4, "just support VecSize be 4 now, 32 * 4!");
   constexpr int NUM_WARPS = 4;
@@ -2810,8 +3110,9 @@ __global__ void append_decode_cache_int4_neox_rope_kernel(
                            &right_out_scale_vec);
       // q rope
       const uint32_t emb_idx = write_seq_id * HeadDim + head_bias;
-      Load<float, VecSize>(&cos_emb[emb_idx], &cos_emb_vec);
-      Load<float, VecSize>(&sin_emb[emb_idx], &sin_emb_vec);
+      uint32_t new_emb_idx = rope_3d ? emb_idx + bid * max_seq_len * HeadDim * 2 : emb_idx;
+      Load<float, VecSize>(&cos_emb[new_emb_idx], &cos_emb_vec);
+      Load<float, VecSize>(&sin_emb[new_emb_idx], &sin_emb_vec);
 #pragma unroll
       for (int i = 0; i < VecSize; i++) {
         // dequant + add_bias + rope
@@ -2920,10 +3221,11 @@ __global__ void append_decode_cache_int4_neox_rope_kernel(
                                      &right_out_scale_vec2);
 
         const uint32_t emb_idx = write_seq_id * HeadDim + head_bias;
-        Load<float, HALF_K_VEC_SIZE>(&cos_emb[emb_idx], &cos_emb_vec1);
-        Load<float, HALF_K_VEC_SIZE>(&cos_emb[emb_idx + 8], &cos_emb_vec2);
-        Load<float, HALF_K_VEC_SIZE>(&sin_emb[emb_idx], &sin_emb_vec1);
-        Load<float, HALF_K_VEC_SIZE>(&sin_emb[emb_idx + 8], &sin_emb_vec2);
+        uint32_t new_emb_idx = rope_3d ? emb_idx + bid * max_seq_len * HeadDim * 2 : emb_idx;
+        Load<float, HALF_K_VEC_SIZE>(&cos_emb[new_emb_idx], &cos_emb_vec1);
+        Load<float, HALF_K_VEC_SIZE>(&cos_emb[new_emb_idx + 8], &cos_emb_vec2);
+        Load<float, HALF_K_VEC_SIZE>(&sin_emb[new_emb_idx], &sin_emb_vec1);
+        Load<float, HALF_K_VEC_SIZE>(&sin_emb[new_emb_idx + 8], &sin_emb_vec2);
         Load<T, HALF_K_VEC_SIZE>(&cache_k_scale[left_cache_idx],
                                  &left_scale_vec1);
         Load<T, HALF_K_VEC_SIZE>(&cache_k_scale[left_cache_idx + 8],

custom_ops/gpu_ops/append_attn/decoder_write_cache_with_rope_kernel.cu

@@ -15,6 +15,69 @@
 #include "decoder_write_cache_with_rope_kernel.h"
 #include "utils.cuh"
 
+template <typename T, typename QKV_TYPE>
+void append_decode_cache_rope_qk_norm(const QKV_TYPE* qkv,
+                              T* key_cache,
+                              T* value_cache,
+                              T* qkv_out,
+                              const int* block_tables,
+                              const int* batch_id_per_token,
+                              const int* cu_seqlens_q,
+                              const int* seq_lens,
+                              const int* seq_lens_encoder,
+                              const float* cos_emb,
+                              const float* sin_emb,
+                              const float* qkv_out_scales,
+                              const T* qkv_biases,
+                              const int max_seq_len,
+                              const int max_blocks_per_seq,
+                              const int num_heads,
+                              const int kv_num_heads,
+                              const int dim_head,
+                              const int block_size,
+                              const int bsz,
+                              const cudaStream_t& stream,
+                              const bool use_neox_style,
+                              const bool rope_3d,
+                              const float* q_norm_weight,
+                              const float* k_norm_weight,
+                              const float rms_norm_eps) {
+  const uint32_t elem_nums =
+      use_neox_style ? bsz * (num_heads + 2 * kv_num_heads) * dim_head / 2
+                     : bsz * (num_heads + 2 * kv_num_heads) * dim_head;
+  constexpr int HEAD_DIM = 128;
+
+  constexpr int PackSize = HEAD_DIM / kWarpSize;
+  const int pack_num = elem_nums / PackSize;
+  const int blocksize = 128;
+  int grid_size = 1;
+  GetNumBlocks<128>(pack_num, &grid_size);
+  dim3 block_dim(kWarpSize, blocksize / kWarpSize, 1);
+  append_decode_cache_T_rope_qk_norm_kernel<T, PackSize>
+      <<<grid_size, block_dim, 0, stream>>>(reinterpret_cast<const T*>(qkv),
+                                            key_cache,
+                                            value_cache,
+                                            qkv_out,
+                                            block_tables,
+                                            batch_id_per_token,
+                                            cu_seqlens_q,
+                                            seq_lens,
+                                            seq_lens_encoder,
+                                            cos_emb,
+                                            sin_emb,
+                                            max_seq_len,
+                                            max_blocks_per_seq,
+                                            num_heads,
+                                            dim_head,
+                                            block_size,
+                                            elem_nums,
+                                            kv_num_heads,
+                                            rope_3d,
+                                            q_norm_weight,
+                                            k_norm_weight,
+                                            rms_norm_eps);
+}
+
 template <typename T, typename QKV_TYPE>
 void append_decode_cache_rope(const QKV_TYPE* qkv,
                               T* key_cache,
@@ -34,6 +97,7 @@ void append_decode_cache_rope(const QKV_TYPE* qkv,
                               const int num_heads,
                               const int kv_num_heads,
                               const int dim_head,
+                              const int rotary_dim,
                               const int block_size,
                               const int bsz,
                               const cudaStream_t& stream,
@@ -71,9 +135,32 @@ void append_decode_cache_rope(const QKV_TYPE* qkv,
               dim_head,
               block_size,
               elem_nums,
-              kv_num_heads);
+              kv_num_heads,
+              rope_3d);
     } else {
-      append_decode_cache_T_neox_rope_kernel<T, PackSize>
+      if (rotary_dim < dim_head){
+        append_decode_cache_T_neox_partial_rope_kernel<T, PackSize>
+          <<<grid_size, blocksize, 0, stream>>>(reinterpret_cast<const T*>(qkv),
+                                                key_cache,
+                                                value_cache,
+                                                qkv_out,
+                                                block_tables,
+                                                cu_seqlens_q,
+                                                seq_lens,
+                                                seq_lens_encoder,
+                                                cos_emb,
+                                                sin_emb,
+                                                max_seq_len,
+                                                max_blocks_per_seq,
+                                                num_heads,
+                                                dim_head,
+                                                rotary_dim,
+                                                block_size,
+                                                elem_nums,
+                                                kv_num_heads,
+                                                rope_3d);
+      }else{
+        append_decode_cache_T_neox_rope_kernel<T, PackSize>
           <<<grid_size, blocksize, 0, stream>>>(reinterpret_cast<const T*>(qkv),
                                                 key_cache,
                                                 value_cache,
@@ -91,7 +178,9 @@ void append_decode_cache_rope(const QKV_TYPE* qkv,
                                                 dim_head,
                                                 block_size,
                                                 elem_nums,
-                                                kv_num_heads);
+                                                kv_num_heads,
+                                                rope_3d);
+      }
     }
   } else {
     if (qkv_out_scales) {
@@ -198,7 +287,8 @@ void append_decode_cache_int8_rope(const QKV_TYPE* qkv,
               block_size,
               127.0f,
               -127.0f,
-              kv_num_heads);
+              kv_num_heads,
+              rope_3d);
     } else {
       append_decode_cache_int8_neox_rope_kernel<T, 4>
           <<<grids, num_warps * 32, 0, stream>>>(
@@ -221,7 +311,8 @@ void append_decode_cache_int8_rope(const QKV_TYPE* qkv,
               block_size,
               127.0f,
               -127.0f,
-              kv_num_heads);
+              kv_num_heads,
+              rope_3d);
     }
   } else {
     if (qkv_out_scales) {
@@ -248,7 +339,8 @@ void append_decode_cache_int8_rope(const QKV_TYPE* qkv,
               block_size,
               127.0f,
               -127.0f,
-              kv_num_heads);
+              kv_num_heads,
+              rope_3d);
     } else {
       append_decode_cache_int8_rope_kernel<T, 4, 0, 128, is_scale_channel_wise, IsFP8>
           <<<grids, num_warps * 32, 0, stream>>>(
@@ -271,7 +363,8 @@ void append_decode_cache_int8_rope(const QKV_TYPE* qkv,
               block_size,
               127.0f,
               -127.0f,
-              kv_num_heads);
+              kv_num_heads,
+              rope_3d);
     }
   }
 }
@@ -335,7 +428,8 @@ void append_decode_cache_int4_rope(const QKV_TYPE* qkv,
               block_size,
               7.0f,
               -8.0f,
-              kv_num_heads);
+              kv_num_heads,
+              rope_3d);
     } else {
       append_decode_cache_int4_neox_rope_kernel<T, 4>
           <<<grids, num_warps * 32, 0, stream>>>(
@@ -360,7 +454,8 @@ void append_decode_cache_int4_rope(const QKV_TYPE* qkv,
               block_size,
               7.0f,
               -8.0f,
-              kv_num_heads);
+              kv_num_heads,
+              rope_3d);
     }
   } else {
     if (qkv_out_scales) {
@@ -389,7 +484,8 @@ void append_decode_cache_int4_rope(const QKV_TYPE* qkv,
               block_size,
               7.0f,
               -8.0f,
-              kv_num_heads);
+              kv_num_heads,
+              rope_3d);
     } else {
       append_decode_cache_int4_rope_kernel<T, 4>
           <<<grids, num_warps * 32, 0, stream>>>(
@@ -414,7 +510,8 @@ void append_decode_cache_int4_rope(const QKV_TYPE* qkv,
               block_size,
               7.0f,
               -8.0f,
-              kv_num_heads);
+              kv_num_heads,
+              rope_3d);
     }
   }
 }
@@ -441,7 +538,10 @@ void DecoderWriteCacheWithRoPEKernel(
     cudaStream_t& stream,
     paddle::Tensor* qkv_out,
     paddle::Tensor* key_cache_out,
-    paddle::Tensor* value_cache_out) {
+    paddle::Tensor* value_cache_out,
+    const paddle::optional<paddle::Tensor>& q_norm_weight,
+    const paddle::optional<paddle::Tensor>& k_norm_weight,
+    const float rms_norm_eps) {
   typedef cascade_attn_type_traits<T> traits_;
   typedef cascade_attn_type_traits<QKV_TYPE> qkt_nv_type_;
   typedef typename traits_::type DataType_;
@@ -458,113 +558,93 @@ void DecoderWriteCacheWithRoPEKernel(
   const float* cos_emb =
       rotary_embs ? rotary_embs.get().data<float>() : nullptr;
   const float* sin_emb;
+  int rotary_dim = dim_head;
   if (rotary_embs) {
     sin_emb =
         use_neox_rotary_style
             ? rotary_embs.get().data<float>() + max_seq_len * dim_head
             : rotary_embs.get().data<float>() + max_seq_len * dim_head / 2;
+    rotary_dim = rotary_embs.get().dims()[rotary_embs.get().dims().size()-1] * 2;
+    if(rotary_dim < dim_head){
+      if (!use_neox_rotary_style || qkv_out_scales || q_norm_weight || k_norm_weight|| cache_quant_type_str != "none"){
+        PADDLE_THROW(phi::errors::Fatal(
+          "partial_rotary_factor < 1.0 only supports neox_rotary_style=True, qkv_out_scales is None, q_norm_weight/k_norm_weight) is None, and cache_quant_type_str is 'none'."));
+      }
+      sin_emb = rotary_embs.get().data<float>() + max_seq_len * rotary_dim / 2;
+    }
   }
-  if (cache_quant_type_str == "none") {
-    append_decode_cache_rope(
-        reinterpret_cast<const QKV_TYPE*>(qkv_ptr),
-        reinterpret_cast<DataType_*>(key_cache_out->data<T>()),
-        reinterpret_cast<DataType_*>(value_cache_out->data<T>()),
-        reinterpret_cast<DataType_*>(qkv_out->data<T>()),
-        block_tables.data<int>(),
-        batch_id_per_token.data<int>(),
-        cu_seqlens_q.data<int>(),
-        seq_lens.data<int>(),
-        seq_lens_encoder.data<int>(),
-        cos_emb,
-        sin_emb,
-        qkv_out_scales ? qkv_out_scales.get().data<float>() : nullptr,
-        qkv_biases ? reinterpret_cast<DataType_*>(
-                         const_cast<T*>(qkv_biases.get().data<T>()))
-                   : nullptr,
-        max_seq_len,
-        max_blocks_per_seq,
-        num_heads,
-        kv_num_heads,
-        dim_head,
-        block_size,
-        bsz,
-        stream,
-        use_neox_rotary_style,
-        rope_3d);
-  } else if (cache_quant_type_str == "cache_int8") {
-    bool is_scale_channel_wise = false;
-    if (cache_k_scale && cache_k_scale.get().dims()[0] == dim_head * kv_num_heads) {
-      is_scale_channel_wise = true;
-    }
-    if (is_scale_channel_wise) {
-      append_decode_cache_int8_rope<DataType_, QKV_TYPE, true>(
-        reinterpret_cast<const QKV_TYPE*>(qkv_ptr),
-        key_cache_out->data<uint8_t>(),
-        value_cache_out->data<uint8_t>(),
-        reinterpret_cast<DataType_*>(qkv_out->data<T>()),
-        block_tables.data<int>(),
-        batch_id_per_token.data<int>(),
-        cu_seqlens_q.data<int>(),
-        seq_lens.data<int>(),
-        seq_lens_encoder.data<int>(),
-        cos_emb,
-        sin_emb,
-        qkv_out_scales ? qkv_out_scales.get().data<float>() : nullptr,
-        qkv_biases ? reinterpret_cast<DataType_*>(
-                         const_cast<T*>(qkv_biases.get().data<T>()))
-                   : nullptr,
-        cache_k_scale ? reinterpret_cast<DataType_*>(
-                            const_cast<T*>(cache_k_scale.get().data<T>()))
-                      : nullptr,
-        cache_v_scale ? reinterpret_cast<DataType_*>(
-                            const_cast<T*>(cache_v_scale.get().data<T>()))
-                      : nullptr,
-        max_seq_len,
-        max_blocks_per_seq,
-        num_heads,
-        kv_num_heads,
-        dim_head,
-        block_size,
-        bsz,
-        stream,
-        use_neox_rotary_style,
-        rope_3d);
+
+  if (q_norm_weight && k_norm_weight) {
+    if (cache_quant_type_str == "none") {
+      append_decode_cache_rope_qk_norm(
+          reinterpret_cast<const QKV_TYPE*>(qkv_ptr),
+          reinterpret_cast<DataType_*>(key_cache_out->data<T>()),
+          reinterpret_cast<DataType_*>(value_cache_out->data<T>()),
+          reinterpret_cast<DataType_*>(qkv_out->data<T>()),
+          block_tables.data<int>(),
+          batch_id_per_token.data<int>(),
+          cu_seqlens_q.data<int>(),
+          seq_lens.data<int>(),
+          seq_lens_encoder.data<int>(),
+          cos_emb,
+          sin_emb,
+          qkv_out_scales ? qkv_out_scales.get().data<float>() : nullptr,
+          qkv_biases ? reinterpret_cast<DataType_*>(
+                          const_cast<T*>(qkv_biases.get().data<T>()))
+                    : nullptr,
+          max_seq_len,
+          max_blocks_per_seq,
+          num_heads,
+          kv_num_heads,
+          dim_head,
+          block_size,
+          bsz,
+          stream,
+          use_neox_rotary_style,
+          rope_3d,
+          q_norm_weight ? q_norm_weight.get().data<float>() : nullptr,
+          k_norm_weight ? k_norm_weight.get().data<float>() : nullptr,
+          rms_norm_eps);
     } else {
-      append_decode_cache_int8_rope<DataType_, QKV_TYPE, false>(
-          reinterpret_cast<const QKV_TYPE*>(qkv_ptr),
-          key_cache_out->data<uint8_t>(),
-          value_cache_out->data<uint8_t>(),
-          reinterpret_cast<DataType_*>(qkv_out->data<T>()),
-          block_tables.data<int>(),
-          batch_id_per_token.data<int>(),
-          cu_seqlens_q.data<int>(),
-          seq_lens.data<int>(),
-          seq_lens_encoder.data<int>(),
-          cos_emb,
-          sin_emb,
-          qkv_out_scales ? qkv_out_scales.get().data<float>() : nullptr,
-          qkv_biases ? reinterpret_cast<DataType_*>(
-                          const_cast<T*>(qkv_biases.get().data<T>()))
-                    : nullptr,
-          cache_k_scale ? reinterpret_cast<DataType_*>(
-                              const_cast<T*>(cache_k_scale.get().data<T>()))
-                        : nullptr,
-          cache_v_scale ? reinterpret_cast<DataType_*>(
-                              const_cast<T*>(cache_v_scale.get().data<T>()))
-                        : nullptr,
-          max_seq_len,
-          max_blocks_per_seq,
-          num_heads,
-          kv_num_heads,
-          dim_head,
-          block_size,
-          bsz,
-          stream,
-          use_neox_rotary_style,
-          rope_3d);
+      PD_THROW(
+          "append_decode_cache_rope_qk_norm not support cachekv quant yet");
     }
-  } else if (cache_quant_type_str == "cache_fp8")  {
-      append_decode_cache_int8_rope<DataType_, QKV_TYPE, false, true>(
+  } else {
+    if (cache_quant_type_str == "none") {
+      append_decode_cache_rope(
+          reinterpret_cast<const QKV_TYPE*>(qkv_ptr),
+          reinterpret_cast<DataType_*>(key_cache_out->data<T>()),
+          reinterpret_cast<DataType_*>(value_cache_out->data<T>()),
+          reinterpret_cast<DataType_*>(qkv_out->data<T>()),
+          block_tables.data<int>(),
+          batch_id_per_token.data<int>(),
+          cu_seqlens_q.data<int>(),
+          seq_lens.data<int>(),
+          seq_lens_encoder.data<int>(),
+          cos_emb,
+          sin_emb,
+          qkv_out_scales ? qkv_out_scales.get().data<float>() : nullptr,
+          qkv_biases ? reinterpret_cast<DataType_*>(
+                          const_cast<T*>(qkv_biases.get().data<T>()))
+                    : nullptr,
+          max_seq_len,
+          max_blocks_per_seq,
+          num_heads,
+          kv_num_heads,
+          dim_head,
+          rotary_dim,
+          block_size,
+          bsz,
+          stream,
+          use_neox_rotary_style,
+          rope_3d);
+    } else if (cache_quant_type_str == "cache_int8") {
+      bool is_scale_channel_wise = false;
+      if (cache_k_scale && cache_k_scale.get().dims()[0] == dim_head * kv_num_heads) {
+        is_scale_channel_wise = true;
+      }
+      if (is_scale_channel_wise) {
+        append_decode_cache_int8_rope<DataType_, QKV_TYPE, true>(
           reinterpret_cast<const QKV_TYPE*>(qkv_ptr),
           key_cache_out->data<uint8_t>(),
           value_cache_out->data<uint8_t>(),
@@ -596,49 +676,117 @@ void DecoderWriteCacheWithRoPEKernel(
           stream,
           use_neox_rotary_style,
           rope_3d);
-  } else if (cache_quant_type_str == "cache_int4_zp") {
-    append_decode_cache_int4_rope(
-        reinterpret_cast<const QKV_TYPE*>(qkv_ptr),
-        key_cache_out->data<uint8_t>(),
-        value_cache_out->data<uint8_t>(),
-        reinterpret_cast<DataType_*>(const_cast<T*>(qkv_out->data<T>())),
-        block_tables.data<int>(),
-        batch_id_per_token.data<int>(),
-        cu_seqlens_q.data<int>(),
-        seq_lens.data<int>(),
-        seq_lens_encoder.data<int>(),
-        cos_emb,
-        sin_emb,
-        qkv_out_scales ? qkv_out_scales.get().data<float>() : nullptr,
-        qkv_biases ? reinterpret_cast<DataType_*>(
-                         const_cast<T*>(qkv_biases.get().data<T>()))
-                   : nullptr,
-        cache_k_scale ? reinterpret_cast<DataType_*>(
-                            const_cast<T*>(cache_k_scale.get().data<T>()))
+      } else {
+        append_decode_cache_int8_rope<DataType_, QKV_TYPE, false>(
+            reinterpret_cast<const QKV_TYPE*>(qkv_ptr),
+            key_cache_out->data<uint8_t>(),
+            value_cache_out->data<uint8_t>(),
+            reinterpret_cast<DataType_*>(qkv_out->data<T>()),
+            block_tables.data<int>(),
+            batch_id_per_token.data<int>(),
+            cu_seqlens_q.data<int>(),
+            seq_lens.data<int>(),
+            seq_lens_encoder.data<int>(),
+            cos_emb,
+            sin_emb,
+            qkv_out_scales ? qkv_out_scales.get().data<float>() : nullptr,
+            qkv_biases ? reinterpret_cast<DataType_*>(
+                            const_cast<T*>(qkv_biases.get().data<T>()))
                       : nullptr,
-        cache_v_scale ? reinterpret_cast<DataType_*>(
-                            const_cast<T*>(cache_v_scale.get().data<T>()))
+            cache_k_scale ? reinterpret_cast<DataType_*>(
+                                const_cast<T*>(cache_k_scale.get().data<T>()))
+                          : nullptr,
+            cache_v_scale ? reinterpret_cast<DataType_*>(
+                                const_cast<T*>(cache_v_scale.get().data<T>()))
+                          : nullptr,
+            max_seq_len,
+            max_blocks_per_seq,
+            num_heads,
+            kv_num_heads,
+            dim_head,
+            block_size,
+            bsz,
+            stream,
+            use_neox_rotary_style,
+            rope_3d);
+      }
+    } else if (cache_quant_type_str == "cache_fp8")  {
+        append_decode_cache_int8_rope<DataType_, QKV_TYPE, false, true>(
+            reinterpret_cast<const QKV_TYPE*>(qkv_ptr),
+            key_cache_out->data<uint8_t>(),
+            value_cache_out->data<uint8_t>(),
+            reinterpret_cast<DataType_*>(qkv_out->data<T>()),
+            block_tables.data<int>(),
+            batch_id_per_token.data<int>(),
+            cu_seqlens_q.data<int>(),
+            seq_lens.data<int>(),
+            seq_lens_encoder.data<int>(),
+            cos_emb,
+            sin_emb,
+            qkv_out_scales ? qkv_out_scales.get().data<float>() : nullptr,
+            qkv_biases ? reinterpret_cast<DataType_*>(
+                            const_cast<T*>(qkv_biases.get().data<T>()))
                       : nullptr,
-        cache_k_zp ? reinterpret_cast<DataType_*>(
-                         const_cast<T*>(cache_k_zp.get().data<T>()))
-                   : nullptr,
-        cache_v_zp ? reinterpret_cast<DataType_*>(
-                         const_cast<T*>(cache_v_zp.get().data<T>()))
-                   : nullptr,
-        max_seq_len,
-        max_blocks_per_seq,
-        num_heads,
-        kv_num_heads,
-        dim_head,
-        block_size,
-        bsz,
-        stream,
-        use_neox_rotary_style,
-        rope_3d);
-  } else {
-    PD_THROW(
-        "cache_quant_type_str should be one of [none, cache_int8, cache_fp8 "
-        "cache_int4_zp]");
+            cache_k_scale ? reinterpret_cast<DataType_*>(
+                                const_cast<T*>(cache_k_scale.get().data<T>()))
+                          : nullptr,
+            cache_v_scale ? reinterpret_cast<DataType_*>(
+                                const_cast<T*>(cache_v_scale.get().data<T>()))
+                          : nullptr,
+            max_seq_len,
+            max_blocks_per_seq,
+            num_heads,
+            kv_num_heads,
+            dim_head,
+            block_size,
+            bsz,
+            stream,
+            use_neox_rotary_style,
+            rope_3d);
+    } else if (cache_quant_type_str == "cache_int4_zp") {
+      append_decode_cache_int4_rope(
+          reinterpret_cast<const QKV_TYPE*>(qkv_ptr),
+          key_cache_out->data<uint8_t>(),
+          value_cache_out->data<uint8_t>(),
+          reinterpret_cast<DataType_*>(const_cast<T*>(qkv_out->data<T>())),
+          block_tables.data<int>(),
+          batch_id_per_token.data<int>(),
+          cu_seqlens_q.data<int>(),
+          seq_lens.data<int>(),
+          seq_lens_encoder.data<int>(),
+          cos_emb,
+          sin_emb,
+          qkv_out_scales ? qkv_out_scales.get().data<float>() : nullptr,
+          qkv_biases ? reinterpret_cast<DataType_*>(
+                          const_cast<T*>(qkv_biases.get().data<T>()))
+                    : nullptr,
+          cache_k_scale ? reinterpret_cast<DataType_*>(
+                              const_cast<T*>(cache_k_scale.get().data<T>()))
+                        : nullptr,
+          cache_v_scale ? reinterpret_cast<DataType_*>(
+                              const_cast<T*>(cache_v_scale.get().data<T>()))
+                        : nullptr,
+          cache_k_zp ? reinterpret_cast<DataType_*>(
+                          const_cast<T*>(cache_k_zp.get().data<T>()))
+                    : nullptr,
+          cache_v_zp ? reinterpret_cast<DataType_*>(
+                          const_cast<T*>(cache_v_zp.get().data<T>()))
+                    : nullptr,
+          max_seq_len,
+          max_blocks_per_seq,
+          num_heads,
+          kv_num_heads,
+          dim_head,
+          block_size,
+          bsz,
+          stream,
+          use_neox_rotary_style,
+          rope_3d);
+    } else {
+      PD_THROW(
+          "cache_quant_type_str should be one of [none, cache_int8, cache_fp8 "
+          "cache_int4_zp]");
+    }
   }
 }
 
@@ -667,7 +815,10 @@ template void DecoderWriteCacheWithRoPEKernel<paddle::bfloat16, int>(
     cudaStream_t& stream,
     paddle::Tensor* qkv_out,
     paddle::Tensor* key_cache_out,
-    paddle::Tensor* value_cache_out);
+    paddle::Tensor* value_cache_out,
+    const paddle::optional<paddle::Tensor>& q_norm_weight,
+    const paddle::optional<paddle::Tensor>& k_norm_weight,
+    const float rms_norm_eps);
 
 template void
 DecoderWriteCacheWithRoPEKernel<paddle::bfloat16, paddle::bfloat16>(
@@ -694,7 +845,10 @@ DecoderWriteCacheWithRoPEKernel<paddle::bfloat16, paddle::bfloat16>(
     cudaStream_t& stream,
     paddle::Tensor* qkv_out,
     paddle::Tensor* key_cache_out,
-    paddle::Tensor* value_cache_out);
+    paddle::Tensor* value_cache_out,
+    const paddle::optional<paddle::Tensor>& q_norm_weight,
+    const paddle::optional<paddle::Tensor>& k_norm_weight,
+    const float rms_norm_eps);
 
 template void DecoderWriteCacheWithRoPEKernel<paddle::float16, int>(
     const AppendAttnMetaData& meta_data,
@@ -720,7 +874,10 @@ template void DecoderWriteCacheWithRoPEKernel<paddle::float16, int>(
     cudaStream_t& stream,
     paddle::Tensor* qkv_out,
     paddle::Tensor* key_cache_out,
-    paddle::Tensor* value_cache_out);
+    paddle::Tensor* value_cache_out,
+    const paddle::optional<paddle::Tensor>& q_norm_weight,
+    const paddle::optional<paddle::Tensor>& k_norm_weight,
+    const float rms_norm_eps);
 
 template void DecoderWriteCacheWithRoPEKernel<paddle::float16, paddle::float16>(
     const AppendAttnMetaData& meta_data,
@@ -746,4 +903,7 @@ template void DecoderWriteCacheWithRoPEKernel<paddle::float16, paddle::float16>(
     cudaStream_t& stream,
     paddle::Tensor* qkv_out,
     paddle::Tensor* key_cache_out,
-    paddle::Tensor* value_cache_out);
+    paddle::Tensor* value_cache_out,
+    const paddle::optional<paddle::Tensor>& q_norm_weight,
+    const paddle::optional<paddle::Tensor>& k_norm_weight,
+    const float rms_norm_eps);

custom_ops/gpu_ops/append_attn/decoder_write_cache_with_rope_kernel.h

@@ -40,4 +40,6 @@ void DecoderWriteCacheWithRoPEKernel(
     cudaStream_t& stream,
     paddle::Tensor* qkv_out,
     paddle::Tensor* key_cache_out,
-    paddle::Tensor* value_cache_out);
+    paddle::Tensor* value_cache_out,
+    const paddle::optional<paddle::Tensor>& q_norm_weight,
+    const paddle::optional<paddle::Tensor>& k_norm_weight,    const float rms_norm_eps);

custom_ops/gpu_ops/append_attn/encoder_write_cache_with_rope_impl.cuh

@@ -33,7 +33,8 @@ __global__ void VariableLengthRotaryKernel(
     const int64_t elem_cnt,
     const int num_head,
     const int seq_len,
-    const int last_dim) {
+    const int last_dim,
+    const bool rope_3d) {
   using LoadT = AlignedVector<int, VecSize>;
   using LoadBiasT = AlignedVector<T, VecSize>;
   using LoadScaleT = AlignedVector<float, VecSize>;
@@ -64,6 +65,7 @@ __global__ void VariableLengthRotaryKernel(
     const int ori_seq_id = (token_idx - cu_seqlens_q[ori_bi]) + seq_lens_decoder[ori_bi];
 
     const int emb_idx = ori_seq_id * half_lastdim + h_bias / 2;
+    int new_emb_idx = rope_3d ? emb_idx + ori_bi * last_dim * seq_len : emb_idx;
     const int bias_idx = qkv_id * hidden_size + hi * last_dim + h_bias;
     const int64_t base_idx = token_idx * 3 * hidden_size + bias_idx;
     Load<int, VecSize>(&qkv[base_idx], &src_vec);
@@ -72,8 +74,8 @@ __global__ void VariableLengthRotaryKernel(
     }
     Load<float, VecSize>(&qkv_out_scales[bias_idx], &out_scale_vec);
     if (qkv_id < 2) {
-      Load<float, HalfVecSize>(&cos_emb[emb_idx], &cos_emb_vec);
-      Load<float, HalfVecSize>(&sin_emb[emb_idx], &sin_emb_vec);
+      Load<float, HalfVecSize>(&cos_emb[new_emb_idx], &cos_emb_vec);
+      Load<float, HalfVecSize>(&sin_emb[new_emb_idx], &sin_emb_vec);
     }
 #pragma unroll
     for (int i = 0; i < HalfVecSize; i++) {
@@ -115,7 +117,8 @@ __global__ void VariableLengthRotaryKernel(
     const int64_t elem_cnt,
     const int num_head,
     const int seq_len,
-    const int last_dim) {
+    const int last_dim,
+    const bool rope_3d) {
   using LoadT = AlignedVector<T, VecSize>;
   constexpr int HalfVecSize = VecSize / 2;
   using LoadEmbT = AlignedVector<float, HalfVecSize>;
@@ -142,11 +145,12 @@ __global__ void VariableLengthRotaryKernel(
     const int ori_seq_id = (token_idx - cu_seqlens_q[ori_bi]) + seq_lens_decoder[ori_bi];
 
     const int emb_idx = ori_seq_id * half_lastdim + h_bias / 2;
+    int new_emb_idx = rope_3d ? emb_idx + ori_bi * last_dim * seq_len : emb_idx;
     const int64_t base_idx = token_idx * 3 * hidden_size +
                              qkv_id * hidden_size + hi * last_dim + h_bias;
     Load<T, VecSize>(&qkv[base_idx], &src_vec);
-    Load<float, HalfVecSize>(&cos_emb[emb_idx], &cos_emb_vec);
-    Load<float, HalfVecSize>(&sin_emb[emb_idx], &sin_emb_vec);
+    Load<float, HalfVecSize>(&cos_emb[new_emb_idx], &cos_emb_vec);
+    Load<float, HalfVecSize>(&sin_emb[new_emb_idx], &sin_emb_vec);
 #pragma unroll
     for (int i = 0; i < HalfVecSize; i++) {
       const float input_left = static_cast<float>(src_vec[2 * i]);
@@ -177,7 +181,8 @@ __global__ void NeoxVariableLengthRotaryKernel(
     const int64_t elem_cnt,
     const int num_head,
     const int seq_len,
-    const int last_dim) {
+    const int last_dim,
+    const bool rope_3d) {
   using LoadT = AlignedVector<int, VecSize>;
   using LoadBiasT = AlignedVector<T, VecSize>;
   using LoadScaleT = AlignedVector<float, VecSize>;
@@ -211,6 +216,7 @@ __global__ void NeoxVariableLengthRotaryKernel(
     const int ori_seq_id = (token_idx - cu_seqlens_q[ori_bi]) + seq_lens_decoder[ori_bi];
 
     const int emb_idx = ori_seq_id * last_dim + h_bias;
+    int new_emb_idx = rope_3d ? emb_idx + ori_bi * last_dim * seq_len * 2 : emb_idx;
     const int bias_idx_left =
         qkv_id * full_hidden_size + hi * last_dim + h_bias;
     const int bias_idx_right = bias_idx_left + half_lastdim;
@@ -225,8 +231,8 @@ __global__ void NeoxVariableLengthRotaryKernel(
     Load<float, VecSize>(&qkv_out_scales[bias_idx_left], &left_out_scale_vec);
     Load<float, VecSize>(&qkv_out_scales[bias_idx_right], &right_out_scale_vec);
     if (qkv_id < 2) {
-      Load<float, VecSize>(&cos_emb[emb_idx], &cos_emb_vec);
-      Load<float, VecSize>(&sin_emb[emb_idx], &sin_emb_vec);
+      Load<float, VecSize>(&cos_emb[new_emb_idx], &cos_emb_vec);
+      Load<float, VecSize>(&sin_emb[new_emb_idx], &sin_emb_vec);
     }
 #pragma unroll
     for (int i = 0; i < VecSize; i++) {
@@ -269,7 +275,8 @@ __global__ void NeoxVariableLengthRotaryKernel(
     const int64_t elem_cnt,
     const int num_head,
     const int seq_len,
-    const int last_dim) {
+    const int last_dim,
+    const bool rope_3d) {
   using LoadT = AlignedVector<T, VecSize>;
   using LoadEmbT = AlignedVector<float, VecSize>;
   LoadT left_vec;
@@ -297,6 +304,7 @@ __global__ void NeoxVariableLengthRotaryKernel(
     const int ori_seq_id = (token_idx - cu_seqlens_q[ori_bi]) + seq_lens_decoder[ori_bi];
 
     const int emb_idx = ori_seq_id * last_dim + h_bias;
+    int new_emb_idx = rope_3d ? emb_idx + ori_bi * last_dim * seq_len * 2 : emb_idx;
     const int base_idx_left = token_idx * 3 * full_hidden_size +
                               qkv_id * full_hidden_size + hi * last_dim +
                               h_bias;
@@ -304,8 +312,8 @@ __global__ void NeoxVariableLengthRotaryKernel(
 
     Load<T, VecSize>(&qkv[base_idx_left], &left_vec);
     Load<T, VecSize>(&qkv[base_idx_right], &right_vec);
-    Load<float, VecSize>(&cos_emb[emb_idx], &cos_emb_vec);
-    Load<float, VecSize>(&sin_emb[emb_idx], &sin_emb_vec);
+    Load<float, VecSize>(&cos_emb[new_emb_idx], &cos_emb_vec);
+    Load<float, VecSize>(&sin_emb[new_emb_idx], &sin_emb_vec);
 #pragma unroll
     for (int i = 0; i < VecSize; i++) {
       const float input_left = static_cast<float>(left_vec[i]);
@@ -358,7 +366,7 @@ __global__ void GQAVariableLengthRotaryKernel(
        linear_index < elem_cnt;
        linear_index += step) {
     const int token_idx = linear_index / offset;
-    const int ori_bi = batch_id_per_token[token_idx];;
+    const int ori_bi = batch_id_per_token[token_idx];
     if (seq_lens[ori_bi] == 0) continue;
     const int bias = linear_index % offset;
     const int hi = bias / last_dim;
@@ -367,6 +375,7 @@ __global__ void GQAVariableLengthRotaryKernel(
     const int ori_seq_id = (token_idx - cu_seqlens_q[ori_bi]) + seq_lens_decoder[ori_bi];
 
     const int64_t emb_idx = ori_seq_id * half_lastdim + h_bias / 2;
+    int64_t new_emb_idx = rope_3d ? emb_idx + ori_bi * last_dim * seq_len : emb_idx;
     const int64_t bias_idx = hi * last_dim + h_bias;
     const int64_t base_idx = token_idx * offset + bias_idx;
     Load<int, VecSize>(&qkv[base_idx], &src_vec);
@@ -375,8 +384,8 @@ __global__ void GQAVariableLengthRotaryKernel(
     }
     Load<float, VecSize>(&qkv_out_scales[bias_idx], &out_scale_vec);
     if (hi < q_num_head + kv_num_head) {
-      Load<float, HalfVecSize>(&cos_emb[emb_idx], &cos_emb_vec);
-      Load<float, HalfVecSize>(&sin_emb[emb_idx], &sin_emb_vec);
+      Load<float, HalfVecSize>(&cos_emb[new_emb_idx], &cos_emb_vec);
+      Load<float, HalfVecSize>(&sin_emb[new_emb_idx], &sin_emb_vec);
     }
 #pragma unroll
     for (int i = 0; i < HalfVecSize; i++) {
@@ -405,6 +414,97 @@ __global__ void GQAVariableLengthRotaryKernel(
   }
 }
 
+
+template <typename T, int VecSize = 1>
+__global__ void GQAVariableLengthRotaryQKNormKernel(
+    const T *qkv,
+    const float *cos_emb,
+    const float *sin_emb,
+    const int *batch_id_per_token,
+    const int *cu_seqlens_q,
+    const int *seq_lens,
+    const int *seq_lens_decoder,
+    T *qkv_out,
+    const int64_t elem_cnt,
+    const int q_num_head,
+    const int kv_num_head,
+    const int seq_len,
+    const int last_dim,
+    const bool rope_3d,
+    const float* q_norm_weight,
+    const float* k_norm_weight,
+    const float rms_norm_eps
+) {
+  using LoadT = AlignedVector<T, VecSize>;
+  constexpr int HalfVecSize = VecSize / 2;
+  using LoadEmbT = AlignedVector<float, HalfVecSize>;
+  using LoadFloat = AlignedVector<float, VecSize>;
+  LoadT src_vec;
+  LoadEmbT cos_emb_vec;
+  LoadEmbT sin_emb_vec;
+  LoadFloat tmp_vec;
+  LoadFloat q_norm_vec, k_norm_vec;
+  int64_t global_warp_idx = blockDim.y * blockIdx.x + threadIdx.y;
+  int64_t all_warp_num = gridDim.x * blockDim.y;
+  const int half_lastdim = last_dim / 2;
+  const int offset = (q_num_head + kv_num_head) * last_dim;
+  const int all_head_num = elem_cnt / last_dim;
+  for (int gloabl_hi = global_warp_idx; gloabl_hi < all_head_num; gloabl_hi += all_warp_num) {
+    int64_t linear_index = gloabl_hi * last_dim + threadIdx.x * VecSize;
+    const int token_idx = linear_index / offset;
+    const int ori_bi = batch_id_per_token[token_idx];
+    if (seq_lens[ori_bi] == 0) continue;
+    const int bias = linear_index % offset;
+    const int hi = bias / last_dim;
+    const int h_bias = bias % last_dim;
+
+    const int ori_seq_id = (token_idx - cu_seqlens_q[ori_bi]) + seq_lens_decoder[ori_bi];
+    const int64_t emb_idx = ori_seq_id * half_lastdim + h_bias / 2;
+    const int64_t base_idx =
+        token_idx * (q_num_head + 2 * kv_num_head) * last_dim + hi * last_dim +
+        h_bias;
+    Load<T, VecSize>(&qkv[base_idx], &src_vec);
+
+    int64_t new_emb_idx = rope_3d ? emb_idx + ori_bi * last_dim * seq_len : emb_idx;
+    Load<float, HalfVecSize>(&cos_emb[new_emb_idx], &cos_emb_vec);
+    Load<float, HalfVecSize>(&sin_emb[new_emb_idx], &sin_emb_vec);
+
+    float thread_m2 = 0.0f;
+    float warp_m2 = 0.0f;
+
+#pragma unroll
+    for (int i = 0; i < HalfVecSize; i++) {
+      const float input_left = static_cast<float>(src_vec[2 * i]);
+      const float input_right = static_cast<float>(src_vec[2 * i + 1]);
+      const float cos_tmp = cos_emb_vec[i];
+      const float sin_tmp = sin_emb_vec[i];
+      float tmp1 = input_left * cos_tmp - input_right * sin_tmp;
+      float tmp2 = input_right * cos_tmp + input_left * sin_tmp;
+      tmp_vec[2 * i] = tmp1;
+      tmp_vec[2 * i + 1] = tmp2;
+      thread_m2 += tmp1 * tmp1 + tmp2 * tmp2;
+    }
+    WelfordWarpAllReduce<float, 32>(thread_m2, &warp_m2);
+    float row_variance =
+        max(warp_m2 / last_dim, 0.0f);
+    float row_inv_var = Rsqrt(row_variance + rms_norm_eps);
+
+    if (hi < q_num_head) {
+      Load<float, VecSize>(&q_norm_weight[threadIdx.x * VecSize], &q_norm_vec);
+      #pragma unroll
+      for (int i = 0; i < VecSize; i++) {
+        src_vec[i] = static_cast<T>(tmp_vec[i] * row_inv_var * q_norm_vec[i]);
+      }
+    } else {
+      Load<float, VecSize>(&k_norm_weight[threadIdx.x * VecSize], &k_norm_vec);
+      for (int i = 0; i < VecSize; i++) {
+        src_vec[i] = static_cast<T>(tmp_vec[i] * row_inv_var * k_norm_vec[i]);
+      }
+    }
+    Store<T, VecSize>(src_vec, &qkv_out[base_idx]);
+  }
+}
+
 template <typename T, int VecSize = 1>
 __global__ void GQAVariableLengthRotaryKernel(
     const T *qkv,
@@ -514,6 +614,7 @@ __global__ void GQAVariableLengthRotaryQuantKVKernel(const int *qkv,
     int ori_seq_id = (token_idx - cu_seqlens_q[ori_bi]) + seq_lens_decoder[ori_bi];
 
     const int64_t emb_idx = ori_seq_id * half_lastdim + h_bias / 2;
+    int64_t new_emb_idx = rope_3d ? emb_idx + ori_bi * last_dim * seq_len : emb_idx;
     const int64_t bias_idx = hi * last_dim + h_bias;
     const int64_t base_idx = token_idx * offset + bias_idx;
     Load<int, VecSize>(&qkv[base_idx], &src_vec);
@@ -521,8 +622,8 @@ __global__ void GQAVariableLengthRotaryQuantKVKernel(const int *qkv,
       Load<T, VecSize>(&qkv_biases[bias_idx], &bias_vec);
     }
     Load<float, VecSize>(&qkv_out_scales[bias_idx], &out_scale_vec);
-    Load<float, HalfVecSize>(&cos_emb[emb_idx], &cos_emb_vec);
-    Load<float, HalfVecSize>(&sin_emb[emb_idx], &sin_emb_vec);
+    Load<float, HalfVecSize>(&cos_emb[new_emb_idx], &cos_emb_vec);
+    Load<float, HalfVecSize>(&sin_emb[new_emb_idx], &sin_emb_vec);
 #pragma unroll
     for (int i = 0; i < HalfVecSize; i++) {
       float input_left = static_cast<float>(src_vec[2 * i]);
@@ -599,14 +700,15 @@ __global__ void GQAVariableLengthRotaryQuantKVKernel(const T *qkv,
     int ori_seq_id = (token_idx - cu_seqlens_q[ori_bi]) + seq_lens_decoder[ori_bi];
 
     const int64_t emb_idx = ori_seq_id * half_lastdim + h_bias / 2;
+    int64_t new_emb_idx = rope_3d ? emb_idx + ori_bi * last_dim * seq_len : emb_idx;
     const int64_t bias_idx = hi * last_dim + h_bias;
     const int64_t base_idx = token_idx * offset + bias_idx;
     Load<T, VecSize>(&qkv[base_idx], &src_vec);
     if (qkv_biases) {
       Load<T, VecSize>(&qkv_biases[bias_idx], &bias_vec);
     }
-    Load<float, HalfVecSize>(&cos_emb[emb_idx], &cos_emb_vec);
-    Load<float, HalfVecSize>(&sin_emb[emb_idx], &sin_emb_vec);
+    Load<float, HalfVecSize>(&cos_emb[new_emb_idx], &cos_emb_vec);
+    Load<float, HalfVecSize>(&sin_emb[new_emb_idx], &sin_emb_vec);
 #pragma unroll
     for (int i = 0; i < HalfVecSize; i++) {
       const float input_left = qkv_biases ? static_cast<float>(src_vec[2 * i]+ bias_vec[2 * i]) : static_cast<float>(src_vec[2 * i]);
@@ -654,7 +756,8 @@ __global__ void GQANeoxVariableLengthRotaryKernel(
     const int q_num_head,
     const int kv_num_head,
     const int seq_len,
-    const int last_dim) {
+    const int last_dim,
+    const bool rope_3d) {
   using LoadT = AlignedVector<int, VecSize>;
   using LoadBiasT = AlignedVector<T, VecSize>;
   using LoadScaleT = AlignedVector<float, VecSize>;
@@ -684,6 +787,7 @@ __global__ void GQANeoxVariableLengthRotaryKernel(
     const int ori_seq_id = (token_idx - cu_seqlens_q[ori_bi]) + seq_lens_decoder[ori_bi];
 
     const int emb_idx = ori_seq_id * last_dim + h_bias;
+    int new_emb_idx = rope_3d ? emb_idx + ori_bi * last_dim * seq_len * 2 : emb_idx;
     const int bias_idx_left = hi * last_dim + h_bias;
     const int bias_idx_right = bias_idx_left + half_lastdim;
     const int base_idx_left =
@@ -698,8 +802,8 @@ __global__ void GQANeoxVariableLengthRotaryKernel(
     Load<float, VecSize>(&qkv_out_scales[bias_idx_left], &left_out_scale_vec);
     Load<float, VecSize>(&qkv_out_scales[bias_idx_right], &right_out_scale_vec);
     if (hi < (q_num_head + kv_num_head)) {
-      Load<float, VecSize>(&cos_emb[emb_idx], &cos_emb_vec);
-      Load<float, VecSize>(&sin_emb[emb_idx], &sin_emb_vec);
+      Load<float, VecSize>(&cos_emb[new_emb_idx], &cos_emb_vec);
+      Load<float, VecSize>(&sin_emb[new_emb_idx], &sin_emb_vec);
     }
 #pragma unroll
     for (int i = 0; i < VecSize; i++) {
@@ -745,7 +849,8 @@ __global__ void GQANeoxVariableLengthRotaryKernel(
     const int q_num_head,
     const int kv_num_head,
     const int seq_len,
-    const int last_dim) {
+    const int last_dim,
+    const bool rope_3d) {
   using LoadT = AlignedVector<T, VecSize>;
   using LoadEmbT = AlignedVector<float, VecSize>;
   LoadT left_vec;
@@ -769,6 +874,7 @@ __global__ void GQANeoxVariableLengthRotaryKernel(
     const int ori_seq_id = (token_idx - cu_seqlens_q[ori_bi]) + seq_lens_decoder[ori_bi];
 
     const int emb_idx = ori_seq_id * last_dim + h_bias;
+    int64_t new_emb_idx = rope_3d ? emb_idx + ori_bi * last_dim * seq_len * 2 : emb_idx;
     const int base_idx_left =
         token_idx * (q_num_head + 2 * kv_num_head) * last_dim + hi * last_dim +
         h_bias;
@@ -776,8 +882,76 @@ __global__ void GQANeoxVariableLengthRotaryKernel(
 
     Load<T, VecSize>(&qkv[base_idx_left], &left_vec);
     Load<T, VecSize>(&qkv[base_idx_right], &right_vec);
-    Load<float, VecSize>(&cos_emb[emb_idx], &cos_emb_vec);
-    Load<float, VecSize>(&sin_emb[emb_idx], &sin_emb_vec);
+    Load<float, VecSize>(&cos_emb[new_emb_idx], &cos_emb_vec);
+    Load<float, VecSize>(&sin_emb[new_emb_idx], &sin_emb_vec);
+#pragma unroll
+    for (int i = 0; i < VecSize; i++) {
+      const float input_left = static_cast<float>(left_vec[i]);
+      const float input_right = static_cast<float>(right_vec[i]);
+      const float cos_tmp = cos_emb_vec[i];
+      const float sin_tmp = sin_emb_vec[i];
+      left_vec[i] =
+          static_cast<T>(input_left * cos_tmp - input_right * sin_tmp);
+      right_vec[i] =
+          static_cast<T>(input_right * cos_tmp + input_left * sin_tmp);
+    }
+    Store<T, VecSize>(left_vec, &qkv_out[base_idx_left]);
+    Store<T, VecSize>(right_vec, &qkv_out[base_idx_right]);
+  }
+}
+
+template <typename T, int VecSize = 1>
+__global__ void GQANeoxVariableLengthPartialRotaryKernel(
+    const T *qkv,
+    const float *cos_emb,
+    const float *sin_emb,
+    const int *batch_id_per_token,
+    const int *cu_seqlens_q,
+    const int *seq_lens,
+    const int *seq_lens_decoder,
+    const float *qkv_out_scales,
+    const T *qkv_biases,
+    T *qkv_out,
+    const int64_t elem_cnt,
+    const int q_num_head,
+    const int kv_num_head,
+    const int seq_len,
+    const int head_dim,
+    const int rotary_dim,
+    const bool rope_3d) {
+  using LoadT = AlignedVector<T, VecSize>;
+  using LoadEmbT = AlignedVector<float, VecSize>;
+  LoadT left_vec;
+  LoadT right_vec;
+  LoadEmbT cos_emb_vec;
+  LoadEmbT sin_emb_vec;
+  int64_t global_thread_idx = blockDim.x * blockIdx.x + threadIdx.x;
+  const int rotary_dim_half = rotary_dim / 2;
+  const int offset = (q_num_head + kv_num_head) * rotary_dim_half;
+  for (int64_t linear_index = global_thread_idx * VecSize,
+               step = gridDim.x * blockDim.x * VecSize;
+       linear_index < elem_cnt;
+       linear_index += step) {
+    const int token_idx = linear_index / offset;
+    const int ori_bi = batch_id_per_token[token_idx];
+    if (seq_lens && seq_lens[ori_bi] == 0) continue;
+    const int bias = linear_index % offset;
+    const int hi = bias / rotary_dim_half;
+    const int h_bias = bias % rotary_dim_half;
+
+    const int ori_seq_id = (token_idx - cu_seqlens_q[ori_bi]) + seq_lens_decoder[ori_bi];
+
+    const int emb_idx = ori_seq_id * rotary_dim_half + h_bias;
+    int64_t new_emb_idx = rope_3d ? emb_idx + ori_bi * head_dim * seq_len * 2 : emb_idx;
+    const int base_idx_left =
+        token_idx * (q_num_head + 2 * kv_num_head) * head_dim + hi * head_dim +
+        h_bias;
+    const int base_idx_right = base_idx_left + rotary_dim_half;
+
+    Load<T, VecSize>(&qkv[base_idx_left], &left_vec);
+    Load<T, VecSize>(&qkv[base_idx_right], &right_vec);
+    Load<float, VecSize>(&cos_emb[new_emb_idx], &cos_emb_vec);
+    Load<float, VecSize>(&sin_emb[new_emb_idx], &sin_emb_vec);
 #pragma unroll
     for (int i = 0; i < VecSize; i++) {
       const float input_left = static_cast<float>(left_vec[i]);
@@ -1512,7 +1686,8 @@ void rotary_qk_variable(
               elem_nums,
               head_num,
               seq_len,
-              dim_head);
+              dim_head,
+              rope_3d);
     } else {
       VariableLengthRotaryKernel<T, PackSize>
           <<<grid_size, blocksize, 0, stream>>>(
@@ -1527,7 +1702,8 @@ void rotary_qk_variable(
               elem_nums,
               head_num,
               seq_len,
-              dim_head);
+              dim_head,
+              rope_3d);
     }
   } else {
     const float *cos_emb = rotary_emb;
@@ -1548,7 +1724,8 @@ void rotary_qk_variable(
               elem_nums,
               head_num,
               seq_len,
-              dim_head);
+              dim_head,
+              rope_3d);
     } else {
       NeoxVariableLengthRotaryKernel<T, PackSize>
           <<<grid_size, blocksize, 0, stream>>>(
@@ -1563,11 +1740,72 @@ void rotary_qk_variable(
               elem_nums,
               head_num,
               seq_len,
-              dim_head);
+              dim_head,
+              rope_3d);
     }
   }
 }
 
+template <typename T, typename QKV_TYPE>
+void gqa_rotary_qk_norm_variable(
+    T *qkv_out,                   // [token_num, 3, num_head, dim_head]
+    const QKV_TYPE *qkv_input,    // qkv
+    const float *qkv_out_scales,  // [3, num_head, dim_head]
+    const T *qkv_bias,
+    const float *rotary_emb,  // [2, 1, 1, seq_len, dim_head / 2]
+    const int *batch_id_per_token,
+    const int *cu_seqlens_q,
+    const int *seq_lens,
+    const int *seq_lens_decoder,
+    const int token_num,
+    const int num_heads,
+    const int kv_num_heads,
+    const int seq_len,
+    const int input_output_len,
+    const int dim_head,
+    const cudaStream_t &stream,
+    bool use_neox_style = false,
+    bool rope_3d = false,
+    const float *q_norm_weight = nullptr,
+    const float *k_norm_weight = nullptr,
+    const float rms_norm_eps = 1e-6) {
+  int64_t elem_nums =
+      qkv_out_scales
+          ? token_num * (num_heads + 2 * kv_num_heads) * dim_head
+          : token_num * (num_heads + kv_num_heads) * dim_head;  // for all q k v
+  assert(dim_head == 128 && "dim_head must be 128");
+  constexpr int HEAD_DIM = 128;
+  constexpr int PackSize = HEAD_DIM / kWarpSize;
+  const int pack_num = elem_nums / PackSize;
+  const int blocksize = 128;
+  int grid_size = 1;
+  GetNumBlocks<128>(pack_num, &grid_size);
+  dim3 Block_Size(kWarpSize, blocksize/kWarpSize, 1);
+
+  const float *cos_emb = rotary_emb;
+  const float *sin_emb = rotary_emb + input_output_len * dim_head / 2;
+
+  GQAVariableLengthRotaryQKNormKernel<T, PackSize>
+      <<<grid_size, Block_Size, 0, stream>>>(
+          reinterpret_cast<const T *>(qkv_input),
+          cos_emb,
+          sin_emb,
+          batch_id_per_token,
+          cu_seqlens_q,
+          seq_lens,
+          seq_lens_decoder,
+          qkv_out,
+          elem_nums,
+          num_heads,
+          kv_num_heads,
+          seq_len,
+          dim_head,
+          rope_3d,
+          q_norm_weight,
+          k_norm_weight,
+          rms_norm_eps);
+}
+
 template <typename T, typename QKV_TYPE>
 void gqa_rotary_qk_variable(
     T *qkv_out,                   // [token_num, 3, num_head, dim_head]
@@ -1585,6 +1823,7 @@ void gqa_rotary_qk_variable(
     const int seq_len,
     const int input_output_len,
     const int dim_head,
+    const int rotary_dim,
     const cudaStream_t &stream,
     bool use_neox_style = false,
     bool rope_3d = false) {
@@ -1662,9 +1901,41 @@ void gqa_rotary_qk_variable(
               num_heads,
               kv_num_heads,
               seq_len,
-              dim_head);
+              dim_head,
+              rope_3d);
     } else {
-      GQANeoxVariableLengthRotaryKernel<T, PackSize>
+      if (rotary_dim < dim_head){
+        PD_CHECK((rotary_dim / 2) % PackSize == 0);
+        elem_nums =
+            qkv_out_scales
+                ? token_num * (num_heads + 2 * kv_num_heads) * rotary_dim
+                : token_num * (num_heads + kv_num_heads) * rotary_dim;  // for all q k v
+        if (use_neox_style) {
+          elem_nums /= 2;
+        }
+        const int pack_num_new = elem_nums / PackSize;
+        GetNumBlocks<128>(pack_num_new, &grid_size);
+        GQANeoxVariableLengthPartialRotaryKernel<T, PackSize>
+            <<<grid_size, blocksize, 0, stream>>>(
+                reinterpret_cast<const T *>(qkv_input),
+                cos_emb,
+                rotary_emb + input_output_len * rotary_dim / 2,
+                batch_id_per_token,
+                cu_seqlens_q,
+                seq_lens,
+                seq_lens_decoder,
+                qkv_out_scales,
+                qkv_bias,
+                qkv_out,
+                elem_nums,
+                num_heads,
+                kv_num_heads,
+                seq_len,
+                dim_head,
+                rotary_dim,
+                rope_3d);
+      }else{
+        GQANeoxVariableLengthRotaryKernel<T, PackSize>
           <<<grid_size, blocksize, 0, stream>>>(
               reinterpret_cast<const T *>(qkv_input),
               cos_emb,
@@ -1680,7 +1951,9 @@ void gqa_rotary_qk_variable(
               num_heads,
               kv_num_heads,
               seq_len,
-              dim_head);
+              dim_head,
+              rope_3d);
+      }
     }
   }
 }

custom_ops/gpu_ops/append_attn/encoder_write_cache_with_rope_kernel.h

@@ -46,38 +46,32 @@ void EncoderWriteCacheWithRopeKernel(
     cudaStream_t& stream,
     paddle::Tensor* qkv_out,
     paddle::Tensor* key_cache_out,
-    paddle::Tensor* value_cache_out) {
+    paddle::Tensor* value_cache_out,
+    const paddle::optional<paddle::Tensor>& q_norm_weight,
+    const paddle::optional<paddle::Tensor>& k_norm_weight,
+    const float rms_norm_eps) {
   auto token_num = meta_data.token_nums;
   auto num_heads = meta_data.q_num_heads;
   auto kv_num_heads = meta_data.kv_num_heads;
   auto head_dim = meta_data.head_dims;
   bool is_scale_channel_wise = false;
+  int rotary_dim = head_dim;
   if (cache_k_scale && cache_k_scale.get().dims()[0] == head_dim * kv_num_heads) {
     is_scale_channel_wise = true;
   }
+  if (rotary_embs){
+    rotary_dim = rotary_embs.get().dims()[rotary_embs.get().dims().size()-1] * 2;
+    if(rotary_dim < head_dim){
+      if (!use_neox_style || q_norm_weight || k_norm_weight || num_heads == kv_num_heads || is_scale_channel_wise){
+        PADDLE_THROW(phi::errors::Fatal(
+          "partial_rotary_factor < 1.0 only supports use_neox_rotary_style=True, q_norm_weight/k_norm_weight) is None, GQA and is_scale_channel_wise=false."));
+      }
+    }
+  }
 
-  if (num_heads == kv_num_heads) {
-    rotary_qk_variable(
-        qkv_out->data<T>(),
-        qkv.data<QKV_TYPE>(),
-        qkv_out_scales ? qkv_out_scales.get().data<float>() : nullptr,
-        qkv_biases ? qkv_biases.get().data<T>() : nullptr,
-        rotary_embs.get().data<float>(),
-        batch_id_per_token.data<int>(),
-        cu_seqlens_q.data<int>(),
-        seq_lens_encoder.data<int>(),
-        seq_lens_decoder.data<int>(),
-        token_num,
-        num_heads,
-        max_seq_len,
-        rotary_embs.get().dims()[2],
-        head_dim,
-        stream,
-        use_neox_style,
-        rope_3d);
-  } else {
-    if (!is_scale_channel_wise) {
-      gqa_rotary_qk_variable(
+  if (q_norm_weight && k_norm_weight) {
+    if (num_heads != kv_num_heads && !is_scale_channel_wise && !use_neox_style) {
+      gqa_rotary_qk_norm_variable(
         qkv_out->data<T>(),
         qkv.data<QKV_TYPE>(),
         qkv_out_scales ? qkv_out_scales.get().data<float>() : nullptr,
@@ -95,31 +89,81 @@ void EncoderWriteCacheWithRopeKernel(
         head_dim,
         stream,
         use_neox_style,
-        rope_3d);
+        rope_3d,
+        q_norm_weight ? q_norm_weight.get().data<float>() : nullptr,
+        k_norm_weight ? k_norm_weight.get().data<float>() : nullptr,
+        rms_norm_eps);
     } else {
-      gqa_rotary_qk_quant_variable(
-        qkv_out->data<T>(),
-        qkv.data<QKV_TYPE>(),
-        qkv_out_scales ? qkv_out_scales.get().data<float>() : nullptr,
-        qkv_biases ? qkv_biases.get().data<T>() : nullptr,
-        cache_k_scale ? cache_k_scale.get().data<T>() : nullptr,
-        cache_v_scale ? cache_v_scale.get().data<T>() : nullptr,
-        rotary_embs.get().data<float>(),
-        batch_id_per_token.data<int>(),
-        cu_seqlens_q.data<int>(),
-        seq_lens_encoder.data<int>(),
-        seq_lens_decoder.data<int>(),
-        token_num,
-        num_heads,
-        kv_num_heads,
-        max_seq_len,
-        rotary_embs.get().dims()[2],
-        head_dim,
-        stream,
-        use_neox_style,
-        rope_3d);
+      PD_THROW(
+          "gqa_rotary_qk_norm_variable only support gqa mode. channel wise scale and neox style are not supported");
     }
+  } else {
+    if (num_heads == kv_num_heads) {
+      rotary_qk_variable(
+          qkv_out->data<T>(),
+          qkv.data<QKV_TYPE>(),
+          qkv_out_scales ? qkv_out_scales.get().data<float>() : nullptr,
+          qkv_biases ? qkv_biases.get().data<T>() : nullptr,
+          rotary_embs.get().data<float>(),
+          batch_id_per_token.data<int>(),
+          cu_seqlens_q.data<int>(),
+          seq_lens_encoder.data<int>(),
+          seq_lens_decoder.data<int>(),
+          token_num,
+          num_heads,
+          max_seq_len,
+          rotary_embs.get().dims()[2],
+          head_dim,
+          stream,
+          use_neox_style,
+          rope_3d);
+    } else {
+      if (!is_scale_channel_wise) {
+        gqa_rotary_qk_variable(
+          qkv_out->data<T>(),
+          qkv.data<QKV_TYPE>(),
+          qkv_out_scales ? qkv_out_scales.get().data<float>() : nullptr,
+          qkv_biases ? qkv_biases.get().data<T>() : nullptr,
+          rotary_embs.get().data<float>(),
+          batch_id_per_token.data<int>(),
+          cu_seqlens_q.data<int>(),
+          seq_lens_encoder.data<int>(),
+          seq_lens_decoder.data<int>(),
+          token_num,
+          num_heads,
+          kv_num_heads,
+          max_seq_len,
+          rope_3d ? rotary_embs.get().dims()[3] : rotary_embs.get().dims()[2],
+          head_dim,
+          rotary_dim,
+          stream,
+          use_neox_style,
+          rope_3d);
+      } else {
+        gqa_rotary_qk_quant_variable(
+          qkv_out->data<T>(),
+          qkv.data<QKV_TYPE>(),
+          qkv_out_scales ? qkv_out_scales.get().data<float>() : nullptr,
+          qkv_biases ? qkv_biases.get().data<T>() : nullptr,
+          cache_k_scale ? cache_k_scale.get().data<T>() : nullptr,
+          cache_v_scale ? cache_v_scale.get().data<T>() : nullptr,
+          rotary_embs.get().data<float>(),
+          batch_id_per_token.data<int>(),
+          cu_seqlens_q.data<int>(),
+          seq_lens_encoder.data<int>(),
+          seq_lens_decoder.data<int>(),
+          token_num,
+          num_heads,
+          kv_num_heads,
+          max_seq_len,
+          rotary_embs.get().dims()[2],
+          head_dim,
+          stream,
+          use_neox_style,
+          rope_3d);
+      }
 
+    }
   }
   const uint32_t block_size = meta_data.block_size;
   if (cache_quant_type_str == "none") {

custom_ops/gpu_ops/append_attn/get_block_shape_and_split_kv_block.cu

@@ -195,22 +195,25 @@ std::vector<paddle::Tensor> GetBlockShapeAndSplitKVBlock(
     const paddle::Tensor &seq_lens_encoder,
     const paddle::Tensor &seq_lens_decoder,
     const paddle::Tensor &seq_lens_this_time,
-    const int encoder_block_shape_q, const int decoder_block_shape_q,
-    const int group_size, const int block_size,
-    const int decoder_step_token_num) {
+    paddle::Tensor &decoder_batch_ids,          // Inplace
+    paddle::Tensor &decoder_tile_ids_per_batch, // Inplace
+    paddle::Tensor &decoder_num_blocks_x_cpu,   // Inplace, Pinned Memory
+    paddle::Tensor &max_len_tensor_cpu,         // Inplace, Pinned Memory
+    const int encoder_block_shape_q,
+    const int decoder_block_shape_q,
+    const int group_size,
+    const int block_size,
+    const int decoder_step_token_num)
+{
   auto stream = seq_lens_encoder.stream();
   int bsz = seq_lens_this_time.shape()[0];
-  auto max_len_tensor =
-      GetEmptyTensor({8}, paddle::DataType::INT32, seq_lens_encoder.place());
-  GetMaxLen(seq_lens_decoder, seq_lens_this_time, seq_lens_encoder,
-            max_len_tensor, bsz);
 
-  // max_len_this_time, max_enc_len_this_time, max_dec_len_this_time,
-  // max_enc_dec_len_this_time, max_just_dec_len_this_time,
-  // max_just_dec_merged_len_this_time, max_system_len,
-  // max_just_dec_len_without_system
-  auto max_len_cpu = max_len_tensor.copy_to(paddle::CPUPlace(), false);
-  auto max_len_cpu_ptr = max_len_cpu.data<int>();
+  paddle::Tensor max_len_tensor_gpu = GetEmptyTensor({max_len_tensor_cpu.shape()[0]}, paddle::DataType::INT32, seq_lens_this_time.place());
+  GetMaxLen(seq_lens_decoder, seq_lens_this_time, seq_lens_encoder,
+            max_len_tensor_gpu, bsz);
+  max_len_tensor_cpu.copy_(max_len_tensor_gpu, max_len_tensor_cpu.place(), false);
+
+  auto max_len_cpu_ptr = max_len_tensor_cpu.data<int>();
   int max_len_this_time = max_len_cpu_ptr[0];
   int max_enc_len_this_time = max_len_cpu_ptr[1];
   int max_dec_len_this_time = max_len_cpu_ptr[2];
@@ -222,14 +225,11 @@ std::vector<paddle::Tensor> GetBlockShapeAndSplitKVBlock(
 
   paddle::Tensor encoder_batch_ids;
   paddle::Tensor encoder_tile_ids_per_batch;
-  paddle::Tensor encoder_num_blocks_x_cpu; /*cpu*/
+  paddle::Tensor encoder_num_blocks_x_cpu;  /*cpu*/
   paddle::Tensor kv_batch_ids;
   paddle::Tensor kv_tile_ids_per_batch;
-  paddle::Tensor kv_num_blocks_x_cpu; /*cpu*/
-  paddle::Tensor decoder_batch_ids;
-  paddle::Tensor decoder_tile_ids_per_batch;
-  paddle::Tensor decoder_num_blocks_x_cpu; /*cpu*/
-  paddle::Tensor max_len_kv_cpu;           /*cpu*/
+  paddle::Tensor kv_num_blocks_x_cpu;       /*cpu*/
+  paddle::Tensor max_len_kv_cpu;            /*cpu*/
 
   auto max_len_kv =
       GetEmptyTensor({1}, paddle::DataType::INT32, seq_lens_decoder.place());
@@ -289,94 +289,66 @@ std::vector<paddle::Tensor> GetBlockShapeAndSplitKVBlock(
     kv_tile_ids_per_batch =
         GetEmptyTensor({0}, paddle::DataType::INT32, seq_lens_encoder.place());
     kv_num_blocks_x_cpu =
-        GetEmptyTensor({0}, paddle::DataType::INT32, seq_lens_encoder.place());
-  }
-  if (max_just_dec_len_this_time > 0) {
-    const uint32_t decoder_max_tile_size_per_bs_q =
-        div_up((decoder_step_token_num * group_size), decoder_block_shape_q);
-
-    decoder_batch_ids =
-        GetEmptyTensor({bsz * decoder_max_tile_size_per_bs_q},
-                       paddle::DataType::INT32, seq_lens_encoder.place());
-    decoder_tile_ids_per_batch =
-        GetEmptyTensor({bsz * decoder_max_tile_size_per_bs_q},
-                       paddle::DataType::INT32, seq_lens_encoder.place());
-    auto decoder_num_blocks_x =
-        GetEmptyTensor({1}, paddle::DataType::INT32, seq_lens_encoder.place());
-    split_q_block<<<1, 32, 0, stream>>>(
-        seq_lens_this_time.data<int>(), seq_lens_encoder.data<int>(),
-        decoder_batch_ids.data<int>(), decoder_tile_ids_per_batch.data<int>(),
-        decoder_num_blocks_x.data<int>(), bsz, decoder_block_shape_q,
-        group_size);
-    decoder_num_blocks_x_cpu =
-        decoder_num_blocks_x.copy_to(paddle::CPUPlace(), false);
-  } else {
-    decoder_batch_ids =
-        GetEmptyTensor({0}, paddle::DataType::INT32, seq_lens_encoder.place());
-    decoder_tile_ids_per_batch =
-        GetEmptyTensor({0}, paddle::DataType::INT32, seq_lens_encoder.place());
-    decoder_num_blocks_x_cpu =
         GetEmptyTensor({0}, paddle::DataType::INT32, paddle::CPUPlace());
   }
 
-  return {encoder_batch_ids,
-          encoder_tile_ids_per_batch,
-          encoder_num_blocks_x_cpu, /*cpu*/
-          kv_batch_ids,
-          kv_tile_ids_per_batch,
-          kv_num_blocks_x_cpu, /*cpu*/
-          decoder_batch_ids,
-          decoder_tile_ids_per_batch,
-          decoder_num_blocks_x_cpu, /*cpu*/
-          max_len_kv_cpu /*cpu*/,
-          max_len_cpu};
-}
+  if (max_just_dec_len_this_time > 0) {
+    // Clear buffer
+    const uint32_t decoder_max_tile_size_per_bs_q = div_up((decoder_step_token_num * group_size), decoder_block_shape_q);
+    const uint32_t decoder_batch_shape = bsz * decoder_max_tile_size_per_bs_q;
+    PADDLE_ENFORCE_GPU_SUCCESS(cudaMemsetAsync(decoder_batch_ids.data<int>(), 0, decoder_batch_shape * sizeof(int32_t), stream));
+    PADDLE_ENFORCE_GPU_SUCCESS(cudaMemsetAsync(decoder_tile_ids_per_batch.data<int>(), 0, decoder_batch_shape * sizeof(int32_t), stream));
+    PADDLE_ENFORCE_GPU_SUCCESS(cudaMemsetAsync(decoder_num_blocks_x_cpu.data<int>(), 0, sizeof(int32_t), stream));
+
+    auto decoder_num_blocks_x =
+        GetEmptyTensor({1}, paddle::DataType::INT32, seq_lens_encoder.place());
+    split_q_block<<<1, 32, 0, stream>>>(
+        seq_lens_this_time.data<int>(),
+        seq_lens_encoder.data<int>(),
+        decoder_batch_ids.data<int>(),
+        decoder_tile_ids_per_batch.data<int>(),
+        decoder_num_blocks_x.data<int>(),
+        bsz,
+        decoder_block_shape_q,
+        group_size);
+    decoder_num_blocks_x_cpu.copy_(decoder_num_blocks_x, decoder_num_blocks_x_cpu.place(), false);
+  }
 
-std::vector<paddle::DataType> GetBlockShapeAndSplitKVBlockInferDtype(
-    const paddle::DataType &seq_lens_encoder_dtype,
-    const paddle::DataType &seq_lens_decoder_dtype,
-    const paddle::DataType &seq_lens_this_time_dtype) {
   return {
-      paddle::DataType::INT32, paddle::DataType::INT32, paddle::DataType::INT32,
-      paddle::DataType::INT32, paddle::DataType::INT32, paddle::DataType::INT32,
-      paddle::DataType::INT32, paddle::DataType::INT32, paddle::DataType::INT32,
-      paddle::DataType::INT32, paddle::DataType::INT32};
-}
-
-std::vector<std::vector<int64_t>> GetBlockShapeAndSplitKVBlockInferShape(
-    const std::vector<int64_t> &seq_lens_encoder_shape,
-    const std::vector<int64_t> &seq_lens_decoder_shape,
-    const std::vector<int64_t> &seq_lens_this_time_shape) {
-  std::vector<int64_t> dynamic_shape = {-1};
-
-  return {dynamic_shape,
-          dynamic_shape,
-          {1},
-          dynamic_shape,
-          dynamic_shape,
-          {1},
-          dynamic_shape,
-          dynamic_shape,
-          {1},
-          {1},
-          {8}};
+    encoder_batch_ids,
+    encoder_tile_ids_per_batch,
+    encoder_num_blocks_x_cpu, /*cpu*/
+    kv_batch_ids,
+    kv_tile_ids_per_batch,
+    kv_num_blocks_x_cpu,      /*cpu*/
+    max_len_kv_cpu,           /*cpu*/
+  };
 }
 
 PD_BUILD_STATIC_OP(get_block_shape_and_split_kv_block)
-    .Inputs({"seq_lens_encoder", "seq_lens_decoder", "seq_lens_this_time"})
-    .Outputs({paddle::Optional("encoder_batch_ids"),
-              paddle::Optional("encoder_tile_ids_per_batch"),
-              paddle::Optional("encoder_num_blocks"),
-              paddle::Optional("kv_batch_ids"),
-              paddle::Optional("kv_tile_ids_per_batch"),
-              paddle::Optional("kv_num_blocks"),
-              paddle::Optional("decoder_batch_ids"),
-              paddle::Optional("decoder_tile_ids_per_batch"),
-              paddle::Optional("decoder_num_blocks"),
-              paddle::Optional("max_len_kv"), "set_max_lengths"})
-    .Attrs({"encoder_block_shape_q: int", "decoder_block_shape_q: int",
-            "group_size: int", "block_size: int",
-            "decoder_step_token_num: int"})
-    .SetKernelFn(PD_KERNEL(GetBlockShapeAndSplitKVBlock))
-    .SetInferShapeFn(PD_INFER_SHAPE(GetBlockShapeAndSplitKVBlockInferShape))
-    .SetInferDtypeFn(PD_INFER_DTYPE(GetBlockShapeAndSplitKVBlockInferDtype));
+    .Inputs({
+      "seq_lens_encoder",
+      "seq_lens_decoder",
+      "seq_lens_this_time",
+      "decoder_batch_ids",
+      "decoder_tile_ids_per_batch",
+      "decoder_num_blocks_x_cpu",
+      "max_len_tensor_cpu"
+    })
+    .Outputs({
+      paddle::Optional("encoder_batch_ids"),
+      paddle::Optional("encoder_tile_ids_per_batch"),
+      paddle::Optional("encoder_num_blocks_x_cpu"),
+      paddle::Optional("kv_batch_ids"),
+      paddle::Optional("kv_tile_ids_per_batch"),
+      paddle::Optional("kv_num_blocks_x_cpu"),
+      "max_len_kv_cpu"
+    })
+    .Attrs({
+      "encoder_block_shape_q: int",
+      "decoder_block_shape_q: int",
+      "group_size: int",
+      "block_size: int",
+      "decoder_step_token_num: int"
+    })
+    .SetKernelFn(PD_KERNEL(GetBlockShapeAndSplitKVBlock));

custom_ops/gpu_ops/append_attn/gqa_rope_write_cache.cu

@@ -37,7 +37,8 @@ __global__ void GQAVariableLengthRotarySplitKernel(
     const int q_num_head,
     const int kv_num_head,
     const int seq_len,
-    const int last_dim) {
+    const int last_dim,
+    const bool rope_3d) {
   using LoadT = AlignedVector<T, VecSize>;
   constexpr int HalfVecSize = VecSize / 2;
   using LoadEmbT = AlignedVector<float, HalfVecSize>;
@@ -62,6 +63,7 @@ __global__ void GQAVariableLengthRotarySplitKernel(
     const int kv_write_idx = cu_seqlens_k[ori_bi] + ori_seq_id;
 
     const int64_t emb_idx = ori_seq_id * half_lastdim + h_bias / 2;
+    int64_t new_emb_idx = rope_3d ? emb_idx + ori_bi * last_dim * seq_len : emb_idx;
     const int64_t base_idx =
         token_idx * (q_num_head + 2 * kv_num_head) * last_dim + hi * last_dim +
         h_bias;
@@ -80,8 +82,8 @@ __global__ void GQAVariableLengthRotarySplitKernel(
     Load<T, VecSize>(&qkv[base_idx], &src_vec);
     // do rope
     if (hi < q_num_head + kv_num_head) {
-      Load<float, HalfVecSize>(&cos_emb[emb_idx], &cos_emb_vec);
-      Load<float, HalfVecSize>(&sin_emb[emb_idx], &sin_emb_vec);
+      Load<float, HalfVecSize>(&cos_emb[new_emb_idx], &cos_emb_vec);
+      Load<float, HalfVecSize>(&sin_emb[new_emb_idx], &sin_emb_vec);
 #pragma unroll
       for (int i = 0; i < HalfVecSize; i++) {
         const float input_left = static_cast<float>(src_vec[2 * i]);
@@ -118,6 +120,7 @@ void gqa_rotary_qk_split_variable(
     const int seq_len,
     const int input_output_len,
     const int dim_head,
+    const bool rope_3d,
     const cudaStream_t &stream) {
   int64_t elem_nums = token_num * (num_heads + 2 * kv_num_heads) * dim_head;
   constexpr int PackSize = 16 / sizeof(T);
@@ -146,7 +149,8 @@ void gqa_rotary_qk_split_variable(
             num_heads,
             kv_num_heads,
             seq_len,
-            dim_head);
+            dim_head,
+            rope_3d);
 }
 
 template <typename T,
@@ -586,9 +590,9 @@ __global__ void append_cache_kv_c4(
 #pragma unroll
   for (uint32_t i = wid * 32 + tid; i < HEAD_DIM; i += 128) {
     cache_k_scale_smem[i] = cache_k_scale_now[i];
-    cache_k_zero_point_smem[i] = cache_k_zp_now[i] - static_cast<T>(136.f);
+    cache_k_zero_point_smem[i] = cache_k_zp_now[i] + static_cast<T>(136.f);
     cache_v_scale_smem[i] = cache_v_scale_now[i];
-    cache_v_zero_point_smem[i] = cache_v_zp_now[i] - static_cast<T>(136.f);
+    cache_v_zero_point_smem[i] = cache_v_zp_now[i] + static_cast<T>(136.f);
   }
 
   smem_t k_smem(smem);
@@ -640,25 +644,25 @@ __global__ void append_cache_kv_c4(
         convert_int4(frag_dq_T + 8, k_frag[2 * i + 1]);
 
         if (row_idx < end_idx) {
-          k_tile_ptr0[0] = frag_dq_T[0] * cache_k_scale_smem[col_idx] + cache_k_zero_point_smem[col_idx];
-          k_tile_ptr0[1] = frag_dq_T[1] * cache_k_scale_smem[col_idx + 1] + cache_k_zero_point_smem[col_idx + 1];
-          k_tile_ptr0[8] = frag_dq_T[2] * cache_k_scale_smem[col_idx + 8] + cache_k_zero_point_smem[col_idx + 8];
-          k_tile_ptr0[9] = frag_dq_T[3] * cache_k_scale_smem[col_idx + 9] + cache_k_zero_point_smem[col_idx + 9];
-          k_tile_ptr0[16] = frag_dq_T[8] * cache_k_scale_smem[col_idx + 16] + cache_k_zero_point_smem[col_idx + 16];
-          k_tile_ptr0[17] = frag_dq_T[9] * cache_k_scale_smem[col_idx + 17] + cache_k_zero_point_smem[col_idx + 17];
-          k_tile_ptr0[24] = frag_dq_T[10] * cache_k_scale_smem[col_idx + 24] + cache_k_zero_point_smem[col_idx + 24];
-          k_tile_ptr0[25] = frag_dq_T[11] * cache_k_scale_smem[col_idx + 25] + cache_k_zero_point_smem[col_idx + 25];
+          k_tile_ptr0[0] = (frag_dq_T[0] - cache_k_zero_point_smem[col_idx]) * cache_k_scale_smem[col_idx];
+          k_tile_ptr0[1] = (frag_dq_T[1] - cache_k_zero_point_smem[col_idx + 1]) * cache_k_scale_smem[col_idx + 1];
+          k_tile_ptr0[8] = (frag_dq_T[2] - cache_k_zero_point_smem[col_idx + 8]) * cache_k_scale_smem[col_idx + 8];
+          k_tile_ptr0[9] = (frag_dq_T[3] - cache_k_zero_point_smem[col_idx + 9]) * cache_k_scale_smem[col_idx + 9];
+          k_tile_ptr0[16] = (frag_dq_T[8] - cache_k_zero_point_smem[col_idx + 16]) * cache_k_scale_smem[col_idx + 16];
+          k_tile_ptr0[17] = (frag_dq_T[9] - cache_k_zero_point_smem[col_idx + 17]) * cache_k_scale_smem[col_idx + 17];
+          k_tile_ptr0[24] = (frag_dq_T[10] - cache_k_zero_point_smem[col_idx + 24]) * cache_k_scale_smem[col_idx + 24];
+          k_tile_ptr0[25] = (frag_dq_T[11] - cache_k_zero_point_smem[col_idx + 25]) * cache_k_scale_smem[col_idx + 25];
         }
 
         if (row_idx + 8 < end_idx) {
-          k_tile_ptr1[0] = frag_dq_T[4] * cache_k_scale_smem[col_idx] + cache_k_zero_point_smem[col_idx];
-          k_tile_ptr1[1] = frag_dq_T[5] * cache_k_scale_smem[col_idx + 1] + cache_k_zero_point_smem[col_idx + 1];
-          k_tile_ptr1[8] = frag_dq_T[6] * cache_k_scale_smem[col_idx + 8] + cache_k_zero_point_smem[col_idx + 8];
-          k_tile_ptr1[9] = frag_dq_T[7] * cache_k_scale_smem[col_idx + 9] + cache_k_zero_point_smem[col_idx + 9];
-          k_tile_ptr1[16] = frag_dq_T[12] * cache_k_scale_smem[col_idx + 16] + cache_k_zero_point_smem[col_idx + 16];
-          k_tile_ptr1[17] = frag_dq_T[13] * cache_k_scale_smem[col_idx + 17] + cache_k_zero_point_smem[col_idx + 17];
-          k_tile_ptr1[24] = frag_dq_T[14] * cache_k_scale_smem[col_idx + 24] + cache_k_zero_point_smem[col_idx + 24];
-          k_tile_ptr1[25] = frag_dq_T[15] * cache_k_scale_smem[col_idx + 25] + cache_k_zero_point_smem[col_idx + 25];
+          k_tile_ptr1[0] = (frag_dq_T[4] - cache_k_zero_point_smem[col_idx]) * cache_k_scale_smem[col_idx];
+          k_tile_ptr1[1] = (frag_dq_T[5] - cache_k_zero_point_smem[col_idx + 1]) * cache_k_scale_smem[col_idx + 1];
+          k_tile_ptr1[8] = (frag_dq_T[6] - cache_k_zero_point_smem[col_idx + 8]) * cache_k_scale_smem[col_idx + 8];
+          k_tile_ptr1[9] = (frag_dq_T[7] - cache_k_zero_point_smem[col_idx + 9]) * cache_k_scale_smem[col_idx + 9];
+          k_tile_ptr1[16] = (frag_dq_T[12] - cache_k_zero_point_smem[col_idx + 16]) * cache_k_scale_smem[col_idx + 16];
+          k_tile_ptr1[17] = (frag_dq_T[13] - cache_k_zero_point_smem[col_idx + 17]) * cache_k_scale_smem[col_idx + 17];
+          k_tile_ptr1[24] = (frag_dq_T[14] - cache_k_zero_point_smem[col_idx + 24]) * cache_k_scale_smem[col_idx + 24];
+          k_tile_ptr1[25] = (frag_dq_T[15] - cache_k_zero_point_smem[col_idx + 25]) * cache_k_scale_smem[col_idx + 25];
         }
         col_idx += 32;
       }
@@ -711,36 +715,36 @@ __global__ void append_cache_kv_c4(
         convert_int4(frag_dq_T, v_frag[2 * i]);
         convert_int4(frag_dq_T + 8, v_frag[2 * i + 1]);
         if (kv_idx < end_idx) {
-          v_tile_ptr0[0] = frag_dq_T[0] * cache_v_scale_smem[dim_idx] + cache_v_zero_point_smem[dim_idx];
-          v_tile_ptr1[0] = frag_dq_T[4] * cache_v_scale_smem[dim_idx + 8] + cache_v_zero_point_smem[dim_idx + 8];
+          v_tile_ptr0[0] = (frag_dq_T[0] - cache_v_zero_point_smem[dim_idx]) * cache_v_scale_smem[dim_idx];
+          v_tile_ptr1[0] = (frag_dq_T[4] - cache_v_zero_point_smem[dim_idx + 8]) * cache_v_scale_smem[dim_idx + 8];
         }
         if (kv_idx + 1 < end_idx) {
-          v_tile_ptr0[kv_t_stride] = frag_dq_T[1] * cache_v_scale_smem[dim_idx] + cache_v_zero_point_smem[dim_idx];
-          v_tile_ptr1[kv_t_stride] = frag_dq_T[5] * cache_v_scale_smem[dim_idx + 8] + cache_v_zero_point_smem[dim_idx + 8];
+          v_tile_ptr0[kv_t_stride] = (frag_dq_T[1] - cache_v_zero_point_smem[dim_idx]) * cache_v_scale_smem[dim_idx];
+          v_tile_ptr1[kv_t_stride] = (frag_dq_T[5] - cache_v_zero_point_smem[dim_idx + 8]) * cache_v_scale_smem[dim_idx + 8];
         }
         if (kv_idx + 8 < end_idx) {
-          v_tile_ptr0[8 * kv_t_stride] = frag_dq_T[2] * cache_v_scale_smem[dim_idx] + cache_v_zero_point_smem[dim_idx];
-          v_tile_ptr1[8 * kv_t_stride] = frag_dq_T[6] * cache_v_scale_smem[dim_idx + 8] + cache_v_zero_point_smem[dim_idx + 8];
+          v_tile_ptr0[8 * kv_t_stride] = (frag_dq_T[2] - cache_v_zero_point_smem[dim_idx]) * cache_v_scale_smem[dim_idx];
+          v_tile_ptr1[8 * kv_t_stride] = (frag_dq_T[6] - cache_v_zero_point_smem[dim_idx + 8]) * cache_v_scale_smem[dim_idx + 8];
         }
         if (kv_idx + 9 < end_idx) {
-          v_tile_ptr0[9 * kv_t_stride] = frag_dq_T[3] * cache_v_scale_smem[dim_idx] + cache_v_zero_point_smem[dim_idx];
-          v_tile_ptr1[9 * kv_t_stride] = frag_dq_T[7] * cache_v_scale_smem[dim_idx + 8] + cache_v_zero_point_smem[dim_idx + 8];
+          v_tile_ptr0[9 * kv_t_stride] = (frag_dq_T[3] - cache_v_zero_point_smem[dim_idx]) * cache_v_scale_smem[dim_idx];
+          v_tile_ptr1[9 * kv_t_stride] = (frag_dq_T[7] - cache_v_zero_point_smem[dim_idx + 8]) * cache_v_scale_smem[dim_idx + 8];
         }
         if (kv_idx + 16 < end_idx) {
-          v_tile_ptr0[16 * kv_t_stride] = frag_dq_T[8] * cache_v_scale_smem[dim_idx] + cache_v_zero_point_smem[dim_idx];
-          v_tile_ptr1[16 * kv_t_stride] = frag_dq_T[12] * cache_v_scale_smem[dim_idx + 8] + cache_v_zero_point_smem[dim_idx + 8];
+          v_tile_ptr0[16 * kv_t_stride] = (frag_dq_T[8] - cache_v_zero_point_smem[dim_idx]) * cache_v_scale_smem[dim_idx];
+          v_tile_ptr1[16 * kv_t_stride] = (frag_dq_T[12] - cache_v_zero_point_smem[dim_idx + 8]) * cache_v_scale_smem[dim_idx + 8];
         }
         if (kv_idx + 17 < end_idx) {
-          v_tile_ptr0[17 * kv_t_stride] = frag_dq_T[9] * cache_v_scale_smem[dim_idx] + cache_v_zero_point_smem[dim_idx];
-          v_tile_ptr1[17 * kv_t_stride] = frag_dq_T[13] * cache_v_scale_smem[dim_idx + 8] + cache_v_zero_point_smem[dim_idx + 8];
+          v_tile_ptr0[17 * kv_t_stride] = (frag_dq_T[9] - cache_v_zero_point_smem[dim_idx]) * cache_v_scale_smem[dim_idx];
+          v_tile_ptr1[17 * kv_t_stride] = (frag_dq_T[13] - cache_v_zero_point_smem[dim_idx + 8]) * cache_v_scale_smem[dim_idx + 8];
         }
         if (kv_idx + 24 < end_idx) {
-          v_tile_ptr0[24 * kv_t_stride] = frag_dq_T[10] * cache_v_scale_smem[dim_idx] + cache_v_zero_point_smem[dim_idx];
-          v_tile_ptr1[24 * kv_t_stride] = frag_dq_T[14] * cache_v_scale_smem[dim_idx + 8] + cache_v_zero_point_smem[dim_idx + 8];
+          v_tile_ptr0[24 * kv_t_stride] = (frag_dq_T[10] - cache_v_zero_point_smem[dim_idx]) * cache_v_scale_smem[dim_idx];
+          v_tile_ptr1[24 * kv_t_stride] = (frag_dq_T[14] - cache_v_zero_point_smem[dim_idx + 8]) * cache_v_scale_smem[dim_idx + 8];
         }
         if (kv_idx + 25 < end_idx) {
-          v_tile_ptr0[25 * kv_t_stride] = frag_dq_T[11] * cache_v_scale_smem[dim_idx] + cache_v_zero_point_smem[dim_idx];
-          v_tile_ptr1[25 * kv_t_stride] = frag_dq_T[15] * cache_v_scale_smem[dim_idx + 8] + cache_v_zero_point_smem[dim_idx + 8];
+          v_tile_ptr0[25 * kv_t_stride] = (frag_dq_T[11] - cache_v_zero_point_smem[dim_idx]) * cache_v_scale_smem[dim_idx];
+          v_tile_ptr1[25 * kv_t_stride] = (frag_dq_T[15] - cache_v_zero_point_smem[dim_idx + 8]) * cache_v_scale_smem[dim_idx + 8];
         }
         kv_idx += 32;
       }
@@ -890,7 +894,8 @@ std::vector<paddle::Tensor> GQARopeWriteCacheKernel(
     const paddle::optional<paddle::Tensor>& kv_signal_data,
     const int kv_token_num,
     const int max_seq_len,
-    const std::string& cache_quant_type) {
+    const std::string& cache_quant_type,
+    const bool rope_3d) {
   typedef PDTraits<paddle::DataType::BFLOAT16> traits_;
   typedef typename traits_::DataType DataType_;
   typedef typename traits_::data_t data_t;
@@ -953,9 +958,34 @@ std::vector<paddle::Tensor> GQARopeWriteCacheKernel(
         num_heads,
         kv_num_heads,
         max_seq_len,
-        rotary_embs.dims()[2],
+        rope_3d ? rotary_embs.dims()[3] : rotary_embs.dims()[2],
         head_dim,
+        rope_3d,
         stream);
+
+  if (token_num < kv_token_num) {
+    AppendCacheKV<data_t, 128, 64>(
+      key_cache,
+      value_cache,
+      cache_k_dequant_scales.get(),
+      cache_v_dequant_scales.get(),
+      cache_k_zp.get(),
+      cache_v_zp.get(),
+      seq_lens_this_time,
+      seq_lens_decoder,
+      cu_seqlens_k,
+      block_tables,
+      cache_batch_ids,
+      cache_tile_ids,
+      cache_num_blocks,
+      max_blocks_per_seq,
+      kv_num_heads,
+      cache_quant_type,
+      &k,
+      &v,
+      stream
+    );
+  }
   // write cache
   if (cache_quant_type == "none") {
     CascadeAppendWriteCacheKVQKV<data_t>(
@@ -1038,30 +1068,6 @@ std::vector<paddle::Tensor> GQARopeWriteCacheKernel(
         }
       }
   }
-
-  if (token_num < kv_token_num) {
-    AppendCacheKV<data_t, 128, 64>(
-      key_cache,
-      value_cache,
-      cache_k_dequant_scales.get(),
-      cache_v_dequant_scales.get(),
-      cache_k_zp.get(),
-      cache_v_zp.get(),
-      seq_lens_this_time,
-      seq_lens_decoder,
-      cu_seqlens_k,
-      block_tables,
-      cache_batch_ids,
-      cache_tile_ids,
-      cache_num_blocks,
-      max_blocks_per_seq,
-      kv_num_heads,
-      cache_quant_type,
-      &k,
-      &v,
-      stream
-    );
-  }
   return {q, k, v, qkv_out};
 }
 

custom_ops/gpu_ops/append_attn/template_instantiation/encoder_write_cache_with_rope_bfloat16_bfloat16_kernel.cu

@@ -43,4 +43,7 @@ EncoderWriteCacheWithRopeKernel<paddle::bfloat16, paddle::bfloat16>(
     cudaStream_t& stream,
     paddle::Tensor* qkv_out,
     paddle::Tensor* key_cache_out,
-    paddle::Tensor* value_cache_out);
+    paddle::Tensor* value_cache_out,
+    const paddle::optional<paddle::Tensor>& q_norm_weight,
+    const paddle::optional<paddle::Tensor>& k_norm_weight,
+    const float rms_norm_eps);

custom_ops/gpu_ops/append_attn/template_instantiation/encoder_write_cache_with_rope_bfloat16_int_kernel.cu

@@ -42,4 +42,7 @@ template void EncoderWriteCacheWithRopeKernel<paddle::bfloat16, int>(
     cudaStream_t& stream,
     paddle::Tensor* qkv_out,
     paddle::Tensor* key_cache_out,
-    paddle::Tensor* value_cache_out);
+    paddle::Tensor* value_cache_out,
+    const paddle::optional<paddle::Tensor>& q_norm_weight,
+    const paddle::optional<paddle::Tensor>& k_norm_weight,
+    const float rms_norm_eps);

custom_ops/gpu_ops/append_attn/template_instantiation/encoder_write_cache_with_rope_float16_float16_kernel.cu

@@ -42,4 +42,7 @@ template void EncoderWriteCacheWithRopeKernel<paddle::float16, paddle::float16>(
     cudaStream_t& stream,
     paddle::Tensor* qkv_out,
     paddle::Tensor* key_cache_out,
-    paddle::Tensor* value_cache_out);
+    paddle::Tensor* value_cache_out,
+    const paddle::optional<paddle::Tensor>& q_norm_weight,
+    const paddle::optional<paddle::Tensor>& k_norm_weight,
+    const float rms_norm_eps);

custom_ops/gpu_ops/append_attn/template_instantiation/encoder_write_cache_with_rope_float16_int_kernel.cu

@@ -42,4 +42,7 @@ template void EncoderWriteCacheWithRopeKernel<paddle::float16, int>(
     cudaStream_t& stream,
     paddle::Tensor* qkv_out,
     paddle::Tensor* key_cache_out,
-    paddle::Tensor* value_cache_out);
+    paddle::Tensor* value_cache_out,
+    const paddle::optional<paddle::Tensor>& q_norm_weight,
+    const paddle::optional<paddle::Tensor>& k_norm_weight,
+    const float rms_norm_eps);

custom_ops/gpu_ops/append_attn/utils.cuh

@@ -27,6 +27,7 @@ struct AppendAttnMetaData {
   int head_dims;
   int head_dims_v;
   int max_blocks_per_seq;
+  const int *mask_offset = nullptr;
 };
 
 __forceinline__ __host__ __device__ int div_up(int a, int b) {
@@ -430,6 +431,9 @@ __forceinline__ __host__ __device__ void vec_cast<nv_bfloat16, float>(
   } else if (group_size == 12) {                             \
     constexpr size_t GROUP_SIZE = 12;                        \
     __VA_ARGS__                                              \
+  } else if (group_size == 14) {                             \
+    constexpr size_t GROUP_SIZE = 14;                        \
+    __VA_ARGS__                                              \
   } else if (group_size == 16) {                             \
     constexpr size_t GROUP_SIZE = 16;                        \
     __VA_ARGS__                                              \
@@ -474,6 +478,9 @@ __forceinline__ __host__ __device__ void vec_cast<nv_bfloat16, float>(
   if (causal) {                              \
     constexpr bool CAUSAL = true;            \
     __VA_ARGS__                              \
+  } else {                                   \
+    constexpr bool CAUSAL = false;           \
+    __VA_ARGS__                              \
   }
 
 #define DISPATCH_ENABLE_PREFILL(enable_prefill, ENABLE_PREFILL, ...) \
@@ -559,3 +566,37 @@ template <typename T, bool IsFP8>inline __device__ static void convert_c8(T * re
     convert_int8(result, source);
   }
 }
+
+constexpr int kWarpSize = 32;
+
+template<typename T>
+inline __device__ void WelfordCombine1(T b_m2, T* m2) {
+  *m2 += b_m2;
+}
+
+template<typename T, int thread_group_width = kWarpSize>
+__inline__ __device__ void WelfordWarpReduce(T thread_m2, T* m2) {
+  *m2 = thread_m2;
+  for (int mask = thread_group_width / 2; mask > 0; mask >>= 1) {
+    T b_m2 = __shfl_xor_sync(0xffffffff, *m2, mask);
+    WelfordCombine1(b_m2, m2);
+  }
+}
+
+template<typename T, int thread_group_width = kWarpSize>
+__inline__ __device__ void WelfordWarpAllReduce(T thread_m2, T* m2) {
+  WelfordWarpReduce<T, thread_group_width>(thread_m2, m2);
+}
+
+template <typename T>
+__inline__ __device__ T Rsqrt(T x);
+
+template <>
+__inline__ __device__ float Rsqrt<float>(float x) {
+  return rsqrt(x);
+}
+
+template <>
+__inline__ __device__ double Rsqrt<double>(double x) {
+  return rsqrt(x);
+}

custom_ops/gpu_ops/cpp_extensions.cc

@@ -77,7 +77,54 @@ std::vector<paddle::Tensor> AppendAttention(
     const paddle::optional<paddle::Tensor> &cache_v_zp,
     const paddle::optional<paddle::Tensor> &out_linear_shifts,
     const paddle::optional<paddle::Tensor> &out_linear_smooths,
+    const paddle::optional<paddle::Tensor> &mask_offset,
     const paddle::optional<paddle::Tensor> &kv_signal_data,
+    const paddle::optional<paddle::Tensor>& q_norm_weight,
+    const paddle::optional<paddle::Tensor>& k_norm_weight,
+    const float rms_norm_eps,
+    const std::string &compute_dtype, const std::string &cache_quant_type_str,
+    const bool use_neox_rotary_style, const bool rope_3d,
+    const int max_input_length, const float quant_max_bound,
+    const float quant_min_bound, const float out_linear_in_scale,
+    const int encoder_block_shape_q, const int decoder_block_shape_q,
+    const int max_partition_size, const int encoder_max_partition_size,
+    const int speculate_max_draft_token_num, const bool causal,
+    const bool speculate_decoder);
+
+void AppendAttentionWithOutput(
+    const paddle::Tensor &qkv, const paddle::Tensor &key_cache,
+    const paddle::Tensor &value_cache, const paddle::Tensor &seq_lens_encoder,
+    const paddle::Tensor &seq_lens_decoder,
+    const paddle::Tensor &seq_lens_this_time,
+    const paddle::Tensor &batch_id_per_token, const paddle::Tensor &cu_seqlens_q,
+    const paddle::Tensor &block_tables, const paddle::Tensor &encoder_batch_ids,
+    const paddle::Tensor &encoder_tile_ids_per_batch,
+    const paddle::Tensor &encoder_num_blocks,
+    const paddle::Tensor &kv_batch_ids,
+    const paddle::Tensor &kv_tile_ids_per_batch,
+    const paddle::Tensor &kv_num_blocks,
+    const paddle::Tensor &decoder_batch_ids,
+    const paddle::Tensor &decoder_tile_ids_per_batch,
+    const paddle::Tensor &decoder_num_blocks,
+    const paddle::Tensor &set_max_lengths, const paddle::Tensor &max_len_kv,
+    paddle::Tensor &fmha_out,
+    const paddle::optional<paddle::Tensor> &rotary_embs,
+    const paddle::optional<paddle::Tensor> &attn_mask,
+    const paddle::optional<paddle::Tensor> &qkv_bias,
+    const paddle::optional<paddle::Tensor> &qkv_out_scales,
+    const paddle::optional<paddle::Tensor> &cache_k_quant_scales,
+    const paddle::optional<paddle::Tensor> &cache_v_quant_scales,
+    const paddle::optional<paddle::Tensor> &cache_k_dequant_scales,
+    const paddle::optional<paddle::Tensor> &cache_v_dequant_scales,
+    const paddle::optional<paddle::Tensor> &cache_k_zp,
+    const paddle::optional<paddle::Tensor> &cache_v_zp,
+    const paddle::optional<paddle::Tensor> &out_linear_shifts,
+    const paddle::optional<paddle::Tensor> &out_linear_smooths,
+    const paddle::optional<paddle::Tensor> &mask_offset,
+    const paddle::optional<paddle::Tensor> &kv_signal_data,
+    const paddle::optional<paddle::Tensor>& q_norm_weight,
+    const paddle::optional<paddle::Tensor>& k_norm_weight,
+    const float rms_norm_eps,
     const std::string &compute_dtype, const std::string &cache_quant_type_str,
     const bool use_neox_rotary_style, const bool rope_3d,
     const int max_input_length, const float quant_max_bound,
@@ -107,7 +154,8 @@ std::vector<paddle::Tensor> GQARopeWriteCacheKernel(
     const paddle::optional<paddle::Tensor> &cache_v_zp,
     const paddle::optional<paddle::Tensor> &kv_signal_data,
     const int kv_token_num, const int max_seq_len,
-    const std::string &cache_quant_type);
+    const std::string &cache_quant_type,
+    const bool rope_3d);
 
 std::vector<paddle::Tensor>
 PreCacheLenConcat(const paddle::Tensor &seq_lens_decoder,
@@ -124,11 +172,29 @@ paddle::Tensor FusedExpertMoeFunc(
     const std::string &quant_method, const int moe_topk,
     const bool norm_topk_prob, const bool group_moe);
 
+std::vector<paddle::Tensor> MacheteMMKernel(
+    paddle::Tensor const& A, paddle::Tensor const& B,
+    paddle::optional<paddle::Tensor> const& maybe_group_scales,
+    paddle::optional<paddle::Tensor> const& maybe_group_zeros,
+    paddle::optional<paddle::Tensor> const& maybe_channel_scales,
+    paddle::optional<paddle::Tensor> const& maybe_token_scales,
+    std::string const& b_type_str,
+    std::string const& maybe_out_type_str,
+    int64_t const& maybe_group_size,
+    std::string const& maybe_schedule);
+
+std::vector<paddle::Tensor> MachetePrepackBKernel(
+    paddle::Tensor const& B, std::string const& a_type_str, std::string const& b_type_str,
+    std::string const& maybe_group_scales_type_str);
+
+std::vector<std::string> MacheteSupportedSchedules(
+    std::string const& a_type_str, std::string const& b_type_str);
+
 std::vector<paddle::Tensor> MoeExpertDispatch(
     const paddle::Tensor &input, const paddle::Tensor &gating_output,
     const paddle::optional<paddle::Tensor> &gating_correction_bias,
     const paddle::optional<paddle::Tensor> &w4a8_in_scale, const int moe_topk,
-    const bool group_moe, const bool topk_only_mode);
+    const bool group_moe, const std::string &moe_quant_type, const bool topk_only_mode);
 
 std::vector<paddle::Tensor>
 MoETopKSelectKernel(const paddle::Tensor &gating_logits,
@@ -188,7 +254,8 @@ paddle::Tensor MoeExpertFFNFunc(
     const paddle::optional<paddle::Tensor>& down_proj_scale,
     const paddle::optional<paddle::Tensor>& down_proj_in_scale,
     const paddle::optional<paddle::Tensor>& expert_idx_per_token,
-    const std::string& quant_method, const bool used_in_ep_low_latency);
+    const std::string& quant_method, const bool used_in_ep_low_latency,
+    const int estimate_total_token_nums);
 
 paddle::Tensor MoeExpertFFNWint2Func(
     const paddle::Tensor& permute_input,
@@ -235,8 +302,14 @@ std::vector<paddle::Tensor> GetBlockShapeAndSplitKVBlock(
     const paddle::Tensor &seq_lens_encoder,
     const paddle::Tensor &seq_lens_decoder,
     const paddle::Tensor &seq_lens_this_time,
-    const int encoder_block_shape_q, const int decoder_block_shape_q,
-    const int group_size, const int block_size,
+    paddle::Tensor &decoder_batch_ids,          // Inplace
+    paddle::Tensor &decoder_tile_ids_per_batch, // Inplace
+    paddle::Tensor &decoder_num_blocks_x_cpu,   // Inplace, Pinned Memory
+    paddle::Tensor &max_len_tensor_cpu,         // Inplace, Pinned Memory
+    const int encoder_block_shape_q,
+    const int decoder_block_shape_q,
+    const int group_size,
+    const int block_size,
     const int decoder_step_token_num);
 
 std::vector<paddle::Tensor> GetPaddingOffset(const paddle::Tensor &input_ids,
@@ -266,13 +339,12 @@ void GetStopFlagsMulti(const paddle::Tensor &topk_ids,
                        const paddle::Tensor &seq_lens,
                        const paddle::Tensor &end_ids,
                        const paddle::Tensor &next_tokens,
+                       const paddle::Tensor &pre_ids,
+                       const paddle::Tensor &step_idx,
+                       const paddle::Tensor &stop_seqs,
+                       const paddle::Tensor &stop_seqs_len,
                        const bool beam_search);
 
-void GetStopFlagsMultiSeqs(
-    const paddle::Tensor &topk_ids, const paddle::Tensor &pre_ids,
-    const paddle::Tensor &step_idx, const paddle::Tensor &stop_flags,
-    const paddle::Tensor &seq_lens, const paddle::Tensor &stop_seqs,
-    const paddle::Tensor &stop_seqs_len, const paddle::Tensor &end_ids);
 
 void UpdateInputes(const paddle::Tensor &stop_flags,
                    const paddle::Tensor &not_need_stop, // only on cpu
@@ -318,7 +390,7 @@ std::vector<paddle::Tensor> ExtractTextTokenOutput(
     const paddle::Tensor &max_seq_len, const paddle::Tensor &max_seq_len_index,
     const paddle::Tensor &mm_token_num_len,
     const paddle::Tensor &seq_lens_this_time,
-    const paddle::Tensor &cu_seqlens_q, const paddle::Tensor &score_text);
+    const paddle::Tensor &cu_seqlens_q, const paddle::Tensor &hidden_states);
 
 std::vector<paddle::Tensor> MoEDeepGEMMPermute(const paddle::Tensor &x,
                                                const paddle::Tensor &topk_idx,
@@ -492,6 +564,7 @@ std::vector<paddle::Tensor> NoauxTc(
       int n_group,
       int topk_group,
       int topk,
+      bool renormalize,
       float routed_scaling_factor);
 
 #ifdef ENABLE_FP8
@@ -521,7 +594,7 @@ paddle::Tensor FusedHadamardQuantFp8Func(
 int64_t init_custom_all_reduce(const std::vector<int64_t>& fake_ipc_ptrs,
                       paddle::Tensor& rank_data, int64_t rank, bool full_nvlink);
 
-void all_reduce(int64_t _fa, paddle::Tensor& inp, paddle::Tensor& out,
+void all_reduce(paddle::Tensor& inp, paddle::Tensor& out, int64_t _fa,
                 int64_t reg_buffer, int64_t reg_buffer_sz_bytes);
 
 void dispose(int64_t _fa);
@@ -543,6 +616,8 @@ int64_t open_mem_handle(paddle::Tensor& mem_handle);
 
 void free_shared_buffer(int64_t buffer);
 
+void clear_ipc_handles(int64_t _fa);
+
 // speculative decoding Kernel
 std::vector<paddle::Tensor> SpeculateGetPaddingOffset(
     const paddle::Tensor& input_ids,
@@ -604,7 +679,7 @@ void SpeculateVerify(
     const paddle::Tensor &actual_draft_token_nums, const paddle::Tensor &topp,
     int max_seq_len, int verify_window, bool enable_topp, bool benchmark_mode);
 
-void SpeculateUpdateV3(const paddle::Tensor &seq_lens_encoder,
+void SpeculateUpdate(const paddle::Tensor &seq_lens_encoder,
                        const paddle::Tensor &seq_lens_decoder,
                        const paddle::Tensor &not_need_stop,
                        const paddle::Tensor &draft_tokens,
@@ -649,6 +724,20 @@ void NgramMatch(const paddle::Tensor &input_ids,
         const int max_draft_tokens);
 
 
+void HybridMtpNgram(const paddle::Tensor &input_ids,
+        const paddle::Tensor &input_ids_len,
+        const paddle::Tensor &pre_ids,
+        const paddle::Tensor &step_idx,
+        const paddle::Tensor &draft_token_num,
+        const paddle::Tensor &draft_tokens,
+        const paddle::Tensor &seq_lens_this_time,
+        const paddle::Tensor &seq_lens_decoder,
+        const paddle::Tensor &max_dec_len,
+        const int max_ngram_size,
+        const int min_ngram_size,
+        const int max_draft_tokens);
+
+
 // MTP
 void DraftModelPostprocess(const paddle::Tensor& base_model_draft_tokens,
                            const paddle::Tensor& base_model_seq_lens_this_time,
@@ -665,8 +754,10 @@ void DraftModelPreprocess(const paddle::Tensor& draft_tokens,
                           const paddle::Tensor& step_idx,
                           const paddle::Tensor& not_need_stop,
                           const paddle::Tensor& batch_drop,
+                          const paddle::Tensor& pre_ids,
                           const paddle::Tensor& accept_tokens,
                           const paddle::Tensor& accept_num,
+                          const paddle::Tensor& base_model_seq_lens_this_time,
                           const paddle::Tensor& base_model_seq_lens_encoder,
                           const paddle::Tensor& base_model_seq_lens_decoder,
                           const paddle::Tensor& base_model_step_idx,
@@ -756,6 +847,33 @@ void SpeculateStepPaddle(
     const int encoder_decoder_block_num,
     const int max_draft_tokens);
 
+void MergePrefillDecodeOutput(
+    const paddle::Tensor &encoder_res,
+    const paddle::Tensor &decoder_res,
+    const paddle::Tensor &seq_lens_encoder,
+    const paddle::Tensor &seq_lens_decoder,
+    const paddle::Tensor &seq_lens_this_time,
+    const paddle::Tensor &cu_seq_q,
+    const int head_num,
+    const int head_dim,
+    const int max_token);
+
+std::vector<paddle::Tensor> TopPSamplingReject(const paddle::Tensor &probs,
+                                               const paddle::Tensor &top_p,
+                                               const paddle::optional<paddle::Tensor> &top_k,
+                                               int64_t seed);
+
+std::vector<paddle::Tensor> TopKRenorm(const paddle::Tensor &probs,
+                                       const paddle::Tensor &top_k);
+
+std::vector<paddle::Tensor> MinPSamplingFromProbs(const paddle::Tensor &probs,
+                                               const paddle::Tensor &min_p);
+
+void SaveOutMmsgStatic(const paddle::Tensor& x,
+                       const paddle::Tensor& not_need_stop,
+                       int64_t rank_id,
+                       bool save_each_rank);
+
 PYBIND11_MODULE(fastdeploy_ops, m) {
 
   m.def("get_expert_token_num", &GetExpertTokenNum, py::arg("topk_ids"),
@@ -809,6 +927,7 @@ PYBIND11_MODULE(fastdeploy_ops, m) {
    * append_attention
    */
   m.def("append_attention", &AppendAttention, "append attention function");
+  m.def("append_attention_with_output", &AppendAttentionWithOutput, "append attention with output function");
   /**
    * gqa_rope_write_cache.cu
    * gqa_rope_write_cache
@@ -840,7 +959,7 @@ PYBIND11_MODULE(fastdeploy_ops, m) {
   m.def("moe_expert_dispatch", &MoeExpertDispatch, py::arg("input"),
         py::arg("gating_output"), py::arg("gating_correction_bias"),
         py::arg("w4a8_in_scale"), py::arg("moe_topk"), py::arg("group_moe"),
-        py::arg("topk_only_mode"), "moe export dispatch function");
+        py::arg("moe_quant_type"), py::arg("topk_only_mode"), "moe export dispatch function");
 
   /**
    * moe/fused_moe/ep_moe_prefill_func.cu
@@ -870,6 +989,27 @@ PYBIND11_MODULE(fastdeploy_ops, m) {
         py::arg("recv_expert_count"), py::arg("block_size"),
         "per token per block quant");
 
+#ifdef ENABLE_MACHETE
+  /*machete/machete_mm.cu
+   * machete_mm
+   */
+  m.def("machete_mm", &MacheteMMKernel, py::arg("A"), py::arg("B"), py::arg("maybe_group_scale"),
+        py::arg("maybe_group_zeros"), py::arg("maybe_channel_scales"), py::arg("maybe_token_scales"),
+        py::arg("b_type_str"), py::arg("maybe_out_type_str"), py::arg("maybe_group_size"),
+        py::arg("maybe_schedule"),
+        "machete mm function");
+
+  /*machete/machete_prepack_B.cu
+   * machete_prepack_B
+   */
+  m.def("machete_prepack_B", &MachetePrepackBKernel, "machete prepacked B function");
+
+  /*machete/machete_supported_schedules.cu
+   * machete_supported_schedules
+   */
+  m.def("machete_supported_schedules", &MacheteSupportedSchedules, "machete supported schedules function");
+#endif
+
   /**
    * moe/fused_moe/moe_topk_select.cu
    * moe_topk_select
@@ -954,12 +1094,6 @@ PYBIND11_MODULE(fastdeploy_ops, m) {
   m.def("set_stop_value_multi_ends", &GetStopFlagsMulti,
         "update_inputs function");
 
-  /**
-   * stop_generation_multi_stop_seqs.cu
-   * set_stop_value_multi_seqs
-   */
-  m.def("set_stop_value_multi_seqs", &GetStopFlagsMultiSeqs,
-        "update_inputs function");
 
   /**
    * update_inputs.cu
@@ -1072,6 +1206,8 @@ PYBIND11_MODULE(fastdeploy_ops, m) {
 
   m.def("free_shared_buffer", &free_shared_buffer, "free_shared_buffer");
 
+  m.def("clear_ipc_handles", &clear_ipc_handles, "clear_ipc_handles");
+
   m.def("open_mem_handle", &open_mem_handle, "open_mem_handle");
 
   m.def("get_graph_buffer_ipc_meta", &get_graph_buffer_ipc_meta, "get_graph_buffer_ipc_meta");
@@ -1089,7 +1225,7 @@ PYBIND11_MODULE(fastdeploy_ops, m) {
 
   m.def("speculate_verify",&SpeculateVerify, "speculate_verify function");
 
-  m.def("speculate_update_v3",&SpeculateUpdateV3, "noaux_tc for Deepseekv3 MoE compute function");
+  m.def("speculate_update",&SpeculateUpdate, "Speculate Update Kernel");
 
   m.def("speculate_set_value_by_flags_and_idx",&SpeculateSetValueByFlagsAndIdx, "speculate_set_value_by_flags_and_idx function");
 
@@ -1099,6 +1235,8 @@ PYBIND11_MODULE(fastdeploy_ops, m) {
 
   m.def("ngram_match", &NgramMatch, "ngram_match function");
 
+  m.def("hybird_mtp_ngram", &HybridMtpNgram, "ngram_match_mixed function");
+
   m.def("draft_model_postprocess",&DraftModelPostprocess, "draft_model_postprocess function");
 
   m.def("draft_model_preprocess",&DraftModelPreprocess, "draft_model_preprocess function");
@@ -1112,4 +1250,14 @@ PYBIND11_MODULE(fastdeploy_ops, m) {
   m.def("mtp_step_paddle",&MTPStepPaddle, "mtp_step_paddle function");
 
   m.def("speculate_step_paddle",&SpeculateStepPaddle, "speculate_step_paddle function");
+
+  m.def("merge_prefill_decode_output", &MergePrefillDecodeOutput, "merge_prefill_decode_output function");
+
+  m.def("rejection_top_p_sampling", &TopPSamplingReject, "rejection_top_p_sampling function");
+
+  m.def("top_k_renorm_probs", &TopKRenorm, "top_k_renorm_probs function");
+
+  m.def("min_p_sampling", &MinPSamplingFromProbs, "min_p_sampling function");
+
+  m.def("save_output", &SaveOutMmsgStatic, "save_output function");
 }

custom_ops/gpu_ops/custom_all_reduce/all_reduce.cu

@@ -49,7 +49,7 @@ fptr_t init_custom_all_reduce(const std::vector<fptr_t>& fake_ipc_ptrs,
  * Otherwise, _reg_buffer is assumed to be IPC-registered and inp is first
  * copied into _reg_buffer.
  */
-void all_reduce(fptr_t _fa, paddle::Tensor& inp, paddle::Tensor& out,
+void all_reduce(paddle::Tensor& inp, paddle::Tensor& out, fptr_t _fa,
                 fptr_t _reg_buffer, int64_t reg_buffer_sz_bytes) {
   auto fa = reinterpret_cast<paddle::CustomAllreduce*>(_fa);
   auto stream = inp.stream();
@@ -122,10 +122,14 @@ void register_graph_buffers(fptr_t _fa,
   for (int i = 0; i < handles.size(); i++) {
     bytes.emplace_back(handles[i].begin(), handles[i].end());
   }
-  bytes.reserve(handles.size());
   fa->register_graph_buffers(bytes, offsets);
 }
 
+void clear_ipc_handles(fptr_t _fa) {
+  auto fa = reinterpret_cast<paddle::CustomAllreduce*>(_fa);
+  fa->clear_ipc_handles();
+}
+
 std::tuple<fptr_t, paddle::Tensor> allocate_shared_buffer_and_handle(
     int64_t size) {
 
@@ -163,3 +167,12 @@ fptr_t open_mem_handle(paddle::Tensor& mem_handle) {
 void free_shared_buffer(fptr_t buffer) {
   CUDACHECK(cudaFree(reinterpret_cast<void*>(buffer)));
 }
+
+
+PD_BUILD_STATIC_OP(all_reduce)
+    .Inputs({"inp",
+             "out"})
+    .Outputs({"new_out"})
+    .Attrs({"_fa: int64_t", "_reg_buffer: int64_t", "reg_buffer_sz_bytes: int64_t"})
+    .SetInplaceMap({{"out", "new_out"}})
+    .SetKernelFn(PD_KERNEL(all_reduce));

custom_ops/gpu_ops/custom_all_reduce/all_reduce.cuh

@@ -517,10 +517,15 @@ class CustomAllreduce {
 #undef KL
   }
 
-  ~CustomAllreduce() {
+  void clear_ipc_handles(){
     for (auto [_, ptr] : ipc_handles_) {
       CUDACHECK(cudaIpcCloseMemHandle(ptr));
     }
+    ipc_handles_.clear();
+  }
+
+  ~CustomAllreduce() {
+    clear_ipc_handles();
   }
 };
 }  // namespace paddle

custom_ops/gpu_ops/cutlass_extensions/epilogue/scaled_mm_epilogues_c3x.hpp

@@ -6,6 +6,8 @@
 // clang-format off
 #include "cutlass_extensions/epilogue/broadcast_load_epilogue_c3x.hpp"
 #include "cutlass_extensions/epilogue/broadcast_load_epilogue_array_c3x.hpp"
+
+#include "helper.h"
 // clang-format on
 
 /*

custom_ops/gpu_ops/cutlass_extensions/gemm/kernel/mixed_gemm_B_layout.h

@@ -133,10 +133,18 @@ public:
 template <typename TypeA, typename Arch>
 struct LayoutDetailsB<TypeA, uint2b_t, Arch, typename platform::enable_if<Arch::kMinComputeCapability >= 75>::type>
 {
-    static constexpr int ThreadblockK = 128 * 8 / cutlass::sizeof_bits<TypeA>::value;
-    using Layout = layout::RowMajor;
-    static constexpr int ElementsPerAccess = 128 / cutlass::sizeof_bits<TypeA>::value;
-    using Operator = cutlass::arch::OpMultiplyAdd;
+    static constexpr int ThreadblockK = 128 * 8 / cutlass::sizeof_bits<TypeA>::value; // 64
+
+private:
+    static constexpr int ElementsPerCacheLine = 128 * 8 / sizeof_bits<uint2b_t>::value;
+    static constexpr int ColumnsInterleaved = ElementsPerCacheLine / ThreadblockK; // 8
+
+public:
+    // using Layout = layout::ColumnMajor;
+    // static constexpr int ElementsPerAccess = 16; // at least 4-bytes
+    using Layout = layout::ColumnMajorTileInterleave<ThreadblockK, ColumnsInterleaved>;
+    static constexpr int ElementsPerAccess = 128 / cutlass::sizeof_bits<uint2b_t>::value; // 64
+    using Operator = cutlass::arch::OpMultiplyAddDequantizeInterleavedBToA;
 };
 
 template <typename TypeA, typename Arch>

custom_ops/gpu_ops/cutlass_extensions/gemm/threadblock/default_mma.h

@@ -18,14 +18,12 @@
 
 #include "cutlass_extensions/gemm/threadblock/default_dq_mma_multistage.h"
 #include "cutlass_extensions/gemm/threadblock/default_dq_mma_pipelined.h"
+#include "cutlass_extensions/gemm/threadblock/default_wint2x_mma.h"
 #include "cutlass_extensions/gemm/threadblock/default_mma_bf16.h"
 
-namespace cutlass
-{
-namespace gemm
-{
-namespace threadblock
-{
+namespace cutlass {
+namespace gemm {
+namespace threadblock {
 
 ////////////////////////////////////////////////////////////////////////////////
 
@@ -378,38 +376,23 @@ template <
 struct DefaultMma<cutlass::half_t, LayoutA, kAlignmentA, uint2b_t, LayoutB, kAlignmentB, ElementAccumulator,
     layout::RowMajor, arch::OpClassTensorOp, ArchTag, ThreadblockShape, WarpShape, InstructionShape, 2, Operator>
 {
-    static cutlass::arch::CacheOperation::Kind const CacheOpA =
-        ((sizeof_bits<half_t>::value * kAlignmentA) == 128) ? cutlass::arch::CacheOperation::Global
-            : cutlass::arch::CacheOperation::Always;
-
-    static cutlass::arch::CacheOperation::Kind const CacheOpB =
-        ((sizeof_bits<half_t>::value * kAlignmentB) == 128) ? cutlass::arch::CacheOperation::Global
-            : cutlass::arch::CacheOperation::Always;
+private:
+    using Mma = DefaultWint2xMma<half_t, LayoutA, kAlignmentA, uint2b_t, LayoutB, kAlignmentB,
+        ElementAccumulator, layout::RowMajor, arch::OpClassTensorOp, ArchTag, ThreadblockShape,
+        WarpShape, InstructionShape, 2, Operator>;
 
+public:
     // Define the MmaCore components
-    using MmaCore =
-        typename cutlass::gemm::threadblock::DefaultMmaCore<ThreadblockShape, WarpShape, InstructionShape, half_t,
-            LayoutA, half_t, LayoutB, ElementAccumulator, layout::RowMajor, arch::OpClassTensorOp, 3, Operator,
-            false, CacheOpA, CacheOpB>;
+    using MmaCore = typename Mma::MmaCore;
 
     // Define iterators over tiles from the A operand
-    using ThreadMapA = typename MmaCore::IteratorThreadMapA;
-    using AccessTypeA = cutlass::Array<half_t, kAlignmentA>;
-    using IteratorA = cutlass::transform::threadblock::PredicatedTileAccessIterator<
-          cutlass::MatrixShape<ThreadblockShape::kM, ThreadblockShape::kK>, half_t, LayoutA, 1, ThreadMapA,
-        AccessTypeA>;
+    using IteratorA = typename Mma::IteratorA;
 
     // Define iterators over tiles from the B operand
-    using ThreadMapB = typename MmaCore::IteratorThreadMapB;
-    using AccessTypeB = cutlass::Array<half_t, kAlignmentB>;
-    using IteratorB = cutlass::transform::threadblock::PredicatedTileAccessIterator<
-        cutlass::MatrixShape<ThreadblockShape::kK, ThreadblockShape::kN>, half_t, LayoutB, 0, ThreadMapB,
-        AccessTypeB>;
+    using IteratorB = typename Mma::IteratorB;
 
     // Define the threadblock-scoped multistage matrix multiply
-    using ThreadblockMma = cutlass::gemm::threadblock::Wint2xMmaMultistage<typename MmaCore::Shape, IteratorA,
-        typename MmaCore::SmemIteratorA, MmaCore::kCacheOpA, IteratorB, typename MmaCore::SmemIteratorB,
-        MmaCore::kCacheOpB, ElementAccumulator, layout::RowMajor, typename MmaCore::MmaPolicy, 2>;
+    using ThreadblockMma = typename Mma::ThreadblockMma;
 };
 
 template <
@@ -441,38 +424,23 @@ struct DefaultMma<half_t, LayoutA, kAlignmentA, uint2b_t, LayoutB, kAlignmentB,
     layout::RowMajor, arch::OpClassTensorOp, ArchTag, ThreadblockShape, WarpShape, InstructionShape, kStages, Operator,
     false, SharedMemoryClear>
 {
-    static cutlass::arch::CacheOperation::Kind const CacheOpA =
-        ((sizeof_bits<half_t>::value * kAlignmentA) == 128) ? cutlass::arch::CacheOperation::Global
-            : cutlass::arch::CacheOperation::Always;
-
-    static cutlass::arch::CacheOperation::Kind const CacheOpB =
-        ((sizeof_bits<half_t>::value * kAlignmentB) == 128) ? cutlass::arch::CacheOperation::Global
-            : cutlass::arch::CacheOperation::Always;
+private:
+    using Mma = DefaultWint2xMma<half_t, LayoutA, kAlignmentA, uint2b_t, LayoutB, kAlignmentB,
+        ElementAccumulator, layout::RowMajor, arch::OpClassTensorOp, ArchTag, ThreadblockShape,
+        WarpShape, InstructionShape, kStages, Operator, SharedMemoryClear>;
 
+public:
     // Define the MmaCore components
-    using MmaCore =
-        typename cutlass::gemm::threadblock::DefaultMmaCore<ThreadblockShape, WarpShape, InstructionShape, half_t,
-            LayoutA, half_t, LayoutB, ElementAccumulator, layout::RowMajor, arch::OpClassTensorOp, kStages, Operator,
-            false, CacheOpA, CacheOpB>;
+    using MmaCore = typename Mma::MmaCore;
 
     // Define iterators over tiles from the A operand
-    using ThreadMapA = typename MmaCore::IteratorThreadMapA;
-    using AccessTypeA = cutlass::Array<half_t, kAlignmentA>;
-    using IteratorA = cutlass::transform::threadblock::PredicatedTileAccessIterator<
-          cutlass::MatrixShape<ThreadblockShape::kM, ThreadblockShape::kK>, half_t, LayoutA, 1, ThreadMapA,
-        AccessTypeA>;
+    using IteratorA = typename Mma::IteratorA;
 
     // Define iterators over tiles from the B operand
-    using ThreadMapB = typename MmaCore::IteratorThreadMapB;
-    using AccessTypeB = cutlass::Array<half_t, kAlignmentB>;
-    using IteratorB = cutlass::transform::threadblock::PredicatedTileAccessIterator<
-        cutlass::MatrixShape<ThreadblockShape::kK, ThreadblockShape::kN>, half_t, LayoutB, 0, ThreadMapB,
-        AccessTypeB>;
+    using IteratorB = typename Mma::IteratorB;
 
     // Define the threadblock-scoped multistage matrix multiply
-    using ThreadblockMma = cutlass::gemm::threadblock::Wint2xMmaMultistage<typename MmaCore::Shape, IteratorA,
-        typename MmaCore::SmemIteratorA, MmaCore::kCacheOpA, IteratorB, typename MmaCore::SmemIteratorB,
-        MmaCore::kCacheOpB, ElementAccumulator, layout::RowMajor, typename MmaCore::MmaPolicy, kStages, SharedMemoryClear>;
+    using ThreadblockMma = typename Mma::ThreadblockMma;
 };
 
 } // namespace threadblock

custom_ops/gpu_ops/cutlass_extensions/gemm/threadblock/default_mma_bf16.h

@@ -19,7 +19,7 @@
 #include "cutlass/gemm/threadblock/default_mma.h"
 #include "cutlass_extensions/gemm/threadblock/default_dq_mma_multistage.h"
 #include "cutlass_extensions/gemm/threadblock/default_dq_mma_pipelined.h"
-#include "cutlass_extensions/gemm/threadblock/wint2x_mma_multistage.h"
+#include "cutlass_extensions/gemm/threadblock/default_wint2x_mma.h"
 
 namespace cutlass {
 namespace gemm {
@@ -379,38 +379,23 @@ template <
 struct DefaultMma<cutlass::bfloat16_t, LayoutA, kAlignmentA, uint2b_t, LayoutB, kAlignmentB, ElementAccumulator,
     layout::RowMajor, arch::OpClassTensorOp, ArchTag, ThreadblockShape, WarpShape, InstructionShape, 2, Operator>
 {
-    static cutlass::arch::CacheOperation::Kind const CacheOpA =
-        ((sizeof_bits<bfloat16_t>::value * kAlignmentA) == 128) ? cutlass::arch::CacheOperation::Global
-            : cutlass::arch::CacheOperation::Always;
-
-    static cutlass::arch::CacheOperation::Kind const CacheOpB =
-        ((sizeof_bits<bfloat16_t>::value * kAlignmentB) == 128) ? cutlass::arch::CacheOperation::Global
-            : cutlass::arch::CacheOperation::Always;
+private:
+    using Mma = DefaultWint2xMma<bfloat16_t, LayoutA, kAlignmentA, uint2b_t, LayoutB, kAlignmentB,
+        ElementAccumulator, layout::RowMajor, arch::OpClassTensorOp, ArchTag, ThreadblockShape,
+        WarpShape, InstructionShape, 2, Operator>;
 
+public:
     // Define the MmaCore components
-    using MmaCore =
-        typename cutlass::gemm::threadblock::DefaultMmaCore<ThreadblockShape, WarpShape, InstructionShape, bfloat16_t,
-            LayoutA, bfloat16_t, LayoutB, ElementAccumulator, layout::RowMajor, arch::OpClassTensorOp, 3, Operator,
-            false, CacheOpA, CacheOpB>;
+    using MmaCore = typename Mma::MmaCore;
 
     // Define iterators over tiles from the A operand
-    using ThreadMapA = typename MmaCore::IteratorThreadMapA;
-    using AccessTypeA = cutlass::Array<bfloat16_t, kAlignmentA>;
-    using IteratorA = cutlass::transform::threadblock::PredicatedTileAccessIterator<
-          cutlass::MatrixShape<ThreadblockShape::kM, ThreadblockShape::kK>, bfloat16_t, LayoutA, 1, ThreadMapA,
-        AccessTypeA>;
+    using IteratorA = typename Mma::IteratorA;
 
     // Define iterators over tiles from the B operand
-    using ThreadMapB = typename MmaCore::IteratorThreadMapB;
-    using AccessTypeB = cutlass::Array<bfloat16_t, kAlignmentB>;
-    using IteratorB = cutlass::transform::threadblock::PredicatedTileAccessIterator<
-        cutlass::MatrixShape<ThreadblockShape::kK, ThreadblockShape::kN>, bfloat16_t, LayoutB, 0, ThreadMapB,
-        AccessTypeB>;
+    using IteratorB = typename Mma::IteratorB;
 
     // Define the threadblock-scoped multistage matrix multiply
-    using ThreadblockMma = cutlass::gemm::threadblock::Wint2xMmaMultistage<typename MmaCore::Shape, IteratorA,
-        typename MmaCore::SmemIteratorA, MmaCore::kCacheOpA, IteratorB, typename MmaCore::SmemIteratorB,
-        MmaCore::kCacheOpB, ElementAccumulator, layout::RowMajor, typename MmaCore::MmaPolicy, 2>;
+    using ThreadblockMma = typename Mma::ThreadblockMma;
 };
 
 template <
@@ -442,38 +427,23 @@ struct DefaultMma<bfloat16_t, LayoutA, kAlignmentA, uint2b_t, LayoutB, kAlignmen
     layout::RowMajor, arch::OpClassTensorOp, ArchTag, ThreadblockShape, WarpShape, InstructionShape, kStages, Operator,
     false, SharedMemoryClear>
 {
-    static cutlass::arch::CacheOperation::Kind const CacheOpA =
-        ((sizeof_bits<bfloat16_t>::value * kAlignmentA) == 128) ? cutlass::arch::CacheOperation::Global
-            : cutlass::arch::CacheOperation::Always;
-
-    static cutlass::arch::CacheOperation::Kind const CacheOpB =
-        ((sizeof_bits<bfloat16_t>::value * kAlignmentB) == 128) ? cutlass::arch::CacheOperation::Global
-            : cutlass::arch::CacheOperation::Always;
+private:
+    using Mma = DefaultWint2xMma<bfloat16_t, LayoutA, kAlignmentA, uint2b_t, LayoutB, kAlignmentB,
+        ElementAccumulator, layout::RowMajor, arch::OpClassTensorOp, ArchTag, ThreadblockShape,
+        WarpShape, InstructionShape, kStages, Operator, SharedMemoryClear>;
 
+public:
     // Define the MmaCore components
-    using MmaCore =
-        typename cutlass::gemm::threadblock::DefaultMmaCore<ThreadblockShape, WarpShape, InstructionShape, bfloat16_t,
-            LayoutA, bfloat16_t, LayoutB, ElementAccumulator, layout::RowMajor, arch::OpClassTensorOp, kStages, Operator,
-            false, CacheOpA, CacheOpB>;
+    using MmaCore = typename Mma::MmaCore;
 
     // Define iterators over tiles from the A operand
-    using ThreadMapA = typename MmaCore::IteratorThreadMapA;
-    using AccessTypeA = cutlass::Array<bfloat16_t, kAlignmentA>;
-    using IteratorA = cutlass::transform::threadblock::PredicatedTileAccessIterator<
-          cutlass::MatrixShape<ThreadblockShape::kM, ThreadblockShape::kK>, bfloat16_t, LayoutA, 1, ThreadMapA,
-        AccessTypeA>;
+    using IteratorA = typename Mma::IteratorA;
 
     // Define iterators over tiles from the B operand
-    using ThreadMapB = typename MmaCore::IteratorThreadMapB;
-    using AccessTypeB = cutlass::Array<bfloat16_t, kAlignmentB>;
-    using IteratorB = cutlass::transform::threadblock::PredicatedTileAccessIterator<
-        cutlass::MatrixShape<ThreadblockShape::kK, ThreadblockShape::kN>, bfloat16_t, LayoutB, 0, ThreadMapB,
-        AccessTypeB>;
+    using IteratorB = typename Mma::IteratorB;
 
     // Define the threadblock-scoped multistage matrix multiply
-    using ThreadblockMma = cutlass::gemm::threadblock::Wint2xMmaMultistage<typename MmaCore::Shape, IteratorA,
-        typename MmaCore::SmemIteratorA, MmaCore::kCacheOpA, IteratorB, typename MmaCore::SmemIteratorB,
-        MmaCore::kCacheOpB, ElementAccumulator, layout::RowMajor, typename MmaCore::MmaPolicy, kStages, SharedMemoryClear>;
+    using ThreadblockMma = typename Mma::ThreadblockMma;
 };
 
 } // namespace threadblock

custom_ops/gpu_ops/cutlass_extensions/gemm/threadblock/default_mma_core.h

@@ -0,0 +1,182 @@
+/***************************************************************************************************
+ * Copyright (c) 2017 - 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * SPDX-License-Identifier: BSD-3-Clause
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions are met:
+ *
+ * 1. Redistributions of source code must retain the above copyright notice, this
+ * list of conditions and the following disclaimer.
+ *
+ * 2. Redistributions in binary form must reproduce the above copyright notice,
+ * this list of conditions and the following disclaimer in the documentation
+ * and/or other materials provided with the distribution.
+ *
+ * 3. Neither the name of the copyright holder nor the names of its
+ * contributors may be used to endorse or promote products derived from
+ * this software without specific prior written permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+ * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+ * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+ * DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+ * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+ * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+ * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+ * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+
+#pragma once
+
+#include "cutlass/gemm/threadblock/default_mma_core_sm80.h"
+
+namespace cutlass {
+namespace gemm {
+namespace threadblock {
+
+/// Partial specialization:
+///
+///   A: row-major
+///   B: uint2b_t, column-major
+///   Operator: tensor op class
+///
+/// This uses the default warp-level operator given tile sizes
+template <
+    /// Shape of threadblock-scoped matrix multiply operator (concept:
+    /// GemmShape)
+    typename Shape_,
+    /// Shape of warp-level matrix multiply operator (concept: GemmShape)
+    typename WarpShape_,
+    /// Shape of one matrix production operation (concept: GemmShape)
+    typename InstructionShape_,
+    /// Data type of A operand
+    typename ElementA_,
+    /// Data type of accumulator
+    typename ElementC_,
+    /// Layout of accumulator
+    typename LayoutC_,
+    /// Number of stages
+    int Stages,
+    /// Operation performed by MMA
+    typename Operator_,
+    /// Cache operation of operand A
+    cutlass::arch::CacheOperation::Kind CacheOpA,
+    /// Cache operation of operand B
+    cutlass::arch::CacheOperation::Kind CacheOpB>
+struct DefaultMmaCore<Shape_, WarpShape_, InstructionShape_, ElementA_,
+                      layout::RowMajor, uint2b_t, layout::ColumnMajor,
+                      ElementC_, LayoutC_, arch::OpClassTensorOp, Stages,
+                      Operator_, false, CacheOpA, CacheOpB> {
+  using Shape = Shape_;
+  using WarpShape = WarpShape_;
+  using InstructionShape = InstructionShape_;
+  using ElementA = ElementA_;
+  using LayoutA = layout::RowMajor;
+  using ElementB = uint2b_t;
+  using LayoutB = layout::ColumnMajor;
+  using ElementC = ElementC_;
+  using LayoutC = LayoutC_;
+  static int const kStages = Stages;
+  static cutlass::arch::CacheOperation::Kind const kCacheOpA = CacheOpA;
+  static cutlass::arch::CacheOperation::Kind const kCacheOpB = CacheOpB;
+
+  /// Number of warps present
+  using WarpCount = GemmShape<Shape::kM / WarpShape::kM,
+                              Shape::kN / WarpShape::kN,
+                              Shape::kK / WarpShape::kK>;
+
+  // Divisility requirements
+  static_assert(
+      !(Shape::kM % WarpShape::kM) && !(Shape::kN % WarpShape::kN),
+      "Threadblock-scoped GEMM should be divisible by warp-scoped GEMM size.");
+
+  /// Number of threads per warp
+  static int const kWarpSize = warp::WarpSize<arch::OpClassTensorOp>::value;
+
+  /// Size of a threadblock-scoped access
+  static int const kAccessSizeInBits = 128;
+
+  /// Number of threads total
+  static int const kThreads = WarpCount::kCount * kWarpSize;
+
+  /// Size of a threadblock-scoped access of B
+  static constexpr int kMaxThreadsForB =
+      (Shape::kK * Shape::kN * sizeof_bits<ElementB>::value) / kAccessSizeInBits;
+  static constexpr int kThreadsForB =
+      kMaxThreadsForB > kThreads ? kThreads : kMaxThreadsForB;
+
+  /// Default Operator
+  using Operator = Operator_;
+
+  // Warp thread arrangement
+  static int const kWarpThreadArrangementContiguousA =
+      Shape::kK / (kAccessSizeInBits / sizeof_bits<ElementA>::value);
+
+  static int const kWarpThreadArrangementStridedA =
+      kWarpSize / kWarpThreadArrangementContiguousA;
+
+  static int const kWarpThreadArrangementContiguousB =
+      Shape::kK / (kAccessSizeInBits / sizeof_bits<ElementB>::value);
+
+  static int const kWarpThreadArrangementStridedB =
+      kWarpSize / kWarpThreadArrangementContiguousB;
+
+  //
+  // Shared memory layouts
+  //
+
+  using SmemLayoutA = layout::RowMajorTensorOpMultiplicandCrosswise<
+      sizeof_bits<ElementA>::value, Shape::kK>;
+
+  // Shared memory layout
+  using SmemLayoutB = layout::ColumnMajorTensorOpMultiplicandCrosswise<
+      sizeof_bits<ElementB>::value, Shape::kK>;
+
+  //
+  // Iterators to write to shared memory
+  //
+
+  /// ThreadMap of iterator A
+  using IteratorThreadMapA = transform::PitchLinearWarpRakedThreadMap<
+      layout::PitchLinearShape<Shape::kK, Shape::kM>, kThreads,
+      layout::PitchLinearShape<kWarpThreadArrangementContiguousA,
+                               kWarpThreadArrangementStridedA>,
+      kAccessSizeInBits / sizeof_bits<ElementA>::value>;
+
+  /// Shared memory iterator to A operand
+  using SmemIteratorA = transform::threadblock::RegularTileAccessIterator<
+      MatrixShape<Shape::kM, Shape::kK>, ElementA, SmemLayoutA, 0,
+      IteratorThreadMapA>;
+
+  /// ThreadMap of iterator B
+  using IteratorThreadMapB = transform::PitchLinearWarpRakedThreadMap<
+      layout::PitchLinearShape<Shape::kK, Shape::kN>, kThreadsForB,
+      layout::PitchLinearShape<kWarpThreadArrangementContiguousB,
+                               kWarpThreadArrangementStridedB>,
+      kAccessSizeInBits / sizeof_bits<ElementB>::value>;
+
+  /// Shared memory iterator to B operand
+  using SmemIteratorB = transform::threadblock::RegularTileAccessIterator<
+      MatrixShape<Shape::kK, Shape::kN>, ElementB, SmemLayoutB, 1,
+      IteratorThreadMapB>;
+
+  //
+  // Warp-level matrix multiply operator
+  //
+
+  // Define the warp-level tensor op
+  using MmaTensorOp = typename cutlass::gemm::warp::DefaultMmaTensorOp<
+      WarpShape, InstructionShape, ElementA, SmemLayoutA, ElementB, SmemLayoutB,
+      ElementC, LayoutC, Operator, WarpCount::kK>::Type;
+
+  /// Policy used to define MmaPipelined
+  using MmaPolicy = MmaPolicy<MmaTensorOp, MatrixShape<0, 0>,
+                                        MatrixShape<0, 0>, WarpCount::kK>;
+};
+
+}  // namespace threadblock
+}  // namespace gemm
+}  // namespace cutlass

custom_ops/gpu_ops/cutlass_extensions/gemm/threadblock/default_wint2x_mma.h

@@ -0,0 +1,246 @@
+/*
+ * SPDX-FileCopyrightText: Copyright (c) 2022-2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#pragma once
+
+#include "cutlass_extensions/arch/mma.h"
+#include "cutlass_extensions/gemm/threadblock/default_dq_mma.h"
+#include "cutlass_extensions/gemm/threadblock/default_mma_core.h"
+#include "cutlass_extensions/gemm/threadblock/wint2x_mma_multistage.h"
+#include "cutlass_extensions/gemm/threadblock/wint2x_params_accessor.h"
+
+namespace cutlass {
+namespace gemm {
+namespace threadblock {
+
+////////////////////////////////////////////////////////////////////////////////
+
+template <typename ThreadblockShape, typename ElementT, int GroupSize>
+struct DefaultQuantParamsIterators {
+private:
+    static constexpr int kAlignment = 128 / sizeof_bits<ElementT>::value;
+    static_assert((ThreadblockShape::kN % kAlignment) == 0, "");
+
+    static constexpr int kRows =
+        (GroupSize == -1) ? 1 : (ThreadblockShape::kK + GroupSize - 1) / GroupSize;
+    static constexpr int kColumns = ThreadblockShape::kN;
+
+    using IteratorThreadMap = transform::PitchLinearStripminedThreadMap<
+        layout::PitchLinearShape<kColumns, kRows>,
+        kColumns / kAlignment, kAlignment>;
+
+public:
+    using Iterator = cutlass::transform::threadblock::PredicatedTileIterator<
+        MatrixShape<kRows, kColumns>, ElementT, layout::RowMajor, 0,
+        IteratorThreadMap, kAlignment>;
+    using SmemIterator = Iterator;
+};
+
+template <typename ThreadblockShape, int GroupSize>
+struct DefaultQuantParamsIterators<ThreadblockShape, uint4b_t, GroupSize> {
+private:
+    static constexpr int kAlignment = 32 / sizeof_bits<uint4b_t>::value;
+    static_assert((ThreadblockShape::kN % kAlignment) == 0, "");
+
+    static constexpr int kRows =
+        (GroupSize == -1) ? 1 : (ThreadblockShape::kK + 2 * GroupSize - 1) / (2 * GroupSize);
+    static constexpr int kColumns =
+        (GroupSize == -1) ? ThreadblockShape::kN : ThreadblockShape::kN * 2;
+
+    using IteratorThreadMap = transform::PitchLinearStripminedThreadMap<
+        layout::PitchLinearShape<kColumns, kRows>,
+        kColumns / kAlignment, kAlignment>;
+
+public:
+    using AccessType = cutlass::Array<uint4b_t, kAlignment>;
+    using Iterator = cutlass::transform::threadblock::PredicatedTileAccessIterator<
+        MatrixShape<kRows, kColumns>, uint4b_t, layout::RowMajor,
+        0, IteratorThreadMap, AccessType>;
+
+    using SmemIterator = Iterator;
+};
+
+template <
+    /// Element type for A matrix operand
+    typename ElementA_,
+    /// Layout type for A matrix operand
+    typename LayoutA_,
+    /// Access granularity of A matrix in units of elements
+    int kAlignmentA,
+    /// Element type for B matrix operand
+    typename ElementB_,
+    /// Layout type for B matrix operand
+    typename LayoutB_,
+    /// Access granularity of B matrix in units of elements
+    int kAlignmentB,
+    /// Element type for internal accumulation
+    typename ElementAccumulator_,
+    /// Layout type for C and D matrix operands
+    typename LayoutC_,
+    /// Operator class tag
+    typename OperatorClass_,
+    /// Tag indicating architecture to tune for
+    typename ArchTag_,
+    /// Threadblock-level tile size (concept: GemmShape)
+    typename ThreadblockShape_,
+    /// Warp-level tile size (concept: GemmShape)
+    typename WarpShape_,
+    /// Instruction-level tile size (concept: GemmShape)
+    typename InstructionShape_,
+    /// Number of stages used in the pipelined mainloop
+    int Stages,
+    /// Operation performed by GEMM
+    typename Operator_,
+    /// Use zfill or predicate for out-of-bound cp.async
+    SharedMemoryClearOption SharedMemoryClear = SharedMemoryClearOption::kNone>
+struct DefaultWint2xMma;
+
+////////////////////////////////////////////////////////////////////////////////
+
+template <
+    /// Type for element A
+    typename ElementA,
+    /// Layout type for A matrix operand
+    typename LayoutA,
+    /// Access granularity of A matrix in units of elements
+    int kAlignmentA,
+    /// Type for element B
+    typename ElementB,
+    /// Layout type for B matrix operand
+    typename LayoutB,
+    /// Access granularity of B matrix in units of elements
+    int kAlignmentB,
+    /// Element type for internal accumulation
+    typename ElementAccumulator,
+    /// Operator class tag
+    typename OperatorClass,
+    /// Tag indicating architecture to tune for
+    typename ArchTag,
+    /// Threadblock-level tile size (concept: GemmShape)
+    typename ThreadblockShape,
+    /// Warp-level tile size (concept: GemmShape)
+    typename WarpShape,
+    /// Instruction-level tile size (concept: GemmShape)
+    typename InstructionShape,
+    /// Stages in GEMM
+    int kStages,
+    /// Operator performed by GEMM
+    typename Operator,
+    /// Use zfill or predicate for out-of-bound cp.async
+    SharedMemoryClearOption SharedMemoryClear>
+struct DefaultWint2xMma<ElementA, LayoutA, kAlignmentA, ElementB, LayoutB, kAlignmentB, ElementAccumulator,
+    layout::RowMajor, OperatorClass, ArchTag, ThreadblockShape, WarpShape, InstructionShape,
+    kStages, Operator, SharedMemoryClear>
+{
+public:
+    static_assert(platform::is_same<ElementA, half_t>::value || platform::is_same<ElementA, bfloat16_t>::value,
+        "Element A must be fp16 or bf16");
+
+    static_assert(platform::is_same<ElementB, uint2b_t>::value,
+        "Element B must be uint2b_t");
+
+    static_assert(platform::is_same<Operator, arch::OpMultiplyAddDequantizeInterleavedBToA>::value,
+        "Mma multistage must dequantize after ldsm");
+
+    using ElementSuperScale = ElementA;
+    using ElementLocalScale = uint4b_t;
+    using ElementCodeScaleZp = float;
+
+    static constexpr int kGroupSize = 64;
+
+    static cutlass::arch::CacheOperation::Kind const CacheOpA = ((sizeof_bits<ElementA>::value * kAlignmentA) == 128)
+        ? cutlass::arch::CacheOperation::Global
+        : cutlass::arch::CacheOperation::Always;
+
+    static cutlass::arch::CacheOperation::Kind const CacheOpB = ((sizeof_bits<ElementB>::value * kAlignmentB) == 128)
+        ? cutlass::arch::CacheOperation::Global
+        : cutlass::arch::CacheOperation::Always;
+
+    // Define the MmaCore components
+    // Mma core does not depend on stages, so pass in at least 3 here to mma multistage pieces are created
+    using MmaCore = typename cutlass::gemm::threadblock::DefaultMmaCore<ThreadblockShape, WarpShape, InstructionShape,
+        ElementA, LayoutA, ElementB, layout::ColumnMajor, ElementAccumulator, layout::RowMajor, OperatorClass,
+        std::max(kStages, 3), Operator, false, CacheOpA, CacheOpB>;
+
+    // Define iterators over tiles from the A operand
+    using ThreadMapA = typename MmaCore::IteratorThreadMapA;
+    using AccessTypeA = cutlass::Array<ElementA, kAlignmentA>;
+    using IteratorA = cutlass::transform::threadblock::PredicatedTileAccessIterator<
+        cutlass::MatrixShape<ThreadblockShape::kM, ThreadblockShape::kK>, ElementA, LayoutA, 1, ThreadMapA,
+        AccessTypeA>;
+
+private:
+    static constexpr int kColumnsInterleaved = LayoutB::kColumnsInterleaved;
+    static constexpr int kRowsPerTile = LayoutB::kRowsPerTile;
+    static_assert(!(MmaCore::Shape::kN % kColumnsInterleaved), "ThreadblockShape must be disivle by kColumnsInterleaved");
+    static_assert(kRowsPerTile == MmaCore::Shape::kK, "");
+
+    using ThreadMapB = typename MmaCore::IteratorThreadMapB;
+    using WarpArrangement = typename ThreadMapB::Detail::WarpThreadArrangement;
+    static_assert(!(WarpArrangement::kStrided % kColumnsInterleaved), "");
+
+    using IteratorShapeB = MatrixShape<
+        MmaCore::Shape::kK * kColumnsInterleaved, MmaCore::Shape::kN / kColumnsInterleaved>;
+    using InterleavedThreadMapB = transform::PitchLinearWarpRakedThreadMap<
+        layout::PitchLinearShape<IteratorShapeB::kRow, IteratorShapeB::kColumn>,
+        ThreadMapB::kThreads,
+        layout::PitchLinearShape<WarpArrangement::kContiguous * kColumnsInterleaved,
+            WarpArrangement::kStrided / kColumnsInterleaved>,
+        MmaCore::kAccessSizeInBits / sizeof_bits<ElementB>::value>;
+
+public:
+    // Define iterators over tiles from the B operand
+    using AccessTypeB = cutlass::Array<ElementB, kAlignmentB>;
+    using IteratorB = cutlass::transform::threadblock::PredicatedTileAccessIterator<
+        IteratorShapeB, ElementB, layout::ColumnMajor, 0, InterleavedThreadMapB,
+        AccessTypeB>;
+
+private:
+    // Define iterators over tiles from extra quant params for B operand
+    using IteratorSuperScale = typename DefaultQuantParamsIterators<
+        ThreadblockShape, ElementSuperScale, -1>::Iterator;
+    using SmemIteratorSuperScale = typename DefaultQuantParamsIterators<
+        ThreadblockShape, ElementSuperScale, -1>::SmemIterator;
+
+    using IteratorLocalScale = typename DefaultQuantParamsIterators<
+        ThreadblockShape, ElementLocalScale, kGroupSize>::Iterator;
+    using SmemIteratorLocalScale = typename DefaultQuantParamsIterators<
+        ThreadblockShape, ElementLocalScale, kGroupSize>::SmemIterator;
+
+    using IteratorCodeScaleZp = typename DefaultQuantParamsIterators<
+        ThreadblockShape, ElementCodeScaleZp, -1>::Iterator;
+    using SmemIteratorCodeScaleZp = typename DefaultQuantParamsIterators<
+        ThreadblockShape, ElementCodeScaleZp, -1>::Iterator;
+
+public:
+    using QuantParamsAccessor = Wint2ParamsAccessor<
+        ElementA, ThreadblockShape, IteratorSuperScale, SmemIteratorSuperScale,
+        IteratorLocalScale, SmemIteratorLocalScale,
+        IteratorCodeScaleZp, SmemIteratorCodeScaleZp, kStages, kGroupSize>;
+
+    // Define the threadblock-scoped multistage matrix multiply
+    using ThreadblockMma = cutlass::gemm::threadblock::Wint2xMmaMultistage<
+        typename MmaCore::Shape,
+        IteratorA, typename MmaCore::SmemIteratorA, MmaCore::kCacheOpA,
+        IteratorB, typename MmaCore::SmemIteratorB, MmaCore::kCacheOpB,
+        ElementAccumulator, layout::RowMajor, typename MmaCore::MmaPolicy,
+        kStages, QuantParamsAccessor, SharedMemoryClear>;
+};
+
+} // namespace threadblock
+} // namespace gemm
+} // namespace cutlass

custom_ops/gpu_ops/cutlass_extensions/gemm/threadblock/wint2x_mma_base.h

@@ -63,8 +63,8 @@ template <
     typename Policy_,
     /// Number of stages,
     int Stages,
-    /// Used for partial specialization
-    typename Enable = bool>
+    /// Size of extra quantized params
+    typename QuantParamsShape>
 class Wint2xMmaBase {
 public:
   ///< Size of the Gemm problem - concept: gemm::GemmShape<>
@@ -93,6 +93,14 @@ public:
   static int const kWarpGemmIterations =
       (WarpGemm::kK / Operator::Policy::MmaShape::kK);
 
+  /// Number of warp-level GEMM oeprations per load for B
+  static constexpr int kWarpGemmIterationsPerLoadForB =
+      Operator::IteratorB::InstructionShape::kRow / Operator::InstructionShape::kK;
+  static_assert(!(kWarpGemmIterations % kWarpGemmIterationsPerLoadForB), "");
+
+  static constexpr int kWarpLoadIterationsForB =
+      kWarpGemmIterations / kWarpGemmIterationsPerLoadForB;
+
   /// Number of stages
   static int const kStages = Stages;
 
@@ -104,8 +112,6 @@ public:
   using TensorRefB =
       TensorRef<typename Operator::ElementB, typename Operator::LayoutB>;
 
-  // using TensorRefZippedB = TensorRef<uint8_t, typename Operator::LayoutB>;
-
   static_assert(kWarpGemmIterations > 1,
                 "The pipelined structure requires at least two warp-level "
                 "GEMM operations.");
@@ -130,20 +136,11 @@ public:
                     Shape::kK * kStages + Policy::SmemPaddingA::kColumn>;
 
     /// Shape of the B matrix operand in shared memory
-    using ShapeB = MatrixShape<Shape::kK + Policy::SmemPaddingB::kRow,
+    using ShapeB = MatrixShape<Shape::kK * kStages + Policy::SmemPaddingB::kRow,
                                Shape::kN + Policy::SmemPaddingB::kColumn>;
 
-    // w uint8; local_scale uint8;
-    constexpr static int kZippedRowsPerStages =
-	Shape::kK / 4 + (Shape::kK + 127) / 128;
-
-    // code_scale float; code_zp float; super_scale ElementB
-    constexpr static int kColumnWiseParamsRows = 2 * sizeof(float) +
-        sizeof_bits<typename Operator::ElementB>::value / 8;
-
-    using ZippedShapeB = MatrixShape<kColumnWiseParamsRows + kZippedRowsPerStages * kStages, Shape::kN>;
-
-    using NopaddingShapeB = MatrixShape<Shape::kK, Shape::kN>;
+    /// Shape of all quant params in shared memory
+    using QuantParamsShapeB = QuantParamsShape;
 
   public:
     //
@@ -156,12 +153,8 @@ public:
     /// Buffer for B operand
     AlignedBuffer<typename Operator::ElementB, ShapeB::kCount> operand_B;
 
-    /// Buffer for quanted B operand
-    AlignedBuffer<uint8_t, ZippedShapeB::kCount> operand_zipped_B;
-
-    /// Buffer for unzip B operand
-    AlignedBuffer<typename Operator::ElementB, NopaddingShapeB::kCount>
-        operand_unzip_B;
+    /// Buffer for extra quant params of B operand
+    AlignedBuffer<uint8_t, QuantParamsShapeB::kCount> operand_quant_params_B;
 
   public:
     //
@@ -191,14 +184,6 @@ public:
     TensorRefB operand_B_ref() {
       return TensorRefB{operand_B.data(), LayoutB()};
     }
-
-    CUTLASS_HOST_DEVICE
-    uint8_t *operand_zipped_B_ptr() { return operand_zipped_B.data(); }
-
-    CUTLASS_HOST_DEVICE
-    typename Operator::ElementB *operand_unzip_B_ptr() {
-      return operand_unzip_B.data();
-    }
   };
 
 protected:

custom_ops/gpu_ops/cutlass_extensions/gemm/threadblock/wint2x_mma_multistage.h

@@ -45,7 +45,8 @@
 
 #include "cutlass_extensions/arch/memory_copy_sm80.h"
 #include "cutlass_extensions/gemm/threadblock/wint2x_mma_base.h"
-#include "cutlass_extensions/gemm/threadblock/wint2x_tile_dequanter.h"
+#include "cutlass_extensions/gemm/threadblock/wint2x_params_accessor.h"
+#include "cutlass_extensions/gemm/warp/mma_tensorop_wint2x_dequantizer.h"
 
 /////////////////////////////////////////////////////////////////////////////////////////////////
 
@@ -86,15 +87,15 @@ template <
     typename Policy_,
     /// Number of stages,
     int Stages,
+    /// Accessor for extra quantized params
+    typename QuantParamsAccessor_,
     /// Use zfill or predicate for out-of-bound cp.async
-    SharedMemoryClearOption SharedMemoryClear = SharedMemoryClearOption::kNone,
-    /// Used for partial specialization
-    typename Enable = bool>
+    SharedMemoryClearOption SharedMemoryClear = SharedMemoryClearOption::kNone>
 class Wint2xMmaMultistage :
-  public Wint2xMmaBase<Shape_, Policy_, Stages> {
+  public Wint2xMmaBase<Shape_, Policy_, Stages, typename QuantParamsAccessor_::QuantParamsShape> {
 public:
   ///< Base class
-  using Base = Wint2xMmaBase<Shape_, Policy_, Stages>;
+  using Base = Wint2xMmaBase<Shape_, Policy_, Stages, typename QuantParamsAccessor_::QuantParamsShape>;
   ///< Size of the Gemm problem - concept: gemm::GemmShape<>
   using Shape = Shape_;
   ///< Iterates over tiles of A operand in global memory
@@ -107,8 +108,11 @@ public:
   using LayoutC = LayoutC_;
   ///< Policy describing tuning details
   using Policy = Policy_;
+  /// Accessor for extra quantized params
+  using QuantParamsAccessor = QuantParamsAccessor_;
+  using QuantArguments = typename QuantParamsAccessor::Arguments;
 
-  using ZippedShapeB = typename Base::SharedStorage::ZippedShapeB;
+  static constexpr int kInterleave = IteratorB::Shape::kRow / Shape::kK;
 
   using SmemIteratorA = SmemIteratorA_;
   using SmemIteratorB = SmemIteratorB_;
@@ -129,6 +133,18 @@ public:
   /// Minimum architecture is Sm80 to support cp.async
   using ArchTag = arch::Sm80;
 
+  //using LayoutScale = typename QuantParamsAccessor::IteratorSuperScale::Layout;
+  using LayoutScale = layout::RowMajor;
+  using WarpTransformedFragmentB = typename Operator::TransformedFragmentB;
+  using WarpDequantizer =
+      warp::MmaTensorOpWin2xDequantizer<Operator,
+                                        typename Base::WarpGemm,
+                                        Operand::kB,
+                                        typename WarpTransformedFragmentB::Element,
+                                        LayoutScale,
+                                        QuantParamsAccessor::kGroupSize>;
+  static_assert(sizeof(WarpDequantizer) > 0, "WarpDequantizer template instantiation failed");
+
   /// Complex transform on A operand
   static ComplexTransform const kTransformA = Operator::kTransformA;
 
@@ -174,18 +190,37 @@ public:
     using WarpTransformedFragmentA = typename Operator::TransformedFragmentA;
     using WarpTransformedFragmentB = typename Operator::TransformedFragmentB;
 
+    using FragmentSuperScale = typename WarpDequantizer::FragmentSuperScale;
+    using FragmentCodeScaleZp = typename WarpDequantizer::FragmentCodeScaleZp;
+    using FragmentLocalScale = typename WarpDequantizer::FragmentLocalScale;
+
     /// Temporary accumulator to facilitate staged-accumulation
     FragmentC tmp_accum_;
 
     /// Pair of A fragments used to overlap shared memory loads and math instructions
-    WarpLoadedFragmentA warp_loaded_frag_A_[2];
-    WarpTransformedFragmentA warp_transformed_frag_A_[2];
+    WarpTransformedFragmentA warp_frag_A_[2];
 
     /// Pair of B fragments used to overlap shared memory loads and math instructions
-    WarpLoadedFragmentB warp_loaded_frag_B_[2];
-    WarpTransformedFragmentB warp_transformed_frag_B_[2];
+    WarpLoadedFragmentB warp_loaded_frag_B_;
+    WarpTransformedFragmentB warp_frag_B_[2];
+
+    /// channel-wise quant params
+    FragmentCodeScaleZp warp_frag_code_scale_;
+    FragmentCodeScaleZp warp_frag_code_zp_;
+    FragmentSuperScale warp_frag_super_scale_;
+
+    /// group-wise quant params
+    FragmentLocalScale warp_frag_local_scale_;
   };
 
+  using ElementA = typename IteratorA::Element;
+  using ElementB = typename IteratorB::Element;
+  using LayoutDetailsForB = kernel::LayoutDetailsB<ElementA, ElementB, ArchTag>;
+
+  static constexpr bool IsTileInterleaveLayout =
+      layout::IsColumnMajorTileInterleave<typename LayoutDetailsForB::Layout>::value;
+  static_assert(!IsTileInterleaveLayout || (IsTileInterleaveLayout && (Shape::kK == LayoutDetailsForB::ThreadblockK)),
+      "Layout K must match threadblockK");
 
  private:
 
@@ -202,17 +237,18 @@ public:
   /// Iterator to write threadblock-scoped tile of B operand to shared memory
   SmemIteratorB smem_iterator_B_;
 
+  /// Accessor for extra quant params for B
+  QuantParamsAccessor quant_params_accessor_B_;
+
+  // Wint2 unzip operator
+  WarpDequantizer warp_dequantizer_;
+
   /// Shared memory write stage index
   int smem_write_stage_idx_;
 
   /// Shared memory read stage index
   int smem_read_stage_idx_;
 
-  uint8_t* column_wise_smem_ptr_B_;
-
-  uint8_t* smem_zipped_ptr_B_;
-  int smem_zipped_bytes_per_stage_B_;
-
 public:
 
   /// Construct from tensor references
@@ -226,10 +262,15 @@ public:
       int warp_idx,
       ///< ID of each thread within a warp
       int lane_idx
-    ):
-      Base(shared_storage, thread_idx, warp_idx, lane_idx),
+  ) : Base(shared_storage, thread_idx, warp_idx, lane_idx),
       smem_iterator_A_(shared_storage.operand_A_ref(), thread_idx),
       smem_iterator_B_(shared_storage.operand_B_ref(), thread_idx),
+      quant_params_accessor_B_(shared_storage.operand_quant_params_B.data(), thread_idx, warp_idx, lane_idx),
+      warp_dequantizer_(quant_params_accessor_B_.super_scale_ref(),
+                        quant_params_accessor_B_.local_scale_ref(),
+                        quant_params_accessor_B_.code_scale_ref(),
+                        quant_params_accessor_B_.code_zp_ref(),
+                        (warp_idx % (Base::WarpCount::kM * Base::WarpCount::kN)) / Base::WarpCount::kM, lane_idx),
       smem_write_stage_idx_(0),
       smem_read_stage_idx_(0)
   {
@@ -250,11 +291,6 @@ public:
         {warp_idx_m, Base::kWarpGemmIterations * warp_idx_k});
     this->warp_tile_iterator_B_.add_tile_offset(
         {Base::kWarpGemmIterations * warp_idx_k, warp_idx_n});
-
-    column_wise_smem_ptr_B_ = shared_storage.operand_zipped_B_ptr();
-
-    smem_zipped_ptr_B_ = column_wise_smem_ptr_B_ + Base::SharedStorage::kColumnWiseParamsRows * ZippedShapeB::kColumn;
-    smem_zipped_bytes_per_stage_B_ = Base::SharedStorage::kZippedRowsPerStages * ZippedShapeB::kColumn;
   }
 
   /// Advance shared memory read-iterators to the next stage
@@ -266,28 +302,22 @@ public:
     if (smem_read_stage_idx_ == Base::kStages) {
       // Wrap back around to the 'start' of the circular buffer in shared memory
       this->warp_tile_iterator_A_.add_tile_offset({0, -Base::kStages * Policy::kPartitionsK * Base::kWarpGemmIterations});
-      // this->warp_tile_iterator_B_.add_tile_offset({-Base::kStages * Policy::kPartitionsK * Base::kWarpGemmIterations, 0});
+      this->warp_tile_iterator_B_.add_tile_offset({-Base::kStages * Policy::kPartitionsK * Base::kWarpLoadIterationsForB, 0});
       smem_read_stage_idx_ = 0;
     }
-    this->warp_tile_iterator_B_.add_tile_offset({-Policy::kPartitionsK * Base::kWarpGemmIterations, 0});
   }
 
   /// Advance global memory read-iterators and shared memory write-iterators to the stage
-  template <typename TileDequanterB>
   CUTLASS_DEVICE
-  void advance_smem_write_stage(
-    IteratorA &iterator_A,
-    IteratorB &iterator_B,
-    TileDequanterB &tile_dequanter_B)
+  void advance_smem_write_stage(IteratorA &iterator_A, IteratorB &iterator_B)
   {
     // Advance global iterators
     iterator_A.add_tile_offset({0, 1});
-    //iterator_B.add_tile_offset({1, 0});
-    tile_dequanter_B.AddTileOffset({1, 0});
+    iterator_B.add_tile_offset({1, 0});
 
     // Advance shared iterators
     smem_iterator_A_.add_tile_offset({0, 1});
-    //smem_iterator_B_.add_tile_offset({1, 0});
+    smem_iterator_B_.add_tile_offset({1, 0});
 
     // Increment shared memory write stage index
     ++smem_write_stage_idx_;
@@ -295,7 +325,7 @@ public:
     if (smem_write_stage_idx_ == Base::kStages) {
       // Wrap back around to the 'start' of the circular buffer in shared memory
       smem_iterator_A_.add_tile_offset({0, -Base::kStages});
-      //smem_iterator_B_.add_tile_offset({-Base::kStages, 0});
+      smem_iterator_B_.add_tile_offset({-Base::kStages, 0});
       smem_write_stage_idx_ = 0;
     }
   }
@@ -338,9 +368,14 @@ public:
     }
   }
 
-  template <bool GlobalToSharedB>
   CUTLASS_DEVICE
   void copy_tiles_and_advance_B(IteratorB &iterator_B, int group_start_B = 0) {
+    if constexpr (SharedMemoryClear == SharedMemoryClearOption::kZfill) {
+      if (threadIdx.x >= IteratorB::ThreadMap::kThreads) {
+        return;
+      }
+    }
+
     iterator_B.set_iteration_index(group_start_B *
                                    IteratorB::kAccessesPerVector);
     this->smem_iterator_B_.set_iteration_index(group_start_B);
@@ -360,13 +395,14 @@ public:
         CUTLASS_PRAGMA_UNROLL
         for (int v = 0; v < IteratorB::kAccessesPerVector; ++v) {
           auto gmem_ptr = iterator_B.get();
+          bool is_valid = (threadIdx.x < IteratorB::ThreadMap::kThreads) ? iterator_B.valid() : false;
 
           if (SharedMemoryClear == SharedMemoryClearOption::kZfill) {
-            cutlass::arch::copy_zfill<kSrcBytes, kCacheOpB, GlobalToSharedB>(
-                dst_ptr + v, gmem_ptr, iterator_B.valid());
+            cutlass::arch::cp_async_zfill<kSrcBytes, kCacheOpB>(
+                dst_ptr + v, gmem_ptr, is_valid);
           } else {
-            cutlass::arch::copy<kSrcBytes, kCacheOpB, GlobalToSharedB>(
-                dst_ptr + v, gmem_ptr, iterator_B.valid());
+            cutlass::arch::cp_async<kSrcBytes, kCacheOpB>(
+                dst_ptr + v, gmem_ptr, is_valid);
           }
 
           ++iterator_B;
@@ -375,7 +411,6 @@ public:
         ++this->smem_iterator_B_;
       }
     }
-    __syncthreads();
   }
 
   CUTLASS_DEVICE
@@ -399,8 +434,6 @@ public:
             IteratorA::ThreadMap::kElementsPerAccess /
             IteratorA::kAccessesPerVector / 8;
 
-        int src_bytes = (iterator_A.valid() ? kSrcBytes : 0);
-
         cutlass::arch::cp_async_zfill<kSrcBytes, kCacheOpA>(
             dst_ptr + v, iterator_A.get(), iterator_A.valid());
 
@@ -411,9 +444,12 @@ public:
     }
   }
 
-  template <bool GlobalToSharedB, bool InitStage>
   CUTLASS_DEVICE
   void copy_tiles_and_advance_per_stage_B(IteratorB &iterator_B) {
+    if (threadIdx.x >= IteratorB::ThreadMap::kThreads) {
+      return;
+    }
+
     iterator_B.set_iteration_index(0);
     this->smem_iterator_B_.set_iteration_index(0);
 
@@ -433,35 +469,23 @@ public:
             IteratorB::ThreadMap::kElementsPerAccess /
             IteratorB::kAccessesPerVector / 8;
 
-        if (InitStage) {
-          cutlass::arch::copy_zfill<kSrcBytes, kCacheOpB, GlobalToSharedB>(
-              dst_ptr + v, iterator_B.get(), iterator_B.valid());
-        } else {
-          if (SharedMemoryClear == SharedMemoryClearOption::kZfill) {
-            cutlass::arch::copy_zfill<kSrcBytes, kCacheOpB, GlobalToSharedB>(
-                dst_ptr + v, gmem_ptr, iterator_B.valid());
-          } else {
-            cutlass::arch::copy<kSrcBytes, kCacheOpB, GlobalToSharedB>(
-                dst_ptr + v, gmem_ptr, iterator_B.valid());
-          }
-        }
+        cutlass::arch::cp_async_zfill<kSrcBytes, kCacheOpB>(
+            dst_ptr + v, iterator_B.get(), iterator_B.valid());
 
         ++iterator_B;
       }
 
       ++this->smem_iterator_B_;
     }
-    __syncthreads();
   }
 
   /// GEMM prologue.  Bootstrap the global->shared memory pipeline by fetching
   /// the global fragments needed by the first kStages-1 threadblock mainloop iterations
-  template <typename TileDequanterB>
   CUTLASS_DEVICE
   void prologue(
     IteratorA &iterator_A,      ///< [in|out] iterator over A operand in global memory
     IteratorB &iterator_B,      ///< [in|out] iterator over B operand in global memory
-    TileDequanterB &tile_dequanter_B,
+    QuantArguments &mma_quant_args, ///< iterators for extra quant params for B
     int &gemm_k_iterations)     ///< [in|out] number of threadblock mainloop iterations remaining
   {
     // Issue several complete stages
@@ -476,11 +500,18 @@ public:
       copy_tiles_and_advance_per_stage_A(iterator_A);
 
       // Async copy zipped B to shared memory.
-      tile_dequanter_B.Load(smem_zipped_ptr_B_ + (stage % Base::kStages) * smem_zipped_bytes_per_stage_B_,
-                            column_wise_smem_ptr_B_, stage);
+      copy_tiles_and_advance_per_stage_B(iterator_B);
+
+      // Async copy other quantized params to shared memory, local_scale, code_scale, code_zp, super_scale.
+      if (stage == 0) {
+        quant_params_accessor_B_.copy_tiles_and_advance_per_stage<true>(mma_quant_args, stage);
+      } else {
+        quant_params_accessor_B_.copy_tiles_and_advance_per_stage<false>(mma_quant_args, stage);
+      }
 
       // Move to the next write stage
-      advance_smem_write_stage(iterator_A, iterator_B, tile_dequanter_B);
+      advance_smem_write_stage(iterator_A, iterator_B);
+      quant_params_accessor_B_.advance_smem_write_stage(mma_quant_args);
 
       // Defines the boundary of a stage of cp.async.
       cutlass::arch::cp_async_fence();
@@ -510,6 +541,10 @@ public:
         ++last_smem_iterator_A;
       }
 
+      if (threadIdx.x >= IteratorB::ThreadMap::kThreads) {
+        return;
+      }
+
       /// Iterator to write threadblock-scoped tile of B operand to shared memory
       SmemIteratorB last_smem_iterator_B(this->smem_iterator_B_);
       typename IteratorB::AccessType zero_B;
@@ -542,57 +577,57 @@ public:
   }
 
   /// Perform a threadblock mainloop iteration of matrix multiply-accumulate
-  template <typename TileDequanterB>
   CUTLASS_DEVICE
   void mac_loop_iter(
     PipeState &pipe_state,          ///< [in|out] loop-carried pipeline state
     FragmentC &accum,               ///< [in|out] destination accumulator tile
     IteratorA &iterator_A,          ///< [in|out] iterator over A operand in global memory
     IteratorB &iterator_B,          ///< [in|out] iterator over B operand in global memory
-    TileDequanterB &tile_dequanter_B, ///< [in|out] tile dequantizer for B operand
-    int &gemm_k_iterations, ///< [in|out] number of threadblock mainloop iterations remaining
+    QuantArguments &mma_quant_args, ///< iterators for extra quant params for B
+    int &gemm_k_iterations,         ///< [in|out] number of threadblock mainloop iterations remaining
     int stage)
   {
+    const int mma_stage = stage - Base::kStages + 1;
+
     // Unroll the warp-level MMA tiles of a threadblock's mainloop iteration
     CUTLASS_PRAGMA_UNROLL
     for (int warp_mma_k = 0; warp_mma_k < Base::kWarpGemmIterations; ++warp_mma_k) {
-      // CUTLASS_TRACE_DEVICE(" [MMa] stage=%d, warp_mma_k=%d", stage, warp_mma_k);
+
+      int warp_k_compute_offset_B = warp_mma_k % Base::kWarpGemmIterationsPerLoadForB;
+
+      if (warp_k_compute_offset_B == Base::kWarpGemmIterationsPerLoadForB - 1) {
+        // Load the next warp-tile's B fragment from shared memory
+        this->warp_tile_iterator_B_.set_kgroup_index(((warp_mma_k + 1) % Base::kWarpGemmIterations) / Base::kWarpLoadIterationsForB);
+        this->warp_tile_iterator_B_.load(pipe_state.warp_loaded_frag_B_);
+        ++this->warp_tile_iterator_B_;
+      }
+
+      // load next-tile of group-wise local_scale from shared memory
+      if (warp_mma_k == Base::kWarpGemmIterations - 1) {
+        warp_dequantizer_.load(pipe_state.warp_frag_local_scale_);
+      }
 
       // Load the next warp-tile's A fragment from shared memory
       this->warp_tile_iterator_A_.set_kgroup_index((warp_mma_k + 1) % Base::kWarpGemmIterations);
-      this->warp_tile_iterator_A_.load(pipe_state.warp_loaded_frag_A_[(warp_mma_k + 1) % 2]);
+      this->warp_tile_iterator_A_.load(pipe_state.warp_frag_A_[(warp_mma_k + 1) % 2]);
       ++this->warp_tile_iterator_A_;
 
-      if (warp_mma_k + 1 == Base::kWarpGemmIterations) {
-        // Unpack and dequant the first stage of B.
-        int unpack_stage = stage - Base::kStages + 2;
-        tile_dequanter_B.UnpackAndDequant(smem_zipped_ptr_B_ + (unpack_stage % Base::kStages) * smem_zipped_bytes_per_stage_B_,
-                                          column_wise_smem_ptr_B_, unpack_stage);
-
-        // Copy dequatized data to shared memory used by mma core.
-        copy_tiles_and_advance_per_stage_B<false, false>(iterator_B);
-      }
-
-      // Load the next warp-tile's B fragment from shared memory
-      this->warp_tile_iterator_B_.set_kgroup_index((warp_mma_k + 1) % Base::kWarpGemmIterations);
-      this->warp_tile_iterator_B_.load(pipe_state.warp_loaded_frag_B_[(warp_mma_k + 1) % 2]);
-      ++this->warp_tile_iterator_B_;
-
-      // Except for the first warp-tile, all warp-tiles convert their incoming shared memory fragments as necessary
-      if (warp_mma_k > 0) {
-        warp_mma_.transform(
-          pipe_state.warp_transformed_frag_A_[warp_mma_k % 2],
-          pipe_state.warp_transformed_frag_B_[warp_mma_k % 2],
-          pipe_state.warp_loaded_frag_A_[warp_mma_k % 2],
-          pipe_state.warp_loaded_frag_B_[warp_mma_k % 2]);
-      }
+      // dequantizes next warp-tile
+      warp_dequantizer_.dequantize(pipe_state.warp_frag_local_scale_,
+                                   pipe_state.warp_frag_code_scale_,
+                                   pipe_state.warp_frag_code_zp_,
+                                   pipe_state.warp_frag_super_scale_,
+                                   pipe_state.warp_loaded_frag_B_,
+                                   pipe_state.warp_frag_B_[(warp_mma_k + 1) % 2],
+                                   ((warp_mma_k == Base::kWarpGemmIterations - 1) ? (mma_stage + 1) : mma_stage) * Shape::kK,
+                                   (warp_mma_k + 1) % Base::kWarpGemmIterationsPerLoadForB);
 
       // Execute the current warp-tile of MMA operations
-      if (Detail::kStagedAccumulation) {
+      if constexpr (Detail::kStagedAccumulation) {
         warp_mma_(
           pipe_state.tmp_accum_,
-          pipe_state.warp_transformed_frag_A_[warp_mma_k % 2],
-          pipe_state.warp_transformed_frag_B_[warp_mma_k % 2],
+          pipe_state.warp_frag_A_[warp_mma_k % 2],
+          pipe_state.warp_frag_B_[warp_mma_k % 2],
           pipe_state.tmp_accum_
         );
 
@@ -604,22 +639,22 @@ public:
       } else {
         warp_mma_(
           accum,
-          pipe_state.warp_transformed_frag_A_[warp_mma_k % 2],
-          pipe_state.warp_transformed_frag_B_[warp_mma_k % 2],
-          accum
-        );
+          pipe_state.warp_frag_A_[warp_mma_k % 2],
+          pipe_state.warp_frag_B_[warp_mma_k % 2],
+          accum);
       }
 
       // Except for the last warp-tile, all warp-tiles issue their share of
       // global->shared fragment copies
       if (warp_mma_k < Base::kWarpGemmIterations - 1) {
         int group_start_iteration_A = warp_mma_k * Detail::kAccessesPerGroupA;
+        int group_start_iteration_B = warp_mma_k * Detail::kAccessesPerGroupB;
 
         copy_tiles_and_advance_A(iterator_A, group_start_iteration_A);
+        copy_tiles_and_advance_B(iterator_B, group_start_iteration_B);
 
         if (warp_mma_k == 0) {
-          tile_dequanter_B.Load(smem_zipped_ptr_B_ + (stage % Base::kStages) * smem_zipped_bytes_per_stage_B_,
-                                column_wise_smem_ptr_B_, stage);
+          quant_params_accessor_B_.copy_tiles_and_advance_per_stage<false>(mma_quant_args, stage);
         }
       }
 
@@ -628,9 +663,15 @@ public:
       //   - moves to the next global fetch stage
       if (warp_mma_k + 2 == Base::kWarpGemmIterations) {
         // Performs the last warp-tile's share of global->shared fragment copies
-        int group_start_iteration_A = (warp_mma_k + 1) * Detail::kAccessesPerGroupA;
+        if constexpr (Detail::AsyncCopyIterationsPerStageA >= Base::kWarpGemmIterations) {
+          int group_start_iteration_A = (warp_mma_k + 1) * Detail::kAccessesPerGroupA;
+          copy_tiles_and_advance_A(iterator_A, group_start_iteration_A);
+        }
 
-        copy_tiles_and_advance_A(iterator_A, group_start_iteration_A);
+        if constexpr (Detail::AsyncCopyIterationsPerStageB >= Base::kWarpGemmIterations) {
+          int group_start_iteration_B = (warp_mma_k + 1) * Detail::kAccessesPerGroupB;
+          copy_tiles_and_advance_B(iterator_B, group_start_iteration_B);
+        }
 
         // Inserts a memory fence between stages of cp.async instructions.
         cutlass::arch::cp_async_fence();
@@ -639,69 +680,66 @@ public:
         gmem_wait();
 
         // Move to the next global fetch stage
-        advance_smem_write_stage(iterator_A, iterator_B, tile_dequanter_B);
+        advance_smem_write_stage(iterator_A, iterator_B);
+        quant_params_accessor_B_.advance_smem_write_stage(mma_quant_args);
+
         advance_smem_read_stage();
+        int byte_offset = quant_params_accessor_B_.advance_smem_read_stage();
+        warp_dequantizer_.add_pointer_offset(byte_offset);
 
         // Disable global fetching when done with global fetch iterations
         --gemm_k_iterations;
         iterator_A.clear_mask(gemm_k_iterations == 0);
-        iterator_B.clear_mask(gemm_k_iterations == (-Base::kStages + 1));
-      }
-
-      // The last warp-tile also converts the shared memory fragments used by
-      // the first warp-tile of the next iteration, if necessary (so we can
-      // immediately start issuing MMA instructions at the top of the loop )
-      if (warp_mma_k + 1 == Base::kWarpGemmIterations) {
-        warp_mma_.transform(
-          pipe_state.warp_transformed_frag_A_[(warp_mma_k + 1) % 2],
-          pipe_state.warp_transformed_frag_B_[(warp_mma_k + 1) % 2],
-          pipe_state.warp_loaded_frag_A_[(warp_mma_k + 1) % 2],
-          pipe_state.warp_loaded_frag_B_[(warp_mma_k + 1) % 2]);
+        iterator_B.clear_mask(gemm_k_iterations == 0);
+        quant_params_accessor_B_.clear_mask(mma_quant_args, gemm_k_iterations == 0);
       }
     }
   }
 
   /// Perform the specified number of threadblock mainloop iterations of matrix
   /// multiply-accumulate.  Assumes prologue has been initiated.
-  template <typename TileDequanterB>
   CUTLASS_DEVICE
   void gemm_iters(
       int gemm_k_iterations,        ///< number of threadblock mainloop iterations
       FragmentC &accum,             ///< [in|out] accumulator tile
       IteratorA &iterator_A,        ///< [in|out] iterator over A operand in global memory
-      IteratorB &iterator_B,
-      TileDequanterB &tile_dequanter_B)        ///< [in|out] iterator over B operand in global memory
+      IteratorB &iterator_B,        ///< [in|out] iterator over B operand in global memory
+      QuantArguments &mma_quant_args)
   {
     PipeState pipe_state;
 
-    // Unpack and dequant the first stage of B.
-    tile_dequanter_B.UnpackAndDequant(smem_zipped_ptr_B_, column_wise_smem_ptr_B_, 0);
-
     // Disable global fetching if done with global fetch iterations
     iterator_A.clear_mask(gemm_k_iterations == 0);
-    iterator_B.clear_mask(gemm_k_iterations == (-Base::kStages + 1));
-
-    // Load first warp-tile's A fragment from shared memory
-    this->warp_tile_iterator_A_.set_kgroup_index(0);
-    this->warp_tile_iterator_A_.load(pipe_state.warp_loaded_frag_A_[0]);
-    ++this->warp_tile_iterator_A_;
-
-    // Copy dequatized data to shared memory used by mma core.
-    copy_tiles_and_advance_per_stage_B<false, true>(iterator_B);
+    iterator_B.clear_mask(gemm_k_iterations == 0);
+    quant_params_accessor_B_.clear_mask(mma_quant_args, gemm_k_iterations == 0);
 
     // Load first warp-tile's B fragment from shared memory
     this->warp_tile_iterator_B_.set_kgroup_index(0);
-    this->warp_tile_iterator_B_.load(pipe_state.warp_loaded_frag_B_[0]);
+    this->warp_tile_iterator_B_.load(pipe_state.warp_loaded_frag_B_);
     ++this->warp_tile_iterator_B_;
 
-    // Transform, if necessary, the first warp-tile's shared memory fragments
-    warp_mma_.transform(
-      pipe_state.warp_transformed_frag_A_[0],
-      pipe_state.warp_transformed_frag_B_[0],
-      pipe_state.warp_loaded_frag_A_[0],
-      pipe_state.warp_loaded_frag_B_[0]);
+    warp_dequantizer_.load(pipe_state.warp_frag_code_scale_,
+                           pipe_state.warp_frag_code_zp_,
+                           pipe_state.warp_frag_super_scale_);
 
-    if (Detail::kStagedAccumulation) {
+    warp_dequantizer_.load(pipe_state.warp_frag_local_scale_);
+
+    // Load first warp-tile's A fragment from shared memory
+    this->warp_tile_iterator_A_.set_kgroup_index(0);
+    this->warp_tile_iterator_A_.load(pipe_state.warp_frag_A_[0]);
+    ++this->warp_tile_iterator_A_;
+
+    // Dequantize B to in register
+    warp_dequantizer_.dequantize(pipe_state.warp_frag_local_scale_,
+                                 pipe_state.warp_frag_code_scale_,
+                                 pipe_state.warp_frag_code_zp_,
+                                 pipe_state.warp_frag_super_scale_,
+                                 pipe_state.warp_loaded_frag_B_,
+                                 pipe_state.warp_frag_B_[0],
+                                 0,
+                                 0);
+
+    if constexpr (Detail::kStagedAccumulation) {
       pipe_state.tmp_accum_.clear();
     }
 
@@ -715,13 +753,13 @@ public:
         accum,
         iterator_A,
         iterator_B,
-        tile_dequanter_B,
+        mma_quant_args,
         gemm_k_iterations,
         stage);
       stage += 1;
     }
 
-    if (Detail::kStagedAccumulation) {
+    if constexpr (Detail::kStagedAccumulation) {
       plus<FragmentC> plus_accum;
       accum = plus_accum(accum, pipe_state.tmp_accum_);
     }
@@ -761,14 +799,12 @@ public:
     else
     {
       this->warp_tile_iterator_A_.add_tile_offset({0, ((Base::kStages - 2) * kStageIters)});
-      //this->warp_tile_iterator_B_.add_tile_offset({((Base::kStages - 2) * kStageIters), 0});
-      this->warp_tile_iterator_B_.add_tile_offset({(-2 * kStageIters), 0});
+      this->warp_tile_iterator_B_.add_tile_offset({((Base::kStages - 2) * kStageIters), 0});
     }
     smem_read_stage_idx_ = smem_write_stage_idx_;
   }
 
   /// Perform a threadblock-scoped matrix multiply-accumulate, pre-load B to shared memory.
-  template <typename TileDequanterB>
   CUTLASS_DEVICE
   void operator()(
       ///< problem size of GEMM
@@ -779,13 +815,13 @@ public:
       IteratorA iterator_A,
       ///< iterator over B operand in global memory
       IteratorB iterator_B,
-      ///< pre-load and dequantize B to shared memory
-      TileDequanterB tile_dequanter_B,
+      ///< iterators for extra quant params for B
+      QuantArguments mma_quant_args,
       ///< initial value of accumulator
       FragmentC const &src_accum) {
 
     // Prologue (start fetching iterations of global fragments into shared memory)
-    prologue(iterator_A, iterator_B, tile_dequanter_B, gemm_k_iterations);
+    prologue(iterator_A, iterator_B, mma_quant_args, gemm_k_iterations);
 
     // Wait until we have at least one completed global fetch stage
     gmem_wait();
@@ -794,7 +830,7 @@ public:
     accum = src_accum;
 
     // Perform the MAC-iterations
-    gemm_iters(gemm_k_iterations, accum, iterator_A, iterator_B, tile_dequanter_B);
+    gemm_iters(gemm_k_iterations, accum, iterator_A, iterator_B, mma_quant_args);
   }
 };
 

custom_ops/gpu_ops/cutlass_extensions/gemm/threadblock/wint2x_params_accessor.h

@@ -0,0 +1,315 @@
+// Copyright (c) 2025 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#pragma once
+
+#include "cutlass/arch/memory_sm80.h"
+#include "cutlass/cutlass.h"
+#include "cutlass/gemm/gemm.h"
+#include "cutlass/matrix_shape.h"
+#include "cutlass/trace.h"
+
+namespace cutlass {
+namespace gemm {
+namespace threadblock {
+
+template <
+    /// Original data type
+    typename T,
+    /// Size of the Gemm problem - concept: gemm::GemmShape<>
+    typename Shape_,
+    /// Iterators over super scales in global memory
+    typename IteratorSuperScale_,
+    /// Iterators over super scales in shared memory
+    typename SmemIteratorSuperScale_,
+    /// Iterators over local scales in global memory
+    typename IteratorLocalScale_,
+    /// Iterators over local scales in shared memory
+    typename SmemIteratorLocalScale_,
+    /// Iterators over code scales and zps in global memory
+    typename IteratorCodeScaleZp_,
+    /// Iterators over code scales and zps in shared memory
+    typename SmemIteratorCodeScaleZp_,
+    /// Number of stages,
+    int Stages_,
+    /// Group size for quantization
+    int GroupSize_>
+class Wint2ParamsAccessor {
+public:
+  static_assert(platform::is_same<T, half_t>::value || platform::is_same<T, bfloat16_t>::value,
+        "T must be fp16 or bf16");
+
+  using ElementType = T;
+  using Shape = Shape_;
+
+  using IteratorSuperScale = IteratorSuperScale_;
+  using SmemIteratorSuperScale = SmemIteratorSuperScale_;
+
+  using IteratorLocalScale = IteratorLocalScale_;
+  using SmemIteratorLocalScale = SmemIteratorLocalScale_;
+
+  using IteratorCodeScaleZp = IteratorCodeScaleZp_;
+  using SmemIteratorCodeScaleZp = SmemIteratorCodeScaleZp_;
+
+  constexpr static int kStages = Stages_;
+  constexpr static int kGroupSize = GroupSize_;
+
+  using ElementSuperScale = typename IteratorSuperScale::Element;
+  using LayoutSuperScale = typename IteratorSuperScale::Layout;
+
+  /// local_scale uint4 and group-wise
+  using ElementLocalScale = typename IteratorLocalScale::Element;
+  using LayoutLocalScale = typename IteratorLocalScale::Layout;
+  static_assert(platform::is_same<ElementLocalScale, uint4b_t>::value,
+        "local_scale's type must be uint4b_t.");
+
+  using ElementCodeScaleZp = typename IteratorCodeScaleZp::Element;
+  using LayoutCodeScaleZp = typename IteratorCodeScaleZp::Layout;
+
+  /// 2 uint4b_t values are stored in a single uint8_t
+  constexpr static int kStagesPerLocalScaleLoad = 2 * kGroupSize / Shape::kK;
+  constexpr static int kLocalScaleRows =
+      IteratorLocalScale::Shape::kRow * IteratorLocalScale::Shape::kColumn * sizeof_bits<ElementLocalScale>::value / 8 / Shape::kN;
+
+  using SmemElement = uint8_t;
+  constexpr static int kSmemRows =
+      kLocalScaleRows * kStages + sizeof(ElementSuperScale) + sizeof(ElementCodeScaleZp) * 2;
+  constexpr static int kSmemColumns = Shape::kN;
+
+  using QuantParamsShape = MatrixShape<kSmemRows, kSmemColumns>;
+
+  constexpr static int kSuperScaleSmemOffset = 0;
+  constexpr static int kCodeScaleSmemOffset = kSmemColumns * sizeof(ElementSuperScale);
+  constexpr static int kCodeZpSmemOffset = kCodeScaleSmemOffset + kSmemColumns * sizeof(ElementCodeScaleZp);
+  constexpr static int kLocalScaleSmemOffset = kCodeZpSmemOffset + kSmemColumns * sizeof(ElementCodeScaleZp);
+
+  /// TensorRef type for loading element from a tensor
+  using SuperTensorRef = cutlass::TensorRef<ElementSuperScale, LayoutSuperScale>;
+  using LocalTensorRef = cutlass::TensorRef<ElementLocalScale, LayoutLocalScale>;
+  using CodeTensorRef = cutlass::TensorRef<ElementCodeScaleZp, LayoutCodeScaleZp>;
+
+  struct Arguments {
+    IteratorSuperScale iterator_super_scale;
+    IteratorLocalScale iterator_local_scale;
+    IteratorCodeScaleZp iterator_code_scale;
+    IteratorCodeScaleZp iterator_code_zp;
+
+    int local_scale_pointer_offset;
+
+    CUTLASS_DEVICE
+    Arguments(IteratorSuperScale iterator_super_scale,
+              IteratorLocalScale iterator_local_scale,
+              IteratorCodeScaleZp iterator_code_scale,
+              IteratorCodeScaleZp iterator_code_zp,
+              int local_scale_pointer_offset)
+      : iterator_super_scale(iterator_super_scale),
+        iterator_local_scale(iterator_local_scale),
+        iterator_code_scale(iterator_code_scale),
+        iterator_code_zp(iterator_code_zp),
+        local_scale_pointer_offset(local_scale_pointer_offset) {}
+  };
+
+private:
+  //
+  // Data members
+  //
+
+  /// Begin address of shared memory
+  uint8_t* smem_pointer_;
+
+  /// Iterator to write threadblock-scoped tile of super scale operand to shared memory
+  SmemIteratorSuperScale smem_iterator_super_scale_;
+  /// Iterator to write threadblock-scoped tile of local scale operand to shared memory
+  SmemIteratorLocalScale smem_iterator_local_scale_;
+  /// Iterator to write threadblock-scoped tile of code scale operand to shared memory
+  SmemIteratorCodeScaleZp smem_iterator_code_scale_;
+  /// Iterator to write threadblock-scoped tile of code zp operand to shared memory
+  SmemIteratorCodeScaleZp smem_iterator_code_zp_;
+
+  /// Shared memory write stage index
+  int smem_write_stage_idx_;
+
+  /// Shared memory read stage index
+  int smem_read_stage_idx_;
+
+  CUTLASS_DEVICE
+  ElementSuperScale* get_super_scale_smem_ptr() {
+    return reinterpret_cast<ElementSuperScale*>(smem_pointer_ + kSuperScaleSmemOffset);
+  }
+
+  CUTLASS_DEVICE
+  ElementLocalScale* get_local_scale_smem_ptr() {
+    return reinterpret_cast<ElementLocalScale*>(smem_pointer_ + kLocalScaleSmemOffset);
+  }
+
+  CUTLASS_DEVICE
+  ElementCodeScaleZp* get_code_scale_smem_ptr() {
+    return reinterpret_cast<ElementCodeScaleZp*>(smem_pointer_ + kCodeScaleSmemOffset);
+  }
+
+  CUTLASS_DEVICE
+  ElementCodeScaleZp* get_code_zp_smem_ptr() {
+    return reinterpret_cast<ElementCodeScaleZp*>(smem_pointer_ + kCodeZpSmemOffset);
+  }
+
+public:
+  /// Construct from tensor references
+  CUTLASS_DEVICE
+  Wint2ParamsAccessor(
+      ///< prointer of shared memory
+      uint8_t* smem_pointer,
+      ///< ID within the threadblock
+      int thread_idx,
+      ///< ID of warp
+      int warp_idx,
+      ///< ID of each thread within a warp
+      int lane_idx)
+    : smem_pointer_(smem_pointer),
+      smem_iterator_super_scale_(LayoutSuperScale(IteratorSuperScale::Shape::kColumn),
+          get_super_scale_smem_ptr(), {1, IteratorSuperScale::Shape::kColumn}, thread_idx),
+      smem_iterator_local_scale_(LayoutLocalScale(IteratorLocalScale::Shape::kColumn),
+          get_local_scale_smem_ptr(), {1, IteratorLocalScale::Shape::kColumn}, thread_idx),
+      smem_iterator_code_scale_(LayoutCodeScaleZp(IteratorCodeScaleZp::Shape::kColumn),
+          get_code_scale_smem_ptr(), {1, IteratorCodeScaleZp::Shape::kColumn}, thread_idx),
+      smem_iterator_code_zp_(LayoutCodeScaleZp(IteratorCodeScaleZp::Shape::kColumn),
+          get_code_zp_smem_ptr(), {1, IteratorCodeScaleZp::Shape::kColumn}, thread_idx),
+      smem_write_stage_idx_(0),
+      smem_read_stage_idx_(0) {}
+
+  CUTLASS_DEVICE
+  SuperTensorRef super_scale_ref() {
+    return {get_super_scale_smem_ptr(), LayoutSuperScale(IteratorSuperScale::Shape::kColumn)};
+  }
+
+  CUTLASS_DEVICE
+  LocalTensorRef local_scale_ref() {
+    return {get_local_scale_smem_ptr(), LayoutLocalScale(IteratorLocalScale::Shape::kColumn)};
+  }
+
+  CUTLASS_DEVICE
+  CodeTensorRef code_scale_ref() {
+    return {get_code_scale_smem_ptr(), LayoutCodeScaleZp(IteratorCodeScaleZp::Shape::kColumn)};
+  }
+
+  CUTLASS_DEVICE
+  CodeTensorRef code_zp_ref() {
+    return {get_code_zp_smem_ptr(), LayoutCodeScaleZp(IteratorCodeScaleZp::Shape::kColumn)};
+  }
+
+  template <bool IsFirstStage>
+  CUTLASS_DEVICE
+  void copy_tiles_and_advance_per_stage(Arguments &quant_args, int stage) {
+    if constexpr (IsFirstStage) {
+      // Load channel-wise super_scale to shared memory, which only needs to be done once.
+      typename IteratorSuperScale::Fragment tb_frag_super_scale;
+      tb_frag_super_scale.clear();
+      quant_args.iterator_super_scale.load(tb_frag_super_scale);
+      this->smem_iterator_super_scale_.store(tb_frag_super_scale);
+
+      // Load channel-wise code_scale to shared memory, which only needs to be done once.
+      typename IteratorCodeScaleZp::Fragment tb_frag_code_scale;
+      tb_frag_code_scale.clear();
+      quant_args.iterator_code_scale.load(tb_frag_code_scale);
+      this->smem_iterator_code_scale_.store(tb_frag_code_scale);
+
+      // Load channel-wise code_zp to shared memory, which only needs to be done once.
+      typename IteratorCodeScaleZp::Fragment tb_frag_code_zp;
+      tb_frag_code_zp.clear();
+      quant_args.iterator_code_zp.load(tb_frag_code_zp);
+      this->smem_iterator_code_zp_.store(tb_frag_code_zp);
+    }
+
+    if ((stage % kStagesPerLocalScaleLoad) == 0) {
+      // Load group-wise local_scale to shared memory, which only needs to be done at each stage.
+      // Since 2 uint4b_t values of local_scale are saved in a single uint8_t, local_scale needs to be loaded once every two stages.
+      using AccessType = typename IteratorLocalScale::AccessType;
+      cutlass::arch::CacheOperation::Kind const kCacheOp = (sizeof_bits<AccessType>::value == 128)
+          ? cutlass::arch::CacheOperation::Global : cutlass::arch::CacheOperation::Always;
+
+      quant_args.iterator_local_scale.set_iteration_index(0);
+      this->smem_iterator_local_scale_.set_iteration_index(0);
+
+      // Async Copy for local_scale
+      CUTLASS_PRAGMA_UNROLL
+      for (int j = 0; j < IteratorLocalScale::ThreadMap::Iterations::kCount; ++j) {
+        AccessType *dst_ptr =
+            reinterpret_cast<AccessType *>(this->smem_iterator_local_scale_.get());
+
+        CUTLASS_PRAGMA_UNROLL
+        for (int v = 0; v < IteratorLocalScale::kAccessesPerVector; ++v) {
+          auto gmem_ptr = quant_args.iterator_local_scale.get();
+
+          int const kSrcBytes =
+              sizeof_bits<typename IteratorLocalScale::Element>::value *
+              IteratorLocalScale::ThreadMap::kElementsPerAccess /
+              IteratorLocalScale::kAccessesPerVector / 8;
+
+              cutlass::arch::cp_async<kSrcBytes, kCacheOp>(
+                  dst_ptr + v, gmem_ptr, quant_args.iterator_local_scale.valid());
+        }
+        ++quant_args.iterator_local_scale;
+      }
+      ++this->smem_iterator_local_scale_;
+    }
+  }
+
+  CUTLASS_DEVICE
+  void advance_smem_write_stage(Arguments &quant_args) {
+    if (smem_write_stage_idx_ % kStagesPerLocalScaleLoad == 0) {
+      // Advance global iterators
+      quant_args.iterator_local_scale.add_pointer_offset(quant_args.local_scale_pointer_offset);
+
+      // Advance shared iterators
+      int smem_pointer_offset = IteratorLocalScale::Shape::kRow * IteratorLocalScale::Shape::kColumn;
+      smem_iterator_local_scale_.add_pointer_offset(smem_pointer_offset);
+    }
+
+    // Increment shared memory write stage index
+    ++smem_write_stage_idx_;
+
+    if (smem_write_stage_idx_ == kStagesPerLocalScaleLoad * kStages) {
+      // Wrap back around to the 'start' of the circular buffer in shared memory
+      int pointer_offset = - kStages * IteratorLocalScale::Shape::kRow * IteratorLocalScale::Shape::kColumn;
+      smem_iterator_local_scale_.add_pointer_offset(pointer_offset);
+      smem_write_stage_idx_ = 0;
+    }
+  }
+
+  CUTLASS_DEVICE
+  int advance_smem_read_stage() {
+    int byte_offset = 0;
+
+    ++smem_read_stage_idx_;
+
+    if (smem_read_stage_idx_ % kStagesPerLocalScaleLoad == 0) {
+      byte_offset = kLocalScaleRows * kSmemColumns;
+    }
+
+    if (smem_read_stage_idx_ == kStagesPerLocalScaleLoad * kStages) {
+      smem_read_stage_idx_ = 0;
+      byte_offset = - (kStages - 1) * kLocalScaleRows * kSmemColumns;
+    }
+
+    return byte_offset;
+  }
+
+  CUTLASS_DEVICE
+  int clear_mask(Arguments &quant_args, bool cond) {
+    quant_args.iterator_local_scale.clear_mask(cond);
+  }
+};
+
+}  // namespace threadblock
+}  // namespace gemm
+}  // namespace cutlass

custom_ops/gpu_ops/cutlass_extensions/gemm/threadblock/wint2x_tile_dequanter.h

@@ -1,130 +0,0 @@
-// Copyright (c) 2025 PaddlePaddle Authors. All Rights Reserved.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#pragma once
-
-#include "cutlass/gemm_coord.h"
-#include "cutlass/trace.h"
-
-#include "cutlass_extensions/gemm/threadblock/wint2x_unzip.h"
-
-namespace cutlass {
-namespace gemm {
-namespace threadblock {
-
-template <typename ElementT, typename ScaleElementT, int Rows, int Columns,
-          int Stages, int NumThreads, WintQuantMethod Method>
-struct TileDequanter {
-  using WeightQuantTraits = WintQuantTraits<ElementT, Method>;
-  using MmaElementT = typename WeightQuantTraits::MmaWeightType;
-  using QuantArguments = typename WeightQuantTraits::Arguments;
-
-  using UnzipAndDequantFunctor =
-      UnzipAndDequantFunctor<MmaElementT, Method, Rows, Columns, NumThreads>;
-
-  static constexpr bool kUseSharedMemory = true;
-
-  static constexpr int kRows = Rows;
-  static constexpr int kColumns = Columns;
-  static constexpr int kStages = Stages;
-
-  MmaElementT *out_smem_ptr{nullptr};
-
-  char *pointer{nullptr};
-  int64_t ldm{0};
-  cutlass::MatrixCoord tb_offset;
-  cutlass::MatrixCoord extent;
-
-  ScaleElementT *super_scale_ptr{nullptr};
-  cutlass::MatrixCoord tb_offset_scale;
-
-  QuantArguments quant_args;
-
-  int64_t block_start_rows[kStages];
-  bool need_preload{true};
-  UnzipAndDequantFunctor unzip_functor;
-
-  CUTLASS_DEVICE
-  TileDequanter(MmaElementT *out_smem_ptr, char *pointer, int64_t ldm,
-                const cutlass::MatrixCoord &extent,
-                const cutlass::MatrixCoord &tb_offset,
-                ScaleElementT *super_scale_ptr,
-                const cutlass::MatrixCoord &tb_offset_scale,
-                const QuantArguments &quant_args)
-      : out_smem_ptr(out_smem_ptr), pointer(pointer), ldm(ldm), extent(extent),
-        tb_offset(tb_offset), super_scale_ptr(super_scale_ptr),
-        tb_offset_scale(tb_offset_scale), quant_args(quant_args) {}
-
-  CUTLASS_DEVICE
-  MmaElementT *GetOutPtr() { return out_smem_ptr; }
-
-  CUTLASS_DEVICE
-  void AddTileOffset(const cutlass::MatrixCoord &tile_offset) {
-    tb_offset.row() += tile_offset.row() * kRows;
-    tb_offset.column() += tile_offset.column() * kColumns;
-    tb_offset_scale.column() += tile_offset.column() * kColumns;
-  }
-
-  CUTLASS_DEVICE
-  void Load(uint8_t *zipped_smem_ptr, uint8_t *column_wise_smem_ptr, int stage) {
-    int zipped_row = WeightQuantTraits::CaclPackedDim(tb_offset.row());
-    if (tb_offset.row() >= extent.row() ||
-        tb_offset.column() >= extent.column()) {
-      return;
-    }
-
-    block_start_rows[stage % kStages] = tb_offset.row();
-
-    using ZippedT = typename WeightQuantTraits::WeightType;
-    ZippedT *in_ptr = reinterpret_cast<ZippedT *>(pointer) + zipped_row * ldm +
-                      tb_offset.column();
-    ScaleElementT *scale_ptr = super_scale_ptr + tb_offset_scale.column();
-
-    if constexpr (Method == WintQuantMethod::kWeightOnlyInt2) {
-      const uint8_t *local_scale_ptr = quant_args.local_scale_ptr +
-                                       (tb_offset.row() / 128) * ldm +
-                                       tb_offset_scale.column();
-      const float *code_scale_ptr =
-          quant_args.code_scale_ptr + tb_offset_scale.column();
-      const float *code_zp_ptr =
-          quant_args.code_zp_ptr + tb_offset_scale.column();
-
-      typename UnzipAndDequantFunctor::Arguments args(zipped_smem_ptr, column_wise_smem_ptr);
-      unzip_functor.LoadAsync(in_ptr, local_scale_ptr, code_scale_ptr, code_zp_ptr,
-                              scale_ptr, &args, ldm, need_preload);
-      need_preload = false;
-    } else {
-      // CUTLASS_TRACE_DEVICE("Not Supported!");
-    }
-  }
-
-  CUTLASS_DEVICE
-  void UnpackAndDequant(uint8_t *zipped_smem_ptr, uint8_t *column_wise_smem_ptr, int stage) {
-    int64_t block_start_row = block_start_rows[stage % kStages];
-    if (block_start_row >= extent.row()) {
-      return;
-    }
-
-    if constexpr (Method == WintQuantMethod::kWeightOnlyInt2) {
-      typename UnzipAndDequantFunctor::Arguments args(zipped_smem_ptr, column_wise_smem_ptr);
-      unzip_functor.ComputeVectorized(args, out_smem_ptr, block_start_row);
-    } else {
-      // CUTLASS_TRACE_DEVICE("Not Supported!");
-    }
-  }
-};
-
-}  // namespace threadblock
-}  // namespace gemm
-}  // namespace cutlass

custom_ops/gpu_ops/cutlass_extensions/gemm/warp/default_mma_tensor_op.h

@@ -41,12 +41,9 @@
 #include "cutlass_extensions/arch/mma.h"
 #include "cutlass_extensions/gemm/warp/mma_tensorop_compute_B_with_f16.h"
 
-namespace cutlass
-{
-namespace gemm
-{
-namespace warp
-{
+namespace cutlass {
+namespace gemm {
+namespace warp {
 
 /////////////////////////////////////////////////////////////////////////////////////////////////
 
@@ -81,7 +78,7 @@ private:
     // Shape for computing the FP16s
     using ComputeInstructionShape = InstructionShape_;
 
-    // Chosen so we get K=16 for int8 and K=32 for int4.
+    // Chosen so we get K=16 for int8, K=32 for int4, K=64 for int2.
     static constexpr int LoadInstructionK = 128 / sizeof_bits<ElementB>::value;
 
     // Shape for loading the narrow data type from shared memory

custom_ops/gpu_ops/cutlass_extensions/gemm/warp/mma_tensorop_compute_B_with_f16.h

@@ -58,15 +58,12 @@
 
 /////////////////////////////////////////////////////////////////////////////////////////////////
 
-namespace cutlass
-{
-namespace gemm
-{
-namespace warp
-{
+namespace cutlass {
+namespace gemm {
+namespace warp {
 
 /////////////////////////////////////////////////////////////////////////////////////////////////
-/// Structure to compute the matrix product targeting CUDA cores and SIMT math instructions.
+/// Structure to compute the matrix product targeting Tensor Cores, for the case when A is floating point and B is quantized integer.
 template <
     /// Size of the Gemm problem - concept: gemm::GemmShape<>
     typename Shape_,
@@ -297,6 +294,235 @@ public:
     }
 };
 
+/////////////////////////////////////////////////////////////////////////////////////////////////
+/// Structure to compute the matrix product targeting Tensor Cores, for the case when A is floating point and B is quantized integer.
+/// Specialization for B of uint2b_t.
+template <
+    /// Size of the Gemm problem - concept: gemm::GemmShape<>
+    typename Shape_,
+    /// Data type of A elements
+    typename ElementA_,
+    /// Layout of A matrix (concept: MatrixLayout)
+    typename LayoutA_,
+    /// Layout of B matrix (concept: MatrixLayout)
+    typename LayoutB_,
+    /// Element type of C matrix
+    typename ElementC_,
+    /// Layout of C matrix (concept: MatrixLayout)
+    typename LayoutC_,
+    /// Policy describing warp-level MmaTensorOp (concept: MmaTensorOp policy)
+    typename Policy_,
+    /// Instruction shape to override shared memory iterators with
+    typename SharedMemoryInstructionShape_,
+    /// Number of partitions along K dimension
+    int PartitionsK_,
+    /// Store the accumulators in row major or column major.  Row major is used
+    /// when output layout is interleaved.
+    bool AccumulatorsInRowMajor>
+class MmaTensorOpComputeBWithF16<
+    Shape_,
+    ElementA_,
+    LayoutA_,
+    uint2b_t,
+    LayoutB_,
+    ElementC_,
+    LayoutC_,
+    Policy_,
+    SharedMemoryInstructionShape_,
+    PartitionsK_,
+    AccumulatorsInRowMajor>
+{
+public:
+    /// Shape of warp-level matrix operation (concept: GemmShape)
+    using Shape = Shape_;
+
+    /// Data type of multiplicand A
+    using ElementA = ElementA_;
+
+    /// Layout of multiplicand A
+    using LayoutA = LayoutA_;
+
+    /// Data type of multiplicand B
+    using ElementB = uint2b_t;
+
+    /// Layout of multiplicand B
+    using LayoutB = LayoutB_;
+
+    /// Data type of accumulator matrix C
+    using ElementC = ElementC_;
+
+    /// Layout of accumulator matrix C
+    using LayoutC = LayoutC_;
+
+    /// Shape of the warp in units of thread (concept: MmaLanePolicySimt)
+    using Policy = Policy_;
+
+    /// Underlying matrix multiply operator (concept: arch::Mma)
+    using ArchMmaOperator = typename Policy::Operator;
+
+    /// Indicates math operator
+    using MathOperator = typename ArchMmaOperator::Operator;
+
+    /// Architecture tag from underlying instruction
+    using ArchTag = typename ArchMmaOperator::ArchTag;
+    static_assert((platform::is_same<typename ArchMmaOperator::ElementA, half_t>::value
+                      && platform::is_same<typename ArchMmaOperator::ElementB, half_t>::value)
+            || (platform::is_same<typename ArchMmaOperator::ElementA, bfloat16_t>::value
+                && platform::is_same<typename ArchMmaOperator::ElementB, bfloat16_t>::value
+                && ArchTag::kMinComputeCapability >= 80),
+        "MmaTensorOpCvtBToA only supports underlying HMMA/QMMA");
+
+    static_assert(platform::is_same<ElementA, half_t>::value
+            || (platform::is_same<ElementA, bfloat16_t>::value && ArchTag::kMinComputeCapability >= 80),
+        "MmaTensorOpCvtBToA only supports Fp16 A or Bf16 A on Ampere+");
+
+    /// Indicates class of matrix operator
+    using OperatorClass = arch::OpClassTensorOp;
+
+    /// Shape of underlying instruction
+    using InstructionShape = typename ArchMmaOperator::Shape;
+
+    /// Instruction shape to override shared memory iterators with
+    using SharedMemoryInstructionShape = SharedMemoryInstructionShape_;
+
+    static_assert(
+        SharedMemoryInstructionShape::kM == InstructionShape::kM, "M dimension of compute instruction must match load");
+    static_assert(
+        SharedMemoryInstructionShape::kN == InstructionShape::kN, "N dimension of compute instruction must match load");
+
+    static constexpr int kExpansionFactor = SharedMemoryInstructionShape::kK / InstructionShape::kK;
+
+    static_assert(!(Shape::kK % SharedMemoryInstructionShape::kK), "");
+
+    /// Complex transform on A operand
+    static ComplexTransform const kTransformA = ComplexTransform::kNone;
+
+    /// Complex transform on B operand
+    static ComplexTransform const kTransformB = ComplexTransform::kNone;
+
+    /// Number of threads participating in warp-level matrix product
+    static int const kThreadCount = 32;
+
+    /// Number of partitions along K dimension
+    static int const kPartitionsK = PartitionsK_;
+
+public:
+    /// Iterates over the A operand in memory
+    using IteratorA
+        = MmaTensorOpMultiplicandTileIterator<MatrixShape<Shape::kM, Shape::kK>, Operand::kA, ElementA, LayoutA,
+            MatrixShape<InstructionShape::kM, InstructionShape::kK>, Policy::OpDelta::kRow, kThreadCount, kPartitionsK>;
+
+    /// Storage for A tile
+    using FragmentA = typename IteratorA::Fragment;
+
+    /// Storage for transformed A tile
+    using TransformedFragmentA = Array<typename ArchMmaOperator::ElementA, FragmentA::kElements>;
+
+    /// Iterates over the B operand in memory
+    using IteratorB = MmaTensorOpMultiplicandTileIterator<MatrixShape<Shape::kK, Shape::kN>, Operand::kB, ElementB,
+        LayoutB, MatrixShape<SharedMemoryInstructionShape::kK, InstructionShape::kN>, Policy::OpDelta::kRow,
+        kThreadCount, kPartitionsK>;
+
+    /// Storage for B tile
+    using FragmentB = typename IteratorB::Fragment;
+
+    /// Storage for transformed B tile
+    using TransformedFragmentB =
+        Array<typename ArchMmaOperator::ElementB, FragmentB::kElements / kExpansionFactor>;
+
+    /// Iterates over the C operand in memory
+    using IteratorC = MmaTensorOpAccumulatorTileIterator<MatrixShape<Shape::kM, Shape::kN>, ElementC, LayoutC,
+        typename ArchMmaOperator::Shape, typename Policy::OpDelta>;
+
+    /// Storage for C tile
+    using FragmentC = typename IteratorC::Fragment;
+
+    /// Number of mma operations performed
+    using MmaIterations = MatrixShape<(Shape::kM + ArchMmaOperator::Shape::kM - 1) / ArchMmaOperator::Shape::kM,
+        (Shape::kN + ArchMmaOperator::Shape::kN - 1) / ArchMmaOperator::Shape::kN>;
+
+public:
+    /// Underlying matrix multiply operator (concept: arch::Mma)
+    ArchMmaOperator mma;
+
+public:
+    //
+    // Methods
+    //
+
+    /// Ctor
+    CUTLASS_DEVICE
+    MmaTensorOpComputeBWithF16() {}
+
+    /// Performs a warp-level matrix multiply-accumulate operation
+    CUTLASS_DEVICE
+    void operator()(FragmentC& D, TransformedFragmentA const& A, TransformedFragmentB const& B, FragmentC const& C) const
+    {
+
+        using MmaOperandA = typename ArchMmaOperator::FragmentA;
+        using MmaOperandB = typename ArchMmaOperator::FragmentB;
+        using MmaOperandC = typename ArchMmaOperator::FragmentC;
+
+        D = C;
+
+        MmaOperandA const* ptr_A = reinterpret_cast<MmaOperandA const*>(&A);
+        MmaOperandB const* ptr_B = reinterpret_cast<MmaOperandB const*>(&B);
+        MmaOperandC* ptr_D = reinterpret_cast<MmaOperandC*>(&D);
+
+#if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ < 800)
+        // Serpentine visitation order maximizing reuse of Rb
+        CUTLASS_PRAGMA_UNROLL
+        for (int n = 0; n < MmaIterations::kColumn; ++n)
+        {
+
+            CUTLASS_PRAGMA_UNROLL
+            for (int m = 0; m < MmaIterations::kRow; ++m)
+            {
+
+                int m_serpentine = ((n % 2) ? (MmaIterations::kRow - 1 - m) : m);
+
+                if (AccumulatorsInRowMajor)
+                { // matrix B is reordered
+                    mma(ptr_D[n + m_serpentine * MmaIterations::kColumn], ptr_A[m_serpentine], ptr_B[n],
+                        ptr_D[n + m_serpentine * MmaIterations::kColumn]);
+                }
+                else
+                {
+                    mma(ptr_D[m_serpentine + n * MmaIterations::kRow], ptr_A[m_serpentine], ptr_B[n],
+                        ptr_D[m_serpentine + n * MmaIterations::kRow]);
+                }
+            }
+        }
+#elif defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 800)
+        // Serpentine visitation order maximizing reuse of Ra
+        CUTLASS_PRAGMA_UNROLL
+        for (int m = 0; m < MmaIterations::kRow; ++m)
+        {
+
+            CUTLASS_PRAGMA_UNROLL
+            for (int n = 0; n < MmaIterations::kColumn; ++n)
+            {
+
+                int n_serpentine = ((m % 2) ? (MmaIterations::kColumn - 1 - n) : n);
+
+                if (AccumulatorsInRowMajor)
+                { // matrix B is reordered
+                    mma(ptr_D[n_serpentine + m * MmaIterations::kColumn], ptr_A[m], ptr_B[n_serpentine],
+                        ptr_D[n_serpentine + m * MmaIterations::kColumn]);
+                }
+                else
+                {
+                    mma(ptr_D[m + n_serpentine * MmaIterations::kRow], ptr_A[m], ptr_B[n_serpentine],
+                        ptr_D[m + n_serpentine * MmaIterations::kRow]);
+                }
+            }
+        }
+#else
+        assert(0);
+#endif
+    }
+};
+
 /////////////////////////////////////////////////////////////////////////////////////////////////
 
 } // namespace warp

custom_ops/gpu_ops/cutlass_extensions/gemm/warp/mma_tensorop_wint2x_dequantizer.h

@@ -0,0 +1,442 @@
+/***************************************************************************************************
+ * Copyright (c) 2017 - 2022 NVIDIA CORPORATION & AFFILIATES. All rights
+ *reserved. SPDX-License-Identifier: BSD-3-Clause
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions are met:
+ *
+ * 1. Redistributions of source code must retain the above copyright notice,
+ *this list of conditions and the following disclaimer.
+ *
+ * 2. Redistributions in binary form must reproduce the above copyright notice,
+ * this list of conditions and the following disclaimer in the documentation
+ * and/or other materials provided with the distribution.
+ *
+ * 3. Neither the name of the copyright holder nor the names of its
+ * contributors may be used to endorse or promote products derived from
+ * this software without specific prior written permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+ * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+ * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ *ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE
+ *LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+ *CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+ *SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+ *INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+ *CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+ *ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+ *POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+/*! \file
+  \brief Defines iterators used by warp-level matrix multiply operations
+  targeting Tensor Cores.
+*/
+
+#pragma once
+
+#include "cutlass/cutlass.h"
+
+#include "cutlass/array.h"
+#include "cutlass/matrix_shape.h"
+#include "cutlass/numeric_types.h"
+#include "cutlass/tensor_ref.h"
+#include "cutlass/arch/arch.h"
+#include "cutlass/arch/memory_sm75.h"
+#include "cutlass/gemm/gemm.h"
+#include "cutlass/layout/matrix.h"
+#include "cutlass/layout/pitch_linear.h"
+#include "cutlass/layout/tensor.h"
+#include "cutlass/functional.h"
+#include "cutlass/platform/platform.h"
+
+#include "cutlass_extensions/interleaved_numeric_conversion.h"
+
+namespace cutlass {
+namespace gemm {
+namespace warp {
+
+namespace detail {
+
+template <typename T>
+struct DataTypeTraits;
+
+template <>
+struct DataTypeTraits<bfloat16_t> {
+    using Type = __nv_bfloat16;
+    using DualType = __nv_bfloat162;
+};
+
+template <>
+struct DataTypeTraits<half_t> {
+    using Type = __half;
+    using DualType = __half2;
+};
+
+template <typename T, int N, typename Enable = void>
+struct LocalScaleConverter {
+    using FragmentSource = Array<uint8_t, N>;
+    using FragmentResult = Array<T, N>;
+
+    CUTLASS_DEVICE
+    static void Apply(FragmentSource const& local_scale_frag,
+                      FragmentResult const& super_scale_frag,
+                      FragmentResult& scale_frag,
+                      int shift_bit) {
+        constexpr uint32_t kLocalScaleMask = 0xf;
+
+        CUTLASS_PRAGMA_UNROLL
+        for (int i = 0; i < N; ++i) {
+            int32_t shifted_value = (static_cast<int32_t>(local_scale_frag[i]) >> shift_bit) & kLocalScaleMask;
+            scale_frag[i] = static_cast<T>(shifted_value) * super_scale_frag[i];
+        }
+    }
+};
+
+template <int N>
+struct LocalScaleConverter<half_t, N, typename platform::enable_if<N % 4 == 0>::type> {
+    using FragmentSource = Array<uint8_t, N>;
+    using FragmentResult = Array<half_t, N>;
+
+    CUTLASS_DEVICE
+    static void Apply(FragmentSource const& local_scale_frag,
+                      FragmentResult const& super_scale_frag,
+                      FragmentResult& scale_frag,
+                      int shift_bit) {
+        constexpr uint32_t immLut = (0xf0 & 0xcc) | 0xaa;
+        constexpr uint32_t MASK = 0x000f000f;
+        // 2^10 = 1024
+        constexpr uint32_t I4s_TO_FP16s_MAGIC_NUM = 0x64006400;
+
+        // -2^10 = -1024
+        constexpr uint32_t FP16_BIAS = 0xE400E400;
+        // 1.0
+        constexpr uint32_t FP16_ONE = 0x3C003C00;
+
+        __half2* scale_ptr = reinterpret_cast<__half2 *>(&scale_frag);
+        __half2 const* super_scale_ptr = reinterpret_cast<__half2 const*>(&super_scale_frag);
+
+        uint32_t const* local_scale_ptr = reinterpret_cast<uint32_t const*>(&local_scale_frag);
+
+        CUTLASS_PRAGMA_UNROLL
+        for (int i = 0; i < N / 4; ++i) {
+            int i4s = local_scale_ptr[i] >> shift_bit;
+
+            // unpack: 0, 1
+            int32_t low = __byte_perm(i4s, i4s, 0xF1F0);
+            int32_t unpack0 = lop3<immLut>(low, MASK, I4s_TO_FP16s_MAGIC_NUM);
+            // unpack: 2, 3
+            int32_t high = __byte_perm(i4s, i4s, 0xF3F2);
+            int32_t unpack1 = lop3<immLut>(high, MASK, I4s_TO_FP16s_MAGIC_NUM);
+
+            __half2 scale0 = __hfma2(*reinterpret_cast<__half2*>(&unpack0),
+                                     *reinterpret_cast<const __half2*>(&FP16_ONE),
+                                     *reinterpret_cast<const __half2*>(&FP16_BIAS));
+            __half2 scale1 = __hfma2(*reinterpret_cast<__half2*>(&unpack1),
+                                     *reinterpret_cast<const __half2*>(&FP16_ONE),
+                                     *reinterpret_cast<const __half2*>(&FP16_BIAS));
+
+            scale_ptr[2 * i] = __hmul2(scale0, super_scale_ptr[2 * i]);
+            scale_ptr[2 * i + 1] = __hmul2(scale1, super_scale_ptr[2 * i + 1]);
+        }
+    }
+};
+
+template <int N>
+struct LocalScaleConverter<bfloat16_t, N, typename platform::enable_if<N % 4 == 0>::type> {
+    using FragmentSource = Array<uint8_t, N>;
+    using FragmentResult = Array<bfloat16_t, N>;
+
+    CUTLASS_DEVICE
+    static void Apply(FragmentSource const& local_scale_frag,
+                      FragmentResult const& super_scale_frag,
+                      FragmentResult& scale_frag,
+                      int shift_bit) {
+#if (defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 800) && defined(ENABLE_BF16))
+        constexpr uint32_t immLut = (0xF0 & 0xCC) | 0xAA;
+        constexpr uint32_t MASK = 0x000F000F;
+        constexpr uint32_t I4s_TO_BF16s_MAGIC_NUM = 0x43004300;
+
+        constexpr uint32_t BF16_BIAS = 0xC300C300;
+        constexpr uint32_t BF16_ONE = 0x3F803F80;
+
+        __nv_bfloat162* scale_ptr = reinterpret_cast<__nv_bfloat162 *>(&scale_frag);
+        __nv_bfloat162 const* super_scale_ptr = reinterpret_cast<__nv_bfloat162 const*>(&super_scale_frag);
+
+        uint32_t const* local_scale_ptr = reinterpret_cast<uint32_t const*>(&local_scale_frag);
+
+        CUTLASS_PRAGMA_UNROLL
+        for (int i = 0; i < N / 4; ++i) {
+            int i4s = local_scale_ptr[i] >> shift_bit;
+
+            // unpack: 0, 1
+            int32_t low = __byte_perm(i4s, i4s, 0xF1F0);
+            int32_t unpack0 = lop3<immLut>(low, MASK, I4s_TO_BF16s_MAGIC_NUM);
+            // unpack: 2, 3
+            int32_t high = __byte_perm(i4s, i4s, 0xF3F2);
+            int32_t unpack1 = lop3<immLut>(high, MASK, I4s_TO_BF16s_MAGIC_NUM);
+
+            nv_bfloat162 scale0 = __hfma2(*reinterpret_cast<nv_bfloat162*>(&unpack0),
+                                          *reinterpret_cast<const nv_bfloat162*>(&BF16_ONE),
+                                          *reinterpret_cast<const nv_bfloat162*>(&BF16_BIAS));
+            nv_bfloat162 scale1 = __hfma2(*reinterpret_cast<nv_bfloat162*>(&unpack1),
+                                          *reinterpret_cast<const nv_bfloat162*>(&BF16_ONE),
+                                          *reinterpret_cast<const nv_bfloat162*>(&BF16_BIAS));
+
+            scale_ptr[2 * i] = __hmul2(scale0, super_scale_ptr[2 * i]);
+            scale_ptr[2 * i + 1] = __hmul2(scale1, super_scale_ptr[2 * i + 1]);
+        }
+#else
+        // Slow path not implemented here on purpose. If we need to do HMMA on older arch, scale conversion should
+        // happen before scales are stored to shared memory and we should use the fp16 dequantizer. This will avoid
+        // numerous conversion instructions in GEMM main loop.
+        arch::device_breakpoint();
+#endif
+    }
+};
+
+} // namespace detail
+
+////////////////////////////////////////////////////////////////////////////////
+
+template <
+    /// Matrix multiply operator
+    typename MmaOperator_,
+    /// Size of the matrix to load (concept: MatrixShape)
+    typename Shape_,
+    /// Operand identity
+    Operand Operand,
+    /// Data type of Scale elements
+    typename ElementOperand_,
+    /// Layout of operand
+    typename Layout_,
+    /// Group size for quantization
+    int GroupSize_,
+    ///
+    typename Enable = void>
+class MmaTensorOpWin2xDequantizer {
+    //static_assert(false, "Not Supported!");
+};
+
+////////////////////////////////////////////////////////////////////////////////
+// Bfloat specialization for Ampere
+template <
+    /// Underlying matrix multiply operator (concept: MmaTensorOp)
+    typename MmaOperator_,
+    /// Shape of the warp level matrix multiply (concept: GemmShape)
+    typename Shape_,
+    /// Data type of Scale elements
+    typename ElementOperand_,
+    /// Group size for quantization
+    int GroupSize_>
+class MmaTensorOpWin2xDequantizer<
+    MmaOperator_,
+    Shape_,
+    Operand::kB,
+    ElementOperand_,
+    layout::RowMajor,
+    GroupSize_>
+    //typename platform::enable_if<MmaOperator_::ArchTag::kMinComputeCapability >= 80
+    //    && platform::is_same<typename MmaOperator_::ArchMmaOperator::LayoutB, layout::ColumnMajor>::value>::type>
+{
+public:
+    static_assert(platform::is_same<ElementOperand_, half_t>::value || platform::is_same<ElementOperand_, bfloat16_t>::value,
+        "T must be fp16 or bf16");
+
+    /// Mma Operator
+    using MmaOperator = MmaOperator_;
+
+    // The architecture specific mma ooperator being used
+    using ArchMmaOperator = typename MmaOperator::ArchMmaOperator;
+
+    // Mma Instruction Shape
+    using InstructionShape = typename ArchMmaOperator::Shape;
+
+    /// Warp mma shape
+    using Shape = Shape_;
+
+    /// Type of mma operand
+    using ElementOperand = ElementOperand_;
+
+    /// Layout of the scales in shared memory
+    using Layout = layout::RowMajor;
+
+    /// Group size for quantization
+    static constexpr int kGroupSize = GroupSize_;
+
+    /// Type of input
+    using ElementB = typename MmaOperator::FragmentB::Element;
+    static_assert(platform::is_same<ElementB, uint2b_t>::value, "ElementB must be uint2b_t");
+
+    /// Type of the scales
+    using ElementLocalScale = uint4b_t;
+    using ElementSuperScale = ElementOperand;
+    using ElementCodeScaleZp = float;
+
+    // Fragment to hold scale data to apply to B before mma
+    // We need 1 fp16 per matrix iteration in the N dimension
+    static constexpr int kWarpIterationsAlongN = MmaOperator::MmaIterations::kColumn;
+
+    // use uint8_t to save 2 4-bits local scales
+    using FragmentLocalScale = Array<uint8_t, kWarpIterationsAlongN>;
+    using FragmentSuperScale = Array<ElementSuperScale, kWarpIterationsAlongN>;
+    using FragmentCodeScaleZp = Array<ElementCodeScaleZp, kWarpIterationsAlongN>;
+
+    /// Fragment to hold B data before Mma
+    using FragmentInput = Array<ElementB, MmaOperator::FragmentB::kElements>;
+
+    // This is the ratio of the load instruction vs the compute instruction.
+    static constexpr int kExpansionFactor = MmaOperator::IteratorB::InstructionShape::kRow / InstructionShape::kK;
+
+    static constexpr int kNumPacks = sizeof_bits<uint8_t>::value / sizeof_bits<ElementB>::value;
+    static constexpr int kUnpackFactor = MmaOperator::FragmentB::kElements / (kWarpIterationsAlongN * kNumPacks);
+    static constexpr int kUnpackInterval = kExpansionFactor / kUnpackFactor;
+
+    /// Unpack 4 uint2b_t values compreseed in a uint8_t to floating points.
+    using Uint2Converter = FastInterleavedAndBiasedNumericArrayConverter<
+        ElementOperand, ElementB, MmaOperator::FragmentB::kElements / kUnpackFactor>;
+    using FragmentInputUnpack = typename Uint2Converter::result_type;
+
+    /// Fragment to hold internal scales before Mma
+    using FragmentScale = Array<ElementOperand, FragmentLocalScale::kElements>;
+
+    /// Fragment of dequantized B
+    using FragmentOutput = Array<ElementOperand, MmaOperator::FragmentB::kElements / kExpansionFactor>;
+
+    /// TensorRef type for loading element from a tensor
+    using SuperTensorRef = cutlass::TensorRef<ElementSuperScale, Layout>;
+    using LocalTensorRef = cutlass::TensorRef<ElementLocalScale, Layout>;
+    using CodeTensorRef = cutlass::TensorRef<ElementCodeScaleZp, Layout>;
+
+private:
+    //
+    // Data members
+    //
+
+    uint8_t* pointer_local_scale_;
+    ElementCodeScaleZp* pointer_code_scale_;
+    ElementCodeScaleZp* pointer_code_zp_;
+    ElementSuperScale* pointer_super_scale_;
+
+    //FragmentInputUnpack unpacked_frag_;
+    FragmentScale scale_frag_;
+
+public:
+    CUTLASS_DEVICE
+    MmaTensorOpWin2xDequantizer(SuperTensorRef smem_super_scale,
+                                LocalTensorRef smem_local_scale,
+                                CodeTensorRef smem_code_scale,
+                                CodeTensorRef smem_code_zp,
+                                int warp_idx_n,
+                                int lane_idx) {
+        int warp_offset = warp_idx_n * Shape::kN;
+        int quad = lane_idx / 4;
+        int thread_offset = warp_offset + quad;
+        pointer_super_scale_ = smem_super_scale.data() + thread_offset;
+        pointer_code_scale_ = smem_code_scale.data() + thread_offset;
+        pointer_code_zp_ = smem_code_zp.data() + thread_offset;
+        pointer_local_scale_ = reinterpret_cast<uint8_t *>(smem_local_scale.data()) + thread_offset;
+    }
+
+    /// Channel-wise params, need to load just once
+    CUTLASS_DEVICE
+    void load(FragmentCodeScaleZp& code_scale_frag,
+              FragmentCodeScaleZp& code_zp_frag,
+              FragmentSuperScale& super_scale_frag) {
+        CUTLASS_PRAGMA_UNROLL
+        for (int mma_n_iter = 0; mma_n_iter < kWarpIterationsAlongN; ++mma_n_iter) {
+            super_scale_frag[mma_n_iter] = pointer_super_scale_[mma_n_iter * InstructionShape::kN]; // bank conflict
+            code_scale_frag[mma_n_iter] = pointer_code_scale_[mma_n_iter * InstructionShape::kN];
+            code_zp_frag[mma_n_iter] = pointer_code_zp_[mma_n_iter * InstructionShape::kN];
+        }
+    }
+
+    /// Group-wise params, need to load multiple times
+    CUTLASS_DEVICE
+    void load(FragmentLocalScale& local_scale_frag) {
+        CUTLASS_PRAGMA_UNROLL
+        for (int mma_n_iter = 0; mma_n_iter < kWarpIterationsAlongN; ++mma_n_iter) {
+            local_scale_frag[mma_n_iter] = pointer_local_scale_[mma_n_iter * InstructionShape::kN]; // bank conflict
+        }
+    }
+
+    CUTLASS_DEVICE
+    void dequantize(const FragmentLocalScale& local_scale_frag,
+                    const FragmentCodeScaleZp& code_scale_frag,
+                    const FragmentCodeScaleZp& code_zp_frag,
+                    const FragmentSuperScale& super_scale_frag,
+                    const FragmentInput& input_frag,
+                    FragmentOutput& output_frag,
+                    int tb_offset_k,
+                    int warp_k_compute_offset) {
+        if constexpr (kUnpackInterval != 1) {
+            // unsupport now
+            arch::device_breakpoint();
+        }
+
+        typename Uint2Converter::source_type source_frag;
+
+        int in_offset = warp_k_compute_offset * kUnpackInterval;
+
+        uint8_t const* ptr_input = reinterpret_cast<uint8_t const*>(&input_frag);
+        uint8_t* ptr_source = reinterpret_cast<uint8_t *>(&source_frag);
+
+        CUTLASS_PRAGMA_UNROLL
+        for (int mma_n_iter = 0; mma_n_iter < kWarpIterationsAlongN; ++mma_n_iter) {
+            ptr_source[mma_n_iter] = ptr_input[mma_n_iter * kUnpackFactor + in_offset];
+        }
+        FragmentInputUnpack unpacked_frag = Uint2Converter::convert(source_frag, code_scale_frag, code_zp_frag);
+
+        // dequantize local_scale
+        if (warp_k_compute_offset == 0) {
+            using LocalScaleConverter = detail::LocalScaleConverter<ElementOperand, FragmentLocalScale::kElements>;
+
+            // special for TileRows = 64
+            int local_scale_shift = (((tb_offset_k / kGroupSize) + 1) & 1) * 4;
+            LocalScaleConverter::Apply(local_scale_frag, super_scale_frag, scale_frag_, local_scale_shift);
+        }
+
+        // unscale
+        // After applying LOP3 optimizations for performance, the B operand requires data rearrangement.
+        // reorder: [0, 4, 1, 5, 2, 6, 3, 7, 8, 12, 9, 13, 10, 14, 11, 15]
+        const int kWarpIterationsAlongK = FragmentOutput::kElements / kWarpIterationsAlongN;
+
+        using Type = typename detail::DataTypeTraits<ElementOperand>::Type;
+        using DualType = typename detail::DataTypeTraits<ElementOperand>::DualType;
+
+        Type* output_ptr = reinterpret_cast<Type *>(&output_frag);
+        DualType const* unpacked_ptr = reinterpret_cast<DualType const*>(&unpacked_frag);
+        DualType const* scale_ptr = reinterpret_cast<DualType const*>(&scale_frag_);
+
+        CUTLASS_PRAGMA_UNROLL
+        for (int mma_n_iter = 0; mma_n_iter < kWarpIterationsAlongN; mma_n_iter += 2) {
+            int mapped_idx_base = (mma_n_iter / 2) * kWarpIterationsAlongK;
+
+            DualType scalex2 = scale_ptr[mma_n_iter / 2];
+
+            CUTLASS_PRAGMA_UNROLL
+            for (int mma_k_iter = 0; mma_k_iter < kWarpIterationsAlongK; ++mma_k_iter) {
+                DualType unpacked_valuex2 = unpacked_ptr[mapped_idx_base + mma_k_iter];
+                DualType scaled_value = __hmul2(unpacked_valuex2, scalex2);
+                output_ptr[mma_n_iter * kWarpIterationsAlongK + mma_k_iter] = scaled_value.x;
+                output_ptr[(mma_n_iter + 1) * kWarpIterationsAlongK + mma_k_iter] = scaled_value.y;
+            }
+        }
+    }
+
+    /// Add an offset to pointer in units of elements.
+    /// Only group-wise params needs.
+    CUTLASS_DEVICE
+    void add_pointer_offset(int64_t const& offset) {
+        pointer_local_scale_ += offset;
+    }
+};
+
+////////////////////////////////////////////////////////////////////////////////
+
+}  // namespace warp
+}  // namespace gemm
+}  // namespace cutlass
+
+////////////////////////////////////////////////////////////////////////////////

custom_ops/gpu_ops/cutlass_extensions/interleaved_numeric_conversion.h

@@ -39,18 +39,25 @@
 #include "cutlass/array.h"
 #include "cutlass/half.h"
 #include "cutlass/numeric_types.h"
+#include "cutlass/trace.h"
 
-namespace cutlass
-{
+namespace cutlass {
+
+template <int lut>
+__device__ inline int lop3(int a, int b, int c) {
+  int res;
+  asm volatile("lop3.b32 %0, %1, %2, %3, %4;\n"
+               : "=r"(res)
+               : "r"(a), "r"(b), "r"(c), "n"(lut));
+  return res;
+}
 
 // This converter is meant to be used with data interleaved in a 32-bit register where the even elements are in the low
 // bits and the odd elemeents are in the high bits of the register. In addition, it assumes elements were originally
 // signed and had a bias of 2**(b-1) added (where b is the number of bits in the type) to make all numbers unsigned.
 // This converter will uninterleave the data and subtract the bias while converting to the result type.
 template <typename T, typename S, int N>
-struct FastInterleavedAndBiasedNumericArrayConverter
-{
-};
+struct FastInterleavedAndBiasedNumericArrayConverter;
 
 template <>
 struct FastInterleavedAndBiasedNumericArrayConverter<half_t, uint8_t, 4>
@@ -440,6 +447,329 @@ struct FastInterleavedAndBiasedNumericArrayConverter<bfloat16_t, uint4b_t, N>
     }
 };
 
+template <>
+struct FastInterleavedAndBiasedNumericArrayConverter<half_t, uint2b_t, 16>
+{
+    using result_type = Array<half_t, 16>;
+    using source_type = Array<uint2b_t, 16>;
+
+    using ScaleComputeT = float;
+    using code_type = Array<ScaleComputeT, 4>;
+
+    CUTLASS_DEVICE
+    static result_type convert(source_type const& source, ScaleComputeT code_scale, ScaleComputeT code_zp)
+    {
+        uint32_t const i8s = reinterpret_cast<uint32_t const&>(source);
+
+        // 2^23 = 8388608
+        static constexpr uint32_t FP32_BASE = 0x4B000000;
+
+        float fp32_intermediates[4];
+        uint32_t* fp32_intermediates_casted = reinterpret_cast<uint32_t*>(fp32_intermediates);
+        fp32_intermediates_casted[0] = __byte_perm(i8s, FP32_BASE, 0x7650);
+        fp32_intermediates_casted[1] = __byte_perm(i8s, FP32_BASE, 0x7651);
+        fp32_intermediates_casted[2] = __byte_perm(i8s, FP32_BASE, 0x7652);
+        fp32_intermediates_casted[3] = __byte_perm(i8s, FP32_BASE, 0x7653);
+
+        asm volatile("sub.f32 %0, %1, %2;\n" : "=r"(fp32_intermediates_casted[0]) : "r"(fp32_intermediates_casted[0]), "r"(FP32_BASE));
+        asm volatile("sub.f32 %0, %1, %2;\n" : "=r"(fp32_intermediates_casted[1]) : "r"(fp32_intermediates_casted[1]), "r"(FP32_BASE));
+        asm volatile("sub.f32 %0, %1, %2;\n" : "=r"(fp32_intermediates_casted[2]) : "r"(fp32_intermediates_casted[2]), "r"(FP32_BASE));
+        asm volatile("sub.f32 %0, %1, %2;\n" : "=r"(fp32_intermediates_casted[3]) : "r"(fp32_intermediates_casted[3]), "r"(FP32_BASE));
+
+        int32_t decode_value[4];
+        ScaleComputeT new_code_zp = code_zp + 0.5f;
+
+        decode_value[0] = __float2int_rd(fmaf(fp32_intermediates[0], code_scale, new_code_zp));
+        decode_value[1] = __float2int_rd(fmaf(fp32_intermediates[1], code_scale, new_code_zp));
+        decode_value[2] = __float2int_rd(fmaf(fp32_intermediates[2], code_scale, new_code_zp));
+        decode_value[3] = __float2int_rd(fmaf(fp32_intermediates[3], code_scale, new_code_zp));
+
+        return convert_impl(decode_value);
+    }
+
+    CUTLASS_DEVICE
+    static result_type convert(source_type const& source, code_type const& code_scale, code_type const& code_zp)
+    {
+        uint32_t const i8s = reinterpret_cast<uint32_t const&>(source);
+
+        // 2^23 = 8388608
+        static constexpr uint32_t FP32_BASE = 0x4B000000;
+
+        float fp32_intermediates[4];
+        uint32_t* fp32_intermediates_casted = reinterpret_cast<uint32_t*>(fp32_intermediates);
+        fp32_intermediates_casted[0] = __byte_perm(i8s, FP32_BASE, 0x7650);
+        fp32_intermediates_casted[1] = __byte_perm(i8s, FP32_BASE, 0x7651);
+        fp32_intermediates_casted[2] = __byte_perm(i8s, FP32_BASE, 0x7652);
+        fp32_intermediates_casted[3] = __byte_perm(i8s, FP32_BASE, 0x7653);
+
+        asm volatile("sub.f32 %0, %1, %2;\n" : "=r"(fp32_intermediates_casted[0]) : "r"(fp32_intermediates_casted[0]), "r"(FP32_BASE));
+        asm volatile("sub.f32 %0, %1, %2;\n" : "=r"(fp32_intermediates_casted[1]) : "r"(fp32_intermediates_casted[1]), "r"(FP32_BASE));
+        asm volatile("sub.f32 %0, %1, %2;\n" : "=r"(fp32_intermediates_casted[2]) : "r"(fp32_intermediates_casted[2]), "r"(FP32_BASE));
+        asm volatile("sub.f32 %0, %1, %2;\n" : "=r"(fp32_intermediates_casted[3]) : "r"(fp32_intermediates_casted[3]), "r"(FP32_BASE));
+
+        int32_t decode_value[4];
+
+        decode_value[0] = __float2int_rd(fmaf(fp32_intermediates[0], code_scale[0], code_zp[0] + 0.5f));
+        decode_value[1] = __float2int_rd(fmaf(fp32_intermediates[1], code_scale[1], code_zp[1] + 0.5f));
+        decode_value[2] = __float2int_rd(fmaf(fp32_intermediates[2], code_scale[2], code_zp[2] + 0.5f));
+        decode_value[3] = __float2int_rd(fmaf(fp32_intermediates[3], code_scale[3], code_zp[3] + 0.5f));
+
+        return convert_impl(decode_value);
+    }
+
+    CUTLASS_DEVICE
+    static result_type convert_impl(int32_t* decode_value)
+    {
+        result_type result;
+        static constexpr uint32_t immLut = (0xF0 & 0xCC) | 0xAA;
+
+        static constexpr uint32_t MASK = 0x003F003F;
+        // 2^10 = 1024
+        static constexpr uint32_t EX = 0x64006400;
+
+        uint32_t* h = reinterpret_cast<uint32_t*>(&result);
+
+        int32_t q0 = __byte_perm(decode_value[0], decode_value[1], 0x5410);
+        int32_t q1 = __byte_perm(decode_value[2], decode_value[3], 0x5410);
+
+        h[0] = lop3<immLut>(q0 >> 9, MASK, EX);
+        h[1] = lop3<immLut>(q0 >> 6, MASK, EX);
+        h[2] = lop3<immLut>(q0 >> 3, MASK, EX);
+        h[3] = lop3<immLut>(q0, MASK, EX);
+
+        h[4] = lop3<immLut>(q1 >> 9, MASK, EX);
+        h[5] = lop3<immLut>(q1 >> 6, MASK, EX);
+        h[6] = lop3<immLut>(q1 >> 3, MASK, EX);
+        h[7] = lop3<immLut>(q1, MASK, EX);
+
+        // 1024 + 32 = 1056
+        static constexpr uint32_t SUB = 0x64206420;
+
+        asm volatile("sub.f16x2 %0, %1, %2;\n" : "=r"(h[0]) : "r"(h[0]), "r"(SUB));
+        asm volatile("sub.f16x2 %0, %1, %2;\n" : "=r"(h[1]) : "r"(h[1]), "r"(SUB));
+        asm volatile("sub.f16x2 %0, %1, %2;\n" : "=r"(h[2]) : "r"(h[2]), "r"(SUB));
+        asm volatile("sub.f16x2 %0, %1, %2;\n" : "=r"(h[3]) : "r"(h[3]), "r"(SUB));
+
+        asm volatile("sub.f16x2 %0, %1, %2;\n" : "=r"(h[4]) : "r"(h[4]), "r"(SUB));
+        asm volatile("sub.f16x2 %0, %1, %2;\n" : "=r"(h[5]) : "r"(h[5]), "r"(SUB));
+        asm volatile("sub.f16x2 %0, %1, %2;\n" : "=r"(h[6]) : "r"(h[6]), "r"(SUB));
+        asm volatile("sub.f16x2 %0, %1, %2;\n" : "=r"(h[7]) : "r"(h[7]), "r"(SUB));
+
+        return result;
+    }
+
+    CUTLASS_DEVICE
+    result_type operator()(source_type const& s, ScaleComputeT code_scale, ScaleComputeT code_zp)
+    {
+        return convert(s, code_scale, code_zp);
+    }
+};
+
+template <>
+struct FastInterleavedAndBiasedNumericArrayConverter<bfloat16_t, uint2b_t, 16>
+{
+    using result_type = Array<bfloat16_t, 16>;
+    using source_type = Array<uint2b_t, 16>;
+
+    using ScaleComputeT = float;
+    using code_type = Array<ScaleComputeT, 4>;
+
+    CUTLASS_DEVICE
+    static result_type convert(source_type const& source, ScaleComputeT code_scale, ScaleComputeT code_zp)
+    {
+        uint32_t const i8s = reinterpret_cast<uint32_t const&>(source);
+
+        // 2^23 = 8388608
+        static constexpr uint32_t FP32_BASE = 0x4B000000;
+
+        float fp32_intermediates[4];
+        uint32_t* fp32_intermediates_casted = reinterpret_cast<uint32_t*>(fp32_intermediates);
+        fp32_intermediates_casted[0] = __byte_perm(i8s, FP32_BASE, 0x7650);
+        fp32_intermediates_casted[1] = __byte_perm(i8s, FP32_BASE, 0x7651);
+        fp32_intermediates_casted[2] = __byte_perm(i8s, FP32_BASE, 0x7652);
+        fp32_intermediates_casted[3] = __byte_perm(i8s, FP32_BASE, 0x7653);
+
+        asm volatile("sub.f32 %0, %1, %2;\n" : "=r"(fp32_intermediates_casted[0]) : "r"(fp32_intermediates_casted[0]), "r"(FP32_BASE));
+        asm volatile("sub.f32 %0, %1, %2;\n" : "=r"(fp32_intermediates_casted[1]) : "r"(fp32_intermediates_casted[1]), "r"(FP32_BASE));
+        asm volatile("sub.f32 %0, %1, %2;\n" : "=r"(fp32_intermediates_casted[2]) : "r"(fp32_intermediates_casted[2]), "r"(FP32_BASE));
+        asm volatile("sub.f32 %0, %1, %2;\n" : "=r"(fp32_intermediates_casted[3]) : "r"(fp32_intermediates_casted[3]), "r"(FP32_BASE));
+
+        int32_t decode_value[4];
+        ScaleComputeT new_code_zp = code_zp + 0.5f;
+
+        decode_value[0] = __float2int_rd(fmaf(fp32_intermediates[0], code_scale, new_code_zp));
+        decode_value[1] = __float2int_rd(fmaf(fp32_intermediates[1], code_scale, new_code_zp));
+        decode_value[2] = __float2int_rd(fmaf(fp32_intermediates[2], code_scale, new_code_zp));
+        decode_value[3] = __float2int_rd(fmaf(fp32_intermediates[3], code_scale, new_code_zp));
+
+        return convert_impl(decode_value);
+    }
+
+    CUTLASS_DEVICE
+    static result_type convert(source_type const& source, code_type const& code_scale, code_type const& code_zp)
+    {
+        uint32_t const i8s = reinterpret_cast<uint32_t const&>(source);
+
+        // 2^23 = 8388608
+        static constexpr uint32_t FP32_BASE = 0x4B000000;
+
+        float fp32_intermediates[4];
+        uint32_t* fp32_intermediates_casted = reinterpret_cast<uint32_t*>(fp32_intermediates);
+        fp32_intermediates_casted[0] = __byte_perm(i8s, FP32_BASE, 0x7650);
+        fp32_intermediates_casted[1] = __byte_perm(i8s, FP32_BASE, 0x7651);
+        fp32_intermediates_casted[2] = __byte_perm(i8s, FP32_BASE, 0x7652);
+        fp32_intermediates_casted[3] = __byte_perm(i8s, FP32_BASE, 0x7653);
+
+        asm volatile("sub.f32 %0, %1, %2;\n" : "=r"(fp32_intermediates_casted[0]) : "r"(fp32_intermediates_casted[0]), "r"(FP32_BASE));
+        asm volatile("sub.f32 %0, %1, %2;\n" : "=r"(fp32_intermediates_casted[1]) : "r"(fp32_intermediates_casted[1]), "r"(FP32_BASE));
+        asm volatile("sub.f32 %0, %1, %2;\n" : "=r"(fp32_intermediates_casted[2]) : "r"(fp32_intermediates_casted[2]), "r"(FP32_BASE));
+        asm volatile("sub.f32 %0, %1, %2;\n" : "=r"(fp32_intermediates_casted[3]) : "r"(fp32_intermediates_casted[3]), "r"(FP32_BASE));
+
+        int32_t decode_value[4];
+
+        decode_value[0] = __float2int_rd(fmaf(fp32_intermediates[0], code_scale[0], code_zp[0] + 0.5f));
+        decode_value[1] = __float2int_rd(fmaf(fp32_intermediates[1], code_scale[1], code_zp[1] + 0.5f));
+        decode_value[2] = __float2int_rd(fmaf(fp32_intermediates[2], code_scale[2], code_zp[2] + 0.5f));
+        decode_value[3] = __float2int_rd(fmaf(fp32_intermediates[3], code_scale[3], code_zp[3] + 0.5f));
+
+        return convert_impl(decode_value);
+    }
+
+    CUTLASS_DEVICE
+    static result_type convert_impl(int32_t* decode_value)
+    {
+        result_type result;
+
+        static constexpr uint32_t immLut = (0xF0 & 0xCC) | 0xAA;
+        static constexpr uint32_t MASK = 0x003F003F;
+        // 2^7 = 128
+        static constexpr uint32_t EX = 0x43004300;
+
+        uint32_t* h = reinterpret_cast<uint32_t*>(&result);
+
+        int32_t q0 = __byte_perm(decode_value[0], decode_value[1], 0x5410);
+        int32_t q1 = __byte_perm(decode_value[2], decode_value[3], 0x5410);
+
+        h[0] = lop3<immLut>(q0 >> 9, MASK, EX);
+        h[1] = lop3<immLut>(q0 >> 6, MASK, EX);
+        h[2] = lop3<immLut>(q0 >> 3, MASK, EX);
+        h[3] = lop3<immLut>(q0, MASK, EX);
+
+        h[4] = lop3<immLut>(q1 >> 9, MASK, EX);
+        h[5] = lop3<immLut>(q1 >> 6, MASK, EX);
+        h[6] = lop3<immLut>(q1 >> 3, MASK, EX);
+        h[7] = lop3<immLut>(q1, MASK, EX);
+
+#if (defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 900) && defined(ENABLE_BF16))
+        // 128 + 32 = 160
+        static constexpr uint32_t SUB = 0x43204320;
+
+        asm volatile("sub.bf16x2 %0, %1, %2;\n" : "=r"(h[0]) : "r"(h[0]), "r"(SUB));
+        asm volatile("sub.bf16x2 %0, %1, %2;\n" : "=r"(h[1]) : "r"(h[1]), "r"(SUB));
+        asm volatile("sub.bf16x2 %0, %1, %2;\n" : "=r"(h[2]) : "r"(h[2]), "r"(SUB));
+        asm volatile("sub.bf16x2 %0, %1, %2;\n" : "=r"(h[3]) : "r"(h[3]), "r"(SUB));
+
+        asm volatile("sub.bf16x2 %0, %1, %2;\n" : "=r"(h[4]) : "r"(h[4]), "r"(SUB));
+        asm volatile("sub.bf16x2 %0, %1, %2;\n" : "=r"(h[5]) : "r"(h[5]), "r"(SUB));
+        asm volatile("sub.bf16x2 %0, %1, %2;\n" : "=r"(h[6]) : "r"(h[6]), "r"(SUB));
+        asm volatile("sub.bf16x2 %0, %1, %2;\n" : "=r"(h[7]) : "r"(h[7]), "r"(SUB));
+#else
+        // 1.0
+        static constexpr uint32_t MUL = 0x3F803F80;
+        // -160
+        static constexpr uint32_t ADD = 0xC320C320;
+
+        asm volatile("fma.rn.bf16x2 %0, %1, %2, %3;\n" : "=r"(h[0]) : "r"(h[0]), "r"(MUL), "r"(ADD));
+        asm volatile("fma.rn.bf16x2 %0, %1, %2, %3;\n" : "=r"(h[1]) : "r"(h[1]), "r"(MUL), "r"(ADD));
+        asm volatile("fma.rn.bf16x2 %0, %1, %2, %3;\n" : "=r"(h[2]) : "r"(h[2]), "r"(MUL), "r"(ADD));
+        asm volatile("fma.rn.bf16x2 %0, %1, %2, %3;\n" : "=r"(h[3]) : "r"(h[3]), "r"(MUL), "r"(ADD));
+
+        asm volatile("fma.rn.bf16x2 %0, %1, %2, %3;\n" : "=r"(h[4]) : "r"(h[4]), "r"(MUL), "r"(ADD));
+        asm volatile("fma.rn.bf16x2 %0, %1, %2, %3;\n" : "=r"(h[5]) : "r"(h[5]), "r"(MUL), "r"(ADD));
+        asm volatile("fma.rn.bf16x2 %0, %1, %2, %3;\n" : "=r"(h[6]) : "r"(h[6]), "r"(MUL), "r"(ADD));
+        asm volatile("fma.rn.bf16x2 %0, %1, %2, %3;\n" : "=r"(h[7]) : "r"(h[7]), "r"(MUL), "r"(ADD));
+#endif
+
+        return result;
+    }
+
+    CUTLASS_DEVICE
+    result_type operator()(source_type const& s, ScaleComputeT code_scale, ScaleComputeT code_zp)
+    {
+        return convert(s, code_scale, code_zp);
+    }
+};
+
+template <typename T, int N>
+struct FastInterleavedAndBiasedNumericArrayConverter<T, uint2b_t, N>
+{
+    static_assert(platform::is_same<T, half_t>::value || platform::is_same<T, bfloat16_t>::value,
+        "T must be fp16 or bf16");
+
+    static constexpr int kVecWidth = 16;
+    static_assert(!(N % kVecWidth), "N must be multiple of 16.");
+
+    using result_type = Array<T, N>;
+    using source_type = Array<uint2b_t, N>;
+    using code_type = Array<float, N / kVecWidth>;
+
+    CUTLASS_DEVICE
+    static result_type convert(source_type const& source, code_type const& code_scale, code_type const& code_zp)
+    {
+        using scalar_result_type = typename result_type::Element;
+        using scalar_source_type = typename source_type::Element;
+        FastInterleavedAndBiasedNumericArrayConverter<scalar_result_type, scalar_source_type, kVecWidth>
+            convert_vector_;
+
+        result_type result;
+        using vec_result = Array<scalar_result_type, kVecWidth>;
+        using vec_source = Array<scalar_source_type, kVecWidth>;
+
+        vec_result* result_ptr = reinterpret_cast<vec_result*>(&result);
+        vec_source const* source_ptr = reinterpret_cast<vec_source const*>(&source);
+
+        CUTLASS_PRAGMA_UNROLL
+        for (int i = 0; i < N / kVecWidth; ++i)
+        {
+            result_ptr[i] = convert_vector_(source_ptr[i], code_scale[i], code_zp[i]);
+        }
+
+        return result;
+    }
+
+    CUTLASS_DEVICE
+    static result_type convert(source_type const& source, Array<float, N / 4> const& code_scale, Array<float, N / 4> const& code_zp)
+    {
+        using scalar_result_type = typename result_type::Element;
+        using scalar_source_type = typename source_type::Element;
+        using Converter = FastInterleavedAndBiasedNumericArrayConverter<scalar_result_type, scalar_source_type, kVecWidth>;
+
+        result_type result;
+        using vec_result = typename Converter::result_type;
+        using vec_source = typename Converter::source_type;
+        using vec_code = typename Converter::code_type;
+
+        vec_result* result_ptr = reinterpret_cast<vec_result*>(&result);
+        vec_source const* source_ptr = reinterpret_cast<vec_source const*>(&source);
+        vec_code const* code_scale_ptr = reinterpret_cast<vec_code const*>(&code_scale);
+        vec_code const* code_zp_ptr = reinterpret_cast<vec_code const*>(&code_zp);
+
+        CUTLASS_PRAGMA_UNROLL
+        for (int i = 0; i < N / kVecWidth; ++i)
+        {
+            result_ptr[i] = Converter::convert(source_ptr[i], code_scale_ptr[i], code_zp_ptr[i]);
+        }
+
+        return result;
+    }
+
+    CUTLASS_DEVICE
+    result_type operator()(source_type const& s, code_type const& code_scale, code_type const& code_zp)
+    {
+        return convert(s, code_scale, code_zp);
+    }
+};
+
 /////////////////////////////////////////////////////////////////////////////////////////////////
 
 } // namespace cutlass

custom_ops/gpu_ops/cutlass_extensions/wint_type_traits.h

@@ -125,10 +125,13 @@ struct WintQuantTraits<ElementT, WintQuantMethod::kWeightOnlyInt2> {
   static constexpr int32_t kNumPackedValues = 4;
   static constexpr int32_t kPackedSize = 16;
 
+  using LocalScaleType = uint4b_t;
+  using CodeScaleZpType = float;
+
   struct Arguments {
-    const uint8_t *local_scale_ptr; // quanted 4-bits
-    const float *code_scale_ptr;
-    const float *code_zp_ptr;
+    uint8_t *local_scale_ptr; // quanted 4-bits
+    float *code_scale_ptr;
+    float *code_zp_ptr;
   };
 
   CUTLASS_DEVICE

custom_ops/gpu_ops/cutlass_kernels/moe_gemm/fused_moe_cutlass_kernel.h

@@ -43,7 +43,6 @@
 #include "cutlass/trace.h"
 
 #include "cutlass_extensions/gemm/kernel/gemm_moe_problem_visitor.h"
-#include "cutlass_extensions/gemm/threadblock/wint2x_tile_dequanter.h"
 #include "cutlass_extensions/tile_interleaved_layout.h"
 
 /////////////////////////////////////////////////////////////////////////////////////////////////
@@ -775,17 +774,54 @@ struct Wint2xMoeFCGemm : public MoeFCGemm<Mma_, Epilogue_, ThreadblockSwizzle_,
   template <WintQuantMethod QuantMethod, typename dummy>
   struct KernelRunner<QuantMethod, true, dummy> {
     using WeightQuantTraits = WintQuantTraits<ElementA, QuantMethod>;
-    using QuantArguments = typename WeightQuantTraits::Arguments;
+    using MmaQuantArguments = typename Mma::QuantParamsAccessor::Arguments;
 
     CUTLASS_DEVICE
-    static QuantArguments get_quant_args(Params const& params, int32_t problem_idx, const int64_t gemm_k, const int64_t gemm_n) {
-      QuantArguments quant_args;
-      if constexpr (QuantMethod == WintQuantMethod::kWeightOnlyInt2) {
-        quant_args.local_scale_ptr = params.local_scale + problem_idx * gemm_k * gemm_n / 128;
-        quant_args.code_scale_ptr = params.code_scale + problem_idx * gemm_n;
-        quant_args.code_zp_ptr = params.code_zp + problem_idx * gemm_n;
-      }
-      return quant_args;
+    static MmaQuantArguments prepare_quant_args(
+        Params const& params, cutlass::gemm::GemmCoord const& threadblock_offset,
+        int64_t problem_idx, const int32_t gemm_k, const int32_t gemm_n, const int thread_idx) {
+      // the begin threadblock_offset of scale, which holds the same column id with C, but with no row id
+      cutlass::MatrixCoord tb_offset_scale{0, threadblock_offset.n()};
+      cutlass::MatrixCoord tb_offset_local_scale{0, threadblock_offset.n() * 2};
+
+      ElementScale* weight_scale_ptr = params.weight_scales + problem_idx * gemm_n;
+      typename Mma::QuantParamsAccessor::IteratorSuperScale iterator_super_scale(
+          Mma::QuantParamsAccessor::LayoutSuperScale(gemm_n),
+          weight_scale_ptr,
+          {1, gemm_n},
+          thread_idx,
+          tb_offset_scale);
+
+      int local_scale_pointer_offset = ((ThreadblockShape::kK + 127) / 128) * (gemm_n * 2);
+      int64_t offset_in_bytes = problem_idx * gemm_k * gemm_n / 128;
+      uint4b_t *local_scale_ptr = reinterpret_cast<uint4b_t *>(params.local_scale + offset_in_bytes);
+
+      typename Mma::QuantParamsAccessor::IteratorLocalScale iterator_local_scale(
+          Mma::QuantParamsAccessor::LayoutLocalScale(gemm_n * 2),
+          local_scale_ptr,
+          {(gemm_k + 127) / 128, gemm_n * 2},
+          thread_idx,
+          tb_offset_local_scale);
+
+      float* code_scale_ptr = params.code_scale + problem_idx * gemm_n;
+      typename Mma::QuantParamsAccessor::IteratorCodeScaleZp iterator_code_scale(
+          Mma::QuantParamsAccessor::LayoutCodeScaleZp(gemm_n),
+          code_scale_ptr,
+          {1, gemm_n},
+          thread_idx,
+          tb_offset_scale);
+
+      float* code_zp_ptr = params.code_zp + problem_idx * gemm_n;
+      typename Mma::QuantParamsAccessor::IteratorCodeScaleZp iterator_code_zp(
+          Mma::QuantParamsAccessor::LayoutCodeScaleZp(gemm_n),
+          code_zp_ptr,
+          {1, gemm_n},
+          thread_idx,
+          tb_offset_scale);
+
+      MmaQuantArguments mma_quant_args(
+          iterator_super_scale, iterator_local_scale, iterator_code_scale, iterator_code_zp, local_scale_pointer_offset);
+      return mma_quant_args;
     }
 
     CUTLASS_DEVICE
@@ -814,9 +850,6 @@ struct Wint2xMoeFCGemm : public MoeFCGemm<Mma_, Epilogue_, ThreadblockSwizzle_,
                   kInterleave >= 1,
           "B must be row major/col major OR col major interleaved.");
 
-      // LayoutB should be RowMajor
-      using TileDequanterB = cutlass::gemm::threadblock::TileDequanter<ElementA, ElementScale, ThreadblockShape::kK, ThreadblockShape::kN, kStages, kThreadCount, QuantMethod>;
-
       //
       // Problem visitor.
       //
@@ -843,12 +876,6 @@ struct Wint2xMoeFCGemm : public MoeFCGemm<Mma_, Epilogue_, ThreadblockSwizzle_,
             int(cta_idx % grid_shape.n()) * Mma::Shape::kN,  // NOLINT
             0);
 
-        // begin address offset for weight_scale.
-        ElementScale* weight_scale_ptr =
-            params.weight_scales ? params.weight_scales + problem_idx * problem_size.n() : nullptr;
-        // the begin threadblock_offset of scale, which holds the same column id with C, but with no row id
-        cutlass::MatrixCoord tb_offset_scale{0, threadblock_offset.n()};
-
         // Load element pointers. Exchange pointers and strides if working on
         // the transpose
         int64_t rows_to_jump = 0;
@@ -866,42 +893,20 @@ struct Wint2xMoeFCGemm : public MoeFCGemm<Mma_, Epilogue_, ThreadblockSwizzle_,
 
         // Compute initial location in logical coordinates
         // the begin threadblock_offset of A, which holds the same row id with C
-        cutlass::MatrixCoord tb_offset_A{
-            threadblock_offset.m(),
-            0,
-        };
+        cutlass::MatrixCoord tb_offset_A{threadblock_offset.m(), 0};
 
         // begin address offset for B for current problem_idx, totally num_experts problems
         char* byte_ptr_B = ((char*)params.ptr_B) +                 // NOLINT
                            problem_idx * bytes_per_expert_matrix;  // NOLINT
-
+        ElementB* ptr_B = reinterpret_cast<ElementB*>(byte_ptr_B);
         typename LayoutB::LongIndex ldm_B =
             platform::is_same<layout::RowMajor, LayoutB>::value
                 ? gemm_n
                 : gemm_k * kInterleave;
-        typename LayoutB::LongIndex ldm_B_shared = TileDequanterB::kColumns;
 
         // the begin threadblock_offset of B, which holds the same column id with C
-        cutlass::MatrixCoord tb_offset_B{0,
-                                         threadblock_offset.n() / kInterleave};
-
+        cutlass::MatrixCoord tb_offset_B{0, threadblock_offset.n() / kInterleave};
         cutlass::MatrixCoord extent_B{problem_size.k() * kInterleave, problem_size.n() / kInterleave};
-        cutlass::MatrixCoord extent_B_shared{TileDequanterB::kRows, TileDequanterB::kColumns};
-
-        MmaElementB* smem_unzip_B_ptr = nullptr;
-        if constexpr (QuantMethod == WintQuantMethod::kWeightOnlyInt2) {
-          smem_unzip_B_ptr = shared_storage.main_loop.operand_unzip_B_ptr();
-        }
-        QuantArguments quant_args = get_quant_args(params, problem_idx, gemm_k, gemm_n);
-        TileDequanterB tile_dequanter_B(smem_unzip_B_ptr,
-                                        byte_ptr_B,
-                                        ldm_B,
-                                        extent_B,
-                                        tb_offset_B,
-                                        weight_scale_ptr,
-                                        tb_offset_scale,
-                                        quant_args);
-        MmaElementB* ptr_B = tile_dequanter_B.GetOutPtr();
 
         // Compute position within threadblock
         int thread_idx = threadIdx.x;
@@ -914,20 +919,21 @@ struct Wint2xMoeFCGemm : public MoeFCGemm<Mma_, Epilogue_, ThreadblockSwizzle_,
                                            tb_offset_A);
 
         typename Mma::IteratorB iterator_B(
-            LayoutB(TileDequanterB::kUseSharedMemory ? ldm_B_shared : ldm_B),
+            LayoutB(ldm_B),
             ptr_B,
-            TileDequanterB::kUseSharedMemory ? extent_B_shared : extent_B,
+            extent_B,
             thread_idx,
-            TileDequanterB::kUseSharedMemory ? cutlass::make_Coord(0, 0) : tb_offset_B);
+            tb_offset_B);
+
+        MmaQuantArguments mma_quant_args = prepare_quant_args(
+            params, threadblock_offset, problem_idx, gemm_k, gemm_n, thread_idx);
 
         typename Mma::FragmentC accumulators;
-
         accumulators.clear();
 
         // Broadcast the warp_id computed by lane 0 to ensure dependent code
         // is compiled as warp-uniform.
         int warp_idx = __shfl_sync(0xffffffff, threadIdx.x / 32, 0);
-
         int lane_idx = threadIdx.x % 32;
 
         //
@@ -950,7 +956,7 @@ struct Wint2xMoeFCGemm : public MoeFCGemm<Mma_, Epilogue_, ThreadblockSwizzle_,
             accumulators,
             iterator_A,
             iterator_B,
-            tile_dequanter_B,
+            mma_quant_args,
             accumulators);
 
         //

custom_ops/gpu_ops/cutlass_kernels/moe_gemm/fused_moe_gemm_kernels_template.h

@@ -205,7 +205,7 @@ void generic_moe_gemm_kernelLauncher(const T* A,
       threadblock_count,
       epilogue_op,
       reinterpret_cast<const ElementType*>(A),
-      reinterpret_cast<const CutlassMmaWeightType*>(B),
+      reinterpret_cast<const CutlassMmaKernelType*>(B),
       reinterpret_cast<const ElementType*>(weight_scales),
       reinterpret_cast<const ElementType*>(biases),
       reinterpret_cast<ElementType*>(C),

custom_ops/gpu_ops/cutlass_kernels/w4a8_moe/w4a8_gemm_grouped.h

@@ -223,14 +223,11 @@ public:
   static Status can_implement(Arguments const &args)
   {
     CUTLASS_TRACE_HOST("W4A8MoeGemmUniversalBase::can_implement()");
-    // printf("--1\n");
     // Initialize static kernel and device properties, if necessary.
     Status result = init_device_props();
-    // printf("--1-2\n");
     if (result != Status::kSuccess) {
       return result;
     }
-    // printf("--2\n");
     dim3 grid = get_grid_shape(args);
     // printf("--grid:%d, %d, %d\n", grid.x, grid.y, grid.z);
     if (!(grid.y <= std::numeric_limits<uint16_t>::max() &&
@@ -238,7 +235,6 @@ public:
     {
       return Status::kErrorInvalidProblem;
     }
-    // printf("--3\n");
     return GemmKernel::can_implement(args);
   }
 
@@ -285,18 +281,50 @@ public:
   }
 
 
+
   /// Returns the maximum number of active thread blocks per multiprocessor
-  static int maximum_active_blocks()
+  static int maximum_active_blocks(int smem_capacity = -1)
   {
     CUTLASS_TRACE_HOST("W4A8MoeGemmUniversalBase::maximum_active_blocks()");
 
-    // Initialize static device properties, if necessary
-    if (init_device_props() != Status::kSuccess) {
+    int smem_size = int(sizeof(typename GemmKernel_::SharedStorage));
+
+    CUTLASS_TRACE_HOST("  smem_size: " << smem_size << " bytes");
+
+    cudaError_t result;
+    if (smem_size > (48 << 10)) {
+      result = cudaFuncSetAttribute(Kernel2<GemmKernel_>,
+                                    cudaFuncAttributeMaxDynamicSharedMemorySize,
+                                    smem_size);
+
+      if (result != cudaSuccess) {
+        // Call cudaGetLastError() to clear the error bit
+        result = cudaGetLastError();
+        CUTLASS_TRACE_HOST(
+          "  cudaFuncSetAttribute() returned error "
+          << cudaGetErrorString(result));
+        return -1;
+      }
+    }
+
+    int max_active_blocks = -1;
+    result = cudaOccupancyMaxActiveBlocksPerMultiprocessor(
+        &max_active_blocks,
+        Kernel2<GemmKernel_>,
+        GemmKernel_::kThreadCount,
+        smem_size);
+
+    if (result != cudaSuccess) {
+      // Call cudaGetLastError() to clear the error bit
+      result = cudaGetLastError();
+      CUTLASS_TRACE_HOST(
+        "  cudaOccupancyMaxActiveBlocksPerMultiprocessor() returned error "
+        << cudaGetErrorString(result));
       return -1;
     }
 
-    CUTLASS_TRACE_HOST("  max_active_blocks: " << sm_occupancy_);
-    return sm_occupancy_;
+    CUTLASS_TRACE_HOST("  max_active_blocks: " << max_active_blocks);
+    return max_active_blocks;
   }
 
 
@@ -341,8 +369,7 @@ public:
 
     // Configure grid and block dimensions
     dim3 block(GemmKernel::kThreadCount, 1, 1);
-    // dim3 grid = params_.get_grid_dims();
-        dim3 grid(216, 1, 1);
+    dim3 grid(params_.threadblock_count, 1, 1);
 
     // Launch kernel
     CUTLASS_TRACE_HOST("  "

custom_ops/gpu_ops/cutlass_kernels/w4a8_moe/w4a8_moe_gemm_config_search.sh

@@ -21,12 +21,12 @@ rm -rf up_gate_proj_7168_8192.log
 rm -rf down_proj_8192_3584.log
 num_experts=8
 
-for tokens_per_expert in 12
+for tokens_per_expert in 1 2 4 8 16 20 24 28 32 36 48 64 96 128 160 192 224 256 384 512 768 1024 2048 3072 4096 8192
 
 do
 wait
-CUDA_VISIBLE_DEVICES=2 ./w4a8_moe_gemm_test ${num_experts} ${up_gate_proj_n} ${up_gate_proj_k} ${tokens_per_expert} 1 0 >> up_gate_proj_${up_gate_proj_n}_${up_gate_proj_k}.log 2>&1 &
-# CUDA_VISIBLE_DEVICES=3 ./w4a8_moe_gemm_test ${num_experts} ${down_proj_n} ${down_proj_k} ${tokens_per_expert} 1 0 >> down_proj_${down_proj_n}_${down_proj_k}.log 2>&1 &
+CUDA_VISIBLE_DEVICES=2 ./w4a8_moe_gemm_test ${num_experts} ${ffn1_n} ${ffn1_k} ${tokens_per_expert} 0 1 >> ffn1_${ffn1_n}_${ffn1_k}.log 2>&1 &
+CUDA_VISIBLE_DEVICES=3 ./w4a8_moe_gemm_test ${num_experts} ${ffn2_n} ${ffn2_k} ${tokens_per_expert} 0 1 >> ffn2_${ffn2_n}_${ffn2_k}.log 2>&1 &
 done
 wait
 echo "#### finish ####"

custom_ops/gpu_ops/cutlass_kernels/w4a8_moe/w4a8_moe_gemm_test.cu

@@ -996,7 +996,6 @@ int main(int argc, char *argv[]) {
         CutlassTileConfig::CtaShape64x256x64_WarpShape64x64x64,
         CutlassTileConfig::CtaShape32x512x64_WarpShape32x128x64,
         CutlassTileConfig::CtaShape128x128x64_WarpShape128x32x64,
-        CutlassTileConfig::CtaShape32x512x64_WarpShape32x128x64,
     };
     std::vector<SplitKStyle> all_split_k_style{SplitKStyle::NO_SPLIT_K};
 

custom_ops/gpu_ops/extract_text_token_output.cu

@@ -20,7 +20,7 @@ __global__ void extract_text_token_output_kernel(int *max_seq_len,
                                 int *mm_token_num_len,
                                 int *seq_lens_this_time,
                                 int *cu_seqlens_q,
-                                float *score_text,
+                                float *hidden_states,
                                 float *output,
                                 const int bsz,
                                 const int hidden_size) {
@@ -32,14 +32,11 @@ __global__ void extract_text_token_output_kernel(int *max_seq_len,
     int max_seq_len_index_data = max_seq_len_index[0];
     int mm_token_num_len_data = mm_token_num_len[0];
     int true_bsz = cu_seqlens_q[bsz_index + 1] - 1;
-    if (bsz_index >= max_seq_len_index_data) {
-        true_bsz = true_bsz - mm_token_num_len_data;
-    }
     if (max_seq_len_data == mm_token_num_len_data && bsz_index == max_seq_len_index_data) {
         output[bsz_index * hidden_size + block_idx] = 0.0;
     } else {
         if (seq_lens_this_time[bsz_index] != 0) {
-            output[bsz_index * hidden_size + block_idx] = score_text[true_bsz * hidden_size + block_idx];
+            output[bsz_index * hidden_size + block_idx] = hidden_states[true_bsz * hidden_size + block_idx];
         }
     }
     __syncthreads();
@@ -51,19 +48,19 @@ std::vector<paddle::Tensor> ExtractTextTokenOutput(
             const paddle::Tensor& mm_token_num_len,
             const paddle::Tensor& seq_lens_this_time,
             const paddle::Tensor& cu_seqlens_q,
-            const paddle::Tensor& score_text) {
+            const paddle::Tensor& hidden_states) {
 
     const int bsz = seq_lens_this_time.shape()[0];
-    const int hidden_size = score_text.shape()[1];
-    paddle::Tensor output = paddle::full({bsz, hidden_size}, 1, paddle::DataType::FLOAT32, score_text.place());
+    const int hidden_size = hidden_states.shape()[1];
+    paddle::Tensor output = paddle::full({bsz, hidden_size}, 1, paddle::DataType::FLOAT32, hidden_states.place());
 
-    extract_text_token_output_kernel<1024><<<hidden_size, 1024, 0, score_text.stream()>>>(
+    extract_text_token_output_kernel<1024><<<hidden_size, 1024, 0, hidden_states.stream()>>>(
       const_cast<int*>(max_seq_len.data<int>()),
       const_cast<int*>(max_seq_len_index.data<int>()),
       const_cast<int*>(mm_token_num_len.data<int>()),
       const_cast<int*>(seq_lens_this_time.data<int>()),
       const_cast<int*>(cu_seqlens_q.data<int>()),
-      const_cast<float*>(score_text.data<float>()),
+      const_cast<float*>(hidden_states.data<float>()),
       output.data<float>(),
       bsz,
       hidden_size
@@ -76,9 +73,9 @@ std::vector<std::vector<int64_t>> ExtractTextTokenOutputInferShape(const std::ve
                                                              const std::vector<int64_t>& mm_token_num_len_shape,
                                                              const std::vector<int64_t>& seq_lens_this_time_shape,
                                                              const std::vector<int64_t>& cu_seqlens_q_shape,
-                                                             const std::vector<int64_t>& score_text_shape) {
+                                                             const std::vector<int64_t>& hidden_states_shape) {
     const int bsz = seq_lens_this_time_shape[0];
-    const int hidden_size = score_text_shape[1];
+    const int hidden_size = hidden_states_shape[1];
     return {{bsz, hidden_size}};
 }
 
@@ -87,8 +84,8 @@ std::vector<paddle::DataType> ExtractTextTokenOutputInferDtype(const paddle::Dat
                                                          const paddle::DataType& mm_token_num_len_dtype,
                                                          const paddle::DataType& seq_lens_this_time_dtype,
                                                          const paddle::DataType& cu_seqlens_q_dtype,
-                                                         const paddle::DataType& score_text_dtype) {
-    return {score_text_dtype};
+                                                         const paddle::DataType& hidden_states_dtype) {
+    return {hidden_states_dtype};
 }
 
 PD_BUILD_STATIC_OP(extract_text_token_output)
@@ -97,7 +94,7 @@ PD_BUILD_STATIC_OP(extract_text_token_output)
              "mm_token_num_len",
              "seq_lens_this_time",
              "cu_seqlens_q",
-             "score_text"})
+             "hidden_states"})
     .Outputs({"output"})
     .SetKernelFn(PD_KERNEL(ExtractTextTokenOutput))
     .SetInferShapeFn(PD_INFER_SHAPE(ExtractTextTokenOutputInferShape))

custom_ops/gpu_ops/flash_mask_attn/flash_mask_attn.cu

@@ -0,0 +1,163 @@
+/******************************************************************************
+ * Copyright (c) 2024, Jay Shah, Ganesh Bikshandi, Ying Zhang, Vijay Thakkar, Pradeep Ramani, Tri Dao.
+ ******************************************************************************/
+
+// Copyright (c) 2024 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "paddle/extension.h"
+#include "kernel_traits.h"
+#include "flash_mask_attn_kernel.hpp"
+
+template <typename paddle_type>
+struct cuteType;
+
+template <>
+struct cuteType<phi::dtype::float16> {
+    using type = cutlass::half_t;
+};
+
+template <>
+struct cuteType<phi::dtype::bfloat16> {
+    using type = cutlass::bfloat16_t;
+};
+
+template <typename T>
+std::vector<paddle::Tensor> DispatchFlashAttentionMask(
+        const paddle::Tensor& q_input,
+        const paddle::Tensor& k_input,
+        const paddle::Tensor& v_input,
+        const paddle::Tensor& cu_seq_q,
+        const paddle::Tensor& cu_seq_k,
+        const paddle::Tensor& seq_len_encoder,
+        const paddle::optional<paddle::Tensor>& mask,
+        const int head_num,
+        const int kv_head_num,
+        const int head_dim,
+        const int max_seq_len,
+        const int max_enc_len_this_time,
+        const int max_dec_len_this_time) {
+
+    constexpr int kBlockM = 128;
+    constexpr int kBlockN = 128;
+    const int batch_size = cu_seq_q.dims()[0];
+
+    paddle::Tensor out = paddle::empty(
+        {q_input.dims()[0], head_num * head_dim}, q_input.dtype(), q_input.place());
+
+    Flash_mask_params params;
+    memset(&params, 0, sizeof(Flash_mask_params));
+
+    params.q_ptr = const_cast<T*>(q_input.data<T>());
+    params.k_ptr = const_cast<T*>(k_input.data<T>());
+    params.v_ptr = const_cast<T*>(v_input.data<T>());
+    params.o_ptr = const_cast<T*>(out.data<T>());
+    params.cu_seq_q = const_cast<int*>(cu_seq_q.data<int>());
+    params.cu_seq_k = const_cast<int*>(cu_seq_k.data<int>());
+    params.seq_len_encoder = const_cast<int*>(seq_len_encoder.data<int>());
+    params.head_num = head_num;
+    params.kv_head_num = kv_head_num;
+    params.max_seq_len_q = max_enc_len_this_time;
+    params.max_seq_len_k = max_enc_len_this_time + max_dec_len_this_time;
+    params.batch_size = batch_size;
+    params.gqa_group_size = head_num / kv_head_num;
+    constexpr float kLog2e = 1.4426950408889634074;
+    params.scale_softmax_log2 = 1.0f / std::sqrt(head_dim) * kLog2e;
+
+    using cute_type = typename cuteType<T>::type;
+
+    if (mask) {
+        params.mask = const_cast<int*>(mask.get().data<int>());
+        flash_attn_headdim128<kBlockM, kBlockN, true, cute_type>(params, 0);
+    } else {
+        flash_attn_headdim128<kBlockM, kBlockN, false, cute_type>(params, 0);
+    }
+
+    return {out};
+}
+
+
+std::vector<paddle::Tensor> FlashAttentionMask(
+        const paddle::Tensor& q_input,
+        const paddle::Tensor& k_input,
+        const paddle::Tensor& v_input,
+        const paddle::Tensor& cu_seq_q,
+        const paddle::Tensor& cu_seq_k,
+        const paddle::Tensor& seq_len_encoder,
+        const paddle::optional<paddle::Tensor> &mask,
+        const int head_num,
+        const int kv_head_num,
+        const int head_dim,
+        const int max_seq_len,
+        const int max_enc_len_this_time,
+        const int max_dec_len_this_time) {
+
+    if (q_input.dtype() == paddle::DataType::FLOAT16) {
+        using T = phi::dtype::float16;
+        return std::move(
+            DispatchFlashAttentionMask<T>(
+                q_input,
+                k_input,
+                v_input,
+                cu_seq_q,
+                cu_seq_k,
+                seq_len_encoder,
+                mask,
+                head_num,
+                kv_head_num,
+                head_dim,
+                max_seq_len,
+                max_enc_len_this_time,
+                max_dec_len_this_time));
+    } else if (q_input.dtype() == paddle::DataType::BFLOAT16) {
+        using T = phi::dtype::bfloat16;
+        return std::move(
+            DispatchFlashAttentionMask<T>(
+                q_input,
+                k_input,
+                v_input,
+                cu_seq_q,
+                cu_seq_k,
+                seq_len_encoder,
+                mask,
+                head_num,
+                kv_head_num,
+                head_dim,
+                max_seq_len,
+                max_enc_len_this_time,
+                max_dec_len_this_time));
+    }
+
+}
+
+
+PD_BUILD_OP(flash_attention_mask)
+    .Inputs({
+        "q_input",
+        "k_input",
+        "v_input",
+        "cu_seq_q",
+        "cu_seq_k",
+        "seq_len_encoder",
+        paddle::Optional("mask")})
+    .Attrs({
+        "head_num: int",
+        "kv_head_num: int",
+        "head_dim: int",
+        "max_seq_len: int",
+        "max_enc_len_this_time: int",
+        "max_dec_len_this_time: int"})
+    .Outputs({
+        "out"})
+    .SetKernelFn(PD_KERNEL(FlashAttentionMask));

custom_ops/gpu_ops/flash_mask_attn/flash_mask_attn_kernel.hpp

@@ -0,0 +1,231 @@
+/******************************************************************************
+ * Copyright (c) 2024, Jay Shah, Ganesh Bikshandi, Ying Zhang, Vijay Thakkar, Pradeep Ramani, Tri Dao.
+ ******************************************************************************/
+
+#pragma once
+#include "cute/algorithm/copy.hpp"
+#include "cute/atom/mma_atom.hpp"
+#include "cutlass/gemm/collective/collective_builder.hpp"
+
+#include "cutlass/cutlass.h"
+#include "cutlass/layout/layout.h"
+#include "cutlass/numeric_types.h"
+#include "cutlass/pipeline/pipeline.hpp"
+#include "cutlass/cluster_launch.hpp"
+#include "cutlass/arch/reg_reconfig.h"
+
+#include "kernel_traits.h"
+#include "mainloop_attn.hpp"
+#include "softmax.hpp"
+
+using namespace cute;
+
+template <int kHeadDim>
+auto get_gmem_layout(int token_num, int head_num) {
+    return make_layout(
+        make_shape(token_num, kHeadDim, head_num),
+        make_stride(head_num * kHeadDim, cute::_1{}, kHeadDim));
+}
+
+template <typename Ktraits>
+__global__ void __launch_bounds__(Ktraits::kNWarps * cutlass::NumThreadsPerWarp, 1)
+    compute_attn_ws(
+        CUTE_GRID_CONSTANT typename CollectiveMainloopAttn<Ktraits>::Params const mainloop_params,
+        CUTE_GRID_CONSTANT Flash_mask_params const data_params) {
+
+    using Element = typename Ktraits::Element;
+    using ElementAccum = typename Ktraits::ElementAccum;
+    using SoftType = ElementAccum;
+    using TileShape_MNK = typename Ktraits::TileShape_MNK;
+    using ClusterShape = typename Ktraits::ClusterShape_MNK;
+
+    static constexpr int NumMmaThreads = size(typename Ktraits::TiledMma0{});
+    static constexpr int NumCopyThreads = cutlass::NumThreadsPerWarpGroup;
+    static constexpr int kBlockM = Ktraits::kBlockM;
+    static constexpr int kBlockN = Ktraits::kBlockN;
+    constexpr int kHeadDim = Ktraits::kHeadDim;
+    constexpr bool NeedMask = Ktraits::NeedMask;
+
+    using CollectiveMainloop = CollectiveMainloopAttn<Ktraits>;
+
+    using MainloopPipeline = typename Ktraits::MainloopPipeline;
+    using PipelineParams = typename MainloopPipeline::Params;
+    using PipelineState = typename MainloopPipeline::PipelineState;
+
+    extern __shared__ char shared_memory[];
+    auto &shared_storage = *reinterpret_cast<typename Ktraits::SharedStorage*>(shared_memory);
+
+    __align__(16) __shared__ int mask[kBlockM];
+
+    const int m_block = blockIdx.x;
+    const int bidh = blockIdx.y;
+    const int bidb = blockIdx.z;
+
+    if constexpr (NeedMask) {
+        const int *mask_this_batch = data_params.mask + data_params.cu_seq_q[bidb] + m_block * kBlockM;
+
+        for (int i = threadIdx.x; i < kBlockM; i += Ktraits::kNWarps * cutlass::NumThreadsPerWarp) {
+            mask[i] = mask_this_batch[i];
+        }
+    }
+
+    const int seq_len_q = data_params.seq_len_encoder[bidb];
+    const int seq_len_k = data_params.cu_seq_k[bidb + 1] - data_params.cu_seq_k[bidb];
+
+    if (m_block * kBlockM >= seq_len_q) {
+        return;
+    }
+
+    int const lane_predicate = cute::elect_one_sync();
+    int const warp_idx = cutlass::canonical_warp_idx_sync();
+
+    if (warp_idx == 0 && lane_predicate) {
+        CollectiveMainloop::prefetch_tma_descriptors(mainloop_params);
+    }
+
+    int const warp_group_thread_idx = threadIdx.x % cutlass::NumThreadsPerWarpGroup;
+
+    PipelineParams pipeline_params;
+    pipeline_params.transaction_bytes = CollectiveMainloop::TmaTransactionBytesK;
+    int warp_group_idx = cutlass::canonical_warp_group_idx();
+    pipeline_params.role = warp_group_idx == 0
+        ? MainloopPipeline::ThreadCategory::Producer
+        : MainloopPipeline::ThreadCategory::Consumer;
+    pipeline_params.is_leader = warp_group_thread_idx == 0;
+    pipeline_params.num_consumers = NumMmaThreads;
+
+    if (warp_idx == 0 && lane_predicate) {
+        shared_storage.barrier_Q.init(1);
+    }
+
+    MainloopPipeline pipeline_k(shared_storage.pipeline_k, pipeline_params, ClusterShape{});
+    MainloopPipeline pipeline_v(shared_storage.pipeline_v, pipeline_params, ClusterShape{});
+
+    __syncthreads();
+
+    CollectiveMainloop collective_mainloop;
+
+    const int real_seq = seq_len_q - m_block * kBlockM;
+
+    const int n_block_max = NeedMask ? cute::ceil_div(mask[min(kBlockM - 1, real_seq - 1)], kBlockN) : cute::ceil_div((m_block + 1) * kBlockM + seq_len_k - seq_len_q, kBlockN);
+
+    if (warp_group_idx == 0) {  // Producer
+        cutlass::arch::warpgroup_reg_dealloc<Ktraits::kNWarps == 8 ? 56 : 24>();
+
+        int warp_idx_in_warpgroup = __shfl_sync(0xffffffff, (threadIdx.x / 32) % 4, 0);
+        if (warp_idx_in_warpgroup == 0) {  // Load Q, K, V
+            PipelineState smem_pipe_write_k = cutlass::make_producer_start_state<MainloopPipeline>();
+            PipelineState smem_pipe_write_v = cutlass::make_producer_start_state<MainloopPipeline>();
+
+            collective_mainloop.load(
+                mainloop_params,
+                pipeline_k,
+                pipeline_v,
+                smem_pipe_write_k,
+                smem_pipe_write_v,
+                shared_storage,
+                n_block_max,
+                m_block,
+                bidh,
+                bidb,
+                data_params.cu_seq_q,
+                data_params.cu_seq_k,
+                seq_len_q,
+                seq_len_k);
+        }
+    } else {  // Consumer
+        cutlass::arch::warpgroup_reg_alloc<Ktraits::kNWarps == 8 ? 256 : 240>();
+        typename Ktraits::TiledMma1 tiled_mma1;
+
+        PipelineState smem_pipe_read_k, smem_pipe_read_v;
+
+        Tensor tOrO = partition_fragment_C(tiled_mma1, select<0, 2>(TileShape_MNK{}));
+        Softmax<2 * (2 * kBlockM / NumMmaThreads)> softmax;
+
+        collective_mainloop.mma(
+            mainloop_params,
+            pipeline_k,
+            pipeline_v,
+            smem_pipe_read_k,
+            smem_pipe_read_v,
+            tOrO,
+            softmax,
+            mask,
+            n_block_max,
+            threadIdx.x - NumCopyThreads,
+            m_block,
+            seq_len_q,
+            seq_len_k,
+            shared_storage);
+
+        const int o_head_stride = data_params.head_num * kHeadDim;
+        const int store_offset = (data_params.cu_seq_q[bidb] + m_block * kBlockM) * o_head_stride + bidh * kHeadDim;
+
+        collective_mainloop.store<NumMmaThreads>(
+            mainloop_params,
+            tOrO,
+            shared_storage,
+            tiled_mma1,
+            threadIdx.x - NumCopyThreads,
+            o_head_stride,
+            real_seq,
+            reinterpret_cast<Element*>(data_params.o_ptr) + store_offset);
+    }
+
+}
+
+
+template<typename Kernel_traits>
+void run_flash_mask(Flash_mask_params &params, cudaStream_t stream) {
+    using Element = typename Kernel_traits::Element;
+    using TileShape_MNK = typename Kernel_traits::TileShape_MNK;
+    using ClusterShape = typename Kernel_traits::ClusterShape_MNK;
+
+    using CollectiveMainloop = CollectiveMainloopAttn<Kernel_traits>;
+    constexpr int kHeadDim = Kernel_traits::kHeadDim;
+
+    typename CollectiveMainloop::Params mainloop_params =
+        CollectiveMainloop::to_underlying_arguments({
+            static_cast<Element const*>(params.q_ptr),
+            get_gmem_layout<kHeadDim>(params.max_seq_len_q, params.head_num),
+            static_cast<Element const*>(params.k_ptr),
+            get_gmem_layout<kHeadDim>(params.max_seq_len_k, params.kv_head_num),
+            static_cast<Element const*>(params.v_ptr),
+            get_gmem_layout<kHeadDim>(params.max_seq_len_k, params.kv_head_num),
+            params.scale_softmax_log2
+        });
+
+    int num_blocks_m = cutlass::ceil_div(params.max_seq_len_q, Kernel_traits::kBlockM);
+
+    num_blocks_m = cutlass::ceil_div(num_blocks_m, size<0>(ClusterShape{})) * size<0>(ClusterShape{});
+
+    void *kernel;
+    kernel = (void *)compute_attn_ws<Kernel_traits>;
+    int smem_size = sizeof(typename Kernel_traits::SharedStorage);
+
+    if (smem_size >= 48 * 1024) {
+       cudaFuncSetAttribute(kernel, cudaFuncAttributeMaxDynamicSharedMemorySize, smem_size);
+    }
+
+    dim3 grid_dims;
+    grid_dims.x = num_blocks_m;
+    grid_dims.y = params.head_num;
+    grid_dims.z = params.batch_size;
+
+    static constexpr int ctaSize = Kernel_traits::kNWarps * 32;
+    dim3 block_dims(ctaSize);
+    dim3 cluster_dims(size<0>(ClusterShape{}), size<1>(ClusterShape{}), size<2>(ClusterShape{}));
+    cutlass::ClusterLaunchParams launch_params{grid_dims, block_dims, cluster_dims, smem_size, stream};
+    cutlass::launch_kernel_on_cluster(launch_params, kernel, mainloop_params, params);
+}
+
+template <int kBlockM, int kBlockN, bool NeedMask, typename InputType>
+void flash_attn_headdim128(Flash_mask_params &params, cudaStream_t stream) {
+
+    constexpr static int Headdim = 128;
+    constexpr static int kNWarps = kBlockM / 16 + 4;
+    constexpr static int kStages = 2;
+
+    using Ktraits = Flash_mask_kernel_traits<Headdim, kBlockM, kBlockN, kNWarps, kStages, NeedMask, InputType>;
+    run_flash_mask<Ktraits>(params, stream);
+}

custom_ops/gpu_ops/flash_mask_attn/kernel_traits.h

@@ -0,0 +1,124 @@
+/******************************************************************************
+ * Copyright (c) 2024, Jay Shah, Ganesh Bikshandi, Ying Zhang, Vijay Thakkar, Pradeep Ramani, Tri Dao.
+ ******************************************************************************/
+
+#pragma once
+
+#include "cute/atom/mma_atom.hpp"
+#include "cutlass/gemm/collective/collective_builder.hpp"
+
+#include "cutlass/cutlass.h"
+#include "cutlass/layout/layout.h"
+#include "cutlass/numeric_types.h"
+#include "cutlass/pipeline/pipeline.hpp"
+
+using namespace cute;
+
+struct Flash_mask_params {
+    void *__restrict__ q_ptr;
+    void *__restrict__ k_ptr;
+    void *__restrict__ v_ptr;
+    void * __restrict__ o_ptr;
+    int * __restrict__ cu_seq_q;
+    int * __restrict__ cu_seq_k;
+    int * __restrict__ mask;
+    int * seq_len_encoder;
+    int head_num;
+    int kv_head_num;
+    int max_seq_len_q;
+    int max_seq_len_k;
+    int batch_size;
+    int gqa_group_size;
+    float scale_softmax_log2;
+};
+
+template <int kStages, class Gemm1Type, class Gemm2Type, class OutputType, class SmemLayoutQ,
+          class SmemLayoutK, class SmemLayoutV, class SmemLayoutO>
+struct SharedStorageQKVO {
+    cute::array_aligned<Gemm1Type, cute::cosize_v<SmemLayoutQ>> smem_q;
+    cute::array_aligned<Gemm1Type, cute::cosize_v<SmemLayoutK>> smem_k;
+    union {
+        cute::array_aligned<Gemm2Type, cute::cosize_v<SmemLayoutV>> smem_v;
+        cute::array_aligned<OutputType, cute::cosize_v<SmemLayoutO>> smem_o;
+    };
+    struct {
+        cutlass::arch::ClusterTransactionBarrier barrier_Q;
+        typename cutlass::PipelineTmaAsync<kStages>::SharedStorage pipeline_k;
+        typename cutlass::PipelineTmaAsync<kStages>::SharedStorage pipeline_v;
+    };
+};
+
+template<int kHeadDim_, int kBlockM_, int kBlockN_, int kNWarps_, int kStages_, bool NeedMask_, typename elem_type=cutlass::half_t>
+struct Flash_mask_kernel_traits {
+    using Element = elem_type;
+    using ElementAccum = float;
+    using index_t = int32_t;
+
+    static constexpr int kNWarps = kNWarps_;
+    static constexpr int kNThreads = kNWarps * cutlass::NumThreadsPerWarp;
+
+    static constexpr int kBlockM = kBlockM_;
+    static constexpr int kBlockN = kBlockN_;
+    static constexpr int kHeadDim = kHeadDim_;
+    static_assert(kHeadDim % 32 == 0);
+    using TileShape_MNK = Shape<Int<kBlockM>, Int<kBlockN>, Int<kHeadDim>>;
+    using ClusterShape_MNK = Shape<Int<1>, Int<1>, Int<1>>;
+    static constexpr int kStages = kStages_;
+    static constexpr int NeedMask = NeedMask_;
+
+    using AtomLayoutMNK = Layout<Shape<Int<kBlockM / 64>, _1, _1>>;
+    using TiledMma0 = decltype(cute::make_tiled_mma(
+        cute::GMMA::ss_op_selector<Element, Element, ElementAccum, TileShape_MNK>(),
+        AtomLayoutMNK{}));
+    using TiledMma1 = decltype(cute::make_tiled_mma(
+        cute::GMMA::rs_op_selector<Element, Element, ElementAccum, decltype(select<0, 2, 1>(TileShape_MNK{})),
+            GMMA::Major::K, GMMA::Major::MN>(),
+        AtomLayoutMNK{}));
+
+    using SmemLayoutAtomQ = decltype(cutlass::gemm::collective::detail::ss_smem_selector<GMMA::Major::K, Element,
+        decltype(cute::get<0>(TileShape_MNK{})), decltype(cute::get<2>(TileShape_MNK{}))>());
+    using SmemLayoutQ = decltype(tile_to_shape(SmemLayoutAtomQ{}, select<0, 2>(TileShape_MNK{})));
+
+    using SmemLayoutAtomK = decltype(cutlass::gemm::collective::detail::ss_smem_selector<GMMA::Major::K, Element,
+        decltype(cute::get<1>(TileShape_MNK{})), decltype(cute::get<2>(TileShape_MNK{}))>());
+    using SmemLayoutK =
+        decltype(tile_to_shape(SmemLayoutAtomK{},
+                 make_shape(shape<1>(TileShape_MNK{}), shape<2>(TileShape_MNK{}), Int<kStages>{})));
+
+    using SmemLayoutAtomV = decltype(cutlass::gemm::collective::detail::ss_smem_selector<GMMA::Major::K, Element,
+        decltype(cute::get<1>(TileShape_MNK{})), decltype(cute::get<2>(TileShape_MNK{}))>());
+    using SmemLayoutV =
+        decltype(tile_to_shape(SmemLayoutAtomV{},
+                 make_shape(shape<1>(TileShape_MNK{}), shape<2>(TileShape_MNK{}), Int<kStages>{})));
+
+    using SmemLayoutAtomO = decltype(cutlass::gemm::collective::detail::ss_smem_selector<GMMA::Major::K, Element,
+        decltype(cute::get<0>(TileShape_MNK{})), decltype(cute::get<2>(TileShape_MNK{}))>());
+    using SmemLayoutO = decltype(tile_to_shape(SmemLayoutAtomO{}, select<0, 2>(TileShape_MNK{})));
+
+    using SmemCopyAtomQ = Copy_Atom<cute::SM75_U32x4_LDSM_N, Element>;
+    using SmemCopyAtomO = Copy_Atom<cute::SM90_U32x4_STSM_N, Element>;
+
+    using SharedStorage = SharedStorageQKVO<kStages, Element, Element, Element, SmemLayoutQ, SmemLayoutK, SmemLayoutV, SmemLayoutO>;
+
+    static constexpr int NumProducerThreads = cutlass::NumThreadsPerWarpGroup;
+    static constexpr int NumMmaThreads = kNThreads - NumProducerThreads;
+    static constexpr int kNumVecElem = ceil_div(128, sizeof_bits_v<Element>);
+    static constexpr int kNumThreadsPerRow = kHeadDim / kNumVecElem;
+    static_assert(NumMmaThreads % kNumThreadsPerRow == 0);
+    static constexpr int kNumRows = NumMmaThreads / kNumThreadsPerRow;
+    using TiledCopyOAtom = cute::Copy_Atom<cute::UniversalCopy<cutlass::uint128_t>, Element>;
+    using TiledCopyOThrLayout = decltype(cute::make_layout(
+        cute::make_shape(Int<kNumRows>{}, Int<kNumThreadsPerRow>{}),
+        LayoutRight{}));
+    using TiledCopyOValLayout = decltype(cute::make_layout(
+        cute::make_shape(_1{}, Int<kNumVecElem>{}),
+        LayoutRight{}));
+    using GmemTiledCopyO = decltype(make_tiled_copy(
+        TiledCopyOAtom{},
+        TiledCopyOThrLayout{},
+        TiledCopyOValLayout{}
+    ));
+
+    using MainloopPipeline = typename cutlass::PipelineTmaAsync<kStages>;
+    using PipelineState = typename cutlass::PipelineState<kStages>;
+};

custom_ops/gpu_ops/flash_mask_attn/mainloop_attn.hpp

@@ -0,0 +1,431 @@
+/******************************************************************************
+ * Copyright (c) 2024, Jay Shah, Ganesh Bikshandi, Ying Zhang, Vijay Thakkar, Pradeep Ramani, Tri Dao.
+ ******************************************************************************/
+
+#pragma once
+
+#include <cutlass/cutlass.h>
+#include <cutlass/array.h>
+#include <cutlass/numeric_types.h>
+#include <cutlass/numeric_conversion.h>
+#include "cutlass/pipeline/pipeline.hpp"
+
+#include "cute/tensor.hpp"
+
+#include "cutlass/gemm/collective/collective_builder.hpp"
+
+#include "utils.hpp"
+
+
+using namespace cute;
+
+template <typename Ktraits>
+struct CollectiveMainloopAttn {
+
+    using Element = typename Ktraits::Element;
+    using TileShape_MNK = typename Ktraits::TileShape_MNK;
+    using ClusterShape = typename Ktraits::ClusterShape_MNK;
+
+    static constexpr int kStages = Ktraits::kStages;
+    static constexpr int kHeadDim = Ktraits::kHeadDim;
+    static constexpr int kBlockM = Ktraits::kBlockM;
+    static constexpr int kBlockN = Ktraits::kBlockN;
+    static constexpr bool NeedMask = Ktraits::NeedMask;
+
+    using ShapeT = cute::Shape<int32_t, int32_t, int32_t>;
+    using StrideT = cute::Shape<int32_t, _1, int32_t>;
+    using LayoutT = cute::Layout<ShapeT, StrideT>;
+
+
+    using GmemTiledCopyQ = cute::SM90_TMA_LOAD;
+    using GmemTiledCopyKV = decltype(cutlass::gemm::collective::detail::sm90_cluster_shape_to_tma_atom(shape<0>(ClusterShape{})));
+    using GmemTiledCopyO = typename Ktraits::GmemTiledCopyO;
+
+    using SmemLayoutAtomQ = decltype(cutlass::gemm::collective::detail::ss_smem_selector<GMMA::Major::K, Element,
+        decltype(cute::get<0>(TileShape_MNK{})), decltype(cute::get<2>(TileShape_MNK{}))>());
+    using SmemLayoutQ = decltype(tile_to_shape(SmemLayoutAtomQ{}, select<0, 2>(TileShape_MNK{})));
+
+    using SmemLayoutAtomK = decltype(cutlass::gemm::collective::detail::ss_smem_selector<GMMA::Major::K, Element,
+        decltype(cute::get<1>(TileShape_MNK{})), decltype(cute::get<2>(TileShape_MNK{}))>());
+    using SmemLayoutK =
+        decltype(tile_to_shape(SmemLayoutAtomK{},
+                 make_shape(shape<1>(TileShape_MNK{}), shape<2>(TileShape_MNK{}), Int<kStages>{})));
+    using SmemLayoutV = SmemLayoutK;
+    // Note this is the transpose in terms of the view, not in terms of memory.
+    using SmemLayoutVt =
+        decltype(cute::composition(SmemLayoutV{},
+                                   make_layout(make_shape(get<2>(TileShape_MNK{}), get<1>(TileShape_MNK{}), Int<kStages>{}),
+                                               make_stride(get<1>(TileShape_MNK{}), _1{}, Int<size(SmemLayoutV{}(_, _, _0{}))>{}))));
+    using SmemLayoutO = typename Ktraits::SmemLayoutO;
+    using SmemCopyAtomO = typename Ktraits::SmemCopyAtomO;
+
+    using TMA_Q = decltype(make_tma_copy(
+        GmemTiledCopyQ{},
+        make_tensor(
+            make_gmem_ptr(static_cast<Element const*>(nullptr)),
+            repeat_like(StrideT{}, int32_t(0)),
+            StrideT{}
+        ),
+        SmemLayoutQ{},
+        select<0, 2>(TileShape_MNK{}),
+        _1{}));  // no mcast for Q
+
+    using TMA_KV = decltype(make_tma_copy(
+        GmemTiledCopyKV{},
+        make_tensor(
+            make_gmem_ptr(static_cast<Element const*>(nullptr)),
+            repeat_like(StrideT{}, int32_t(0)),
+            StrideT{}
+        ),
+        take<0, 2>(SmemLayoutK{}),
+        select<1, 2>(TileShape_MNK{}),
+        size<0>(ClusterShape{}))); // mcast along M mode for this N load, if any
+
+    static constexpr int NumMmaThreads = size(typename Ktraits::TiledMma0{});
+    using MainloopPipeline = typename Ktraits::MainloopPipeline;
+    using PipelineParams = typename MainloopPipeline::Params;
+    using PipelineState = typename MainloopPipeline::PipelineState;
+
+    // Set the bytes transferred in this TMA transaction (may involve multiple issues)
+    static constexpr uint32_t TmaTransactionBytesQ = static_cast<uint32_t>(size(SmemLayoutQ{}) * cutlass::sizeof_bits_v<Element> / 8);
+    static constexpr uint32_t TmaTransactionBytesK = static_cast<uint32_t>(size(take<0, 2>(SmemLayoutK{})) * cutlass::sizeof_bits_v<Element> / 8);
+
+    static constexpr bool UseSchedulerBarrier = kHeadDim <= 128;
+
+    // Host side kernel arguments
+    struct Arguments {
+        Element const* ptr_Q;
+        LayoutT layout_Q;
+        Element const* ptr_K;
+        LayoutT layout_K;
+        Element const* ptr_V;
+        LayoutT layout_V;
+        float const softmax_scale_log2;
+    };
+
+    // Device side kernel params
+    struct Params {
+        LayoutT layout_Q;
+        LayoutT layout_K;
+        LayoutT layout_V;
+        cutlass::FastDivmod qhead_per_khead_divmod;
+        TMA_Q tma_load_Q;
+        TMA_KV tma_load_K, tma_load_V;
+        float const softmax_scale_log2;
+    };
+
+
+    static Params
+    to_underlying_arguments(Arguments const& args) {
+        Tensor mQ = make_tensor(make_gmem_ptr(args.ptr_Q), args.layout_Q);
+        TMA_Q tma_load_Q = make_tma_copy(
+            GmemTiledCopyQ{},
+            mQ,
+            SmemLayoutQ{},
+            select<0, 2>(TileShape_MNK{}),
+            _1{});
+        Tensor mK = make_tensor(make_gmem_ptr(args.ptr_K), args.layout_K);
+        TMA_KV tma_load_K = make_tma_copy(
+            GmemTiledCopyKV{},
+            mK,
+            SmemLayoutK{}(_, _, _0{}),
+            select<1, 2>(TileShape_MNK{}),
+            size<0>(ClusterShape{})); // mcast along M mode for this N load, if any
+        Tensor mV = make_tensor(make_gmem_ptr(args.ptr_V), args.layout_V);
+        TMA_KV tma_load_V = make_tma_copy(
+            GmemTiledCopyKV{},
+            mV,
+            SmemLayoutV{}(_, _, _0{}),
+            select<1, 2>(TileShape_MNK{}),
+            size<0>(ClusterShape{})); // mcast along M mode for this N load, if any
+        return {args.layout_Q, args.layout_K, args.layout_V,
+                cutlass::FastDivmod(cute::ceil_div(get<2>(args.layout_Q.shape()), get<2>(args.layout_K.shape()))),
+                tma_load_Q, tma_load_K, tma_load_V,
+                args.softmax_scale_log2};
+    }
+
+    /// Issue Tma Descriptor Prefetch -- ideally from a single thread for best performance
+    CUTLASS_DEVICE
+    static void prefetch_tma_descriptors(Params const& mainloop_params) {
+        cute::prefetch_tma_descriptor(mainloop_params.tma_load_Q.get_tma_descriptor());
+        cute::prefetch_tma_descriptor(mainloop_params.tma_load_K.get_tma_descriptor());
+        cute::prefetch_tma_descriptor(mainloop_params.tma_load_V.get_tma_descriptor());
+    }
+
+    template <typename MTensor, typename Shape>
+    CUTLASS_DEVICE auto get_local_tile_tensor(
+        const MTensor &m_tensor,
+        const Shape &tile_shape,
+        const int *cu_seq_len,
+        const int bidh,
+        const int bidb,
+        const int actual_seq_len) const {
+        auto g_offset = local_tile(
+            m_tensor(_, _, bidh),
+            cute::make_shape(1, get<1>(tile_shape)),
+            make_coord(cu_seq_len[bidb], _0{}));
+        auto g_sequence = make_tensor(
+            g_offset.data(),
+            make_layout(
+                cute::make_shape(actual_seq_len, get<1>(tile_shape)),
+                g_offset.stride()
+            ));
+        auto g_tensor = local_tile(g_sequence, tile_shape, make_coord(_, _0{}));
+        return g_tensor;
+    }
+
+
+    template <typename SharedStorage>
+    CUTLASS_DEVICE void
+    load(Params const& mainloop_params,
+         MainloopPipeline pipeline_k,
+         MainloopPipeline pipeline_v,
+         PipelineState& smem_pipe_write_k,
+         PipelineState& smem_pipe_write_v,
+         SharedStorage &shared_storage,
+         const int n_block_max,
+         const int m_block,
+         const int bidh,
+         const int bidb,
+         const int *cu_seq_q,
+         const int *cu_seq_k,
+         const int seq_len_q,
+         const int seq_len_k) {
+
+        Tensor sQ = make_tensor(make_smem_ptr(shared_storage.smem_q.data()), SmemLayoutQ{});
+        Tensor sK = make_tensor(make_smem_ptr(shared_storage.smem_k.data()), SmemLayoutK{});
+        Tensor sV = make_tensor(make_smem_ptr(shared_storage.smem_v.data()), SmemLayoutV{});
+
+        Tensor mQ = mainloop_params.tma_load_Q.get_tma_tensor(mainloop_params.layout_Q.shape());
+        Tensor mK = mainloop_params.tma_load_K.get_tma_tensor(mainloop_params.layout_K.shape());
+        Tensor mV = mainloop_params.tma_load_V.get_tma_tensor(mainloop_params.layout_V.shape());
+        int bidh_kv = mainloop_params.qhead_per_khead_divmod.divide(bidh);
+
+        Tensor gQ = get_local_tile_tensor(
+            mQ, select<0, 2>(TileShape_MNK{}), cu_seq_q, bidh, bidb, seq_len_q)(_, _, m_block);
+        Tensor gK = get_local_tile_tensor(
+            mK, select<1, 2>(TileShape_MNK{}), cu_seq_k, bidh_kv, bidb, seq_len_k);
+        Tensor gV = get_local_tile_tensor(
+            mV, select<1, 2>(TileShape_MNK{}), cu_seq_k, bidh_kv, bidb, seq_len_k);
+
+        Tensor sQ_x = make_tensor(sQ.data(), make_layout(sQ.layout(), Layout<_1>{}));
+        Tensor gQ_x = make_tensor(gQ.data(), make_layout(gQ.layout(), Layout<_1>{}));
+        auto [tQgQ, tQsQ] = tma_partition(mainloop_params.tma_load_Q, _0{}, Layout<_1>{},group_modes<0, 2>(sQ_x), group_modes<0, 2>(gQ_x));
+        auto [tKgK, tKsK] = tma_partition(mainloop_params.tma_load_K, _0{}, Layout<_1>{},group_modes<0, 2>(sK), group_modes<0, 2>(gK));
+        auto [tVgV, tVsV] = tma_partition(mainloop_params.tma_load_V, _0{}, Layout<_1>{},group_modes<0, 2>(sV), group_modes<0, 2>(gV));
+
+        uint16_t mcast_mask_kv = 0;
+
+        int n_block = n_block_max - 1;
+
+        int lane_predicate = cute::elect_one_sync();
+
+        if (lane_predicate) {
+            shared_storage.barrier_Q.arrive_and_expect_tx(TmaTransactionBytesQ);
+            copy(mainloop_params.tma_load_Q.with(reinterpret_cast<cutlass::arch::ClusterTransactionBarrier::ValueType&>(shared_storage.barrier_Q), 0 /*mcast_mask*/), tQgQ, tQsQ);
+        }
+
+
+        if (lane_predicate) {
+            pipeline_k.producer_acquire(smem_pipe_write_k);
+            copy(mainloop_params.tma_load_K.with(*pipeline_k.producer_get_barrier(smem_pipe_write_k), mcast_mask_kv),
+                tKgK(_, n_block), tKsK(_, smem_pipe_write_k.index()));
+            ++smem_pipe_write_k;
+        }
+
+        if (lane_predicate) {
+            #pragma unroll 2
+            for (; n_block > 0; --n_block) {
+                pipeline_k.producer_acquire(smem_pipe_write_k);
+                copy(mainloop_params.tma_load_K.with(*pipeline_k.producer_get_barrier(smem_pipe_write_k), mcast_mask_kv),
+                    tKgK(_, n_block - 1), tKsK(_, smem_pipe_write_k.index()));
+                ++smem_pipe_write_k;
+                pipeline_v.producer_acquire(smem_pipe_write_v);
+                copy(mainloop_params.tma_load_V.with(*pipeline_v.producer_get_barrier(smem_pipe_write_v), mcast_mask_kv),
+                    tVgV(_, n_block), tVsV(_, smem_pipe_write_v.index()));
+                ++smem_pipe_write_v;
+            }
+        }
+        if (lane_predicate) {
+            pipeline_v.producer_acquire(smem_pipe_write_v);
+            copy(mainloop_params.tma_load_V.with(*pipeline_v.producer_get_barrier(smem_pipe_write_v), mcast_mask_kv),
+                tVgV(_, n_block), tVsV(_, smem_pipe_write_v.index()));
+            ++smem_pipe_write_v;
+        }
+    }
+
+    template <typename SharedStorage, typename FrgTensorO, typename Softmax>
+    CUTLASS_DEVICE void
+    mma(Params const& mainloop_params,
+        MainloopPipeline pipeline_k,
+        MainloopPipeline pipeline_v,
+        PipelineState& smem_pipe_read_k,
+        PipelineState& smem_pipe_read_v,
+        FrgTensorO& tOrO,
+        Softmax& softmax,
+        const int *mask,
+        const int n_block_max,
+        const int thread_idx,
+        const int m_block,
+        const int seq_len_q,
+        const int seq_len_k,
+        SharedStorage& shared_storage) {
+
+        Tensor sQ = make_tensor(make_smem_ptr(shared_storage.smem_q.data()), SmemLayoutQ{});
+        Tensor sK = make_tensor(make_smem_ptr(shared_storage.smem_k.data()), SmemLayoutK{});
+        Tensor sVt = make_tensor(make_smem_ptr(shared_storage.smem_v.data()), SmemLayoutVt{});
+
+        typename Ktraits::TiledMma0 tiled_mma0;
+        typename Ktraits::TiledMma1 tiled_mma1;
+        auto threadMma0 = tiled_mma0.get_thread_slice(thread_idx);
+        auto threadMma1 = tiled_mma1.get_thread_slice(thread_idx);
+
+        Tensor tSrQ = threadMma0.partition_fragment_A(sQ);
+        Tensor tSrK = threadMma0.partition_fragment_B(sK);
+        Tensor tOrV = threadMma1.partition_fragment_B(sVt);
+
+        auto consumer_wait = [](auto& pipeline, auto& smem_pipe_read) {
+            auto barrier_token = pipeline.consumer_try_wait(smem_pipe_read);
+            pipeline.consumer_wait(smem_pipe_read, barrier_token);
+        };
+
+        tiled_mma1.accumulate_ = GMMA::ScaleOut::Zero;
+
+        int n_block = n_block_max - 1;
+
+        cutlass::ConsumerToken barrier_token = static_cast<cutlass::BarrierStatus>(shared_storage.barrier_Q.try_wait(0));
+        if (barrier_token == cutlass::BarrierStatus::WaitAgain) { shared_storage.barrier_Q.wait(0); }
+
+        Tensor tSrS = partition_fragment_C(tiled_mma0, select<0, 1>(TileShape_MNK{}));
+        consumer_wait(pipeline_k, smem_pipe_read_k);
+        gemm</*zero_init=*/true, /*wg_wait=*/-1>(tiled_mma0, tSrQ, tSrK(_, _, _, smem_pipe_read_k.index()), tSrS);
+        warpgroup_wait<0>();
+        pipeline_k.consumer_release(smem_pipe_read_k);
+        ++smem_pipe_read_k;
+
+        int mask_start_idx;
+        int mask_row_id;
+        int col_base;
+
+        if constexpr (NeedMask) {
+            const int lane_id = thread_idx % 32;
+            mask_start_idx = mask[0] / kBlockN - 1;
+
+            mask_row_id = thread_idx / 32 * 16 + lane_id / 4;
+
+            col_base = thread_idx % 4 * 2;
+
+            app_mask(
+                tSrS,
+                mask,
+                mask_row_id,
+                col_base + n_block * kBlockN);
+        } else {
+            auto col_limit_causal = [&](int row, int n_block) {
+                return row + 1 + seq_len_k - n_block * kBlockN - seq_len_q + m_block * kBlockM;
+            };
+            Tensor cS = cute::make_identity_tensor(select<0, 1>(TileShape_MNK{}));
+            Tensor tScS = threadMma0.partition_C(cS);
+            #pragma unroll
+            for (int i = 0; i < size(tSrS); ++i) {
+                if (int(get<1>(tScS(i))) >=
+                    std::min(seq_len_k - n_block * kBlockN, col_limit_causal(int(get<0>(tScS(i))), n_block))) {
+                    tSrS(i) = -INFINITY;
+                }
+            }
+        }
+
+        softmax.template online_softmax</*Is_first=*/true>(tSrS, mainloop_params.softmax_scale_log2);
+
+        Tensor tOrP = make_tensor(convert_type<Element>(tSrS).data(), convert_layout_acc_Aregs<typename Ktraits::TiledMma1>(tSrS.layout()));
+        Tensor scores_scale = make_fragment_like(softmax.row_max);
+        clear(scores_scale);
+
+        #pragma unroll 1
+        for (; n_block > 0; --n_block) {
+            Tensor tSrS = partition_fragment_C(tiled_mma0, select<0, 1>(TileShape_MNK{}));
+            consumer_wait(pipeline_k, smem_pipe_read_k);
+
+            if constexpr (NeedMask) {
+                if (n_block >= mask_start_idx) {
+                    app_mask(
+                        tSrS,
+                        mask,
+                        mask_row_id,
+                        col_base + n_block * kBlockN);
+                }
+            }
+
+            gemm</*zero_init=*/true, /*wg_wait=*/-1>(tiled_mma0, tSrQ, tSrK(_, _, _, smem_pipe_read_k.index()), tSrS);
+            softmax.rescale_o(tOrO, scores_scale);
+            consumer_wait(pipeline_v, smem_pipe_read_v);
+            gemm</*zero_init=*/false, /*wg_wait=*/-1>(tiled_mma1, tOrP, tOrV(_, _, _, smem_pipe_read_v.index()), tOrO);
+            warpgroup_wait<1>();
+            pipeline_k.consumer_release(smem_pipe_read_k);  // release K
+            cute::copy(softmax.template max</*Is_first=*/false>(tSrS, mainloop_params.softmax_scale_log2), scores_scale);
+            softmax.template online_softmax</*Is_first=*/false>(tSrS, mainloop_params.softmax_scale_log2);
+            warpgroup_wait<0>();
+            pipeline_v.consumer_release(smem_pipe_read_v);  // release V
+            ++smem_pipe_read_k;
+            ++smem_pipe_read_v;
+            cute::copy(make_tensor(convert_type<Element>(tSrS).data(), convert_layout_acc_Aregs<typename Ktraits::TiledMma1>(tSrS.layout())), tOrP);
+        }
+
+        softmax.rescale_o(tOrO, scores_scale);
+        consumer_wait(pipeline_v, smem_pipe_read_v);
+
+        gemm</*zero_init=*/false, /*wg_wait=*/-1>(tiled_mma1, tOrP, tOrV(_, _, _, smem_pipe_read_v.index()), tOrO);
+        cute::copy(softmax.finalize(mainloop_params.softmax_scale_log2), scores_scale);
+        warpgroup_wait<0>();
+        pipeline_v.consumer_release(smem_pipe_read_v);
+        ++smem_pipe_read_v;
+
+        softmax.rescale_o(tOrO, scores_scale);
+        return;
+    }
+
+    template <int NumMmaThreads, typename SharedStorage, typename FrgTensorO, typename TiledMma, typename T>
+    CUTLASS_DEVICE void
+    store(Params const& mainloop_params,
+        FrgTensorO const& tOrO,
+        SharedStorage& shared_storage,
+        TiledMma tiled_mma,
+        int thread_idx,
+        const int o_head_stride,
+        const int real_seq,
+        T * out_ptr) {
+
+        Tensor sO = make_tensor(make_smem_ptr(shared_storage.smem_o.data()), SmemLayoutO{});
+        auto smem_tiled_copy_O = make_tiled_copy_C(SmemCopyAtomO{}, tiled_mma);
+        auto smem_thr_copy_O = smem_tiled_copy_O.get_thread_slice(thread_idx);
+
+        Tensor tOrO_out = convert_type<Element>(tOrO);
+        Tensor taccOrO = smem_thr_copy_O.retile_S(tOrO_out);
+        Tensor taccOsO = smem_thr_copy_O.partition_D(sO);
+
+        cute::copy(smem_tiled_copy_O, taccOrO, taccOsO);
+
+        cutlass::arch::NamedBarrier::sync(NumMmaThreads, 0);
+
+        Tensor gO = make_tensor(make_gmem_ptr(out_ptr),
+            Shape<Int<kBlockM>, Int<kHeadDim>>{},
+            make_stride(o_head_stride, _1{}));
+
+        GmemTiledCopyO gmem_tiled_copy_O;
+        auto gmem_thr_copy_O = gmem_tiled_copy_O.get_thread_slice(thread_idx);
+
+        Tensor tOsO = gmem_thr_copy_O.partition_S(sO);
+        Tensor tOgO = gmem_thr_copy_O.partition_D(gO);
+
+        Tensor cO = make_identity_tensor(Shape<Int<kBlockM>, Int<kHeadDim>>{});
+
+        Tensor tOcO = gmem_thr_copy_O.partition_S(cO);
+
+        if (real_seq >= kBlockM) {
+            copy<true>(gmem_tiled_copy_O, tOsO, tOgO, tOcO);
+        } else {
+            copy<false>(gmem_tiled_copy_O, tOsO, tOgO, tOcO, real_seq);
+        }
+    }
+
+};

custom_ops/gpu_ops/flash_mask_attn/softmax.hpp

@@ -0,0 +1,206 @@
+/******************************************************************************
+ * Copyright (c) 2024, Jay Shah, Ganesh Bikshandi, Ying Zhang, Vijay Thakkar, Pradeep Ramani, Tri Dao.
+ ******************************************************************************/
+
+#pragma once
+
+#include <cmath>
+
+#include <cute/tensor.hpp>
+
+#include <cutlass/numeric_types.h>
+
+#include "utils.hpp"
+
+
+using namespace cute;
+
+
+template<int THREADS>
+struct Allreduce {
+    static_assert(THREADS == 32 || THREADS == 16 || THREADS == 8 || THREADS == 4);
+    template<typename T, typename Operator>
+    static __device__ __forceinline__ T run(T x, Operator &op) {
+        constexpr int OFFSET = THREADS / 2;
+        x = op(x, __shfl_xor_sync(uint32_t(-1), x, OFFSET));
+        return Allreduce<OFFSET>::run(x, op);
+    }
+};
+
+template<>
+struct Allreduce<2> {
+template<typename T, typename Operator>
+static __device__ __forceinline__ T run(T x, Operator &op) {
+    x = op(x, __shfl_xor_sync(uint32_t(-1), x, 1));
+    return x;
+}
+};
+
+template<bool zero_init=true, typename Engine0, typename Layout0, typename Engine1, typename Layout1, typename Operator>
+__device__ __forceinline__ void thread_reduce_(Tensor<Engine0, Layout0> const &tensor, Tensor<Engine1, Layout1> &summary, Operator &op) {
+    static_assert(Layout0::rank == 2, "Only support 2D Tensor");
+    static_assert(Layout1::rank == 1, "Only support 1D Tensor");
+    CUTE_STATIC_ASSERT_V(size<0>(summary) == size<0>(tensor));
+    #pragma unroll
+    for (int mi = 0; mi < size<0>(tensor); mi++) {
+        summary(mi) = zero_init ? tensor(mi, 0) : op(summary(mi), tensor(mi, 0));
+        #pragma unroll
+        for (int ni = 1; ni < size<1>(tensor); ni++) {
+            summary(mi) = op(summary(mi), tensor(mi, ni));
+        }
+    }
+}
+
+template<typename Engine0, typename Layout0, typename Engine1, typename Layout1, typename Operator>
+__device__ __forceinline__ void quad_allreduce_(Tensor<Engine0, Layout0> &dst, Tensor<Engine1, Layout1> &src, Operator &op) {
+    CUTE_STATIC_ASSERT_V(size(dst) == size(src));
+    #pragma unroll
+    for (int i = 0; i < size(dst); i++){
+        dst(i) = Allreduce<4>::run(src(i), op);
+    }
+}
+
+template<bool zero_init=true, typename Engine0, typename Layout0, typename Engine1, typename Layout1, typename Operator>
+__device__ __forceinline__ void reduce_(Tensor<Engine0, Layout0> const& tensor, Tensor<Engine1, Layout1> &summary, Operator &op) {
+    thread_reduce_<zero_init>(tensor, summary, op);
+    quad_allreduce_(summary, summary, op);
+}
+
+template<bool zero_init=true, typename Engine0, typename Layout0, typename Engine1, typename Layout1>
+__device__ __forceinline__ void reduce_max(Tensor<Engine0, Layout0> const& tensor, Tensor<Engine1, Layout1> &max){
+    MaxOp<float> max_op;
+    reduce_<zero_init>(tensor, max, max_op);
+}
+
+template<bool zero_init=true, bool warp_reduce=true, typename Engine0, typename Layout0, typename Engine1, typename Layout1>
+__device__ __forceinline__ void reduce_sum(Tensor<Engine0, Layout0> const& tensor, Tensor<Engine1, Layout1> &sum){
+    SumOp<float> sum_op;
+    thread_reduce_<zero_init>(tensor, sum, sum_op);
+    if constexpr (warp_reduce) { quad_allreduce_(sum, sum, sum_op); }
+}
+
+__forceinline__ __device__ __half2 half_exp(__half2 x) {
+    uint32_t tmp_out, tmp_in;
+    tmp_in = reinterpret_cast<uint32_t&>(x);
+    asm ("ex2.approx.f16x2 %0, %1;\n"
+      : "=r"(tmp_out)
+      : "r"(tmp_in));
+    __half2 out = reinterpret_cast<__half2&>(tmp_out);
+    return out;
+}
+
+// Apply the exp to all the elements.
+template <bool zero_init=false, typename Engine0, typename Layout0, typename Engine1, typename Layout1>
+__forceinline__ __device__ void max_scale_exp2_sum(Tensor<Engine0, Layout0> &tensor, Tensor<Engine1, Layout1> &max, Tensor<Engine1, Layout1> &sum, const float scale) {
+    static_assert(Layout0::rank == 2, "Only support 2D Tensor"); static_assert(Layout1::rank == 1, "Only support 1D Tensor"); CUTE_STATIC_ASSERT_V(size<0>(max) == size<0>(tensor));
+    #pragma unroll
+    for (int mi = 0; mi < size<0>(tensor); ++mi) {
+        MaxOp<float> max_op;
+        max(mi) = zero_init ? tensor(mi, 0) : max_op(max(mi), tensor(mi, 0));
+        #pragma unroll
+        for (int ni = 1; ni < size<1>(tensor); ni++) {
+            max(mi) = max_op(max(mi), tensor(mi, ni));
+        }
+        max(mi) = Allreduce<4>::run(max(mi), max_op);
+        const float max_scaled = max(mi) == -INFINITY ? 0.f : max(mi) * scale;
+        sum(mi) = 0;
+        #pragma unroll
+        for (int ni = 0; ni < size<1>(tensor); ++ni)  {
+            tensor(mi, ni) = exp2f(tensor(mi, ni) * scale - max_scaled);
+            sum(mi) += tensor(mi, ni);
+        }
+    }
+}
+
+
+template <typename Engine0, typename Layout0, typename Engine1, typename Layout1>
+__forceinline__ __device__ void scale_apply_exp2(Tensor<Engine0, Layout0> &tensor, Tensor<Engine1, Layout1> const &max, const float scale) {
+    static_assert(Layout0::rank == 2, "Only support 2D Tensor");
+    static_assert(Layout1::rank == 1, "Only support 1D Tensor");
+    CUTE_STATIC_ASSERT_V(size<0>(max) == size<0>(tensor));
+    #pragma unroll
+    for (int mi = 0; mi < size<0>(tensor); ++mi) {
+        const float max_scaled = max(mi) * scale;
+        #pragma unroll
+        for (int ni = 0; ni < size<1>(tensor); ++ni)  {
+            tensor(mi, ni) = exp2f(tensor(mi, ni) * scale - max_scaled);
+        }
+    }
+}
+
+
+template <int kNRows>
+struct Softmax {
+
+    using TensorT = decltype(make_tensor<float>(Shape<Int<kNRows>>{}));
+    TensorT row_max, row_sum;
+
+    CUTLASS_DEVICE Softmax() {};
+
+    template<bool Is_first, bool Check_inf=false, typename Tensor0>
+    __forceinline__ __device__ TensorT max(Tensor0 &acc_s, float softmax_scale_log2) {
+        Tensor scores = make_tensor(acc_s.data(), convert_layout_acc_rowcol(acc_s.layout()));
+        static_assert(decltype(size<0>(scores))::value == kNRows);
+        TensorT scores_scale;
+        if constexpr (Is_first) {
+            reduce_max</*zero_init=*/true>(scores, row_max);
+            cute::fill(scores_scale, 1.f);
+        } else {
+            Tensor scores_max_prev = make_fragment_like(row_max);
+            cute::copy(row_max, scores_max_prev);
+            reduce_max</*zero_init=*/false>(scores, row_max);
+            #pragma unroll
+            for (int mi = 0; mi < size(row_max); ++mi) {
+                float scores_max_cur = row_max(mi);
+                scores_scale(mi) = exp2f((scores_max_prev(mi) - scores_max_cur) * softmax_scale_log2);
+                row_sum(mi) *= scores_scale(mi);
+            }
+        }
+        return scores_scale;
+    };
+
+    template<bool Is_first, typename Tensor0>
+    __forceinline__ __device__ TensorT online_softmax(Tensor0 &acc_s, float softmax_scale_log2) {
+        Tensor scores = make_tensor(acc_s.data(), convert_layout_acc_rowcol(acc_s.layout()));
+        static_assert(decltype(size<0>(scores))::value == kNRows);
+        TensorT scores_scale;
+        if constexpr (Is_first) {
+            reduce_max</*zero_init=*/true>(scores, row_max);
+            scale_apply_exp2(scores, row_max, softmax_scale_log2);
+            reduce_sum</*zero_init=*/true, /*warp_reduce=*/false>(scores, row_sum);
+            cute::fill(scores_scale, 1.f);
+        } else {
+            scale_apply_exp2(scores, row_max, softmax_scale_log2);
+            reduce_sum</*zero_init=*/false, /*warp_reduce=*/false>(scores, row_sum);
+        }
+        return scores_scale;
+    };
+
+    __forceinline__ __device__ TensorT finalize(float softmax_scale_log2) {
+        SumOp<float> sum_op;
+        quad_allreduce_(row_sum, row_sum, sum_op);
+        TensorT scores_scale;
+        #pragma unroll
+        for (int mi = 0; mi < size(row_max); ++mi) {
+            float sum = row_sum(mi);
+            float inv_sum = 1.0f / sum;
+            row_sum(mi) = row_max(mi) * (softmax_scale_log2 * float(M_LN2)) + __logf(sum);
+            scores_scale(mi) = inv_sum;
+        }
+        return scores_scale;
+    };
+
+    template<typename Tensor1>
+    __forceinline__ __device__ void rescale_o(Tensor1 &acc_o, TensorT const &scores_scale) {
+        Tensor acc_o_rowcol = make_tensor(acc_o.data(), convert_layout_acc_rowcol(acc_o.layout()));
+        static_assert(decltype(size<0>(acc_o_rowcol))::value == kNRows);
+        #pragma unroll
+        for (int mi = 0; mi < size(row_max); ++mi) {
+            #pragma unroll
+            for (int ni = 0; ni < size<1>(acc_o_rowcol); ++ni) {
+                acc_o_rowcol(mi, ni) *= scores_scale(mi);
+            }
+        }
+    };
+
+};

custom_ops/gpu_ops/flash_mask_attn/utils.hpp

@@ -0,0 +1,453 @@
+/******************************************************************************
+ * Copyright (c) 2024, Jay Shah, Ganesh Bikshandi, Ying Zhang, Vijay Thakkar, Pradeep Ramani, Tri Dao.
+ ******************************************************************************/
+
+#pragma once
+
+#include <fstream>
+#include <iostream>
+#include <thrust/host_vector.h>
+#include <thrust/device_vector.h>
+
+#include <assert.h>
+#include <stdint.h>
+#include <stdlib.h>
+
+#include <cuda_fp16.h>
+
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 800
+#include <cuda_bf16.h>
+#endif
+
+#include <cute/tensor.hpp>
+#include <cute/arch/cluster_sm90.hpp>  // For cute::elect_one_sync()
+
+#include <cutlass/array.h>
+#include <cutlass/cutlass.h>
+#include <cutlass/numeric_conversion.h>
+#include <cutlass/numeric_types.h>
+
+using namespace cute;
+
+template<typename T>
+struct PackedHalf;
+
+template<>
+struct PackedHalf<cutlass::half_t> {
+    using Type = __half2;
+};
+
+template<>
+struct PackedHalf<cutlass::bfloat16_t> {
+    using Type = nv_bfloat162;
+};
+
+template<typename T>
+__forceinline__ __device__ auto float_2_half2(const float x) {
+    if constexpr (std::is_same<T, cutlass::half_t>::value) {
+        return __float2half2_rn(x);
+    } else {
+        return __float2bfloat162_rn(x);
+    }
+}
+
+
+struct uint16 {
+    uint4 u;
+    uint4 v;
+    uint4 s;
+    uint4 t;
+};
+
+
+struct uint8 {
+    uint4 u;
+    uint4 v;
+};
+
+template<int BYTES>
+struct BytesToType {};
+
+template<>
+struct BytesToType<64> {
+    using Type = uint16;
+    static_assert(sizeof(Type) == 64);
+};
+
+template<>
+struct BytesToType<32> {
+    using Type = uint8;
+    static_assert(sizeof(Type) == 32);
+};
+
+template<>
+struct BytesToType<16> {
+    using Type = uint4;
+    static_assert(sizeof(Type) == 16);
+};
+
+template<>
+struct BytesToType<8> {
+    using Type = uint64_t;
+    static_assert(sizeof(Type) == 8);
+};
+
+template<>
+struct BytesToType<4> {
+    using Type = uint32_t;
+    static_assert(sizeof(Type) == 4);
+};
+
+template<>
+struct BytesToType<2> {
+    using Type = uint16_t;
+    static_assert(sizeof(Type) == 2);
+};
+
+template<>
+struct BytesToType<1> {
+    using Type = uint8_t;
+    static_assert(sizeof(Type) == 1);
+};
+
+template<typename Elt_type, uint32_t NUM_ELT>
+struct Vec {
+
+    enum { BYTES = NUM_ELT * sizeof(Elt_type) };
+
+    using Vec_type = typename BytesToType<BYTES>::Type;
+
+    using Alias_type = union {
+        Vec_type vec;
+        Elt_type elt[NUM_ELT];
+    };
+
+    Alias_type data;
+
+    inline __device__ Vec() {}
+
+    template<typename S>
+    inline __device__ void to(Vec<S, NUM_ELT> &other) {
+        #pragma unroll
+        for( int it = 0; it < NUM_ELT; it++ ) {
+            other.data.elt[it] = S(this->data.elt[it]);
+        }
+    }
+
+    template<typename Op>
+    inline __device__ void assign(const Op &op) {
+        #pragma unroll
+        for( int it = 0; it < NUM_ELT; it++ ) {
+            this->data.elt[it] = op(it);
+        }
+    }
+
+    inline __device__ void load_from(const void *base_ptr) {
+        this->data.vec = *reinterpret_cast<const Vec_type *>(base_ptr);
+    }
+
+
+    inline __device__ void store_to(void *base_ptr) {
+        *reinterpret_cast<Vec_type *>(base_ptr) = this->data.vec;
+    }
+
+    inline __device__ void add(const Vec<Elt_type, NUM_ELT> &other) {
+        static_assert(NUM_ELT % 2 == 0);
+        using type = typename PackedHalf<Elt_type>::Type;
+        #pragma unroll
+        for (int it = 0; it < NUM_ELT / 2; it++) {
+            type b = *reinterpret_cast<const type *>(other.data.elt + it * 2);
+            *reinterpret_cast<type *>(this->data.elt + it * 2) += b;
+        }
+    }
+
+    inline __device__ void fma(const Vec<Elt_type, NUM_ELT> &scale, const Vec<Elt_type, NUM_ELT> &bias) {
+        static_assert(NUM_ELT % 2 == 0);
+        using type = typename PackedHalf<Elt_type>::Type;
+        #pragma unroll
+        for (int it = 0; it < NUM_ELT / 2; it++) {
+            type a = *reinterpret_cast<const type *>(scale.data.elt + it * 2);
+            type b = *reinterpret_cast<const type *>(bias.data.elt + it * 2);
+            *reinterpret_cast<type *>(this->data.elt + it * 2) += a * b;
+        }
+    }
+
+    inline __device__ void set_zero() {
+        constexpr int size = sizeof(Vec_type) / sizeof(int);
+        #pragma unroll
+        for (int i = 0; i < size; ++i) {
+            (reinterpret_cast<int *>(this->data.elt))[i] = 0;
+        }
+    }
+};
+
+template<typename T, int PackSize>
+inline __device__ void apply_rotary_embedding(Vec<T, PackSize>& vec, Vec<float, PackSize / 2>& cos, Vec<float, PackSize / 2>& sin) {
+    static_assert(PackSize % 2 == 0);
+    #pragma unroll
+    for (int i = 0; i < PackSize / 2; i++) {
+        const float cos_inv_freq = cos.data.elt[i];
+        const float sin_inv_freq = sin.data.elt[i];
+        const float v1 = static_cast<float>(vec.data.elt[2 * i]);
+        const float v2 = static_cast<float>(vec.data.elt[2 * i + 1]);
+        vec.data.elt[2 * i] = static_cast<T>(cos_inv_freq * v1 - sin_inv_freq * v2);
+        vec.data.elt[2 * i + 1] = static_cast<T>(sin_inv_freq * v1 + cos_inv_freq * v2);
+    }
+}
+
+template <typename Tensor>
+__forceinline__ __device__ void app_mask(
+        Tensor &tSrS,
+        const int *mask,
+        const int &mask_row_id,
+        const int &col_base) {
+    const float mask_value = -1000000.0f;
+    for (int i = 0; i < size(tSrS); i+=8) {
+        const int col = i * 2 + col_base;
+        if (col >= mask[mask_row_id]) {
+            tSrS(i) = mask_value;
+        }
+        if (col + 1 >= mask[mask_row_id]) {
+            tSrS(i + 1) = mask_value;
+        }
+        if (col >= mask[mask_row_id + 8]) {
+            tSrS(i + 2) = mask_value;
+        }
+        if (col + 1 >= mask[mask_row_id + 8]) {
+            tSrS(i + 3) = mask_value;
+        }
+        if (col + 8 >= mask[mask_row_id]) {
+            tSrS(i + 4) = mask_value;
+        }
+        if (col + 9 >= mask[mask_row_id]) {
+            tSrS(i + 5) = mask_value;
+        }
+        if (col + 8 >= mask[mask_row_id + 8]) {
+            tSrS(i + 6) = mask_value;
+        }
+        if (col + 9 >= mask[mask_row_id + 8]) {
+            tSrS(i + 7) = mask_value;
+        }
+    }
+}
+
+template<typename T>
+struct HalfMax;
+template<>
+struct HalfMax<cutlass::half_t> {
+    inline __device__ __half2 operator()(const __half2 x, const __half2 y) {
+        __half2 res;
+        asm volatile("max.f16x2 %0, %1, %2;\n" :
+            "=r"(*reinterpret_cast<uint32_t*>(&res)) :
+            "r"(*reinterpret_cast<const uint32_t*>(&x)),
+            "r"(*reinterpret_cast<const uint32_t*>(&y)));
+        return res;
+    }
+};
+
+template<>
+struct HalfMax<cutlass::bfloat16_t> {
+    inline __device__ nv_bfloat162 operator()(const nv_bfloat162 x, const nv_bfloat162 y) {
+        nv_bfloat162 res;
+        asm volatile("max.bf16x2 %0, %1, %2;\n" :
+            "=r"(*reinterpret_cast<uint32_t*>(&res)) :
+            "r"(*reinterpret_cast<const uint32_t*>(&x)),
+            "r"(*reinterpret_cast<const uint32_t*>(&y)));
+        return res;
+    }
+};
+
+template<typename T>
+struct HalfMin;
+template<>
+struct HalfMin<cutlass::half_t> {
+    inline __device__ __half2 operator()(const __half2 x, const __half2 y) {
+        __half2 res;
+        asm volatile("min.f16x2 %0, %1, %2;\n" :
+            "=r"(*reinterpret_cast<uint32_t*>(&res)) :
+            "r"(*reinterpret_cast<const uint32_t*>(&x)),
+            "r"(*reinterpret_cast<const uint32_t*>(&y)));
+        return res;
+    }
+};
+
+template<>
+struct HalfMin<cutlass::bfloat16_t> {
+    inline __device__ nv_bfloat162 operator()(const nv_bfloat162 x, const nv_bfloat162 y) {
+        nv_bfloat162 res;
+        asm volatile("min.bf16x2 %0, %1, %2;\n" :
+            "=r"(*reinterpret_cast<uint32_t*>(&res)) :
+            "r"(*reinterpret_cast<const uint32_t*>(&x)),
+            "r"(*reinterpret_cast<const uint32_t*>(&y)));
+        return res;
+    }
+};
+
+template <bool Is_even_MN=true, typename TiledCopy, typename Engine0, typename Layout0, typename Engine1, typename Layout1, typename Engine2, typename Layout2>
+__forceinline__ __device__ void copy(
+        TiledCopy tiled_copy, Tensor<Engine0, Layout0> const &S,
+        Tensor<Engine1, Layout1> &D,
+        Tensor<Engine2, Layout2> const &identity_MN,
+        const int max_MN = 0) {
+    CUTE_STATIC_ASSERT_V(rank(S) == Int<3>{});
+    CUTE_STATIC_ASSERT_V(rank(D) == Int<3>{});
+    CUTE_STATIC_ASSERT_V(size<0>(S) == size<0>(D));                     // MMA
+    CUTE_STATIC_ASSERT_V(size<1>(S) == size<1>(D));                     // MMA_M
+    CUTE_STATIC_ASSERT_V(size<2>(S) == size<2>(D));                     // MMA_K
+    #pragma unroll
+    for (int m = 0; m < size<1>(S); ++m) {
+        if (Is_even_MN || get<0>(identity_MN(0, m, 0)) < max_MN) {
+            #pragma unroll
+            for (int k = 0; k < size<2>(S); ++k) {
+                cute::copy(tiled_copy, S(_, m, k), D(_, m, k));
+            }
+        }
+    }
+}
+
+template <typename To_type, typename Engine, typename Layout>
+inline __device__ auto convert_type(Tensor<Engine, Layout> const &tensor) {
+    using From_type = typename Engine::value_type;
+    constexpr int numel = decltype(size(tensor))::value;
+    cutlass::NumericArrayConverter<To_type, From_type, numel> convert_op;
+    auto frag = convert_op(*reinterpret_cast<const cutlass::Array<From_type, numel> *>(tensor.data()));
+    return make_tensor(make_rmem_ptr<To_type>(&frag), tensor.layout());
+}
+
+template<typename T, typename ReductionOp, int block_size>
+__inline__ __device__ T BlockAllReduce(T val) {
+    typedef cub::BlockReduce<T, block_size> BlockReduce;
+    __shared__ typename BlockReduce::TempStorage temp_storage;
+    __shared__ T result_broadcast;
+    T result = BlockReduce(temp_storage).Reduce(val, ReductionOp());
+    if (threadIdx.x == 0) { result_broadcast = result; }
+    __syncthreads();
+    return result_broadcast;
+}
+
+template<typename T, int block_size>
+__inline__ __device__ T BlockScanSum(T val) {
+    typedef cub::BlockScan<int, block_size> BlockScanT;
+    __shared__ typename BlockScanT::TempStorage temp_storage;
+
+    int aggregate;
+    BlockScanT(temp_storage).ExclusiveSum(val, val, aggregate);
+    __syncthreads();
+    return val;
+}
+
+
+
+template<typename T>
+struct MaxOp {
+__device__ __forceinline__ T operator()(T const & x, T const & y) { return x > y ? x : y; }
+};
+
+template <>
+struct MaxOp<float> {
+// This is slightly faster
+__device__ __forceinline__ float operator()(float const &x, float const &y) { return max(x, y); }
+};
+
+template<typename T>
+struct MinOp {
+__device__ __forceinline__ T operator()(T const & x, T const & y) { return x < y ? x : y; }
+};
+
+template <>
+struct MinOp<float> {
+// This is slightly faster
+__device__ __forceinline__ float operator()(float const &x, float const &y) { return min(x, y); }
+};
+
+
+template<typename T>
+struct SumOp {
+__device__ __forceinline__ T operator()(T const & x, T const & y) { return x + y; }
+};
+
+template<typename MMA_traits, typename Layout>
+__forceinline__ __device__ auto convert_layout_acc_Aregs(Layout acc_layout) {
+    using X = Underscore;
+    if constexpr (decltype(rank<0>(acc_layout))::value == 3) {  // SM90
+        static_assert(decltype(size<0, 0>(acc_layout))::value == 2);
+        static_assert(decltype(size<0, 1>(acc_layout))::value == 2);
+        static_assert(decltype(rank(acc_layout))::value == 3);
+        static_assert(decltype(rank(get<0>(acc_layout)))::value == 3);
+        auto l = logical_divide(get<0>(acc_layout), Shape<X, X, _2>{});  // (2, 2, (2, N / 16)))
+        return make_layout(make_layout(get<0>(l), get<1>(l), get<2, 0>(l)), get<1>(acc_layout), make_layout(get<2, 1>(l), get<2>(acc_layout)));
+    } else {  // SM80
+        static_assert(decltype(size<0>(acc_layout))::value == 4);
+        static_assert(decltype(rank(acc_layout))::value == 3);
+        constexpr int mma_shape_K = get<2>(typename MMA_traits::Shape_MNK{});
+        static_assert(mma_shape_K == 8 || mma_shape_K == 16);
+        if constexpr (mma_shape_K == 8) {
+            return acc_layout;
+        } else {
+            auto l = logical_divide(acc_layout, Shape<X, X, _2>{});  // (4, MMA_M, (2, MMA_N / 2)))
+            return make_layout(make_layout(get<0>(l), get<2, 0>(l)), get<1>(l), get<2, 1>(l));
+        }
+    }
+};
+
+template <bool zero_init=false, int wg_wait=0, bool arrive=true, bool commit=true, typename Tensor0, typename Tensor1, typename Tensor2,
+          typename TiledMma>
+__forceinline__ __device__ void gemm(TiledMma &tiled_mma, Tensor0 const &tCrA, Tensor1 const &tCrB, Tensor2 &tCrC) {
+    constexpr bool Is_RS = !cute::is_base_of<cute::GMMA::DescriptorIterator, typename TiledMma::FrgTypeA>::value;
+    // Need to cast away const on tCrA since warpgroup_fence_operand doesn't take const
+    if constexpr (Is_RS) { warpgroup_fence_operand(const_cast<Tensor0 &>(tCrA)); }
+    warpgroup_fence_operand(tCrC);
+    if constexpr (arrive) {
+        warpgroup_arrive();
+    }
+    if constexpr (zero_init) {
+        tiled_mma.accumulate_ = GMMA::ScaleOut::Zero;
+        // Unroll the K mode manually to set scale D to 1
+        CUTLASS_PRAGMA_UNROLL
+        for (int k_block = 0; k_block < size<2>(tCrA); ++k_block) {
+          cute::gemm(tiled_mma, tCrA(_,_,k_block), tCrB(_,_,k_block), tCrC);
+          tiled_mma.accumulate_ = GMMA::ScaleOut::One;
+        }
+    } else {
+        // cute::gemm(tiled_mma, tCrA, tCrB, tCrC);
+        // Unroll the K mode manually to set scale D to 1
+        CUTLASS_PRAGMA_UNROLL
+        for (int k_block = 0; k_block < size<2>(tCrA); ++k_block) {
+          cute::gemm(tiled_mma, tCrA(_,_,k_block), tCrB(_,_,k_block), tCrC);
+          tiled_mma.accumulate_ = GMMA::ScaleOut::One;
+        }
+    }
+    if constexpr (commit) {
+        warpgroup_commit_batch();
+    }
+    if constexpr (wg_wait >= 0) { warpgroup_wait<wg_wait>(); }
+    warpgroup_fence_operand(tCrC);
+    if constexpr (Is_RS) { warpgroup_fence_operand(const_cast<Tensor0 &>(tCrA)); }
+}
+
+
+template<typename Layout>
+__forceinline__ __device__ auto convert_layout_acc_rowcol(Layout acc_layout) {
+    if constexpr (decltype(rank<0>(acc_layout))::value == 3) {  // SM90
+        static_assert(decltype(size<0, 0>(acc_layout))::value == 2);
+        static_assert(decltype(size<0, 1>(acc_layout))::value == 2);
+        static_assert(decltype(rank(acc_layout))::value == 3);
+        auto l = acc_layout;
+        return make_layout(make_layout(get<0, 1>(l), get<1>(l)), make_layout(get<0, 0>(l), get<0, 2>(l), get<2>(l)));
+    } else {  // SM80
+        static_assert(decltype(size<0>(acc_layout))::value == 4);
+        static_assert(decltype(rank(acc_layout))::value == 3);
+        auto l = logical_divide(acc_layout, Shape<_2>{});  // ((2, 2), MMA_M, MMA_N)
+        return make_layout(make_layout(get<0, 1>(l), get<1>(l)), make_layout(get<0, 0>(l), get<2>(l)));
+    }
+};
+
+template<typename T, typename ReductionOp, int thread_group_width = 32>
+__inline__ __device__ T WarpAllReduce(T val) {
+    ReductionOp op;
+    #pragma unroll
+    for (int mask = thread_group_width / 2; mask > 0; mask /= 2) {
+        val = op(val, __shfl_xor_sync(0xffffffff, val, mask));
+    }
+    return val;
+}

custom_ops/gpu_ops/get_img_boundaries.cc

@@ -0,0 +1,60 @@
+// Copyright (c) 2025 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "paddle/extension.h"
+
+std::vector<paddle::Tensor> GetImgBoundaries(const paddle::Tensor& task_input_ids,
+                                             const paddle::Tensor& grid_thw,
+                                             const int64_t image_patch_id) {
+    // All tensor in cpu
+    auto input_ids_ptr = task_input_ids.data<int64_t>();
+    int64_t seq_lens_origin = task_input_ids.numel();
+    auto grid_thw_ptr = grid_thw.data<int64_t>();
+
+    int token_times = 4;
+    int token_idx = 0;
+    int image_idx = 0;
+    std::vector<int> img_boundaries, img_nums;
+    img_boundaries.emplace_back(0);
+    img_nums.emplace_back(0);
+    while (token_idx < seq_lens_origin) {
+        if (input_ids_ptr[token_idx] != image_patch_id) {
+            do {
+                token_idx++;
+            } while (token_idx < seq_lens_origin && input_ids_ptr[token_idx] != image_patch_id);
+        } else {
+            int cur_image_token_len = (grid_thw_ptr[image_idx * 3 + 1] * grid_thw_ptr[image_idx * 3 + 2]) / token_times;
+            image_idx++;
+            token_idx += cur_image_token_len;
+        }
+        img_boundaries.emplace_back(token_idx);
+        img_nums.emplace_back(image_idx);
+    }
+
+    int64_t num_img_boundaries = static_cast<int64_t>(img_boundaries.size());
+    auto out = paddle::full({2, num_img_boundaries}, 0, paddle::DataType::INT64, paddle::CPUPlace());
+
+    for (int i = 0; i < num_img_boundaries; i++) {
+        out.data<int64_t>()[i] = img_boundaries[i];
+        out.data<int64_t>()[num_img_boundaries + i] = img_nums[i];
+    }
+
+    return {out};
+}
+
+PD_BUILD_OP(get_img_boundaries)
+    .Inputs({"task_input_ids", "grid_thw"})
+    .Attrs({"image_patch_id: int64_t"})
+    .Outputs({"img_boundaries"})
+    .SetKernelFn(PD_KERNEL(GetImgBoundaries));

custom_ops/gpu_ops/get_output_ep.cc

@@ -109,11 +109,11 @@ void GetOutputEp(const paddle::Tensor& x,
     return;
 }
 
-void GetOutputStatic(const paddle::Tensor& x, int64_t rank_id, bool wait_flag) {
+void GetOutputEPStatic(const paddle::Tensor& x, int64_t rank_id, bool wait_flag) {
     GetOutputEp(x, rank_id, wait_flag, 1);
 }
 
-void GetOutputDynamic(const paddle::Tensor& x,
+void GetOutputEPDynamic(const paddle::Tensor& x,
                       int64_t rank_id,
                       bool wait_flag,
                       int msg_queue_id) {
@@ -125,11 +125,11 @@ PD_BUILD_STATIC_OP(get_output_ep)
     .Attrs({"rank_id: int64_t", "wait_flag: bool"})
     .Outputs({"x_out"})
     .SetInplaceMap({{"x", "x_out"}})
-    .SetKernelFn(PD_KERNEL(GetOutputStatic));
+    .SetKernelFn(PD_KERNEL(GetOutputEPStatic));
 
 PD_BUILD_STATIC_OP(get_output_ep_dynamic)
     .Inputs({"x"})
     .Attrs({"rank_id: int64_t", "wait_flag: bool", "msg_queue_id: int"})
     .Outputs({"x_out"})
     .SetInplaceMap({{"x", "x_out"}})
-    .SetKernelFn(PD_KERNEL(GetOutputDynamic));
+    .SetKernelFn(PD_KERNEL(GetOutputEPDynamic));

custom_ops/gpu_ops/get_padding_offset.cu

@@ -46,7 +46,11 @@ __global__ void GetPaddingOffsetKernel(int *batch_id_per_token,
     const int ti = threadIdx.x;
     int cum_offset = bi == 0 ? 0 : cum_offsets[bi - 1];
     for (int i = ti; i < seq_lens[bi]; i += blockDim.x) {
+#ifdef PADDLE_WITH_HIP
+        batch_id_per_token[bi * max_seq_len - cum_offset + i] = cum_offset;
+#else
         batch_id_per_token[bi * max_seq_len - cum_offset + i] = bi;
+#endif
     }
     if (ti == 0) {
         cum_offsets_out[bi] = cum_offset;
@@ -101,7 +105,6 @@ std::vector<paddle::Tensor> GetPaddingOffset(const paddle::Tensor &input_ids,
         cum_offsets_out.data<int>(),
         seq_length);
     return {x_remove_padding,
-            cum_offsets_out,
             batch_id_per_token,
             cu_seqlens_q,
             cu_seqlens_k};  // , enc_token_num, dec_token_num};
@@ -114,7 +117,7 @@ std::vector<std::vector<int64_t>> GetPaddingOffsetInferShape(
     const std::vector<int64_t> &seq_len_shape) {
     int64_t bsz = seq_len_shape[0];
     int64_t seq_len = input_ids_shape[1];
-    return {{-1}, {bsz}, {-1}, {bsz + 1}, {bsz + 1}};
+    return {{-1}, {-1}, {bsz + 1}, {bsz + 1}};
 }
 
 std::vector<paddle::DataType> GetPaddingOffsetInferDtype(
@@ -123,7 +126,6 @@ std::vector<paddle::DataType> GetPaddingOffsetInferDtype(
     const paddle::DataType &token_num_dtype,
     const paddle::DataType &seq_len_dtype) {
     return {input_ids_dtype,
-            seq_len_dtype,
             seq_len_dtype,
             seq_len_dtype,
             seq_len_dtype};
@@ -132,7 +134,6 @@ std::vector<paddle::DataType> GetPaddingOffsetInferDtype(
 PD_BUILD_STATIC_OP(get_padding_offset)
     .Inputs({"input_ids", "token_num", "cum_offsets", "seq_len"})
     .Outputs({"x_remove_padding",
-              "cum_offsets_out",
               "batch_id_per_token",
               "cu_seqlens_q",
               "cu_seqlens_k"})

custom_ops/gpu_ops/helper.h

@@ -151,6 +151,34 @@ inline int GetGPUComputeCapability(int id) {
 
 #endif
 
+#ifndef FP8_E4M3_MAX
+#define FP8_E4M3_MAX 448.0
+#endif
+
+#ifndef DISPATCH_FLOAT_FP6_DTYPE
+#define DISPATCH_FLOAT_FP6_DTYPE(pd_dtype, c_type, ...)           \
+    switch (pd_dtype) {                                           \
+      case phi::DataType::FLOAT32: {                           \
+        using c_type = float;                                  \
+        __VA_ARGS__                                            \
+        break;                                                 \
+      }                                                        \
+      case phi::DataType::BFLOAT16: {                          \
+        using c_type = phi::dtype::bfloat16;                   \
+        __VA_ARGS__                                            \
+        break;                                                 \
+      }                                                        \
+      case phi::DataType::FLOAT16: {                          \
+        using c_type = phi::dtype::float16;                 \
+        __VA_ARGS__                                            \
+        break;                                                 \
+      }                                                        \
+      default: {                                               \
+        PD_THROW("Only supported attr of input type in [fp32, fp16, bf16].");  \
+      }                                                        \
+    }
+#endif
+
 inline constexpr uint32_t next_pow_2(uint32_t const num) {
   if (num <= 1)
     return num;
@@ -193,6 +221,12 @@ public:
   typedef uint8_t data_t;
 };
 
+template <> class PDTraits<paddle::DataType::FLOAT8_E4M3FN> {
+public:
+  typedef __nv_fp8_e4m3 DataType;
+  typedef paddle::float8_e4m3fn data_t;
+};
+
 template <typename T, int Size> struct alignas(sizeof(T) * Size) AlignedVector {
   T val[Size];
 
@@ -509,6 +543,7 @@ static void PrintMatrix3(const T *mat_d, int num, std::string name) {
 }
 
 #ifndef PADDLE_WITH_HIP
+#ifndef PADDLE_WITH_CUSTOM_DEVICE_METAX_GPU
 __forceinline__ __device__ uint32_t ld_flag_acquire(uint32_t *flag_addr,
                                                     int mode = 0) {
   uint32_t flag;
@@ -541,7 +576,7 @@ __forceinline__ __device__ void st_flag_release(uint32_t *flag_addr,
                  "l"(flag_addr));
   }
 }
-
+#endif
 inline int get_cuda_max_shared_memory_per_block_opt_in(int const device) {
   int max_shared_mem_per_block_opt_in = 0;
   cudaDeviceGetAttribute(&max_shared_mem_per_block_opt_in,
@@ -556,3 +591,28 @@ inline int GetSMVersion() {
   return sm_version;
 
 }
+
+__device__ __forceinline__ float warpReduceMax(float value) {
+  value = fmaxf(value, __shfl_xor_sync(0xffffffff, value, 16));
+  value = fmaxf(value, __shfl_xor_sync(0xffffffff, value, 8));
+  value = fmaxf(value, __shfl_xor_sync(0xffffffff, value, 4));
+  value = fmaxf(value, __shfl_xor_sync(0xffffffff, value, 2));
+  value = fmaxf(value, __shfl_xor_sync(0xffffffff, value, 1));
+  return value;
+}
+
+__device__ __forceinline__ float blockReduceMax(float value) {
+  static __shared__ float warpLevelMaxs[WARP_SIZE];
+  const int laneId = threadIdx.x % WARP_SIZE;
+  const int warpId = threadIdx.x / WARP_SIZE;
+
+  value = warpReduceMax(value);
+
+  if (laneId == 0) warpLevelMaxs[warpId] = value;
+  __syncthreads();
+
+  value = (threadIdx.x < blockDim.x / WARP_SIZE) ? warpLevelMaxs[laneId] : 0;
+  if (warpId == 0) value = warpReduceMax(value);
+
+  return value;
+}

custom_ops/gpu_ops/machete/generate.py

@@ -0,0 +1,574 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import itertools
+import math
+import os
+import shutil
+import sys
+from collections.abc import Iterable
+from copy import deepcopy
+from dataclasses import dataclass, fields
+from functools import reduce
+from typing import Optional, Union
+
+import jinja2
+
+cur_dir = os.path.dirname(os.path.abspath(__file__))
+p = os.path.abspath(os.path.join(cur_dir, "../../third_party/cutlass/python"))
+sys.path.insert(0, p)
+
+from cutlass_library import (
+    EpilogueScheduleTag,
+    EpilogueScheduleType,
+    TileSchedulerTag,
+    TileSchedulerType,
+)
+
+# yapf conflicts with isort for this block
+# yapf: disable
+from machete_cutlass_library_extension import (
+    DataType,
+    MACHETEDataType,
+    MACHETEDataTypeMACHETEScalarTypeTag,
+    MACHETEDataTypeNames,
+    MACHETEDataTypePaddleDataTypeTag,
+    MACHETEDataTypeSize,
+    MACHETEDataTypeTag,
+    MACHETEKernelScheduleTag,
+    MixedInputKernelScheduleType,
+)
+
+# yapf: enable
+
+#
+#   Generator templating
+#
+
+DISPATCH_TEMPLATE = """
+#include "../machete_mm_launcher.cuh"
+
+namespace machete {
+
+{% for impl_config in impl_configs %}
+{% set type_sig = gen_type_sig(impl_config.types) -%}
+{% for s in impl_config.schedules %}
+extern paddle::Tensor impl_{{type_sig}}_sch_{{gen_sch_sig(s)}}(MMArgs);
+{%- endfor %}
+
+paddle::Tensor mm_dispatch_{{type_sig}}(MMArgs args) {
+  [[maybe_unused]] auto M = args.A.shape()[0];
+  [[maybe_unused]] auto N = args.B.shape()[1];
+  [[maybe_unused]] auto K = args.A.shape()[1];
+
+  if (!args.maybe_schedule) {
+    {%- for cond, s in impl_config.heuristic %}
+    {%if cond is not none%}if ({{cond}})
+    {%- else %}else
+    {%- endif %}
+        return impl_{{type_sig}}_sch_{{ gen_sch_sig(s) }}(args);{% endfor %}
+  }
+
+  {%- for s in impl_config.schedules %}
+  if (*args.maybe_schedule == "{{ gen_sch_sig(s) }}")
+    return impl_{{type_sig}}_sch_{{ gen_sch_sig(s) }}(args);
+  {%- endfor %}
+  PADDLE_ENFORCE(false, "machete_gemm(..) is not implemented ");
+}
+{%- endfor %}
+
+
+static inline std::optional<paddle::DataType> maybe_scalartype(
+    std::optional<paddle::Tensor> const& t) {
+    if (!t) {
+      return std::nullopt;
+    } else {
+      return t->dtype();
+    };
+}
+
+paddle::Tensor mm_dispatch(MMArgs args) {
+  auto out_type = args.maybe_out_type.value_or(args.A.dtype());
+  auto a_type = args.A.dtype();
+  auto maybe_g_scales_type = maybe_scalartype(args.maybe_group_scales);
+  auto maybe_g_zeros_type = maybe_scalartype(args.maybe_group_zeros);
+  auto maybe_ch_scales_type = maybe_scalartype(args.maybe_channel_scales);
+  auto maybe_tok_scales_type = maybe_scalartype(args.maybe_token_scales);
+
+  {% for impl_config in impl_configs %}
+  {% set t = impl_config.types -%}
+  {% set type_sig = gen_type_sig(t) -%}
+  if (args.b_type == {{MACHETEScalarTypeTag[t.b]}}
+      && a_type == {{PaddleTypeTag[t.a]}}
+      && out_type == {{PaddleTypeTag[t.out]}}
+      && {%if t.b_group_scale != void -%}
+      maybe_g_scales_type == {{PaddleTypeTag[t.b_group_scale]}}
+      {%- else %}!maybe_g_scales_type{%endif%}
+      && {%if t.b_group_zeropoint != void -%}
+      maybe_g_zeros_type == {{PaddleTypeTag[t.b_group_zeropoint]}}
+      {%- else %}!maybe_g_zeros_type{%endif%}
+      && {%if t.b_channel_scale != void -%}
+      maybe_ch_scales_type == {{PaddleTypeTag[t.b_channel_scale]}}
+      {%- else %}!maybe_ch_scales_type{%endif%}
+      && {%if t.a_token_scale != void -%}
+      maybe_tok_scales_type == {{PaddleTypeTag[t.a_token_scale]}}
+      {%- else %}!maybe_tok_scales_type{%endif%}
+  ) {
+      return mm_dispatch_{{type_sig}}(args);
+  }
+  {%- endfor %}
+
+  PADDLE_ENFORCE(
+    false, "machete_mm(..) is not implemented "
+    "; implemented types are: \\n",
+    {%- for impl_config in impl_configs %}
+    {% set t = impl_config.types -%}
+    "\\t{{gen_type_option_name(t)}}\\n",
+    {%- endfor %}
+    "");
+}
+
+std::vector<std::string> supported_schedules_dispatch(
+    SupportedSchedulesArgs args) {
+    auto out_type = args.maybe_out_type.value_or(args.a_type);
+
+    {% for impl_config in impl_configs %}
+    {% set t = impl_config.types -%}
+    {% set schs = impl_config.schedules -%}
+    if (args.b_type == {{MACHETEScalarTypeTag[t.b]}}
+        && args.a_type == {{PaddleTypeTag[t.a]}}
+        && out_type == {{PaddleTypeTag[t.out]}}
+        && {%if t.b_group_scale != void -%}
+        args.maybe_group_scales_type == {{PaddleTypeTag[t.b_group_scale]}}
+        {%- else %}!args.maybe_group_scales_type{%endif%}
+        && {%if t.b_group_zeropoint != void-%}
+        args.maybe_group_zeros_type == {{PaddleTypeTag[t.b_group_zeropoint]}}
+        {%- else %}!args.maybe_group_zeros_type{%endif%}
+    ) {
+        return {
+            {%- for s in impl_config.schedules %}
+            "{{gen_sch_sig(s)}}"{% if not loop.last %},{% endif %}
+            {%- endfor %}
+        };
+    }
+    {%- endfor %}
+
+    return {};
+};
+
+}; // namespace machete
+"""
+
+IMPL_TEMPLATE = """
+#include "../machete_mm_launcher.cuh"
+
+namespace machete {
+
+{% for sch in unique_schedules(impl_configs) %}
+{% set sch_sig = gen_sch_sig(sch) -%}
+struct sch_{{sch_sig}} {
+  using TileShapeNM = Shape<{{
+      to_cute_constant(sch.tile_shape_mn)|join(', ')}}>;
+  using ClusterShape = Shape<{{
+      to_cute_constant(sch.cluster_shape_mnk)|join(', ')}}>;
+  // TODO: Reimplement
+  // using KernelSchedule   = {{KernelScheduleTag[sch.kernel_schedule]}};
+  using EpilogueSchedule = {{EpilogueScheduleTag[sch.epilogue_schedule]}};
+  using TileScheduler    = {{TileSchedulerTag[sch.tile_scheduler]}};
+  using EpilogueTileType = cutlass::epilogue::collective::EpilogueTileAuto;
+};
+{% endfor %}
+
+{% for impl_config in impl_configs %}
+{% set t = impl_config.types -%}
+{% set schs = impl_config.schedules -%}
+{% set type_sig = gen_type_sig(t) -%}
+
+template<typename Sch>
+using Kernel_{{type_sig}} = MacheteKernelTemplate<
+  {{DataTypeTag[t.a]}},  // ElementA
+  {{DataTypeTag[t.b]}},  // ElementB
+  {{DataTypeTag[t.out]}},  // ElementD
+  {{DataTypeTag[t.accumulator]}}, // Accumulator
+  {{DataTypeTag[t.b_group_scale]}}, // GroupScaleT
+  {{DataTypeTag[t.b_group_zeropoint]}}, // GroupZeroT
+  {{DataTypeTag[t.b_channel_scale]}}, // ChannelScaleT
+  {{DataTypeTag[t.a_token_scale]}}, // TokenScaleT
+  cutlass::gemm::KernelTmaWarpSpecializedCooperative,
+  Sch>;
+
+{% for sch in schs %}
+{% set sch_sig = gen_sch_sig(sch) -%}
+paddle::Tensor
+impl_{{type_sig}}_sch_{{sch_sig}}(MMArgs args) {
+  return run_impl<Kernel_{{type_sig}}<sch_{{sch_sig}}>>(args);
+}
+{%- endfor %}
+{%- endfor %}
+
+}; // namespace machete
+"""
+
+PREPACK_TEMPLATE = """
+#include "../machete_prepack_launcher.cuh"
+
+namespace machete {
+
+paddle::Tensor prepack_B_dispatch(PrepackBArgs args) {
+  auto convert_type = args.maybe_group_scales_type.value_or(args.a_type);
+  {%- for t in types %}
+  {% set b_type = unsigned_type_with_bitwidth(t.b_num_bits) %}
+  if (args.a_type == {{PaddleTypeTag[t.a]}}
+      && args.b_type.size_bits() == {{t.b_num_bits}}
+      && convert_type == {{PaddleTypeTag[t.convert]}}) {
+    return prepack_impl<
+      PrepackedLayoutBTemplate<
+        {{DataTypeTag[t.a]}}, // ElementA
+        {{DataTypeTag[b_type]}}, // ElementB
+        {{DataTypeTag[t.convert]}}, // ElementConvert
+        {{DataTypeTag[t.accumulator]}}, // Accumulator
+        cutlass::layout::ColumnMajor,
+        cutlass::gemm::KernelTmaWarpSpecializedCooperative>
+    >(args.B);
+  }
+  {%- endfor %}
+
+  PADDLE_ENFORCE(false,
+    "prepack_B_dispatch(..) is not implemented");
+}
+
+}; // namespace machete
+"""
+
+TmaMI = MixedInputKernelScheduleType.TmaWarpSpecializedCooperative
+TmaCoop = EpilogueScheduleType.TmaWarpSpecializedCooperative
+
+
+@dataclass(frozen=True)
+class ScheduleConfig:
+    tile_shape_mn: tuple[int, int]
+    cluster_shape_mnk: tuple[int, int, int]
+    kernel_schedule: MixedInputKernelScheduleType
+    epilogue_schedule: EpilogueScheduleType
+    tile_scheduler: TileSchedulerType
+
+
+@dataclass(frozen=True)
+class TypeConfig:
+    a: DataType
+    b: Union[DataType, MACHETEDataType]
+    b_group_scale: DataType
+    b_group_zeropoint: DataType
+    b_channel_scale: DataType
+    a_token_scale: DataType
+    out: DataType
+    accumulator: DataType
+
+
+@dataclass(frozen=True)
+class PrepackTypeConfig:
+    a: DataType
+    b_num_bits: int
+    convert: DataType
+    accumulator: DataType
+
+
+@dataclass
+class ImplConfig:
+    types: TypeConfig
+    schedules: list[ScheduleConfig]
+    heuristic: list[tuple[Optional[str], ScheduleConfig]]
+
+
+def generate_sch_sig(schedule_config: ScheduleConfig) -> str:
+    tile_shape = f"{schedule_config.tile_shape_mn[0]}x{schedule_config.tile_shape_mn[1]}"
+    cluster_shape = (
+        f"{schedule_config.cluster_shape_mnk[0]}"
+        + f"x{schedule_config.cluster_shape_mnk[1]}"
+        + f"x{schedule_config.cluster_shape_mnk[2]}"
+    )
+    kernel_schedule = MACHETEKernelScheduleTag[schedule_config.kernel_schedule].split("::")[-1]
+    epilogue_schedule = EpilogueScheduleTag[schedule_config.epilogue_schedule].split("::")[-1]
+    tile_scheduler = TileSchedulerTag[schedule_config.tile_scheduler].split("::")[-1]
+
+    return f"{tile_shape}_{cluster_shape}_{kernel_schedule}" + f"_{epilogue_schedule}_{tile_scheduler}"
+
+
+# mostly unique shorter sch_sig
+def generate_terse_sch_sig(schedule_config: ScheduleConfig) -> str:
+    kernel_terse_names_replace = {
+        "KernelTmaWarpSpecializedCooperative": "TmaMI_",
+        "TmaWarpSpecializedCooperative_": "TmaCoop_",
+        "StreamKScheduler": "streamK",
+    }
+
+    sch_sig = generate_sch_sig(schedule_config)
+    for orig, terse in kernel_terse_names_replace.items():
+        sch_sig = sch_sig.replace(orig, terse)
+    return sch_sig
+
+
+# unique type_name
+def generate_type_signature(kernel_types: TypeConfig):
+    return str("".join([MACHETEDataTypeNames[getattr(kernel_types, field.name)] for field in fields(TypeConfig)]))
+
+
+def generate_type_option_name(kernel_types: TypeConfig):
+    return ", ".join(
+        [
+            f"{field.name.replace('b_', 'with_')+'_type'}=" + MACHETEDataTypeNames[getattr(kernel_types, field.name)]
+            for field in fields(TypeConfig)
+        ]
+    )
+
+
+def is_power_of_two(n):
+    return (n != 0) and (n & (n - 1) == 0)
+
+
+def to_cute_constant(value: list[int]):
+
+    def _to_cute_constant(value: int):
+        if is_power_of_two(value):
+            return f"_{value}"
+        else:
+            return f"Int<{value}>"
+
+    if isinstance(value, Iterable):
+        return [_to_cute_constant(value) for value in value]
+    else:
+        return _to_cute_constant(value)
+
+
+def unique_schedules(impl_configs: list[ImplConfig]):
+    return list(set(sch for impl_config in impl_configs for sch in impl_config.schedules))
+
+
+def unsigned_type_with_bitwidth(num_bits):
+    return {
+        4: DataType.u4,
+        8: DataType.u8,
+        16: DataType.u16,
+        32: DataType.u32,
+        64: DataType.u64,
+    }[num_bits]
+
+
+template_globals = {
+    "void": DataType.void,
+    "DataTypeTag": MACHETEDataTypeTag,
+    "MACHETEScalarTypeTag": MACHETEDataTypeMACHETEScalarTypeTag,
+    "PaddleTypeTag": MACHETEDataTypePaddleDataTypeTag,
+    "KernelScheduleTag": MACHETEKernelScheduleTag,
+    "EpilogueScheduleTag": EpilogueScheduleTag,
+    "TileSchedulerTag": TileSchedulerTag,
+    "to_cute_constant": to_cute_constant,
+    "gen_sch_sig": generate_terse_sch_sig,
+    "gen_type_sig": generate_type_signature,
+    "unique_schedules": unique_schedules,
+    "unsigned_type_with_bitwidth": unsigned_type_with_bitwidth,
+    "gen_type_option_name": generate_type_option_name,
+}
+
+
+def create_template(template_str):
+    template = jinja2.Template(template_str)
+    template.globals.update(template_globals)
+    return template
+
+
+mm_dispatch_template = create_template(DISPATCH_TEMPLATE)
+mm_impl_template = create_template(IMPL_TEMPLATE)
+prepack_dispatch_template = create_template(PREPACK_TEMPLATE)
+
+
+def create_sources(impl_configs: list[ImplConfig], num_impl_files=8):
+    sources = []
+
+    sources.append(
+        (
+            "machete_mm_dispatch",
+            mm_dispatch_template.render(impl_configs=impl_configs),
+        )
+    )
+
+    prepack_types = []
+    for impl_config in impl_configs:
+        convert_type = (
+            impl_config.types.a
+            if impl_config.types.b_group_scale == DataType.void
+            else impl_config.types.b_group_scale
+        )
+        prepack_types.append(
+            PrepackTypeConfig(
+                a=impl_config.types.a,
+                b_num_bits=MACHETEDataTypeSize[impl_config.types.b],
+                convert=convert_type,
+                accumulator=impl_config.types.accumulator,
+            )
+        )
+
+    def prepacked_type_key(prepack_type: PrepackTypeConfig):
+        # For now we we can just use the first accumulator type seen since
+        # the tensor core shapes/layouts don't vary based on accumulator
+        # type so we can generate less code this way
+        return (prepack_type.a, prepack_type.b_num_bits, prepack_type.convert)
+
+    unique_prepack_types = []
+    prepack_types_seen = set()
+    for prepack_type in prepack_types:
+        key = prepacked_type_key(prepack_type)
+        if key not in prepack_types_seen:
+            unique_prepack_types.append(prepack_type)
+            prepack_types_seen.add(key)
+
+    sources.append(
+        (
+            "machete_prepack",
+            prepack_dispatch_template.render(
+                types=unique_prepack_types,
+            ),
+        )
+    )
+
+    # Split up impls across files
+    num_impls = reduce(lambda x, y: x + len(y.schedules), impl_configs, 0)
+    num_impls_per_file = math.ceil(num_impls / num_impl_files)
+
+    files_impls: list[list[ImplConfig]] = [[]]
+
+    curr_num_impls_assigned = 0
+    curr_impl_in_file = 0
+    curr_impl_configs = deepcopy(list(reversed(impl_configs)))
+
+    while curr_num_impls_assigned < num_impls:
+        room_left_in_file = num_impls_per_file - curr_impl_in_file
+        if room_left_in_file == 0:
+            files_impls.append([])
+            room_left_in_file = num_impls_per_file
+            curr_impl_in_file = 0
+
+        curr_ic = curr_impl_configs[-1]
+        if len(curr_ic.schedules) >= room_left_in_file:
+            # Break apart the current impl config
+            tmp_ic = deepcopy(curr_ic)
+            tmp_ic.schedules = curr_ic.schedules[:room_left_in_file]
+            curr_ic.schedules = curr_ic.schedules[room_left_in_file:]
+            files_impls[-1].append(tmp_ic)
+        else:
+            files_impls[-1].append(curr_ic)
+            curr_impl_configs.pop()
+        curr_num_impls_assigned += len(files_impls[-1][-1].schedules)
+        curr_impl_in_file += len(files_impls[-1][-1].schedules)
+
+    for part, file_impls in enumerate(files_impls):
+        sources.append(
+            (
+                f"machete_mm_impl_part{part+1}",
+                mm_impl_template.render(impl_configs=file_impls),
+            )
+        )
+
+    return sources
+
+
+def generate():
+    # See csrc/quantization/machete/Readme.md, the Codegeneration for more info
+    # about how this works
+    SCRIPT_DIR = os.path.dirname(__file__)
+
+    sch_common_params = dict(
+        kernel_schedule=TmaMI,
+        epilogue_schedule=TmaCoop,
+        tile_scheduler=TileSchedulerType.StreamK,
+    )
+
+    # Stored as "condition": ((tile_shape_mn), (cluster_shape_mnk))
+    default_tile_heuristic_config = {
+        # M = 257+
+        "M > 256 && K <= 16384 && N <= 4096": ((128, 128), (2, 1, 1)),
+        "M > 256": ((128, 256), (2, 1, 1)),
+        # M = 129-256
+        "M > 128 && K <= 4096 && N <= 4096": ((128, 64), (2, 1, 1)),
+        "M > 128 && K <= 8192 && N <= 8192": ((128, 128), (2, 1, 1)),
+        "M > 128": ((128, 256), (2, 1, 1)),
+        # M = 65-128
+        "M > 64 && K <= 4069 && N <= 4069": ((128, 32), (2, 1, 1)),
+        "M > 64 && K <= 4069 && N <= 8192": ((128, 64), (2, 1, 1)),
+        "M > 64 && K >= 8192 && N >= 12288": ((256, 128), (2, 1, 1)),
+        "M > 64": ((128, 128), (2, 1, 1)),
+        # M = 33-64
+        "M > 40 && K <= 6144 && N <= 6144": ((128, 32), (2, 1, 1)),
+        "M > 32 && K <= 6144 && N <= 6144": ((128, 16), (1, 1, 1)),
+        "M > 32 && K >= 16384 && N >= 12288": ((256, 64), (2, 1, 1)),
+        "M > 32": ((128, 64), (2, 1, 1)),
+        # M = 17-32
+        "M > 16 && K <= 12288 && N <= 8192": ((128, 32), (2, 1, 1)),
+        "M > 16": ((256, 32), (2, 1, 1)),
+        # M = 1-16
+        "N >= 26624": ((256, 16), (1, 1, 1)),
+        None: ((128, 16), (1, 1, 1)),
+    }
+
+    # For now we use the same heuristic for all types
+    # Heuristic is currently tuned for H100s
+    default_heuristic = [
+        (cond, ScheduleConfig(*tile_config, **sch_common_params))  # type: ignore
+        for cond, tile_config in default_tile_heuristic_config.items()
+    ]
+
+    def get_unique_schedules(heuristic: dict[str, ScheduleConfig]):
+        # Do not use schedules = list(set(...)) because we need to make sure
+        # the output list is deterministic; otherwise the generated kernel file
+        # will be non-deterministic and causes ccache miss.
+        schedules = []
+        for _, schedule_config in heuristic:
+            if schedule_config not in schedules:
+                schedules.append(schedule_config)
+        return schedules
+
+    impl_configs = []
+
+    GPTQ_kernel_type_configs = list(
+        TypeConfig(
+            a=a,
+            b=b,
+            b_group_scale=a,
+            b_group_zeropoint=DataType.void,
+            b_channel_scale=DataType.void,
+            a_token_scale=DataType.void,
+            out=a,
+            accumulator=DataType.f32,
+        )
+        for b in (MACHETEDataType.u4b8, MACHETEDataType.u8b128)
+        for a in (DataType.f16, DataType.bf16)
+    )
+
+    impl_configs += [
+        ImplConfig(x[0], x[1], x[2])
+        for x in zip(
+            GPTQ_kernel_type_configs,
+            itertools.repeat(get_unique_schedules(default_heuristic)),
+            itertools.repeat(default_heuristic),
+        )
+    ]
+
+    output_dir = os.path.join(SCRIPT_DIR, "generated")
+
+    # Delete the "generated" directory if it exists
+    if os.path.exists(output_dir):
+        shutil.rmtree(output_dir)
+
+    # Create the "generated" directory
+    os.makedirs(output_dir)
+
+    # Render each group of configurations into separate files
+    for filename, code in create_sources(impl_configs):
+        filepath = os.path.join(output_dir, f"{filename}.cu")
+        with open(filepath, "w") as output_file:
+            output_file.write(code)
+        print(f"Rendered template to {filepath}")
+
+
+if __name__ == "__main__":
+    generate()

custom_ops/gpu_ops/machete/machete_collective_builder.cuh

@@ -0,0 +1,31 @@
+#pragma once
+
+#include "utils/machete_collective_builder.cuh"
+#include "machete_mainloop.cuh"
+
+namespace cutlass::gemm::collective {
+using namespace cute;
+
+struct MacheteKernelTag {};
+
+template <class ElementPairA_, class GmemLayoutA_, int AlignmentA,
+          class ElementPairB_, class GmemLayoutB_, int AlignmentB,
+          class ElementAccumulator, class TileShape_MNK, class ClusterShape_MNK,
+          class StageCountType, class KernelScheduleType>
+struct MacheteCollectiveBuilder<
+    MacheteKernelTag, arch::Sm90, arch::OpClassTensorOp, ElementPairA_,
+    GmemLayoutA_, AlignmentA, ElementPairB_, GmemLayoutB_, AlignmentB,
+    ElementAccumulator, TileShape_MNK, ClusterShape_MNK, StageCountType,
+    KernelScheduleType,
+    cute::enable_if_t<(
+        cute::is_same_v<KernelScheduleType, KernelTmaWarpSpecialized> ||
+        cute::is_same_v<KernelScheduleType, KernelTmaWarpSpecializedPingpong> ||
+        cute::is_same_v<KernelScheduleType,
+                        KernelTmaWarpSpecializedCooperative>)>> {
+  using CollectiveOp = machete::MacheteCollectiveMma<
+      ElementPairA_, GmemLayoutA_, AlignmentA, ElementPairB_, GmemLayoutB_,
+      AlignmentB, ElementAccumulator, TileShape_MNK, ClusterShape_MNK,
+      StageCountType, KernelScheduleType>;
+};
+
+};  // namespace cutlass::gemm::collective

custom_ops/gpu_ops/machete/machete_cutlass_library_extension.py

@@ -0,0 +1,85 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import enum
+from typing import Union
+
+from cutlass_library import (
+    DataType,
+    DataTypeNames,
+    DataTypeSize,
+    DataTypeTag,
+    KernelScheduleTag,
+    KernelScheduleType,
+    enum_auto,
+)
+
+#
+#   Extend cutlass library with custom types, and missing values
+#
+
+
+class MACHETEDataType(enum.Enum):
+    u4b8 = enum_auto()
+    u8b128 = enum_auto()
+
+
+class MixedInputKernelScheduleType(enum.Enum):
+    TmaWarpSpecialized = enum_auto()
+    TmaWarpSpecializedPingpong = enum_auto()
+    TmaWarpSpecializedCooperative = enum_auto()
+
+
+MACHETEDataTypeNames: dict[Union[MACHETEDataType, DataType], str] = {
+    **DataTypeNames,  # type: ignore
+    **{
+        MACHETEDataType.u4b8: "u4b8",
+        MACHETEDataType.u8b128: "u8b128",
+    },
+}
+
+MACHETEDataTypeTag: dict[Union[MACHETEDataType, DataType], str] = {
+    **DataTypeTag,  # type: ignore
+    **{
+        MACHETEDataType.u4b8: "cutlass::machete_uint4b8_t",
+        MACHETEDataType.u8b128: "cutlass::machete_uint8b128_t",
+    },
+}
+
+MACHETEDataTypeSize: dict[Union[MACHETEDataType, DataType], int] = {
+    **DataTypeSize,  # type: ignore
+    **{
+        MACHETEDataType.u4b8: 4,
+        MACHETEDataType.u8b128: 8,
+    },
+}
+
+MACHETEDataTypeMACHETEScalarTypeTag: dict[Union[MACHETEDataType, DataType], str] = {
+    MACHETEDataType.u4b8: "machete::kU4B8",
+    MACHETEDataType.u8b128: "machete::kU8B128",
+    DataType.u4: "machete::kU4",
+    DataType.u8: "machete::kU8",
+    DataType.s4: "machete::kS4",
+    DataType.s8: "machete::kS8",
+    DataType.f16: "machete::kFloat16",
+    DataType.bf16: "machete::kBfloat16",
+}
+
+MACHETEDataTypePaddleDataTypeTag: dict[Union[MACHETEDataType, DataType], str] = {
+    DataType.u8: "paddle::DataType::UINT8",
+    DataType.s8: "paddle::DataType::INT8",
+    DataType.e4m3: "paddle::DataType::FLOAT8_E4M3FN",
+    DataType.s32: "paddle::DataType::INT32",
+    DataType.f16: "paddle::DataType::FLOAT16",
+    DataType.bf16: "paddle::DataType::BFLOAT16",
+    DataType.f32: "paddle::DataType::FLOAT32",
+}
+
+MACHETEKernelScheduleTag: dict[Union[MixedInputKernelScheduleType, KernelScheduleType], str] = {
+    **KernelScheduleTag,  # type: ignore
+    **{
+        MixedInputKernelScheduleType.TmaWarpSpecialized: "cutlass::gemm::KernelTmaWarpSpecialized",
+        MixedInputKernelScheduleType.TmaWarpSpecializedPingpong: "cutlass::gemm::KernelTmaWarpSpecializedPingpong",
+        MixedInputKernelScheduleType.TmaWarpSpecializedCooperative: "cutlass::gemm::KernelTmaWarpSpecializedCooperative",
+    },
+}

custom_ops/gpu_ops/machete/machete_interleaving_utils.cuh

@@ -0,0 +1,35 @@
+#pragma once
+
+#include "cutlass/cutlass.h"
+#include "cute/layout.hpp"
+
+namespace machete {
+
+using namespace cute;
+
+// get an interleaved block layout where each element consecutive element has a
+// stride of bit_stride and the block width is blk_bit_width,
+// examples:
+//  size_bits<T> = 8, bit_stride = 8,  blk_bit_width = 32 -> 4:1
+//  size_bits<T> = 8, bit_stride = 16, blk_bit_width = 32 -> (2, 2):(2, 1)
+//  size_bits<T> = 4, bit_stride = 8,  blk_bit_width = 32 -> (4, 2):(2, 1)
+//  size_bits<T> = 4, bit_stride = 16, blk_bit_width = 32 -> (2, 4):(4, 1)
+template <typename T, int bit_stride, int blk_bit_width>
+CUTE_HOST_DEVICE static constexpr auto get_interleaved_blk_layout() {
+  static_assert(blk_bit_width % bit_stride == 0);
+  static_assert(bit_stride % cute::sizeof_bits_v<T> == 0);
+
+  constexpr auto elems_per_blk = blk_bit_width / cute::sizeof_bits_v<T>;
+
+  if constexpr (cute::sizeof_bits_v<T> == bit_stride) {
+    // identity layout
+    return Layout<Shape<Int<elems_per_blk>>>{};
+  } else {
+    constexpr auto elems_per_stride = bit_stride / cute::sizeof_bits_v<T>;
+    constexpr auto num_strides = elems_per_blk / elems_per_stride;
+    return Layout<Shape<Int<num_strides>, Int<elems_per_stride>>,
+                  Stride<Int<elems_per_stride>, Int<1>>>{};
+  }
+}
+
+};  // namespace machete

custom_ops/gpu_ops/machete/machete_mm.cu

@@ -0,0 +1,88 @@
+// Copyright (c) 2025 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "machete_mm_launcher.cuh"
+#include "machete_prepack_launcher.cuh"
+
+template <typename T>
+std::optional<T> ConvertToStdOptional(const paddle::optional<T>& paddle_opt) {
+    return paddle_opt ? std::optional<T>(paddle_opt.get()) : std::nullopt;
+}
+
+paddle::Tensor mm(paddle::Tensor const& A, paddle::Tensor const& B,
+                 int64_t b_type_id,
+                 std::optional<paddle::DataType> const& maybe_out_type,
+                 std::optional<paddle::Tensor> const& maybe_group_scales,
+                 std::optional<paddle::Tensor> const& maybe_group_zeros,
+                 int64_t maybe_group_size,
+                 std::optional<paddle::Tensor> const& maybe_channel_scales,
+                 std::optional<paddle::Tensor> const& maybe_token_scales,
+                 std::string maybe_schedule) {
+  machete::ScalarType const b_type = machete::ScalarType::from_id(b_type_id);
+  std::optional<int64_t> maybe_group_size_opt = std::optional<int64_t>(maybe_group_size);
+  std::optional<std::string> maybe_schedule_opt;
+  if (maybe_schedule == "") {
+    maybe_schedule_opt = std::nullopt;
+  } else {
+    maybe_schedule_opt = std::optional<std::string>(maybe_schedule);
+  }
+  return machete::mm_dispatch({.A = A,
+                      .B = B,
+                      .b_type = b_type,
+                      .maybe_out_type = maybe_out_type,
+                      .maybe_group_scales = maybe_group_scales,
+                      .maybe_group_zeros = maybe_group_zeros,
+                      .maybe_group_size = maybe_group_size_opt,
+                      .maybe_channel_scales = maybe_channel_scales,
+                      .maybe_token_scales = maybe_token_scales,
+                      .maybe_schedule = maybe_schedule_opt});
+}
+
+std::vector<paddle::Tensor> MacheteMMKernel(
+    paddle::Tensor const& A, paddle::Tensor const& B,
+    paddle::optional<paddle::Tensor> const& maybe_group_scales,
+    paddle::optional<paddle::Tensor> const& maybe_group_zeros,
+    paddle::optional<paddle::Tensor> const& maybe_channel_scales,
+    paddle::optional<paddle::Tensor> const& maybe_token_scales,
+    std::string const& b_type_str,
+    std::string const& maybe_out_type_str,
+    int64_t const& maybe_group_size,
+    std::string const& maybe_schedule
+  ) {
+
+  machete::ScalarTypeId b_type_id;
+  paddle::DataType maybe_out_type;
+  if (b_type_str == "uint4b8") {
+    b_type_id = machete::kU4B8.id();
+  } else if (b_type_str == "uint8b128") {
+    b_type_id = machete::kU8B128.id();
+  } else {
+    PADDLE_ENFORCE(false, "b_type_str not supported!");
+  }
+  if (maybe_out_type_str == "float16") {
+    maybe_out_type = paddle::DataType::FLOAT16;
+  } else if (maybe_out_type_str == "bfloat16") {
+    maybe_out_type = paddle::DataType::BFLOAT16;
+  } else {
+    maybe_out_type = A.dtype();
+  }
+  auto out = mm(A, B, b_type_id, maybe_out_type,
+                ConvertToStdOptional<paddle::Tensor>(maybe_group_scales),
+                ConvertToStdOptional<paddle::Tensor>(maybe_group_zeros),
+                maybe_group_size,
+                ConvertToStdOptional<paddle::Tensor>(maybe_channel_scales),
+                ConvertToStdOptional<paddle::Tensor>(maybe_token_scales),
+                maybe_schedule);
+  return {out};
+}

custom_ops/gpu_ops/machete/machete_mm_kernel.cuh

@@ -0,0 +1,305 @@
+#pragma once
+
+// clang-format off
+// The cutlass include order matters (annoyingly)
+#include "cutlass/cutlass.h"
+
+#include "cute/tensor.hpp"
+#include "cutlass/tensor_ref.h"
+#include "cutlass/epilogue/collective/default_epilogue.hpp"
+#include "cutlass/epilogue/thread/linear_combination.h"
+#include "cutlass/gemm/dispatch_policy.hpp"
+#include "cutlass/gemm/collective/collective_builder.hpp"
+#include "cutlass/epilogue/collective/collective_builder.hpp"
+#include "cutlass/gemm/device/gemm_universal_adapter.h"
+#include "cutlass/gemm/kernel/gemm_universal.hpp"
+// clang-format on
+
+#include "utils/cute_utils.cuh"
+#include "utils/machete_numeric_conversion.cuh"
+#include "cutlass_extensions/epilogue/scaled_mm_epilogues_c3x.hpp"
+#include "utils/paddle_utils.hpp"
+#include "machete_collective_builder.cuh"
+#include "machete_prepacked_layout.cuh"
+#include "machete_interleaving_utils.cuh"
+
+namespace machete {
+
+using namespace cute;
+
+// NOTE This kernel computes D = alpha * A * B + beta * C by computing
+//   D^t = alpha * B^t * A^t + beta * C^t, this is because the wgmma
+//   instructions only support sourcing from registers for the left-hand
+//   operand, we want to upconvert/decompress the quantized operand in
+//   register. Since the primary use case we want to support is Y = XW^t where
+//   W is quantized, in this situation or right-hand operand is quantized so
+//   we compute the transpose to move it to the left-hand side.
+template <typename ElementA_, typename ElementB_, typename ElementD_,
+          typename AccumulatorT, typename GroupScaleT, typename GroupZeroT,
+          typename ChannelScaleT, typename TokenScaleT, class KernelSchedule,
+          typename ScheduleConfig>
+struct MacheteKernelTemplate {
+  static constexpr bool with_C = false;  // not ever used
+  static constexpr bool with_group_scales = !std::is_same_v<GroupScaleT, void>;
+  static constexpr bool with_group_zeropoints =
+      !std::is_same_v<GroupZeroT, void>;
+  static constexpr bool with_channel_scales =
+      !std::is_same_v<ChannelScaleT, void>;
+  static constexpr bool with_token_scales = !std::is_same_v<TokenScaleT, void>;
+
+  using MmaType = ElementA_;
+  using ElementA = ElementA_;
+  using ElementB = ElementB_;
+  using ElementD = ElementD_;
+  using ElementC = cute::conditional_t<with_C, ElementD, void>;
+  using ElementAccumulator = AccumulatorT;
+  using ElementCompute = AccumulatorT;  // For Epilogue
+  // Use dummy values when we don't have scales or zeropoints
+  using ElementZGroup =
+      cute::conditional_t<with_group_zeropoints, GroupZeroT, MmaType>;
+  using ElementSGroup =
+      cute::conditional_t<with_group_scales, GroupScaleT, MmaType>;
+  using ElementConvertGroup =
+      cute::conditional_t<with_group_scales, GroupScaleT, MmaType>;
+  using ElementSChannel =
+      cute::conditional_t<with_channel_scales, ChannelScaleT, AccumulatorT>;
+  using ElementSToken =
+      cute::conditional_t<with_token_scales, TokenScaleT, AccumulatorT>;
+
+  using BTypeTuple = cute::conditional_t<
+      with_group_scales,
+      cute::conditional_t<with_group_zeropoints,
+                          cute::tuple<ElementB, ElementSGroup, ElementZGroup>,
+                          cute::tuple<ElementB, ElementSGroup>>,
+      ElementB>;
+
+  using LayoutA = cutlass::layout::RowMajor;
+  using LayoutC = cutlass::layout::RowMajor;
+  using LayoutD = LayoutC;
+  using LayoutScale = cutlass::layout::RowMajor;
+  // not actually used since B has the prepacked layout, but required by cutlass
+  using _LayoutB = cutlass::layout::ColumnMajor;
+
+  // Interface strides expected by create_arguments (will get transposed)
+  using StrideA = cutlass::detail::TagToStrideA_t<LayoutA>;
+  using StrideC = cutlass::detail::TagToStrideA_t<LayoutC>;
+  using StrideD = cutlass::detail::TagToStrideA_t<LayoutD>;
+  using StrideSGroup = cutlass::detail::TagToStrideA_t<LayoutScale>;
+  using StrideZGroup = StrideSGroup;
+
+  using LayoutA_Transpose =
+      typename cutlass::layout::LayoutTranspose<LayoutA>::type;
+  using LayoutC_Transpose =
+      typename cutlass::layout::LayoutTranspose<LayoutC>::type;
+  using LayoutD_Transpose =
+      typename cutlass::layout::LayoutTranspose<LayoutD>::type;
+
+  using ArchTag = cutlass::arch::Sm90;
+  using OperatorClass = cutlass::arch::OpClassTensorOp;
+
+  using PrepackedLayoutB =
+      PrepackedLayoutBTemplate<ElementA_, ElementB_, ElementConvertGroup,
+                               AccumulatorT, LayoutA_Transpose, KernelSchedule>;
+
+  static int constexpr TileShapeK =
+      128 * 8 / cutlass::sizeof_bits<MmaType>::value;
+  static int constexpr AlignmentA = 128 / cutlass::sizeof_bits_v<ElementA>;
+  static int constexpr AlignmentB = 128 / cutlass::sizeof_bits_v<ElementB>;
+  static int constexpr AlignmentC =
+      (with_C) ? 128 / cutlass::sizeof_bits_v<ElementC> : 0;
+  static int constexpr AlignmentD = 128 / cutlass::sizeof_bits_v<ElementD>;
+
+  using TileShape = decltype(append(typename ScheduleConfig::TileShapeNM{},
+                                    cute::Int<TileShapeK>{}));
+  using ClusterShape = typename ScheduleConfig::ClusterShape;
+  using EpilogueSchedule = typename ScheduleConfig::EpilogueSchedule;
+  using EpilogueTileType = typename ScheduleConfig::EpilogueTileType;
+  using TileScheduler = typename ScheduleConfig::TileScheduler;
+
+  static_assert(
+      (!with_channel_scales && !with_token_scales) ||
+          ((with_channel_scales && with_token_scales) &&
+           std::is_same_v<ElementSChannel, ElementSToken>),
+      "Currently token and channel scales (if present) must be the same type");
+
+  // Currently only supports float scales
+  using ChTokScalesEpilogue =
+      typename fastdeploy::c3x::ScaledEpilogue<ElementAccumulator, ElementD,
+                                         TileShape>;
+  static_assert((with_channel_scales || with_token_scales) ||
+                    (std::is_same_v<ElementSChannel, float> &&
+                     std::is_same_v<ElementSToken, float>),
+                "Currently token and channel scales (if present) must be float "
+                "(and if one is present the other must be too)");
+
+  using StoreEpilogueCompute = typename cutlass::epilogue::fusion::Sm90EVT<
+      cutlass::epilogue::fusion::Sm90AccFetch>;
+
+  using EVTCompute =
+      std::conditional_t<with_channel_scales || with_token_scales,
+                         typename ChTokScalesEpilogue::EVTCompute,
+                         StoreEpilogueCompute>;
+
+  // EVTCompute
+  using CollectiveEpilogue =
+      typename cutlass::epilogue::collective::CollectiveBuilder<
+          ArchTag, OperatorClass, TileShape, ClusterShape, EpilogueTileType,
+          ElementAccumulator, ElementSChannel, ElementC, LayoutC_Transpose,
+          AlignmentC, ElementD, LayoutD_Transpose, AlignmentD, EpilogueSchedule,
+          EVTCompute>::CollectiveOp;
+
+  using CollectiveMainloop =
+      typename cutlass::gemm::collective::MacheteCollectiveBuilder<
+          cutlass::gemm::collective::MacheteKernelTag, ArchTag, OperatorClass,
+          BTypeTuple, PrepackedLayoutB, AlignmentB, ElementA, LayoutA_Transpose,
+          AlignmentA, ElementAccumulator, TileShape, ClusterShape,
+          cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(
+              sizeof(typename CollectiveEpilogue::SharedStorage))>,
+          KernelSchedule>::CollectiveOp;
+
+  using GemmKernel = cutlass::gemm::kernel::GemmUniversal<
+      Shape<int, int, int, int>,  // Indicates ProblemShape
+      CollectiveMainloop, CollectiveEpilogue, TileScheduler>;
+  using Gemm = cutlass::gemm::device::GemmUniversalAdapter<GemmKernel>;
+
+  // stride_B is unused (since B is prepacked), but still required by cutlass
+  using _StrideB = cutlass::detail::TagToStrideB_t<_LayoutB>;
+
+  using Arguments = typename Gemm::Arguments;
+  using MainloopArguments = typename GemmKernel::MainloopArguments;
+  using EpilogueArguments = typename GemmKernel::EpilogueArguments;
+
+  static Arguments create_arguments(
+      cudaStream_t stream,
+      paddle::Tensor const& A,  // MxK matrix
+      paddle::Tensor const& B,  // KxN prepacked matrix
+      paddle::Tensor& D,        // MxN matrix
+      std::optional<paddle::Tensor> const& maybe_g_scales,  // scale_KxN matrix
+      std::optional<paddle::Tensor> const& maybe_g_zeros,   // scale_KxN matrix
+      std::optional<int64_t> maybe_group_size,
+      std::optional<paddle::Tensor> const& maybe_ch_scales,   // len N vector
+      std::optional<paddle::Tensor> const& maybe_tok_scales)  // len M vector
+  {
+    static_assert(!with_group_zeropoints || with_group_scales);
+
+    int M = A.shape()[0], N = B.shape()[1], K = A.shape()[1];
+    PD_CHECK(D.shape()[0] == M && D.shape()[1] == N);
+
+    auto layout_A = make_cute_layout<StrideA>(A, "A");
+    auto layout_D = make_cute_layout<StrideD>(D, "D");
+    auto layout_S_group =
+        maybe_make_cute_layout<StrideSGroup>(maybe_g_scales, "group_scales");
+    auto layout_Z_group =
+        maybe_make_cute_layout<StrideZGroup>(maybe_g_zeros, "group_zeros");
+    int64_t numel_S_channel = maybe_ch_scales ? maybe_ch_scales->numel() : 0;
+    int64_t numel_S_token = maybe_tok_scales ? maybe_tok_scales->numel() : 0;
+
+    auto unwrap = [](auto const& t) {
+      return t ? t->data() : nullptr;
+    };
+    auto A_ptr = static_cast<ElementA const*>(A.data());
+    auto B_ptr = static_cast<ElementB const*>(B.data());
+    auto D_ptr = static_cast<ElementD*>(D.data());
+    auto S_group_ptr =
+        static_cast<ElementSGroup const*>(unwrap(maybe_g_scales));
+    auto Z_group_ptr = static_cast<ElementZGroup const*>(unwrap(maybe_g_zeros));
+    auto S_channel_ptr =
+        static_cast<ElementSChannel const*>(unwrap(maybe_ch_scales));
+    auto S_token_ptr =
+        static_cast<ElementSToken const*>(unwrap(maybe_tok_scales));
+
+    int const group_size =
+        maybe_group_size == -1 ? K : maybe_group_size.value_or(K);
+    int const scale_k = (K + group_size - 1) / group_size;
+
+    PD_CHECK(size<0>(layout_A) == M && size<1>(layout_A) == K);
+    PD_CHECK(size<0>(layout_D) == M && size<1>(layout_D) == N);
+
+    if constexpr (with_group_scales) {
+      PD_CHECK(S_group_ptr && layout_S_group);
+      PD_CHECK((size<0>(*layout_S_group) == scale_k &&
+                   size<1>(*layout_S_group) == N));
+    } else {
+      PD_CHECK(!S_group_ptr, "Scales not supported");
+    }
+
+    if constexpr (with_group_zeropoints) {
+      PD_CHECK(Z_group_ptr && layout_Z_group);
+      PD_CHECK((size<0>(*layout_Z_group) == scale_k &&
+                   size<1>(*layout_Z_group) == N));
+      PD_CHECK(layout_S_group && *layout_Z_group == *layout_S_group,
+                  "Scales and zeros must have the same layout");
+    } else {
+      PD_CHECK(!Z_group_ptr, "Zeropoints not supported");
+    }
+
+    if constexpr (with_channel_scales || with_token_scales) {
+      PD_CHECK(
+          (maybe_ch_scales->numel() == N || maybe_ch_scales->numel() == 1) &&
+          (maybe_tok_scales->numel() == M || maybe_tok_scales->numel() == 1));
+    }
+
+    // Transpose A and D
+    // A doesn't need to be transposed since cutlass expects a NxK matrix
+    //  for B (which is At)
+    auto stride_At = layout_A.stride();
+    auto stride_Dt = permute_layout<1, 0, 2>(layout_D).stride();
+
+    MainloopArguments mainloop_arguments{};
+    // {Accum, C, C_layout, D, D}
+    EpilogueArguments epilogue_arguments{};
+
+    if constexpr (with_channel_scales || with_token_scales) {
+      epilogue_arguments =
+          EpilogueArguments{ChTokScalesEpilogue::prepare_args(
+                                *maybe_ch_scales, *maybe_tok_scales),
+                            nullptr,
+                            {},
+                            D_ptr,
+                            stride_Dt};
+    } else {
+      epilogue_arguments = EpilogueArguments{{}, nullptr, {}, D_ptr, stride_Dt};
+    }
+
+    if constexpr (with_group_scales && with_group_zeropoints) {
+      auto stride_S_group = permute_layout<1, 0, 2>(*layout_S_group).stride();
+      mainloop_arguments = MainloopArguments{
+          B_ptr,       _StrideB{},     A_ptr,      stride_At,
+          S_group_ptr, stride_S_group, group_size, Z_group_ptr};
+    } else if constexpr (with_group_scales) {
+      auto stride_S_group = permute_layout<1, 0, 2>(*layout_S_group).stride();
+      mainloop_arguments =
+          MainloopArguments{B_ptr,       _StrideB{},     A_ptr,     stride_At,
+                            S_group_ptr, stride_S_group, group_size};
+    } else {
+      mainloop_arguments =
+          MainloopArguments{B_ptr, _StrideB{}, A_ptr, stride_At};
+    }
+
+    return Arguments{cutlass::gemm::GemmUniversalMode::kGemm,
+                     {N, M, K, 1},
+                     mainloop_arguments,
+                     epilogue_arguments};
+  };
+
+  static size_t get_workspace_size(Arguments const& args) {
+    return Gemm::get_workspace_size(args);
+  }
+
+  static bool can_implement(Arguments const& args) {
+    return Gemm::can_implement(args) == cutlass::Status::kSuccess;
+  }
+
+  static void run(Arguments const& args, void* workspace, cudaStream_t stream) {
+    Gemm gemm_op;
+
+    cutlass::Status status = gemm_op.initialize(args, workspace, stream);
+    PD_CHECK(status == cutlass::Status::kSuccess,
+                "Machete kernel failed to initialize workspace");
+
+    status = gemm_op.run(stream);
+    PD_CHECK(status == cutlass::Status::kSuccess, "Machete kernel failed");
+  }
+};
+
+};  // namespace machete

custom_ops/gpu_ops/machete/machete_mm_launcher.cuh

@@ -0,0 +1,78 @@
+#pragma once
+
+#include <Python.h>
+
+#include "machete_mm_kernel.cuh"
+#include "utils/paddle_utils.hpp"
+#include "utils/scalar_type.h"
+
+namespace machete {
+
+struct MMArgs {
+  paddle::Tensor const& A;
+  paddle::Tensor const& B;
+  machete::ScalarType const& b_type;
+  std::optional<paddle::DataType> const& maybe_out_type;
+  std::optional<paddle::Tensor> const& maybe_group_scales;
+  std::optional<paddle::Tensor> const& maybe_group_zeros;
+  std::optional<int64_t> maybe_group_size;
+  std::optional<paddle::Tensor> const& maybe_channel_scales;
+  std::optional<paddle::Tensor> const& maybe_token_scales;
+  std::optional<std::string> maybe_schedule;
+};
+
+struct SupportedSchedulesArgs {
+  paddle::DataType a_type;
+  machete::ScalarType b_type;
+  std::optional<paddle::DataType> maybe_group_scales_type;
+  std::optional<paddle::DataType> maybe_group_zeros_type;
+  std::optional<paddle::DataType> maybe_channel_scales_type;
+  std::optional<paddle::DataType> maybe_token_scales_type;
+  std::optional<paddle::DataType> maybe_out_type;
+};
+
+paddle::Tensor mm_dispatch(MMArgs args);
+
+std::vector<std::string> supported_schedules_dispatch(
+    SupportedSchedulesArgs args);
+
+template <typename MacheteKernel>
+paddle::Tensor run_impl(MMArgs args) {
+  // const at::cuda::OptionalCUDAGuard device_guard(device_of(args.A));
+
+  // auto device = args.A.device();
+  // auto stream = at::cuda::getCurrentCUDAStream(device.index());
+  auto place = args.A.place();
+  cudaStream_t stream = args.A.stream();
+
+  int M = args.A.shape()[0];
+  int N = args.B.shape()[1];
+  int K = args.A.shape()[1];
+
+  // Allocate output
+  paddle::Tensor D = paddle::empty(
+      {M, N},
+      equivalent_scalar_type_v<typename MacheteKernel::ElementD>,
+      place);
+
+  auto arguments = MacheteKernel::create_arguments(
+      stream,  //
+      args.A, args.B, D, args.maybe_group_scales, args.maybe_group_zeros,
+      args.maybe_group_size, args.maybe_channel_scales,
+      args.maybe_token_scales);
+  PD_CHECK(MacheteKernel::can_implement(arguments),
+              "Machete kernel cannot be run with these arguments");
+
+  size_t workspace_size = MacheteKernel::get_workspace_size(arguments);
+  int S = static_cast<int>(workspace_size);
+  // phi::Allocator* allocator = paddle::GetAllocator(place);
+  // auto workspace = allocator->Allocate(workspace_size);
+  // MacheteKernel::run(arguments, workspace->ptr(), stream);
+  // paddle::Tensor workspace = paddle::empty({S}, paddle::DataType::UINT8, place);
+  paddle::Tensor workspace = GetEmptyTensor({S}, paddle::DataType::UINT8, place);
+  MacheteKernel::run(arguments, workspace.data(), stream);
+
+  return D;
+};
+
+};  // namespace machete

custom_ops/gpu_ops/machete/machete_prepack_B.cu

@@ -0,0 +1,73 @@
+// Copyright (c) 2025 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "machete_mm_launcher.cuh"
+#include "machete_prepack_launcher.cuh"
+
+paddle::Tensor prepack_B(
+    paddle::Tensor const& B, paddle::DataType const& a_type, int64_t b_type_id,
+    std::string const& maybe_group_scales_type_str) {
+  machete::ScalarType const b_type = machete::ScalarType::from_id(b_type_id);
+  std::optional<paddle::DataType> maybe_group_scales_type;
+  if (maybe_group_scales_type_str == "float16") {
+    maybe_group_scales_type = paddle::DataType::FLOAT16;
+  }
+  else if (maybe_group_scales_type_str == "bfloat16") {
+    maybe_group_scales_type = paddle::DataType::BFLOAT16;
+  }
+  else if (maybe_group_scales_type_str == "float32") {
+    maybe_group_scales_type = paddle::DataType::FLOAT32;
+  }
+  else if (maybe_group_scales_type_str == "") {
+    maybe_group_scales_type = std::nullopt;
+  }
+  else {
+    PADDLE_ENFORCE(false, "maybe_group_scales_type_str not supported!");
+  }
+  return machete::prepack_B_dispatch(
+      {.B = B,
+       .a_type = a_type,
+       .b_type = b_type,
+       .maybe_group_scales_type = maybe_group_scales_type});
+}
+
+std::vector<paddle::Tensor> MachetePrepackBKernel(
+    paddle::Tensor const& B, std::string const& a_type_str, std::string const& b_type_str,
+    std::string const& maybe_group_scales_type_str) {
+
+  machete::ScalarTypeId b_type_id;
+  paddle::DataType a_type, maybe_group_scales_type;
+
+  if (b_type_str == "uint4b8") {
+    b_type_id = machete::kU4B8.id();
+  } else if (b_type_str == "uint8b128") {
+    b_type_id = machete::kU8B128.id();
+  } else {
+    PADDLE_ENFORCE(false, "b_type_str not supported!");
+  }
+
+  if (a_type_str == "float16") {
+    a_type = paddle::DataType::FLOAT16;
+  }
+  else if (a_type_str == "bfloat16") {
+    a_type = paddle::DataType::BFLOAT16;
+  }
+  else {
+    PADDLE_ENFORCE(false, "a_type_str not supported!");
+  }
+  auto Bt = paddle::experimental::transpose(B, {1, 0});
+  paddle::Tensor B_prepacked = prepack_B(Bt, a_type, b_type_id, maybe_group_scales_type_str);
+  return {B_prepacked};
+
+}

custom_ops/gpu_ops/machete/machete_prepack_kernel.cuh

@@ -0,0 +1,76 @@
+#pragma once
+
+#include "machete_mm_kernel.cuh"
+#include "utils/cute_utils.cuh"
+#include "utils/paddle_utils.hpp"
+
+namespace machete {
+
+template <int threads, typename PrepackedLayoutB, typename BInTensor,
+          typename ElementB>
+static __global__ void prepack_B_kernel(BInTensor B_in, ElementB* B_out_ptr) {
+  auto constexpr block_size =
+      Int<size(typename PrepackedLayoutB::PPBlockShape_NK{})>{};
+  auto constexpr eles_per_thread = Int<block_size / threads>{};
+  static_assert(block_size % threads == 0,
+                "block_size must be divisible by the number of threads");
+
+  // Which pre-packed are we responsible for
+  auto blk_coord = make_coord(blockIdx.x, blockIdx.y, blockIdx.z);
+  auto tB_in = local_tile(
+      B_in, append(typename PrepackedLayoutB::PPBlockShape_NK{}, _1{}),
+      blk_coord);
+
+  // Find the start offset in the output for this pre-packed block
+  auto bNbKL_to_offset = PrepackedLayoutB::bNbKL_to_offset(shape(B_in));
+
+  // Tensor representing a 1:1 mapping to the output space in 1D
+  auto tB_out_linear =
+      make_tensor(get_logical_ptr(B_out_ptr) + bNbKL_to_offset(blk_coord),
+                  make_layout(make_shape(block_size)));
+  // Mapping from output space (1D) to input space
+  auto tB_in_linear = make_tensor(
+      tB_in.data(),
+      tB_in.layout()
+          .compose(right_inverse(PrepackedLayoutB::ppblock_ilvd_NK_to_offset()))
+          .with_shape(make_shape(block_size)));
+
+  // Tile for this specific thread (could have used a TiledCopy but these work
+  // best with 2d layouts, this is a simple 1d layout so local_tile is enough,
+  // we are also not that concerned with performance for this kernel)
+  auto thr_tB_in_linear =
+      local_tile(tB_in_linear, make_shape(eles_per_thread), threadIdx.x);
+  auto thr_tB_out_linear =
+      local_tile(tB_out_linear, make_shape(eles_per_thread), threadIdx.x);
+
+  // Construct a register-backed Tensor with the same shape as each thread's
+  // partition
+  auto fragment = make_tensor<ElementB>(shape(thr_tB_in_linear));
+
+  copy(thr_tB_in_linear, fragment);
+  copy(Copy_Atom<DefaultCopy, uint8_t>{}, fragment, thr_tB_out_linear);
+}
+
+template <typename PrepackedLayoutB, typename InLayout>
+static void prepack_B_template(
+    cudaStream_t stream, typename PrepackedLayoutB::ElementB const* B_in_ptr,
+    InLayout B_layout, typename PrepackedLayoutB::ElementB* B_out_ptr) {
+  using TileShapeNKL =
+      decltype(append(typename PrepackedLayoutB::PPBlockShape_NK{}, _1{}));
+  auto ilvd_NKbNbKL_to_offset =
+      PrepackedLayoutB::ilvd_NKbNbKL_to_offset(shape(B_layout));
+
+  PD_CHECK(size<0>(B_layout) % size<0>(TileShapeNKL{}) == 0);
+  PD_CHECK(size<1>(B_layout) % size<1>(TileShapeNKL{}) == 0);
+
+  auto N_tiles = size<0>(B_layout) / size<0>(TileShapeNKL{});
+  auto K_tiles = size<1>(B_layout) / size<1>(TileShapeNKL{});
+  auto L_tiles = size<2>(B_layout);
+
+  auto B_in = make_tensor(get_logical_ptr(B_in_ptr), B_layout);
+
+  prepack_B_kernel<128, PrepackedLayoutB>
+      <<<dim3(N_tiles, K_tiles, L_tiles), 128, 0, stream>>>(B_in, B_out_ptr);
+}
+
+};  // namespace machete

custom_ops/gpu_ops/machete/machete_prepack_launcher.cuh

@@ -0,0 +1,77 @@
+#pragma once
+
+#include "machete_prepack_kernel.cuh"
+#include "utils/paddle_utils.hpp"
+#include "utils/scalar_type.h"
+
+namespace machete {
+
+struct PrepackBArgs {
+  paddle::Tensor const& B;
+  paddle::DataType a_type;
+  machete::ScalarType b_type;
+  std::optional<paddle::DataType> maybe_group_scales_type;
+};
+
+template <typename PrepackedLayoutB>
+paddle::Tensor prepack_impl(paddle::Tensor const B) {
+  // const at::cuda::OptionalCUDAGuard device_guard(device_of(B));
+  using ElementB = typename PrepackedLayoutB::ElementB;
+  using PPBlockShape_NK = typename PrepackedLayoutB::PPBlockShape_NK;
+
+  // auto device = B.device();
+  // auto stream = at::cuda::getCurrentCUDAStream(device.index());
+  cudaStream_t stream = B.stream();
+  auto B_ptr = static_cast<ElementB const*>(B.data());
+  // elements per storage item for B
+  auto eles_per_storage =
+      (SizeOf(B.dtype()) * 8) / cute::sizeof_bits_v<ElementB>;
+
+  // paddle B passed in is/should be (packed_K,N), the kernel expects (N,K,L) (to
+  // match cutlass using (N,K,L) for B), so we transpose B to (N,packed_K,L)
+  // auto Bt_packed = B.transpose();
+  auto Bt_packed = paddle::experimental::transpose(B, {1, 0});
+
+  PD_CHECK(
+      (B.shape()[0] * eles_per_storage) % size<1>(PPBlockShape_NK{}) == 0,
+      "B.shape[0] (in terms of unpacked elements) must be a multiple of ",
+      size<1>(PPBlockShape_NK{}));
+  PD_CHECK(B.shape()[1] % size<0>(PPBlockShape_NK{}) == 0,
+              "B.shape[1] must be a multiple of ", size<0>(PPBlockShape_NK{}));
+
+  using StrideB = cutlass::detail::TagToStrideB_t<cutlass::layout::ColumnMajor>;
+  auto const l_Bt_packed = make_cute_layout<StrideB>(Bt_packed, "B");
+  // auto const l_Bt_packed = make_cute_layout<StrideB>(B, "B");
+
+  // convert (N,packed_K,L) layout to (N,K,L) layout
+  //  in effect we want to do: blocked_product(layout_Bt_packed,
+  //      make_ordered_layout(make_shape(_1{}, eles_per_storage, _1{}),
+  //                          Step<_1, _0, _2>{}));
+  // but blocked_product does not support dynamic strides so we implement the
+  // equivalent manually,
+  //   new_shape = (N, packed_K, L) * (1, eles_per_storage, 1) -> (N, K, L)
+  //   new_stride = (s0, s1, s2) * (eles_per_storage, 1, eles_per_storage)
+  //                 when s1 == 1
+  PD_CHECK(stride<1>(l_Bt_packed) == 1, "stride<1>(l_Bt_packed) must be 1");
+  // clang-format off
+  auto const layout_Bt = make_layout(
+      transform_with_idx(l_Bt_packed.shape(), [&](auto ele, auto idx) {
+        return idx == 1 ? ele * eles_per_storage : ele;
+      }),
+      transform_with_idx(l_Bt_packed.stride(), [&](auto ele, auto idx) {
+        return idx != 1 ? ele * eles_per_storage : ele;
+      }));
+  // clang-format on
+
+  // Allocate output
+  paddle::Tensor D = paddle::empty_like(B);
+
+  prepack_B_template<PrepackedLayoutB>(
+      stream, B_ptr, layout_Bt, static_cast<ElementB*>(D.data()));
+
+  return D;
+};
+
+paddle::Tensor prepack_B_dispatch(PrepackBArgs args);
+
+};  // namespace machete

custom_ops/gpu_ops/machete/machete_prepacked_layout.cuh

@@ -0,0 +1,249 @@
+#pragma once
+
+// clang-format off
+// The cutlass include order matters (annoyingly)
+
+#include "cutlass/cutlass.h"
+
+#include "cute/tensor.hpp"
+#include "cutlass/tensor_ref.h"
+#include "cutlass/epilogue/collective/default_epilogue.hpp"
+#include "cutlass/epilogue/thread/linear_combination.h"
+#include "cutlass/gemm/dispatch_policy.hpp"
+#include "cutlass/gemm/collective/collective_builder.hpp"
+#include "cutlass/epilogue/collective/collective_builder.hpp"
+#include "cutlass/gemm/device/gemm_universal_adapter.h"
+#include "cutlass/gemm/kernel/gemm_universal.hpp"
+// clang-format on
+
+#include "utils/cute_utils.cuh"
+#include "machete_collective_builder.cuh"
+#include "machete_interleaving_utils.cuh"
+
+namespace machete {
+
+using namespace cute;
+
+struct IlvBlkLayoutAuto {};
+
+// This defines a prepacked layout for the B matrix, where the matrix is broken
+// up into PPBlockShape_NK blocks. The data within each block is then compactly
+// stored in memory such that when performing a TiledMMA operation with the same
+// shape as prepacked block, all the data for a given thread is contiguous in
+// memory. This allows us to use wider shared memory loads when loading B from
+// shared memory. The values within a thread are also potentially interlaeved
+// inorder to allow for more efficient upconverting.
+//
+// The contract here is that the `TiledMma` determined below matches the one
+// ultimately used in the kernel. (this is also why the other element types are
+// required along with the kernel schedule)
+template <typename ElementA_, typename ElementB_, typename ElementConvert_,
+          typename AccumulatorT, class LayoutB, class KernelSchedule,
+          typename IlvBlkLayout_ = IlvBlkLayoutAuto>
+// clang-format on
+struct PrepackedLayoutBTemplate {
+  using MmaType = ElementA_;
+  using ElementA = ElementA_;
+  using ElementB = ElementB_;
+  using ElementAccumulator = AccumulatorT;
+  using ElementMma = MmaType;
+
+  // Interleave for 4bit bit types when we are not upconverting to fp8 or int8,
+  // in those cases case we use a LUT using prmt instructions to upconvert and
+  // is more efficient if the data is not interleaved For 8bit+ prmt
+  // instructions makes non-interleaved layouts efficient enough we don't need
+  // iterleaved layouts (and can reuse more of the existing cutlass converts)
+  static constexpr bool should_interleave =
+      sizeof_bits_v<ElementB> <= 4 &&
+      !std::is_same_v<ElementConvert_, cutlass::float_e4m3_t> &&
+      !std::is_same_v<ElementConvert_, int8_t>;
+
+  // Only use interleaved layouts for subbyte weights,
+  using IlvdBlkLayout = std::conditional_t<
+      std::is_same_v<IlvBlkLayout_, IlvBlkLayoutAuto>,
+      std::conditional_t<
+          should_interleave,
+          decltype(get_interleaved_blk_layout<
+                   ElementB, sizeof_bits_v<ElementConvert_>, 32>()),
+          void>,
+      IlvBlkLayout_>;
+
+  // TODO (LucasWilkinson): compare the performance for other sizes
+  // Prepacked block shape, smallest layout atom for loading into registers
+  //   (can contain multiple wgmma instructions worth of data in one block)
+  // We ideally want this to be configured such that a thread can perform 128bit
+  // loads, i.e. we amount of data associated with each thread within a
+  // prepacked block is a multiple of 128bits, when using a cooperative sechdule
+  // we have 256 threads working a single block at a time, this means each
+  // thread works on `sizeof_bits_v<ElementB> * (128*64) / 256` bits of data,
+  // for a 4bit type this would be 128bits
+  using PPBlockShape_NK = Shape<_128, _64>;
+
+  // Create the shape of the tile anticipated to be used by the GEMM kernel,
+  //  when the kernel executes we will compute `Ct = Bt * At` since the
+  //  quantized weights (B), must be the lhs operand so the flow through
+  //  registers.
+  // The _128 here doesn't actually impact the shape of the stored tile directly
+  //  but may impact the op selected by rs_op_selector
+  using GemmTileShape = decltype(make_shape(size<0>(PPBlockShape_NK{}), _128{},
+                                            size<1>(PPBlockShape_NK{})));
+
+  static constexpr cute::GMMA::Major GmmaMajorB =
+      gmma_rs_tag_to_major_B<LayoutB>();
+
+  // For coop schedules we have two warp groups cooperatively issuing wgmma
+  // instructions so we use 2 atoms along the M dim (one for each warpgroup)
+  using AtomLayoutMNK = cute::conditional_t<
+      cute::is_same_v<KernelSchedule, KernelTmaWarpSpecializedCooperative>,
+      Layout<Shape<_2, _1, _1>>, Layout<Shape<_1, _1, _1>>>;
+
+  using TiledMma = decltype(cute::make_tiled_mma(
+      cute::GMMA::rs_op_selector<ElementMma, ElementMma, ElementAccumulator,
+                                 GemmTileShape, GMMA::Major::K, GmmaMajorB>(),
+      AtomLayoutMNK{}));
+
+  // Prepacked block, (athrid, val) -> (N,K)
+  // i.e. ((ThrV,(ThrN,ThrK)),(FrgV,(RestN,RestK,...))) -> (N,K)
+  CUTE_HOST_DEVICE static constexpr auto ppblock_TV_to_NK() {
+    return TiledMma{}.thrfrg_A(make_layout(PPBlockShape_NK{}));
+  }
+
+  // Prepacked block, (N,K) -> (athrid, val)
+  // i.e. (N,K) -> ((ThrV,(ThrN,ThrK)),(FrgV,(RestN,RestK,...)))
+  CUTE_HOST_DEVICE static constexpr auto ppblock_NK_to_TV() {
+    return right_inverse(ppblock_TV_to_NK()).with_shape(PPBlockShape_NK{});
+  }
+
+  // Prepacked block, (athrid, val) -> (storage_offset)
+  // i.e. ((ThrV,(ThrN,ThrK)),(FrgV,(RestN,RestK,...))) -> (storage_idx)
+  CUTE_HOST_DEVICE static constexpr auto ppblock_TV_to_offset() {
+    // Return iterleaved layout
+    return make_ordered_layout(shape(ppblock_TV_to_NK()), Step<_1, _0>{});
+  }
+
+  // Prepacked block, (athrid, val) -> (storage_offset)
+  // i.e. ((ThrV,(ThrM,ThrK)),(IlvdFrgV,(RestM,RestK,...))) -> (storage_idx)
+  CUTE_HOST_DEVICE static constexpr auto ppblock_ilvd_TV_to_offset() {
+    auto layout_no_interleave =
+        make_ordered_layout(shape(ppblock_TV_to_NK()), Step<_1, _0>{});
+
+    if constexpr (std::is_same_v<IlvdBlkLayout, void>) {
+      return layout_no_interleave;
+    } else {
+      // interleave by transforming FrgV into interleaved blocks where each
+      // block has the layout IlvdBlkLayout, for example if IlvdBlkLayout is
+      // (2, 2) : (2, 1) then we get: ((2, 2), size(FrgV) / 4) : ((2, 1), 4)
+      //   if FrgV is {A, B, C, D, E, F, G, H}
+      //   then ((IlvBlk), FrgB) is {A, C, B, D, C, G, D, H}
+      auto frgV = get<1, 0>(layout_no_interleave);
+      auto ilvdBlk = IlvdBlkLayout{};
+      static_assert(size(frgV) % size(ilvdBlk) == 0,
+                    "FrgV must be divisible by size(ilvdBlk)");
+      auto ilvd_FrgV = make_layout(
+          make_shape(shape(ilvdBlk), Int<size(frgV) / size(ilvdBlk)>{}),
+          make_stride(stride(ilvdBlk), size(ilvdBlk)));
+
+      // Return iterleaved layout
+      return make_layout(
+          get<0>(layout_no_interleave),
+          make_layout(ilvd_FrgV, get<1, 1>(layout_no_interleave)));
+    }
+  }
+
+  // Prepacked block, (M,K) -> (storage_offset)
+  CUTE_HOST_DEVICE static constexpr auto ppblock_ilvd_NK_to_offset() {
+    // do (M,K) -> (athrid, val) -> (storage_idx)
+    return ppblock_ilvd_TV_to_offset().compose(ppblock_NK_to_TV());
+  }
+
+  // ((athrid, val), (BlocksN, BlocksK), L) -> (storage_idx)
+  template <typename Shape_NKL>
+  CUTE_HOST_DEVICE static constexpr auto TVbNbKL_to_offset(
+      Shape_NKL shape_mkl) {
+    constexpr auto block_layout = ppblock_TV_to_offset();
+
+    // (BlocksN, BlocksK, L)
+    auto blocks_shape =
+        cute::transform(shape_mkl, append(PPBlockShape_NK{}, _1{}),
+                        [](auto x, auto y) { return x / y; });
+
+    // ((athrid, val), (BlocksN, BlocksK, L)) -> (storage_idx)
+    auto result = make_layout(
+        block_layout,
+        make_layout(blocks_shape,
+                    compact_col_major(blocks_shape, size(block_layout))));
+
+    // ((athrid, val), (BlocksN, BlocksK, L))
+    //   => ((athrid, val), (BlocksN, BlocksK), L)
+    return group<1, 3>(result(_, repeat<rank<1>(result)>(_)));
+  }
+
+  // ((athrid_val), (BlocksN, BlocksK, L)) -> (N, K, L)
+  template <typename Shape_NKL>
+  CUTE_HOST_DEVICE static constexpr auto TVbNbKL_to_offset_copy(
+      Shape_NKL shape_mkl) {
+    auto layout = TVbNbKL_to_offset(shape_mkl);
+    // for 4-bit elements, having >= 64 values per column
+    // allows TMA to load full 32-byte sectors
+    auto inner_layout =
+        make_layout(make_shape(_256{}, size<0>(layout) / _256{}));
+
+    return make_layout(inner_layout, get<1>(layout), get<2>(layout));
+  }
+
+  // ((BlockN, BlockK), (BlocksN, BlocksK), L) -> (storage_idx)
+  template <typename Shape_NKL>
+  CUTE_HOST_DEVICE static constexpr auto ilvd_NKbNbKL_to_offset(
+      Shape_NKL shape_mkl) {
+    constexpr auto block_layout = ppblock_ilvd_NK_to_offset();
+
+    // (BlocksN, BlocksK, L)
+    auto blocks_shape =
+        cute::transform(shape_mkl, append(PPBlockShape_NK{}, _1{}),
+                        [](auto x, auto y) { return x / y; });
+
+    // ((athrid, val), (BlocksN, BlocksK, L)) -> (storage_idx)
+    auto result = make_layout(
+        block_layout,
+        make_layout(blocks_shape,
+                    compact_col_major(blocks_shape, size(block_layout))));
+
+    // ((athrid, val), (BlocksN, BlocksK, L)) => ((athrid, val), (BlocksN,
+    // BlocksK), L)
+    return group<1, 3>(result(_, repeat<rank<1>(result)>(_)));
+  }
+
+  // (BlocksN, BlocksK, L) -> (storage_idx)
+  template <typename Shape_NKL>
+  CUTE_HOST_DEVICE static constexpr auto bNbKL_to_offset(Shape_NKL shape_mkl) {
+    // (BlocksN, BlocksK, L)
+    auto blocks_shape =
+        cute::transform(shape_mkl, append(PPBlockShape_NK{}, _1{}),
+                        [](auto x, auto y) { return x / y; });
+    auto stride = size(PPBlockShape_NK{});
+
+    // (BlocksN, BlocksK, L) -> (storage_idx)
+    return make_layout(blocks_shape, compact_col_major(blocks_shape, stride));
+  }
+
+  // ((athrid, val), (BlocksN, BlocksK, L)) -> (N, K, L)
+  template <class Shape_NKL>
+  CUTE_HOST_DEVICE static auto TVbNbK_to_NKL(Shape_NKL shape_mkl) {
+    auto tile = make_tile(make_layout(size<0>(PPBlockShape_NK{})),
+                          make_layout(size<1>(PPBlockShape_NK{})));
+
+    // ((BlockN, BlockK), (BlocksN, BlocksK, L)) -> (N, K, L)
+    auto tiled_A = zipped_divide(make_layout(shape_mkl), tile);
+    return tiled_A.compose(ppblock_TV_to_NK(), _);
+  }
+
+  // (N, K, L) -> ((athrid, val), (BlocksN, BlocksK), L)
+  template <class Shape_NKL>
+  CUTE_HOST_DEVICE static auto NKL_to_TVbNbK(Shape_NKL shape_mkl) {
+    auto TVbNbK_to_NKL_layout = TVbNbK_to_NKL(shape_mkl);
+    return blocked_product(ppblock_NK_to_TV(),
+                           make_layout(shape<1>(TVbNbK_to_NKL_layout)));
+  }
+};
+
+};  // namespace machete

custom_ops/gpu_ops/machete/machete_supported_schedules.cu

@@ -0,0 +1,72 @@
+// Copyright (c) 2025 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "machete_mm_launcher.cuh"
+#include "machete_prepack_launcher.cuh"
+
+template <typename T>
+std::optional<T> ConvertToStdOptional(const paddle::optional<T>& paddle_opt) {
+    return paddle_opt ? std::optional<T>(paddle_opt.get()) : std::nullopt;
+}
+
+std::vector<std::string> supported_schedules(
+    paddle::DataType a_type, int64_t b_type_id,
+    std::optional<paddle::DataType> maybe_group_scales_type,
+    std::optional<paddle::DataType> maybe_group_zeros_type,
+    std::optional<paddle::DataType> maybe_channel_scales_type,
+    std::optional<paddle::DataType> maybe_token_scales_type,
+    std::optional<paddle::DataType> maybe_out_type) {
+  machete::ScalarType const b_type = machete::ScalarType::from_id(b_type_id);
+  auto schedules = machete::supported_schedules_dispatch({
+      .a_type = a_type,
+      .b_type = b_type,
+      .maybe_group_scales_type = maybe_group_scales_type,
+      .maybe_group_zeros_type = maybe_group_zeros_type,
+      .maybe_channel_scales_type = maybe_channel_scales_type,
+      .maybe_token_scales_type = maybe_token_scales_type,
+      .maybe_out_type = maybe_out_type
+  });
+  return schedules;
+}
+
+std::vector<std::string> MacheteSupportedSchedules(
+    std::string const& a_type_str, std::string const& b_type_str) {
+  machete::ScalarTypeId b_type_id;
+  paddle::DataType a_type;
+  if (b_type_str == "uint4b8") {
+    b_type_id = machete::kU4B8.id();
+  } else {
+    PADDLE_ENFORCE(false, "b_type_str not supported!");
+  }
+  if (a_type_str == "bfloat16") {
+    a_type = paddle::DataType::BFLOAT16;
+  } else if (a_type_str == "float16") {
+    a_type = paddle::DataType::FLOAT16;
+  } else {
+    PADDLE_ENFORCE(false, "a_type_str not supported!");
+  }
+  std::optional<paddle::DataType> maybe_group_scales_type = std::optional<paddle::DataType>(a_type);
+  std::optional<paddle::DataType> maybe_out_type = std::optional<paddle::DataType>(a_type);
+  std::optional<paddle::DataType> maybe_group_zeros_type = std::nullopt;
+  std::optional<paddle::DataType> maybe_channel_scales_type = std::nullopt;
+  std::optional<paddle::DataType> maybe_token_scales_type = std::nullopt;
+
+  auto schedules =  supported_schedules(a_type, b_type_id,
+        maybe_group_scales_type,
+        maybe_group_zeros_type,
+        maybe_channel_scales_type,
+        maybe_token_scales_type,
+        maybe_out_type);
+  return schedules;
+}

custom_ops/gpu_ops/machete/utils/cute_utils.cuh

@@ -0,0 +1,69 @@
+// adapted from: https://github.com/vllm-project/vllm/blob/main/csrc/cutlass_extensions/cute_utils.cuh
+#pragma once
+
+#include <cute/tensor.hpp>
+
+namespace cute {
+
+////////////////////////////////////////////////////////////////////
+// layout utils
+////////////////////////////////////////////////////////////////////
+
+// Permute layout based on indices, example:
+//   permute_layout<1, 0>(layout) will swap the two dimensions
+//   permute_layout<0, 2, 1>(layout) will swap the last two dimensions
+template <size_t... I, typename Layout>
+CUTE_HOST_DEVICE static constexpr auto permute_layout(Layout l) {
+  static_assert(rank(l) == sizeof...(I), "Invalid permutation, rank mismatch");
+  return cute::make_layout(cute::get<I>(l)...);
+}
+
+// is the layout f(x) = x
+template <typename Layout>
+CUTE_HOST_DEVICE static constexpr bool is_identity_layout() {
+  if constexpr (std::is_same_v<Layout, void>) {
+    return true;
+  } else {
+    constexpr auto coalesced_layout = coalesce(Layout{});
+    if constexpr (rank(coalesced_layout) == 1 &&
+                  stride<0>(coalesced_layout) == 1) {
+      return true;
+    }
+    return false;
+  }
+}
+
+////////////////////////////////////////////////////////////////////
+// Pointer utils
+////////////////////////////////////////////////////////////////////
+
+template <class PointerType>
+static constexpr auto get_logical_ptr(PointerType* ptr) {
+  if constexpr (cute::sizeof_bits_v<PointerType> < 8) {
+    return cute::subbyte_iterator<PointerType>(ptr);
+  } else {
+    return ptr;
+  }
+}
+
+////////////////////////////////////////////////////////////////////
+// Misc utils
+////////////////////////////////////////////////////////////////////
+
+template <typename T, typename Elements>
+CUTE_HOST_DEVICE static constexpr auto create_auto_vectorizing_copy() {
+  constexpr auto bits = sizeof_bits_v<T> * Elements{};
+  if constexpr (bits % 128 == 0) {
+    return AutoVectorizingCopyWithAssumedAlignment<128>{};
+  } else if constexpr (bits % 64 == 0) {
+    return AutoVectorizingCopyWithAssumedAlignment<64>{};
+  } else if constexpr (bits % 32 == 0) {
+    return AutoVectorizingCopyWithAssumedAlignment<32>{};
+  } else if constexpr (bits % 16 == 0) {
+    return AutoVectorizingCopyWithAssumedAlignment<16>{};
+  } else {
+    return AutoVectorizingCopyWithAssumedAlignment<8>{};
+  }
+}
+
+};  // namespace cute

custom_ops/gpu_ops/machete/utils/machete_collective_builder.cuh

@@ -0,0 +1,44 @@
+// adapted from: https://github.com/vllm-project/vllm/blob/main/csrc/cutlass_extensions/vllm_collective_builder.cuh
+#pragma once
+
+#include "cutlass_extensions/gemm/collective/collective_builder.hpp"
+
+namespace cutlass::gemm::collective {
+using namespace cute;
+
+//
+// MacheteCollectiveBuilder is a wrapper around CollectiveBuilder that allows for
+// for custom kernel tags, allowing you to build custom collectives. Without
+// touching the cutlass library headers, using `CutlassKernelTag` will mean it
+// will resort to using the standard cutlass collective builder.
+//
+
+// Use the default Cutlass collective builder, i.e. use an unmodified cutless
+// collective
+struct CutlassKernelTag {};
+
+template <class KernelTag, class ArchTag, class OpClass, class ElementA,
+          class GmemLayoutA, int AlignmentA, class ElementB, class GmemLayoutB,
+          int AlignmentB, class ElementAccumulator, class TileShape_MNK,
+          class ClusterShape_MNK, class StageCountType,
+          class KernelScheduleType, class Enable = void>
+struct MacheteCollectiveBuilder {
+  static_assert(sizeof(ElementA) == 0,
+                "Could not build a collective for given parameters.");
+};
+
+template <class ArchTag, class OpClass, class ElementA, class GmemLayoutA,
+          int AlignmentA, class ElementB, class GmemLayoutB, int AlignmentB,
+          class ElementAccumulator, class TileShape_MNK, class ClusterShape_MNK,
+          class StageCountType, class KernelScheduleType>
+struct MacheteCollectiveBuilder<
+    CutlassKernelTag, ArchTag, OpClass, ElementA, GmemLayoutA, AlignmentA,
+    ElementB, GmemLayoutB, AlignmentB, ElementAccumulator, TileShape_MNK,
+    ClusterShape_MNK, StageCountType, KernelScheduleType> {
+  using CollectiveOp = typename CollectiveBuilder<
+      ArchTag, OpClass, ElementA, GmemLayoutA, AlignmentA, ElementB,
+      GmemLayoutB, AlignmentB, ElementAccumulator, TileShape_MNK,
+      ClusterShape_MNK, StageCountType, KernelScheduleType>::CollectiveOp;
+};
+
+};  // namespace cutlass::gemm::collective

custom_ops/gpu_ops/machete/utils/machete_custom_types.cuh

@@ -0,0 +1,51 @@
+// adapted from: https://github.com/vllm-project/vllm/blob/main/csrc/cutlass_extensions/vllm_custom_types.cuh
+#pragma once
+
+#include "cutlass/integer_subbyte.h"
+
+namespace cutlass {
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <int Bits, int Bias, bool Signed = false>
+struct machete_biased_integer_subbyte : public integer_subbyte<Bits, Signed> {
+  using Base = integer_subbyte<Bits, Signed>;
+
+  using Storage = typename Base::Storage;
+  using xint_t = typename Base::xint_t;
+
+  using Base::bits_mask_;
+  using Base::sign_mask_;
+  using Base::storage;
+
+  //
+  // Methods
+  //
+
+  /// No operation
+  machete_biased_integer_subbyte() = default;
+
+  /// Conversion from integer type
+  CUTLASS_HOST_DEVICE explicit machete_biased_integer_subbyte(int value)
+      : Base(value) {}
+  CUTLASS_HOST_DEVICE explicit machete_biased_integer_subbyte(unsigned value)
+      : Base(value) {}
+  CUTLASS_HOST_DEVICE explicit machete_biased_integer_subbyte(double value)
+      : Base(value) {}
+};
+///////////////////////////////////////////////////////////////////////////////////////////////////
+
+// "GPTQ" types, i.e. symmetric quantization
+using machete_uint4b8_t = machete_biased_integer_subbyte<4, 8>;      // u4b8
+using machete_uint8b128_t = machete_biased_integer_subbyte<8, 128>;  // u8b128
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <int Bits, int Bias, bool Signed>
+struct sizeof_bits<machete_biased_integer_subbyte<Bits, Bias, Signed>> {
+  static constexpr int value = Bits;
+};
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+
+}  // namespace cutlass

custom_ops/gpu_ops/machete/utils/machete_numeric_conversion.cuh

@@ -0,0 +1,993 @@
+// adapted from: https://github.com/vllm-project/vllm/blob/main/csrc/cutlass_extensions/vllm_numeric_conversion.cuh
+#pragma once
+
+#include "cutlass/numeric_conversion.h"
+#include "machete_custom_types.cuh"
+#include "cute_utils.cuh"
+#include "machete_type_utils.cuh"
+
+// this file extends:
+//   https://github.com/NVIDIA/cutlass/blob/cutlass-3.5.0/include/cutlass/numeric_conversion.h
+// with vllm specific type conversions, namely: machete_uint4b8_t, machete_uint8b128_t
+// as well as adds interleaved numeric array converters for specific types.
+// (interleaved numeric array converters can be more efficient for subbyte
+// types)
+
+namespace cutlass {
+
+// InterleavedNumericArrayConverter is like NumericArrayConverter but also
+// deinterleaves converted elements based on IlvBlkLayout, interleaving can
+// make subbyte converts more efficient by allowing for efficient extraction
+// of subbyte elements from a 32bit register.
+template <typename IlvBlkLayout, typename T, typename S, int N,
+          FloatRoundStyle Round = FloatRoundStyle::round_to_nearest,
+          class Enable = void>
+struct InterleavedNumericArrayConverter {
+  using Converter = NumericArrayConverter<T, S, N, Round>;
+
+  using result_type = typename Converter::result_type;
+  using source_type = typename Converter::source_type;
+
+  CUTLASS_DEVICE
+  static result_type convert(source_type const& source) {
+    if (cute::elect_one_sync()) {
+      if constexpr (std::is_same_v<IlvBlkLayout, void>) {
+        printf(
+            "Convert %s <= %s (N = %d, IlvBlkLayout = void), not implemented\n",
+            nameof_v<T>, nameof_v<S>, N);
+      } else {
+        printf(
+            "Convert %s <= %s (N = %d, size(IlvBlkLayout{}) = %d), not "
+            "implemented\n",
+            nameof_v<T>, nameof_v<S>, N, size(IlvBlkLayout{}));
+      }
+      __brkpt();
+    }
+    return {};
+  }
+
+  CUTLASS_DEVICE
+  result_type operator()(source_type const& s) const { return convert(s); }
+};
+
+template <typename IlvBlkLayout, typename T, typename S, int N,
+          FloatRoundStyle Round>
+struct InterleavedNumericArrayConverter<
+    IlvBlkLayout, T, S, N, Round,
+    std::enable_if_t<is_identity_layout<IlvBlkLayout>()>> {
+  using Converter = NumericArrayConverter<T, S, N, Round>;
+
+  using result_type = typename Converter::result_type;
+  using source_type = typename Converter::source_type;
+
+  CUTLASS_DEVICE
+  static result_type convert(source_type const& source) {
+    return Converter::convert(source);
+  }
+
+  CUTLASS_DEVICE
+  result_type operator()(source_type const& s) const { return convert(s); }
+};
+
+template <typename RegConvert32bit, typename T, typename S, int N>
+struct ArrayConverterPacked32Bit {
+  using result_type = Array<T, N>;
+  using source_type = Array<S, N>;
+
+  using result_packed_8_t = Array<T, 8>;
+  using result_packed_4_t = Array<T, 4>;
+  using result_packed_2_t = Array<T, 2>;
+  using src_packed_8_t = Array<S, 8>;
+  using src_packed_4_t = Array<S, 4>;
+  using src_packed_2_t = Array<S, 2>;
+
+  static_assert(N % 2 == 0, "N must be a multiple of 2");
+  static_assert(cutlass::sizeof_bits_v<S> >= 4);  // TODO: add 16 packed sources
+  static_assert(32 % cutlass::sizeof_bits_v<S> == 0);
+  static constexpr auto src_elems_per_32bit_reg =
+      32 / cutlass::sizeof_bits_v<S>;
+
+  // Maybe not Valid. ScalarConverter will not actually work unless
+  // NumericConverter<T, S, Round> is implemented. However it won't be used
+  // anyways since we assert N % 2 == 0, just here for compliance with
+  // VectorizedConverter.
+  using ScalarConverter = NumericConverter<T, S>;
+
+  template <typename PackedSrc>
+  CUTLASS_DEVICE static auto to_regs(PackedSrc const& src) {
+    if constexpr (sizeof(PackedSrc) == 1) {
+      return Array<uint32_t, 1>{reinterpret_cast<uint8_t const&>(src)};
+    } else if constexpr (sizeof(PackedSrc) == 2) {
+      return Array<uint32_t, 1>{reinterpret_cast<uint16_t const&>(src)};
+    } else if constexpr (sizeof(PackedSrc) == 4) {
+      return Array<uint32_t, 1>{reinterpret_cast<uint32_t const&>(src)};
+    } else {
+      static_assert(sizeof(PackedSrc) == 8);
+      return reinterpret_cast<Array<uint32_t, 2> const&>(src);
+    }
+  }
+
+  // The core converter uses bit tricks to construct a known FP16 number, then
+  // does a subtraction in FP16 for the final result.
+  template <typename PackedResultType, typename PackedSrcType>
+  CUTLASS_DEVICE static PackedResultType packed_convert(
+      PackedSrcType const& source) {
+    static_assert(PackedSrcType::kElements == PackedResultType::kElements);
+    static_assert(PackedResultType::kElements == 2 ||
+                      PackedResultType::kElements == 4 ||
+                      PackedResultType::kElements == 8,
+                  "Invalid PackedResultType must be 2, 4 or 8.");
+    static_assert(std::is_same_v<typename PackedSrcType::Element, S>);
+    static_assert(std::is_same_v<typename PackedResultType::Element, T>);
+
+    return RegConvert32bit::template convert<PackedResultType>(to_regs(source));
+  }
+
+  friend class detail::VectorizedConverter;
+
+ public:
+  CUTLASS_DEVICE static result_type convert(source_type const& source) {
+    result_type result;
+    using ConverterType =
+        ArrayConverterPacked32Bit<RegConvert32bit,
+                                  typename result_type::Element,
+                                  typename source_type::Element, N>;
+
+    if constexpr (src_elems_per_32bit_reg >= 8) {
+      detail::VectorizedConverter::convert<
+          ConverterType, result_packed_8_t, src_packed_8_t, result_packed_4_t,
+          src_packed_4_t, result_packed_2_t, src_packed_2_t>(result, source);
+    } else if constexpr (src_elems_per_32bit_reg >= 4) {
+      detail::VectorizedConverter::convert<ConverterType, result_packed_4_t,
+                                           src_packed_4_t, result_packed_2_t,
+                                           src_packed_2_t>(result, source);
+    } else {
+      detail::VectorizedConverter::convert<ConverterType, result_packed_2_t,
+                                           src_packed_2_t>(result, source);
+    }
+
+    return result;
+  }
+};
+
+// Convert 8 4bit values packed into a 32bit register to 8 8bit values packed
+// into 2 32bit register.
+template <uint8_t LUT0, uint8_t LUT1, uint8_t LUT2, uint8_t LUT3,    //
+          uint8_t LUT4, uint8_t LUT5, uint8_t LUT6, uint8_t LUT7,    //
+          uint8_t LUT8, uint8_t LUT9, uint8_t LUT10, uint8_t LUT11,  //
+          uint8_t LUT12, uint8_t LUT13, uint8_t LUT14, uint8_t LUT15>
+CUTLASS_DEVICE cutlass::AlignedArray<uint32_t, 2> lut_4bit_to_8bit_convert(
+    uint32_t src) {
+  cutlass::AlignedArray<uint32_t, 2> r;
+  // Determines if the value is in the top half of the LUT if set or
+  //  (i.e. LUT[8:15]) in the bottom half (i.e. LUT[0:7]) if not set. Then move
+  //  into bit position 0x4 of each nibble so when or'd with final_prmt_base it
+  //  selects the correct candidate. When elements in final_prmt_base
+  //  are >= 0x4, the high candidate is selected (i.e. LUT[8:15]), when elements
+  //  are  < 0x4, the low candidate is selected (i.e. LUT[0:7])
+  uint32_t high_bit = (src & 0x88888888) >> 1;
+
+  // `high_bit` is OR'd with 0x31203120 to find the correct value in the LUT
+  // (selects correct high or low candidate)
+  const uint32_t final_prmt_base = 0x32103210;
+
+  // Ignore the high bit when indexing into LUT, for each 4bit value
+  //  we index into both the high and low candidates then use
+  //  high_bit | final_prmt_base to select the correct candidate
+  uint32_t lut_idx = (src & 0x77777777);
+
+  auto pack = [](uint8_t a, uint8_t b, uint8_t c, uint8_t d) {
+    return uint32_t(a) | (uint32_t(b) << 8) | (uint32_t(c) << 16) |
+           (uint32_t(d) << 24);
+  };
+
+  static constexpr uint32_t LOW_0 = pack(LUT0, LUT1, LUT2, LUT3);
+  static constexpr uint32_t LOW_1 = pack(LUT4, LUT5, LUT6, LUT7);
+  static constexpr uint32_t HIGH_0 = pack(LUT8, LUT9, LUT10, LUT11);
+  static constexpr uint32_t HIGH_1 = pack(LUT12, LUT13, LUT14, LUT15);
+
+  CUTLASS_PRAGMA_UNROLL
+  for (int ii = 0; ii < 2; ++ii, lut_idx >>= 16, high_bit >>= 16) {
+    uint32_t final_prmt_idx = final_prmt_base | high_bit;
+
+    // This uses a look up table to convert packed int4s to packed int8s,
+    // using the int4 value as the index to prmt. It first select both the
+    // high and low candidates, then uses the high bit (i.e. `high_bit`) to
+    // select the correct candidate.
+    asm volatile(
+        "{\n"
+        "  .reg .b32 low, high;\n"
+        "  prmt.b32 low, %1, %2, %5;\n"
+        "  prmt.b32 high, %3, %4, %5;\n"
+        "  prmt.b32 %0, low, high, %6;\n"
+        "}\n"
+        : "=r"(r[ii])
+        : "n"(LOW_0), "n"(LOW_1), "n"(HIGH_0), "n"(HIGH_1), "r"(lut_idx),
+          "r"(final_prmt_idx));
+  }
+
+  return r;
+};
+
+// for Array<int8_t, N> <= Array<machete_uint4b8_t, N>
+template <FloatRoundStyle Round, int N>
+struct NumericArrayConverter<int8_t, machete_uint4b8_t, N, Round> {
+  using result_type = Array<int8_t, N>;
+  using source_type = Array<machete_uint4b8_t, N>;
+
+  static FloatRoundStyle const round_style = Round;
+
+ private:
+  struct RegConvert {
+    template <typename PackedResultType>
+    CUTLASS_DEVICE static PackedResultType convert(Array<uint32_t, 1> src_) {
+      // [-8, -7, -6, -5, -4, -3, -2, -1, 0, 1, 2, 3, 4, 5, 6, 7] as int8s
+      auto r = lut_4bit_to_8bit_convert<0xF8, 0xF9, 0xFA, 0xFB,  //
+                                        0xFC, 0xFD, 0xFE, 0xFF,  //
+                                        0x00, 0x01, 0x02, 0x03,  //
+                                        0x04, 0x05, 0x06, 0x07>(src_[0]);
+      return reinterpret_cast<PackedResultType&>(r);
+    };
+  };
+
+ public:
+  CUTLASS_DEVICE
+  static result_type convert(source_type const& source) {
+    return ArrayConverterPacked32Bit<RegConvert, typename result_type::Element,
+                                     typename source_type::Element,
+                                     N>::convert(source);
+  }
+
+  CUTLASS_DEVICE
+  result_type operator()(source_type const& s) const { return convert(s); }
+};
+
+// for Array<cutlass::float_e4m3_t, N> <= Array<machete_uint4b8_t, N>
+template <FloatRoundStyle Round, int N>
+struct NumericArrayConverter<cutlass::float_e4m3_t, machete_uint4b8_t, N, Round> {
+  using result_type = Array<cutlass::float_e4m3_t, N>;
+  using source_type = Array<machete_uint4b8_t, N>;
+
+  static FloatRoundStyle const round_style = Round;
+
+ private:
+  struct RegConvert {
+    template <typename PackedResultType>
+    CUTLASS_DEVICE static PackedResultType convert(Array<uint32_t, 1> src_) {
+      // [-8, -7, -6, -5, -4, -3, -2, -1, 0, 1, 2, 3, 4, 5, 6, 7] as fp8s
+      auto r = lut_4bit_to_8bit_convert<0xD0, 0xCE, 0xCC, 0xCA,  //
+                                        0xC8, 0xC4, 0xC0, 0xB8,  //
+                                        0x00, 0x38, 0x40, 0x44,  //
+                                        0x48, 0x4A, 0x4C, 0x4E>(src_[0]);
+      return reinterpret_cast<PackedResultType&>(r);
+    };
+  };
+
+ public:
+  CUTLASS_DEVICE
+  static result_type convert(source_type const& source) {
+    return ArrayConverterPacked32Bit<RegConvert, typename result_type::Element,
+                                     typename source_type::Element,
+                                     N>::convert(source);
+  }
+
+  CUTLASS_DEVICE
+  result_type operator()(source_type const& s) const { return convert(s); }
+};
+
+// for Array<cutlass::half_t, N> <= Array<machete_uint4b8_t, N>
+template <FloatRoundStyle Round, int N>
+struct NumericArrayConverter<cutlass::half_t, machete_uint4b8_t, N, Round> {
+  using result_type = Array<cutlass::half_t, N>;
+  using source_type = Array<machete_uint4b8_t, N>;
+
+  struct RegConvert {
+    template <typename PackedResultType>
+    CUTLASS_DEVICE static PackedResultType convert(Array<uint32_t, 1> src_) {
+      uint32_t src = src_[0];
+      using RegArray =
+          cutlass::AlignedArray<uint32_t, PackedResultType::kElements / 2,
+                                sizeof(PackedResultType)>;
+      RegArray r;
+
+      // Below constructs the following temporary:
+      // fp16s_01 = {0x00, i4_01, 0x00, i4_01}
+      // fp16s_23 = {0x00, i4_23, 0x00, i4_23}
+      // fp16s_45 = {0x00, i4_45, 0x00, i4_45}
+      // fp16s_67 = {0x00, i4_67, 0x00, i4_67}
+      // We use inline asm instead of __byte_perm intrinsic since we don't want
+      // the documented (& 0x7) on the index. NVCC might be able to optimize it
+      // out since the index is a constexpr, but we choose to be safe about it
+      // here.
+      uint32_t prmt_indices[4] = {0x4040, 0x4141, 0x4242, 0x4343};
+      static_assert(RegArray::kElements <= 4,
+                    "Too many inputs for F16 -> I4 vector converter");
+      CUTLASS_PRAGMA_UNROLL
+      for (int ii = 0; ii < RegArray::kElements; ++ii) {
+        asm volatile(
+            "{\n"
+            "  prmt.b32 %0, %1, %2, %3;\n"
+            "}\n"
+            : "=r"(r[ii])
+            : "r"(src), "n"(0), "r"(prmt_indices[ii]));
+      }
+
+      // Since the stored 4bit values are biased by 8 we get stored_val = (x+8)
+      //  we are trying to construct x and a fp16 value
+      // The below XOR does the following:
+      //  1) Sets the exponent bits of the FP16 to the correct value for the
+      //  FP16 magic_num. We will be constructing {1024+16*(x1+8), 1024+(x0+8)},
+      //  where x1 in the high nibble and x0 is the low nibble then using hfma
+      //  to subtract 1032 from that
+      // The AND does the following:
+      //  1) Clear the set bits for the int4 we will ignore.
+      // We use lop3 so that we can use 1 instruction for AND and XOR.
+      static constexpr uint32_t xor_mask = 0x64006400;
+      static constexpr uint32_t and_mask = 0xFFF0FF0F;
+      static constexpr uint32_t immLut = (0xf0 & 0xcc) ^ 0xaa;
+
+      // For each operand, computes:
+      // r[i] = (r[i] & and_mask) ^ xor_mask
+      CUTLASS_PRAGMA_UNROLL
+      for (int ii = 0; ii < RegArray::kElements; ++ii) {
+        asm volatile(
+            "{\n"
+            "  lop3.b32 %0, %0, %1, %2, %3;\n"
+            "}\n"
+            : "+r"(r[ii])
+            : "n"(and_mask), "n"(xor_mask), "n"(immLut));
+      }
+
+      // We will issue 2 hfmas that do the following:
+      // {x1, x0} = {1024+16*(x1+8), 1024+(x0+8)} * {1/16, 1} - {72, 1032}
+      //          = {x1 + 1152, x0 + 1032} * {1/16, 1} - {72, 1032}
+      static constexpr uint32_t hfma_bias_rep = 0xD480E408;   // {72, 1032}
+      static constexpr uint32_t hfma_scale_rep = 0x2C003C00;  // {1 / 16, 1}
+
+      const half2& hfma_bias = reinterpret_cast<const half2&>(hfma_bias_rep);
+      const half2& hfma_scale = reinterpret_cast<const half2&>(hfma_scale_rep);
+      CUTLASS_PRAGMA_UNROLL
+      for (int ii = 0; ii < RegArray::kElements; ++ii) {
+        half2& fp16x2_val = reinterpret_cast<__half2&>(r[ii]);
+        fp16x2_val = __hfma2(hfma_scale, fp16x2_val, hfma_bias);
+      }
+
+      return reinterpret_cast<PackedResultType&>(r);
+    };
+  };
+
+ public:
+  CUTLASS_DEVICE
+  static result_type convert(source_type const& source) {
+    return ArrayConverterPacked32Bit<RegConvert, typename result_type::Element,
+                                     typename source_type::Element,
+                                     N>::convert(source);
+  }
+
+  CUTLASS_DEVICE
+  result_type operator()(source_type const& s) const { return convert(s); }
+};
+
+// for Array<cutlass::half_t, N> <= Array<machete_uint4b8_t, N>
+//   for IlvdLayout: (2, 4):(4, 1)
+template <FloatRoundStyle Round, int N>
+struct InterleavedNumericArrayConverter<Layout<Shape<_2, _4>, Stride<_4, _1>>,
+                                        cutlass::half_t, machete_uint4b8_t, N,
+                                        Round, void> {
+  using IlvdLayout = Layout<Shape<_2, _4>, Stride<_4, _1>>;
+  static_assert(N % size(IlvdLayout{}) == 0);
+
+  using result_type = Array<cutlass::half_t, N>;
+  using source_type = Array<machete_uint4b8_t, N>;
+
+  static FloatRoundStyle const round_style = Round;
+
+ private:
+  struct RegConvert {
+    template <typename PackedResultType>
+    CUTLASS_DEVICE static PackedResultType convert(Array<uint32_t, 1> src_) {
+      uint32_t src = src_[0];
+      using RegArray =
+          cutlass::AlignedArray<uint32_t, PackedResultType::kElements / 2,
+                                sizeof(PackedResultType)>;
+      RegArray r;
+
+      static_assert(PackedResultType::kElements <= size(IlvdLayout{}));
+      static constexpr uint32_t xor_mask = 0x64006400;
+
+      for (int ii = 0; ii < RegArray::kElements; ii += 2) {
+        auto src_ = src >> (4 * (ii));
+        r[ii + 0] = src_;
+        r[ii + 1] = src_;
+
+        static constexpr uint32_t and_xor_imm_lut = (0xf0 & 0xcc) ^ 0xaa;
+
+        static constexpr uint32_t low_nib_mask = 0x000F000F;
+        static constexpr uint32_t high_nib_mask = 0x00F000F0;
+
+        asm volatile(
+            "{\n"
+            "  lop3.b32 %0, %0, %1, %2, %3;\n"
+            "}\n"
+            : "+r"(r[ii + 0])
+            : "n"(low_nib_mask), "n"(xor_mask), "n"(and_xor_imm_lut));
+
+        asm volatile(
+            "{\n"
+            "  lop3.b32 %0, %0, %1, %2, %3;\n"
+            "}\n"
+            : "+r"(r[ii + 1])
+            : "n"(high_nib_mask), "n"(xor_mask), "n"(and_xor_imm_lut));
+
+        // For low nibble:
+        //  {x1, x0} = {1024+(x1+8), 1024+(x0+8)} * {1, 1} - {1032, 1032}
+        // For high nibble:
+        //  {x1, x0} = {1024+16*(x1+8), 1024+16*(x0+8)} * {1/16, 1/16}
+        //             - {72, 72}
+        static constexpr uint32_t low_nib_bias = 0x64086408;    // {1032, 1032}
+        static constexpr uint32_t high_nib_scale = 0x2C002C00;  // {1/16, 1/16}
+        static constexpr uint32_t high_nib_bias = 0xD480D480;   // {-72, -72}
+
+        {
+          half2& fp16x2_val = reinterpret_cast<__half2&>(r[ii + 0]);
+          fp16x2_val =
+              __hsub2(fp16x2_val, reinterpret_cast<const half2&>(low_nib_bias));
+        }
+
+        {
+          half2& fp16x2_val = reinterpret_cast<__half2&>(r[ii + 1]);
+          fp16x2_val = __hfma2(fp16x2_val,
+                               reinterpret_cast<const half2&>(high_nib_scale),
+                               reinterpret_cast<const half2&>(high_nib_bias));
+        }
+      }
+
+      return reinterpret_cast<PackedResultType&>(r);
+    };
+  };
+
+ public:
+  CUTLASS_DEVICE
+  static result_type convert(source_type const& source) {
+    return ArrayConverterPacked32Bit<RegConvert, typename result_type::Element,
+                                     typename source_type::Element,
+                                     N>::convert(source);
+  }
+
+  CUTLASS_DEVICE
+  result_type operator()(source_type const& s) const { return convert(s); }
+};
+
+// for Array<cutlass::half_t, N> <= Array<uint4_t, N>
+//   for IlvdLayout: (2, 4):(4, 1)
+template <FloatRoundStyle Round, int N>
+struct InterleavedNumericArrayConverter<Layout<Shape<_2, _4>, Stride<_4, _1>>,
+                                        cutlass::half_t, uint4_t, N, Round,
+                                        void> {
+  using IlvdLayout = Layout<Shape<_2, _4>, Stride<_4, _1>>;
+  static_assert(N % size(IlvdLayout{}) == 0);
+
+  using result_type = Array<cutlass::half_t, N>;
+  using source_type = Array<uint4_t, N>;
+
+  static FloatRoundStyle const round_style = Round;
+
+ private:
+  struct RegConvert {
+    template <typename PackedResultType>
+    CUTLASS_DEVICE static PackedResultType convert(Array<uint32_t, 1> src_) {
+      uint32_t src = src_[0];
+      using RegArray =
+          cutlass::AlignedArray<uint32_t, PackedResultType::kElements / 2,
+                                sizeof(PackedResultType)>;
+      RegArray r;
+
+      static_assert(PackedResultType::kElements <= size(IlvdLayout{}));
+      static constexpr uint32_t xor_mask = 0x64006400;
+
+      for (int ii = 0; ii < RegArray::kElements; ii += 2) {
+        auto src_ = src >> (4 * (ii));
+        r[ii + 0] = src_;
+        r[ii + 1] = src_;
+
+        static constexpr uint32_t and_xor_imm_lut = (0xf0 & 0xcc) ^ 0xaa;
+
+        static constexpr uint32_t low_nib_mask = 0x000F000F;
+        static constexpr uint32_t high_nib_mask = 0x00F000F0;
+
+        asm volatile(
+            "{\n"
+            "  lop3.b32 %0, %0, %1, %2, %3;\n"
+            "}\n"
+            : "+r"(r[ii + 0])
+            : "n"(low_nib_mask), "n"(xor_mask), "n"(and_xor_imm_lut));
+
+        asm volatile(
+            "{\n"
+            "  lop3.b32 %0, %0, %1, %2, %3;\n"
+            "}\n"
+            : "+r"(r[ii + 1])
+            : "n"(high_nib_mask), "n"(xor_mask), "n"(and_xor_imm_lut));
+
+        // For low nibble:
+        //  {x1, x0} = {1024+x1, 1024+x0} - {1024, 1024}
+        // For high nibble:
+        //  {x1, x0} = {1024+16*x1, 1024+16*x0} * {1/16, 1/16} - {64, 64}
+        static constexpr uint32_t low_nib_bias = 0x64006400;    // {1024, 1024}
+        static constexpr uint32_t high_nib_scale = 0x2C002C00;  // {1/16, 1/16}
+        static constexpr uint32_t high_nib_bias = 0xD400D400;   // {-64, -64}
+
+        {
+          half2& fp16x2_val = reinterpret_cast<__half2&>(r[ii + 0]);
+          fp16x2_val =
+              __hsub2(fp16x2_val, reinterpret_cast<const half2&>(low_nib_bias));
+        }
+
+        {
+          half2& fp16x2_val = reinterpret_cast<__half2&>(r[ii + 1]);
+          fp16x2_val = __hfma2(fp16x2_val,
+                               reinterpret_cast<const half2&>(high_nib_scale),
+                               reinterpret_cast<const half2&>(high_nib_bias));
+        }
+      }
+
+      return reinterpret_cast<PackedResultType&>(r);
+    };
+  };
+
+ public:
+  CUTLASS_DEVICE
+  static result_type convert(source_type const& source) {
+    return ArrayConverterPacked32Bit<RegConvert, typename result_type::Element,
+                                     typename source_type::Element,
+                                     N>::convert(source);
+  }
+
+  CUTLASS_DEVICE
+  result_type operator()(source_type const& s) const { return convert(s); }
+};
+
+// for Array<cutlass::half_t, N> <= Array<machete_uint8b128_t, N>
+template <FloatRoundStyle Round, int N>
+struct NumericArrayConverter<cutlass::half_t, machete_uint8b128_t, N, Round> {
+  using result_type = Array<cutlass::half_t, N>;
+  using source_type = Array<machete_uint8b128_t, N>;
+
+  struct RegConvert {
+    template <typename PackedResultType>
+    CUTLASS_DEVICE static PackedResultType convert(Array<uint32_t, 1> src_) {
+      uint32_t src = src_[0];
+      // Hold output FP16s in reg. We need 1 reg for every 2 elements
+      using RegArray =
+          cutlass::AlignedArray<uint32_t, PackedResultType::kElements / 2,
+                                sizeof(PackedResultType)>;
+      RegArray r;
+
+      uint32_t const prmt_indices[2] = {0x5150, 0x5352};
+      static constexpr uint32_t start_byte_for_fp16 = 0x64646464;
+
+      for (int ii = 0; ii < RegArray::kElements; ++ii) {
+        asm volatile("prmt.b32 %0,%1,%2,%3;\n"
+                     : "=r"(r[ii])
+                     : "r"(src), "n"(start_byte_for_fp16),
+                       "r"(prmt_indices[ii]));
+      }
+
+      // -128 is folded into bias subtraction, i.e. the 0x80 in the low bytes
+      static constexpr uint32_t bias_rep = 0x64806480;
+      const half2& bias = reinterpret_cast<const half2&>(bias_rep);
+      CUTLASS_PRAGMA_UNROLL
+      for (int ii = 0; ii < RegArray::kElements; ++ii) {
+        half2& fp16x2_val = reinterpret_cast<__half2&>(r[ii]);
+        fp16x2_val = __hsub2(fp16x2_val, bias);
+      }
+
+      return reinterpret_cast<PackedResultType&>(r);
+    };
+  };
+
+ public:
+  CUTLASS_DEVICE
+  static result_type convert(source_type const& source) {
+    return ArrayConverterPacked32Bit<RegConvert, typename result_type::Element,
+                                     typename source_type::Element,
+                                     N>::convert(source);
+  }
+
+  CUTLASS_DEVICE
+  result_type operator()(source_type const& s) const { return convert(s); }
+};
+
+// for Array<cutlass::float, N> <= Array<machete_uint8b128_t, N>
+template <FloatRoundStyle Round, int N>
+struct NumericArrayConverter<float, machete_uint8b128_t, N, Round> {
+  using result_type = Array<float, N>;
+  using source_type = Array<machete_uint8b128_t, N>;
+  static FloatRoundStyle const round_style = Round;
+
+ private:
+  struct RegConvert {
+    template <typename PackedResultType>
+    CUTLASS_DEVICE static PackedResultType convert(Array<uint32_t, 1> src_) {
+      uint32_t src = src_[0];
+      PackedResultType r;
+
+      // __byte_perm simulates the add.u32 0x4B000000 to every u8 element of
+      // u8x4 source and stores the result in r (without introducing extra
+      // cvt.u32.u8 instruction)
+      uint32_t const prmt_indices[4] = {0x7650, 0x7651, 0x7652, 0x7653};
+      uint32_t* result_as_int = reinterpret_cast<uint32_t*>(&r);
+      for (int ii = 0; ii < PackedResultType::kElements; ++ii) {
+        result_as_int[ii] = __byte_perm(src, 0x4B000000, prmt_indices[ii]);
+        // Subtract the magic number 0x4B000000 from tmp in floating-point
+        // arithmetic to obtain final result
+        r[ii] -= (8388608.f + 128.f);  // fold in -128 bias
+      }
+
+      return r;
+    };
+  };
+
+ public:
+  CUTLASS_DEVICE
+  static result_type convert(source_type const& source) {
+    return ArrayConverterPacked32Bit<RegConvert, typename result_type::Element,
+                                     typename source_type::Element,
+                                     N>::convert(source);
+  }
+
+  CUTLASS_DEVICE
+  result_type operator()(source_type const& s) const { return convert(s); }
+};
+
+#if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 800)
+
+// for Array<cutlass::bfloat16_t, N> <= Array<machete_uint4b8_t, N>
+template <FloatRoundStyle Round, int N>
+struct NumericArrayConverter<cutlass::bfloat16_t, machete_uint4b8_t, N, Round> {
+  using result_type = Array<cutlass::bfloat16_t, N>;
+  using source_type = Array<machete_uint4b8_t, N>;
+
+  static FloatRoundStyle const round_style = Round;
+
+ private:
+  struct RegConvert {
+    template <typename PackedResultType>
+    CUTLASS_DEVICE static PackedResultType convert(Array<uint32_t, 1> src_) {
+      uint32_t src_reg = src_[0];
+      // Hold output BF16s in reg. We need 1 reg for every 2 elements
+      using RegArray =
+          cutlass::AlignedArray<uint32_t, PackedResultType::kElements / 2,
+                                sizeof(PackedResultType)>;
+      RegArray r;
+      uint32_t src_reg_shifted = src_reg >> 4;
+
+      // Below constructs the following temporary:
+      uint32_t const prmt_indices[4] = {0xF4F0, 0xF5F1, 0xF6F2, 0xF7F3};
+      static_assert(RegArray::kElements <= 4,
+                    "Too many inputs for uint4b8_t -> BF16 vector converter");
+      CUTLASS_PRAGMA_UNROLL
+      for (int ii = 0; ii < RegArray::kElements; ++ii) {
+        asm volatile(
+            "{\n"
+            "  prmt.b32 %0, %1, %2, %3;\n"
+            "}\n"
+            : "=r"(r[ii])
+            : "r"(src_reg), "r"(src_reg_shifted), "r"(prmt_indices[ii]));
+      }
+
+      // Since the stored 4bit values are biased by 8 we get stored_val = (x+8)
+      //  we are trying to construct x and a BF16 value
+      // The below XOR does the following:
+      //  1) Sets the exponent bits of the BF16 to the correct value for the
+      //  BF16 magic_num. We will be constructing {128 + (x1+8), 128 + (x0+8)}
+      //  and subtracting 136 to get {x1, x0}
+      static constexpr uint32_t xor_mask = 0x43004300;
+      static constexpr uint32_t and_mask = 0x000F000F;
+      static constexpr uint32_t immLut = (0xf0 & 0xcc) ^ 0xaa;
+
+      // For each operand, computes:
+      // r[i] = (r[i] & and_mask) ^ xor_mask
+      CUTLASS_PRAGMA_UNROLL
+      for (int ii = 0; ii < RegArray::kElements; ++ii) {
+        asm volatile(
+            "{\n"
+            "  lop3.b32 %0, %0, %1, %2, %3;\n"
+            "}\n"
+            : "+r"(r[ii])
+            : "n"(and_mask), "n"(xor_mask), "n"(immLut));
+      }
+
+      // We will issue 2 bfmas that do the following:
+      // high BF16:
+      // hi_bf16 - 136, lo_bf16 - 136
+
+      // This is the BF16 {136, 136} represented as an integer.
+      static constexpr uint32_t bias_rep = 0x43084308;
+      const __nv_bfloat162& bias =
+          reinterpret_cast<const __nv_bfloat162&>(bias_rep);
+
+      CUTLASS_PRAGMA_UNROLL
+      for (int ii = 0; ii < RegArray::kElements; ++ii) {
+        __nv_bfloat162& bf16x2_val = reinterpret_cast<__nv_bfloat162&>(r[ii]);
+        bf16x2_val = __hsub2(bf16x2_val, bias);
+      }
+
+      return reinterpret_cast<PackedResultType&>(r);
+    }
+  };
+
+ public:
+  CUTLASS_DEVICE
+  static result_type convert(source_type const& source) {
+    return ArrayConverterPacked32Bit<RegConvert, typename result_type::Element,
+                                     typename source_type::Element,
+                                     N>::convert(source);
+  }
+
+  CUTLASS_DEVICE
+  result_type operator()(source_type const& s) const { return convert(s); }
+};
+
+// for Array<cutlass::bfloat16_t, N> <= Array<machete_uint4b8_t, N>
+//   for IlvdLayout: (2, 4):(4, 1)
+template <FloatRoundStyle Round, int N>
+struct InterleavedNumericArrayConverter<Layout<Shape<_2, _4>, Stride<_4, _1>>,
+                                        cutlass::bfloat16_t, machete_uint4b8_t, N,
+                                        Round, void> {
+  using IlvdLayout = Layout<Shape<_2, _4>, Stride<_4, _1>>;
+  static_assert(N % size(IlvdLayout{}) == 0);
+
+  using result_type = Array<cutlass::bfloat16_t, N>;
+  using source_type = Array<machete_uint4b8_t, N>;
+
+ private:
+  struct RegConvert {
+    template <typename PackedResultType>
+    CUTLASS_DEVICE static PackedResultType convert(Array<uint32_t, 1> src_) {
+      uint32_t src = src_[0];
+      using RegArray =
+          cutlass::AlignedArray<uint32_t, PackedResultType::kElements / 2,
+                                sizeof(PackedResultType)>;
+      RegArray r;
+
+      static_assert(PackedResultType::kElements <= size(IlvdLayout{}));
+      static constexpr uint32_t or_mask = 0x43004300;
+
+      // Unlike float16 where the mantissa is large enough to contain 2
+      // nibbles, bfloat16 can only fit one, so we can only convert one
+      // nibble at a time
+      for (int ii = 0; ii < RegArray::kElements; ++ii) {
+        r[ii] = src >> (4 * ii);
+
+        static constexpr uint32_t and_or_imm_lut = (0xf0 & 0xcc) | 0xaa;
+        static constexpr uint32_t low_nib_mask = 0x000F000F;
+
+        asm volatile(
+            "{\n"
+            "  lop3.b32 %0, %0, %1, %2, %3;\n"
+            "}\n"
+            : "+r"(r[ii + 0])
+            : "n"(low_nib_mask), "n"(or_mask), "n"(and_or_imm_lut));
+
+        // For low nibble:
+        //  {x1, x0} = {128+(x1+8), 128+(x0+8)} * {1, 1} - {136, 136}
+        static constexpr uint32_t low_nib_bias = 0x43084308;  // {136, 136}
+
+        {
+          __nv_bfloat162& fp16x2_val = reinterpret_cast<__nv_bfloat162&>(r[ii]);
+          fp16x2_val =
+              __hsub2(fp16x2_val,
+                      reinterpret_cast<const __nv_bfloat162&>(low_nib_bias));
+        }
+      }
+
+      return reinterpret_cast<PackedResultType&>(r);
+    };
+  };
+
+ public:
+  CUTLASS_DEVICE
+  static result_type convert(source_type const& source) {
+    return ArrayConverterPacked32Bit<RegConvert, typename result_type::Element,
+                                     typename source_type::Element,
+                                     N>::convert(source);
+  }
+
+  CUTLASS_DEVICE
+  result_type operator()(source_type const& s) const { return convert(s); }
+};
+
+// for Array<cutlass::bfloat16_t, N> <= Array<uint4_t, N>
+//   for IlvdLayout: (2, 4):(4, 1)
+template <FloatRoundStyle Round, int N>
+struct InterleavedNumericArrayConverter<Layout<Shape<_2, _4>, Stride<_4, _1>>,
+                                        cutlass::bfloat16_t, uint4_t, N, Round,
+                                        void> {
+  using IlvdLayout = Layout<Shape<_2, _4>, Stride<_4, _1>>;
+  static_assert(N % size(IlvdLayout{}) == 0);
+
+  using result_type = Array<cutlass::bfloat16_t, N>;
+  using source_type = Array<uint4_t, N>;
+
+ private:
+  struct RegConvert {
+    template <typename PackedResultType>
+    CUTLASS_DEVICE static PackedResultType convert(Array<uint32_t, 1> src_) {
+      uint32_t src = src_[0];
+      using RegArray =
+          cutlass::AlignedArray<uint32_t, PackedResultType::kElements / 2,
+                                sizeof(PackedResultType)>;
+      RegArray r;
+
+      static_assert(PackedResultType::kElements <= size(IlvdLayout{}));
+      static constexpr uint32_t or_mask = 0x43004300;
+
+      // Unlike float16 where the mantissa is large enough to contain 2
+      // nibbles, bfloat16 can only fit one, so we can only convert one
+      // nibble at a time
+      for (int ii = 0; ii < RegArray::kElements; ++ii) {
+        r[ii] = src >> (4 * ii);
+
+        static constexpr uint32_t and_or_imm_lut = (0xf0 & 0xcc) | 0xaa;
+        static constexpr uint32_t low_nib_mask = 0x000F000F;
+
+        asm volatile(
+            "{\n"
+            "  lop3.b32 %0, %0, %1, %2, %3;\n"
+            "}\n"
+            : "+r"(r[ii])
+            : "n"(low_nib_mask), "n"(or_mask), "n"(and_or_imm_lut));
+
+        // For low nibble:
+        //  {x1, x0} = {128 + x1, 128 + x0} * {1, 1} - {128, 128}
+        static constexpr uint32_t low_nib_bias = 0x43004300;  // {128, 128}
+
+        {
+          __nv_bfloat162& fp16x2_val = reinterpret_cast<__nv_bfloat162&>(r[ii]);
+          fp16x2_val =
+              __hsub2(fp16x2_val,
+                      reinterpret_cast<const __nv_bfloat162&>(low_nib_bias));
+        }
+      }
+
+      return reinterpret_cast<PackedResultType&>(r);
+    };
+  };
+
+ public:
+  CUTLASS_DEVICE
+  static result_type convert(source_type const& source) {
+    return ArrayConverterPacked32Bit<RegConvert, typename result_type::Element,
+                                     typename source_type::Element,
+                                     N>::convert(source);
+  }
+
+  CUTLASS_DEVICE
+  result_type operator()(source_type const& s) const { return convert(s); }
+};
+
+// for Array<cutlass::bfloat16_t, N> <= Array<machete_uint8b128_t, N>
+template <FloatRoundStyle Round, int N>
+struct NumericArrayConverter<cutlass::bfloat16_t, machete_uint8b128_t, N, Round> {
+  using result_type = Array<cutlass::bfloat16_t, N>;
+  using source_type = Array<machete_uint8b128_t, N>;
+  static FloatRoundStyle const round_style = Round;
+
+ private:
+  using result_packed_4_t = Array<cutlass::bfloat16_t, 4>;
+  using result_packed_2_t = Array<cutlass::bfloat16_t, 2>;
+  using src_packed_4_t = Array<machete_uint8b128_t, 4>;
+  using src_packed_2_t = Array<machete_uint8b128_t, 2>;
+
+  // Not Valid, not supported, only here to satisfy the interface and to avoid
+  //  a compile error. ScalarConverter will not actually work until
+  //  NumericConverter<cutlass::bfloat16_t, machete_uint8b128_t, Round> is
+  //  implemented
+  using ScalarConverter =
+      NumericConverter<cutlass::bfloat16_t, machete_uint8b128_t, Round>;
+
+  template <typename PackedResultType, typename PackedSrcType>
+  CUTLASS_DEVICE static PackedResultType packed_convert(
+      PackedSrcType const& source) {
+    static_assert(
+        (platform::is_same<PackedSrcType, src_packed_2_t>::value &&
+         platform::is_same<PackedResultType, result_packed_2_t>::value) ||
+            (platform::is_same<PackedSrcType, src_packed_4_t>::value &&
+             platform::is_same<PackedResultType, result_packed_4_t>::value),
+        "Invalid PackedSrcType/PackedResultType must be 2 or 4 to use private "
+        "convert dispatch.");
+
+    NumericArrayConverter<float, machete_uint8b128_t, PackedResultType::kElements,
+                          Round>
+        convert_uint8_to_f32;
+    Array<float, PackedResultType::kElements> tmp =
+        convert_uint8_to_f32(source);
+    NumericArrayConverter<cutlass::bfloat16_t, float,
+                          PackedResultType::kElements, Round>
+        convert_f32_to_bf16_;
+    return convert_f32_to_bf16_(tmp);
+  }
+
+  friend class detail::VectorizedConverter;
+
+ public:
+  CUTLASS_DEVICE
+  static result_type convert(source_type const& source) {
+    result_type result;
+    using ConverterType =
+        NumericArrayConverter<typename result_type::Element,
+                              typename source_type::Element, N, Round>;
+    detail::VectorizedConverter::convert<ConverterType, result_packed_4_t,
+                                         src_packed_4_t, result_packed_2_t,
+                                         src_packed_2_t>(result, source);
+
+    return result;
+  }
+
+  CUTLASS_DEVICE
+  result_type operator()(source_type const& s) const { return convert(s); }
+};
+
+#endif
+
+// for Array<int8_t, N> <= Array<cutlass::half_t, N>
+//   FastFP16toINT8 from https://arxiv.org/pdf/2406.09904
+template <FloatRoundStyle Round, int N>
+struct NumericArrayConverter<int8_t, cutlass::half_t, N, Round> {
+  using result_type = Array<int8_t, N>;
+  using source_type = Array<cutlass::half_t, N>;
+
+  struct RegConvert {
+    // FastFP16toINT8 from https://arxiv.org/pdf/2406.09904
+    template <typename PackedResultType, int src_regs>
+    CUTLASS_DEVICE static PackedResultType convert(
+        Array<uint32_t, src_regs> src) {
+      // Hold output int8s in reg. We need 1 reg for every 4 elements
+      using RegArray = cutlass::AlignedArray<
+          uint32_t, std::max(PackedResultType::kElements / 4, size_t(1))>;
+      RegArray r;
+
+      static constexpr uint32_t MAGIC_BIAS_ = 0x64806480;
+      auto MAGIC_BIAS = *reinterpret_cast<const half2*>(&MAGIC_BIAS_);
+
+      *reinterpret_cast<half2*>(&src[0]) =
+          __hadd2(*reinterpret_cast<half2*>(&src[0]), MAGIC_BIAS);
+
+      if constexpr (src_regs > 1) {
+        *reinterpret_cast<half2*>(&src[1]) =
+            __hadd2(*reinterpret_cast<half2*>(&src[1]), MAGIC_BIAS);
+      }
+
+      static_assert(PackedResultType::kElements <= 4);
+      uint32_t uint8s;
+      static constexpr uint32_t MASK_0246 = 0x6420;
+      static constexpr uint32_t UINT8s_TO_INT8s_MASK = 0x80808080;
+      asm volatile("prmt.b32 %0,%1,%2,%3;\n"
+                   : "=r"(uint8s)
+                   : "r"(src[0]), "r"((src_regs > 1) ? src[1] : src[0]),
+                     "n"(MASK_0246));
+
+      uint32_t int8s = (uint8s ^ UINT8s_TO_INT8s_MASK);
+
+      return reinterpret_cast<PackedResultType&>(int8s);
+    };
+  };
+
+ public:
+  CUTLASS_DEVICE
+  static result_type convert(source_type const& source) {
+    return ArrayConverterPacked32Bit<RegConvert, typename result_type::Element,
+                                     typename source_type::Element,
+                                     N>::convert(source);
+  }
+
+  CUTLASS_DEVICE
+  result_type operator()(source_type const& s) const { return convert(s); }
+};
+
+/////////////////////////////////////////////////////////////////////////////////////////////////
+
+}  // namespace cutlass
+
+/////////////////////////////////////////////////////////////////////////////////////////////////

custom_ops/gpu_ops/machete/utils/machete_type_utils.cuh

@@ -0,0 +1,43 @@
+// adapted from: https://github.com/vllm-project/vllm/blob/main/csrc/cutlass_extensions/vllm_numeric_conversion.cuh
+#include "cutlass/bfloat16.h"
+#include "cutlass/half.h"
+#include "cuda_bf16.h"
+
+#include "machete_custom_types.cuh"
+
+namespace cutlass {
+
+template <typename T>
+struct nameof {
+  static constexpr char const* value = "unknown";
+};
+
+template <typename T>
+inline constexpr auto nameof_v = nameof<T>::value;
+
+#define NAMEOF_TYPE(T)                       \
+  template <>                                \
+  struct nameof<T> {                         \
+    static constexpr char const* value = #T; \
+  };
+
+NAMEOF_TYPE(float_e4m3_t)
+NAMEOF_TYPE(float_e5m2_t)
+NAMEOF_TYPE(half_t)
+NAMEOF_TYPE(nv_bfloat16)
+NAMEOF_TYPE(bfloat16_t)
+NAMEOF_TYPE(float)
+
+NAMEOF_TYPE(int4b_t)
+NAMEOF_TYPE(int8_t)
+NAMEOF_TYPE(int32_t)
+NAMEOF_TYPE(int64_t)
+
+NAMEOF_TYPE(machete_uint4b8_t)
+NAMEOF_TYPE(uint4b_t)
+NAMEOF_TYPE(uint8_t)
+NAMEOF_TYPE(machete_uint8b128_t)
+NAMEOF_TYPE(uint32_t)
+NAMEOF_TYPE(uint64_t)
+
+};  // namespace cutlass

custom_ops/gpu_ops/machete/utils/paddle_utils.hpp

@@ -0,0 +1,161 @@
+// adapted from: https://github.com/vllm-project/vllm/blob/main/csrc/cutlass_extensions/torch_utils.hpp
+#pragma once
+
+#include "helper.h"
+
+#include "cute/layout.hpp"
+#include "cutlass/layout/matrix.h"
+#include "cutlass/bfloat16.h"
+#include "cutlass/half.h"
+
+using ColumnMajor = typename cutlass::layout::ColumnMajor;
+using RowMajor = typename cutlass::layout::RowMajor;
+
+namespace cute {
+
+namespace detail {
+
+template <class T, class F, class G, int... I>
+CUTE_HOST_DEVICE constexpr auto tapply_with_idx(T&& t, F&& f, G&& g,
+                                                seq<I...>) {
+  return g(f(cute::get<I>(static_cast<T&&>(t)), I)...);
+}
+
+template <class F, int... I>
+CUTE_HOST_DEVICE constexpr auto make_shape_from_idx(F&& f, seq<I...>) {
+  return make_shape(f(I)...);
+}
+
+};  // namespace detail
+
+template <class T, class F>
+CUTE_HOST_DEVICE constexpr auto transform_with_idx(T const& t, F&& f) {
+  if constexpr (cute::is_tuple<T>::value) {
+    return detail::tapply_with_idx(
+        t, f, [](auto const&... a) { return cute::make_tuple(a...); },
+        tuple_seq<T>{});
+  } else {
+    return f(t);
+  }
+
+  CUTE_GCC_UNREACHABLE;
+}
+
+// calls: make_shape(f(0), f(1), ..., f(N-1))
+template <int N, class F>
+CUTE_HOST_DEVICE constexpr auto make_shape_from_idx(F&& f) {
+  return detail::make_shape_from_idx(f, make_seq<N>{});
+}
+
+};  // namespace cute
+
+// Make a layout from a tensor with `rank(Stride{})`, where the shape is the
+// shape of the passed in tensor and the strides are of type `Stride` and
+// contain the strides of the passed in tensor, checking that any static strides
+// in `Stride{}` match the strides of the passed in tensor.
+// If `tensor.shape().size() < rank(Stride{})`, the shape is padded with 1s and the extra
+// strides are set to be 0 or 1.
+template <typename Stride>
+static inline auto make_cute_layout(paddle::Tensor const& tensor,
+                                    std::string_view name = "tensor") {
+  PD_CHECK(tensor.shape().size() <= rank(Stride{}));
+  auto stride = cute::transform_with_idx(
+      Stride{}, [&](auto const& stride_ele, auto const& idx) {
+        using StrideEle = std::decay_t<decltype(stride_ele)>;
+
+        if (idx < tensor.shape().size()) {
+          if constexpr (cute::is_static_v<StrideEle>) {
+            PD_CHECK(StrideEle::value == tensor.strides()[idx], "Expected ",
+                        name, ".strides()[", idx, "] to be ", StrideEle::value, ", but got ", tensor.strides()[idx], ". ");
+            return StrideEle{};
+          } else {
+            if (tensor.shape()[idx] == 1) {
+              // use 0 stride for dims with size 1, this is easier for
+              // cute/cutlass to optimize (helps the TMA code flatten dims)
+              return StrideEle{0};
+            } else {
+              return tensor.strides()[idx];
+            }
+          }
+        } else {
+          // Extra strides are assumed to be 0 or 1
+          if constexpr (cute::is_static_v<StrideEle>) {
+            static_assert(StrideEle::value == 0 || StrideEle::value == 1);
+          }
+          return StrideEle{};
+        }
+      });
+
+  auto shape = cute::make_shape_from_idx<rank(Stride{})>([&](auto const& idx) {
+    if (idx < tensor.shape().size())
+      return tensor.shape()[idx];
+    else
+      return int64_t(1);
+  });
+
+  return make_layout(shape, stride);
+}
+
+template <typename Stride>
+static inline auto maybe_make_cute_layout(
+    std::optional<paddle::Tensor> const& tensor,
+    std::string_view name = "tensor") {
+  using Layout = decltype(make_cute_layout<Stride>(*tensor));
+
+  if (tensor) {
+    return std::optional<Layout>{make_cute_layout<Stride>(*tensor, name)};
+  } else {
+    return std::optional<Layout>{};
+  }
+}
+
+//
+//  Paddle dtype to Cutlass Type (equivalent_cutlass_type)
+//
+
+template <typename T>
+struct equivalent_cutlass_type {
+  using type = T;
+};
+
+template <typename T>
+using equivalent_cutlass_type_t = typename equivalent_cutlass_type<T>::type;
+
+template <>
+struct equivalent_cutlass_type<phi::dtype::float16> {
+  using type = cutlass::half_t;
+};
+
+template <>
+struct equivalent_cutlass_type<phi::dtype::bfloat16> {
+  using type = cutlass::bfloat16_t;
+};
+
+//
+// equivalent_scalar_t (basically inverse of equivalent_cutlass_type)
+//
+
+// Return a `c10::CppTypeToScalarType<T>` compatible type, i.e. get the C++ from
+// c10 that is equivalent to T, e.g.: `cutlass::half_t -> c10::Half`
+template <typename T>
+struct equivalent_scalar_type {
+  using type = T;
+};
+
+template <typename T>
+using equivalent_scalar_type_t = typename equivalent_scalar_type<T>::type;
+
+template <>
+struct equivalent_scalar_type<cutlass::half_t> {
+  using type = phi::dtype::float16;
+};
+
+template <>
+struct equivalent_scalar_type<cutlass::bfloat16_t> {
+  using type = phi::dtype::bfloat16;
+};
+
+// get equivalent c10::ScalarType tag from compile time type
+template <typename T>
+static inline constexpr paddle::DataType equivalent_scalar_type_v =
+    phi::CppTypeToDataType<equivalent_scalar_type_t<T>>::Type();