Fix int32 to uint32 casting issues causing -1 to become large positive number

Co-authored-by: yuanlehome <23653004+yuanlehome@users.noreply.github.com>

.github/workflows/_accuracy_test.yml

@@ -0,0 +1,187 @@
+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 }}"
+            docker pull ${docker_image}
+            # 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 --pre paddlepaddle-gpu -i https://www.paddlepaddle.org.cn/packages/nightly/cu126/
+
+          pip config set global.index-url https://mirrors.tuna.tsinghua.edu.cn/pypi/web/simple
+
+          python -m pip install ${fastdeploy_wheel_url}
+          python -m pip install pytest
+
+          wget https://paddle-qa.bj.bcebos.com/zhengtianyu/tools/llm-deploy-linux-amd64
+          chmod +x ./llm-deploy-linux-amd64
+          ./llm-deploy-linux-amd64 -python python3.10 \
+          -model_name ERNIE-4.5-0.3B-Paddle \
+          -model_path /MODELDATA \
+          --skip install
+
+          git config --global --add safe.directory /workspace/FastDeploy
+          cd FastDeploy
+          pushd tests/ce/deploy
+          ps -ef | grep "${FD_ENGINE_QUEUE_PORT}" | grep -v grep | awk "{print \$2}" | xargs -r kill -9
+          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 +1,230 @@
+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 }}"
+            docker pull ${docker_image}
+            # 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 --pre paddlepaddle-gpu -i https://www.paddlepaddle.org.cn/packages/nightly/cu126/
+
+          pip config set global.index-url https://mirrors.tuna.tsinghua.edu.cn/pypi/web/simple
+
+          python -m pip install ${fastdeploy_wheel_url}
+          python -m pip install pytest
+
+          wget https://paddle-qa.bj.bcebos.com/zhengtianyu/tools/llm-deploy-linux-amd64
+          chmod +x ./llm-deploy-linux-amd64
+          ./llm-deploy-linux-amd64 -python python3.10 \
+          -model_name ERNIE-4.5-0.3B-Paddle \
+          -model_path /MODELDATA \
+          --skip install
+
+          git config --global --add safe.directory /workspace/FastDeploy
+          cd FastDeploy
+          pushd tests/ce/deploy
+          ps -ef | grep "${FD_ENGINE_QUEUE_PORT}" | grep -v grep | awk "{print \$2}" | xargs -r kill -9
+          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
@@ -45,6 +55,7 @@ on:
 jobs:
   fd-build:
     runs-on: [self-hosted, GPU-Build]
+    timeout-minutes: 360
     outputs:
       wheel_path: ${{ steps.set_output.outputs.wheel_path }}
     steps:
@@ -85,6 +96,10 @@ 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 }}"
@@ -109,6 +124,9 @@ jobs:
             -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}
@@ -116,6 +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)
@@ -124,14 +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 --pre paddlepaddle-gpu -i https://www.paddlepaddle.org.cn/packages/nightly/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/_ci_image_build.yml

@@ -0,0 +1,73 @@
+name: Docker Build
+description: "FastDeploy CI Image Build"
+
+on:
+  workflow_call:
+    inputs:
+      CI_DOCKER_IMAGE_NAME:
+        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
+      DOCKER_IMAGE_NAME:
+        description: "Build Images"
+        required: false
+        type: string
+        default: "ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleqa:fastdeploy-ciuse-cuda126-dailyupdate"
+    outputs:
+      docker_name_precheck:
+        description: "Output path of the generated wheel"
+        value: ${{ jobs.docker_build.outputs.docker_name_precheck }}
+
+jobs:
+  docker_build:
+    runs-on: [self-hosted, Docker-Build]
+    outputs:
+      docker_name_precheck: ${{ steps.docker_build.outputs.docker_name_precheck }}
+    steps:
+      - name: Docker Build
+        id: docker_build
+        shell: bash
+        env:
+          docker_image_name: ${{ inputs.CI_DOCKER_IMAGE_NAME }}
+          docker_image: ${{ inputs.DOCKER_IMAGE_NAME }}
+          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
+
+            # Docker Build
+            cd tools/dockerfile/
+            set -e
+            cp ../../requirements.txt ./
+            cp ../../scripts/unittest_requirement.txt ./
+            docker build -t ${docker_image_name} -f Dockerfile.ci . \
+                --network host \
+                --no-cache
+            docker push ${docker_image_name}
+            echo "docker_name_precheck=${docker_image_name}" >> $GITHUB_OUTPUT

.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

@@ -39,6 +39,7 @@ jobs:
           docker_image: ${{ inputs.DOCKER_IMAGE }}
           paddletest_archive_url: ${{ inputs.PADDLETEST_ARCHIVE_URL }}
         run: |
+            docker pull ${docker_image}
             # Clean the repository directory before starting
             docker run --rm --net=host -v $(pwd):/workspace -w /workspace \
             -e "REPO_NAME=${REPO_NAME}" \
@@ -62,18 +63,24 @@ jobs:
           MODEL_CACHE_DIR: ${{ inputs.MODEL_CACHE_DIR }}
         run: |
           runner_name="${{ runner.name }}"
-          last_char="${runner_name: -1}"
+          CARD_ID=$(echo "${runner_name}" | awk -F'-' '{print $NF}')
+          DEVICES=$(echo "$CARD_ID" | fold -w1 | paste -sd,)
+          DEVICE_PORT=$(echo "$DEVICES" | cut -d',' -f1)
 
-          if [[ "$last_char" =~ [0-7] ]]; then
-            DEVICES="$last_char"
-          else
-            DEVICES="0"
-          fi
-
-          FLASK_PORT=$((9160 + DEVICES * 100))
-          FD_API_PORT=$((9180 + DEVICES * 100))
-          FD_ENGINE_QUEUE_PORT=$((9150 + DEVICES * 100))
-          FD_METRICS_PORT=$((9170 + DEVICES * 100))
+          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}"
@@ -85,28 +92,51 @@ jobs:
             exit 1
           fi
 
-          PARENT_DIR=$(dirname "$WORKSPACE")
+          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} ===="
 
-          docker run --ipc=host --pid=host --net=host \
+          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}"'"' ${docker_image} /bin/bash -c '
-          # python -m pip install --pre paddlepaddle-gpu -i https://www.paddlepaddle.org.cn/packages/nightly/cu126/
-          python -m pip install paddlepaddle-gpu==3.0.0.dev20250729 -i https://www.paddlepaddle.org.cn/packages/nightly/cu126/
+          --gpus '"device='"${DEVICES}"'"' ${docker_image} /bin/bash -xc '
+          python -m pip install --pre paddlepaddle-gpu -i https://www.paddlepaddle.org.cn/packages/nightly/cu126/
+
+          pip config set global.index-url https://mirrors.tuna.tsinghua.edu.cn/pypi/web/simple
 
-          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
           python -m pip install ${fastdeploy_wheel_url}
 
           wget https://paddle-qa.bj.bcebos.com/zhengtianyu/tools/llm-deploy-linux-amd64
@@ -117,6 +147,7 @@ jobs:
           --skip install
 
           cd PaddleTest/framework/ServeTest
+          ps -ef | grep "${FD_ENGINE_QUEUE_PORT}" | grep -v grep | awk "{print \$2}" | xargs -r kill -9
           python3.10 deploy.py > dd.log 2>&1 &
           sleep 3
           curl -X POST http://0.0.0.0:${FLASK_PORT}/start \
@@ -124,6 +155,10 @@ jobs:
               -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

.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 }}"
+            docker pull ${docker_image}
+            # 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 --pre paddlepaddle-gpu -i https://www.paddlepaddle.org.cn/packages/nightly/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 }}"
+            docker pull ${docker_image}
+            # 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 --pre paddlepaddle-gpu -i https://www.paddlepaddle.org.cn/packages/nightly/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

@@ -1,4 +1,4 @@
-name: Run FastDeploy Unit Tests and Coverage
+name: Coverage Check
 description: "Run FastDeploy Unit Tests and Coverage"
 
 on:
@@ -22,13 +22,32 @@ on:
         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-h1z1-4Cards]
+    runs-on: [self-hosted, GPU-h1z1-2Cards]
+    timeout-minutes: 90
+    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.unittest_failed.outputs.unittest_failed_url }}
+      unittest_failed_url: ${{ steps.cov_upload.outputs.unittest_failed_url }}
+      diff_cov_result_json_url: ${{ steps.cov_upload.outputs.diff_cov_result_json_url }}
     steps:
       - name: Code Prepare
         shell: bash
@@ -41,7 +60,7 @@ jobs:
             FULL_REPO="${{ github.repository }}"
             REPO_NAME="${FULL_REPO##*/}"
             BASE_BRANCH="${{ github.base_ref }}"
-
+            docker pull ${docker_image}
             # Clean the repository directory before starting
             docker run --rm --net=host -v $(pwd):/workspace -w /workspace \
             -e "REPO_NAME=${REPO_NAME}" \
@@ -66,58 +85,125 @@ jobs:
           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: ${{ github.event_name == 'pull_request' }}
         run: |
-            set -x
-            runner_name="${{ runner.name }}"
-            CARD_ID=$(echo "${runner_name}" | awk -F'-' '{print $NF}')
-            gpu_id=$(echo "$CARD_ID" | fold -w1 | paste -sd,)
+          if [[ "$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)
 
-            CACHE_DIR="${CACHE_DIR:-$(dirname "$(dirname "${{ github.workspace }}")")}"
-            echo "CACHE_DIR is set to ${CACHE_DIR}"
-            if [ ! -f "${CACHE_DIR}/gitconfig" ]; then
-              touch "${CACHE_DIR}/gitconfig"
-            fi
-            PARENT_DIR=$(dirname "$WORKSPACE")
-            echo "PARENT_DIR:$PARENT_DIR"
-            docker run --rm --net=host \
-            --cap-add=SYS_PTRACE --privileged --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" \
-            -e TZ="Asia/Shanghai" \
-            -e "fd_wheel_url=${fd_wheel_url}" \
-            -e "BASE_REF=${BASE_REF}" \
-            --gpus "\"device=${gpu_id}\"" ${docker_image} /bin/bash -c '
+          FLASK_PORT=$((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 "========================================================="
 
-            git config --global --add safe.directory /workspace/FastDeploy
-            cd FastDeploy
-            # python -m pip install --pre paddlepaddle-gpu -i https://www.paddlepaddle.org.cn/packages/nightly/cu126/
-            python -m pip install paddlepaddle-gpu==3.0.0.dev20250729 -i https://www.paddlepaddle.org.cn/packages/nightly/cu126/
+          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
 
-            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
+          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} ===="
 
-            python -m pip install coverage
-            python -m pip install diff-cover
-            python -m pip install ${fd_wheel_url}
-            export COVERAGE_FILE=/workspace/FastDeploy/coveragedata/.coverage
-            export COVERAGE_RCFILE=/workspace/FastDeploy/scripts/.coveragerc
-            TEST_EXIT_CODE=0
-            bash scripts/coverage_run.sh || TEST_EXIT_CODE=8
-            git diff origin/${BASE_REF}..HEAD --unified=0 > diff.txt
-            echo "TEST_EXIT_CODE=${TEST_EXIT_CODE}" >> exit_code.env
-            coverage combine coveragedata/
-            coverage xml -o python_coverage_all.xml
-            COVERAGE_EXIT_CODE=0
-            diff-cover python_coverage_all.xml --diff-file=diff.txt --fail-under=90 || COVERAGE_EXIT_CODE=9
-            echo "COVERAGE_EXIT_CODE=${COVERAGE_EXIT_CODE}" >> exit_code.env
+          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 --pre paddlepaddle-gpu -i https://www.paddlepaddle.org.cn/packages/nightly/cu126/
+          pip config set global.extra-index-url https://mirrors.tuna.tsinghua.edu.cn/pypi/web/simple
+
+          python -m pip install -r scripts/unittest_requirement.txt
+          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
-            '
-            if [ -f FastDeploy/exit_code.env ]; then
-              cat FastDeploy/exit_code.env >> $GITHUB_ENV
-            fi
+          else
+            echo "Not a PR, skipping diff-cover"
+          fi
+          echo "COVERAGE_EXIT_CODE=${COVERAGE_EXIT_CODE}" >> exit_code.env
+          '
+          if [ -f FastDeploy/exit_code.env ]; then
+            cat FastDeploy/exit_code.env >> $GITHUB_ENV
+          fi
       - name: Upload unit resule and diff coverage to bos
         id: cov_upload
         shell: bash
@@ -125,42 +211,97 @@ jobs:
           cd FastDeploy
           commit_id=${{ 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//,/_}/CoverageData
-          wget  -q --no-proxy --no-check-certificate https://paddle-qa.bj.bcebos.com/CodeSync/develop/PaddlePaddle/PaddleTest/tools/bos_tools.py
+          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}
+            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/${target_path_stripped}/${diff_cov_file}
+            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
-      - name: Determine Unit Succ and whether the coverage rate reaches 90%
+          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 "All tests and coverage passed"
+          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
@@ -170,6 +311,9 @@ jobs:
         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: ./diff_coverage.xml
+          files: ./FastDeploy/diff_coverage.xml
           name: python diff coverage
           verbose: true
+          disable_search: true
+          commit_parent: false
+          flags: diff

.github/workflows/approve.yml

@@ -6,6 +6,9 @@ on:
       - develop
       - 'release/*'
 
+env:
+  GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+
 jobs:
   Approval:
     name: Approval
@@ -33,7 +36,6 @@ jobs:
         uses: actions/setup-python@v5
         with:
           python-version: '3.10'
-          cache: 'pip'
 
       - name: Run approval check script
         run: |

.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: CE-Job-${{ github.ref }}-${{ 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/${target_path_stripped}/${filename}
+
+          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/${target_path_stripped_latest}/${filename}
+          echo "commit wheel url is ${WHEEL_PATH}"
+          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_sm8689.outputs.wheel_path }}"
+          wget -q --no-check-certificate ${{ needs.build_sm8689.outputs.wheel_path }}
+          filename=$(basename ${{ needs.build_sm8689.outputs.wheel_path }})
+
+          commit_id=${{ 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/${target_path_stripped}/${filename}
+
+          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/${target_path_stripped_latest}/${filename}
+          echo "commit wheel url is ${WHEEL_PATH}"
+          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

@@ -1,89 +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
-          FD_API_PORT=$((9180 + gpu_id * 100))
-          FD_ENGINE_QUEUE_PORT=$((9150 + gpu_id * 100))
-          FD_METRICS_PORT=$((9170 + gpu_id * 100))
-
-          PARENT_DIR=$(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

@@ -13,7 +13,8 @@ concurrency:
 
 jobs:
   CI_GCU:
-    runs-on: [self-hosted, GCU-S60-8Card]
+    runs-on:
+      group: GCU
     steps:
       - name: Print current runner name
         run: |
@@ -28,7 +29,9 @@ jobs:
           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 \
+          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 '
@@ -39,6 +42,7 @@ jobs:
           '
           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
@@ -49,6 +53,9 @@ jobs:
             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:
@@ -70,19 +77,21 @@ jobs:
           echo "PARENT_DIR:$PARENT_DIR"
           echo "Install drivers..."
           cd /work/deps
-          bash TopsRider_i3x_*_deb_amd64.run --driver --no-auto-load -y
+          sudo bash TopsRider_i3x_*_deb_amd64.run --driver --no-auto-load -y
           cd -
-          docker run --rm --network=host --ipc=host -it --privileged  \
-          -v $(pwd):/workspace -w /workspace \
-          -v "/home:/home" \
-          -v "/work:/work" \
-          -e "MODEL_PATH=/work/models" \
+          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 "
+          ${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_image_update.yml

@@ -0,0 +1,174 @@
+name: CI Images Build
+
+on:
+  workflow_dispatch:
+  schedule:
+    - cron: '0 18 * * *'   # 2:00 AM China Standard Time (UTC+8)
+
+permissions: read-all
+
+concurrency:
+  group: CI-Images-Build-${{ github.ref }}-${{ github.sha }}
+  cancel-in-progress: true
+
+
+jobs:
+  clone:
+    environment: CodeSync
+    name: FD-Clone-Linux
+    runs-on: ubuntu-latest
+    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 }}"
+
+  ci_image_build:
+    name: CI Images Build
+    needs: clone
+    uses: ./.github/workflows/_ci_image_build.yml
+    with:
+      CI_DOCKER_IMAGE_NAME: ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleqa:fastdeploy-ciuse-cuda126-dailyupdate-precheck
+      FASTDEPLOY_ARCHIVE_URL: ${{ needs.clone.outputs.repo_archive_url }}
+
+
+  build_sm8090:
+    name: BUILD_SM8090
+    needs: [clone, ci_image_build]
+    uses: ./.github/workflows/_build_linux.yml
+    with:
+      DOCKER_IMAGE: ${{ needs.ci_image_build.outputs.docker_name_precheck }}
+      FASTDEPLOY_ARCHIVE_URL: ${{ needs.clone.outputs.repo_archive_url }}
+      COMPILE_ARCH: "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 }}
+
+
+  unittest_coverage:
+    name: Run FastDeploy Unit Tests and Coverage
+    needs: [clone,build_sm8090,ci_image_build]
+    uses: ./.github/workflows/_unit_test_coverage.yml
+    with:
+      DOCKER_IMAGE: ${{ needs.ci_image_build.outputs.docker_name_precheck }}
+      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,ci_image_build]
+    uses: ./.github/workflows/_logprob_test_linux.yml
+    with:
+      DOCKER_IMAGE: ${{ needs.ci_image_build.outputs.docker_name_precheck }}
+      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,ci_image_build]
+    uses: ./.github/workflows/_pre_ce_test.yml
+    with:
+      DOCKER_IMAGE: ${{ needs.ci_image_build.outputs.docker_name_precheck }}
+      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,ci_image_build]
+    uses: ./.github/workflows/_base_test.yml
+    with:
+      DOCKER_IMAGE: ${{ needs.ci_image_build.outputs.docker_name_precheck }}
+      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,ci_image_build]
+    uses: ./.github/workflows/_accuracy_test.yml
+    with:
+      DOCKER_IMAGE: ${{ needs.ci_image_build.outputs.docker_name_precheck }}
+      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"
+
+  stable_test:
+    name: Run Stable Tests
+    needs: [clone,build_sm8090,ci_image_build]
+    uses: ./.github/workflows/_stable_test.yml
+    with:
+      DOCKER_IMAGE: ${{ needs.ci_image_build.outputs.docker_name_precheck }}
+      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"
+
+
+  publish_pre_check:
+    name: Publish Docker Images Pre Check
+    needs: [ci_image_build, unittest_coverage,logprob_test,pre_ce_test,base_test,accuracy_test,stable_test]
+    runs-on: [self-hosted, Docker-Build]
+    steps:
+      - name: Images Uploading
+        env:
+          images_name: ${{ needs.ci_image_build.outputs.docker_name_precheck }}
+          ci_image_name: "ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleqa:fastdeploy-ciuse-cuda126-dailyupdate"
+        run: |
+          echo "images_name=${images_name}"
+          docker images ${ci_image_name}
+          docker tag ${images_name} ${ci_image_name}
+          docker push ${ci_image_name}

.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

@@ -19,7 +19,7 @@ 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"
       WITH_NIGHTLY_BUILD: "OFF"
@@ -39,16 +39,59 @@ jobs:
     needs: [clone,build]
     uses: ./.github/workflows/_unit_test_coverage.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 }}
       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:cuda126-py310
+      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/ModelCache"
+      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 +1,331 @@
+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: Publish-Job-${{ github.ref }}-${{ 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"
+
+  stable_test:
+    name: Run Stable Tests
+    needs: [clone,build_sm8090]
+    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"

.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 +1,10 @@
+[submodule "custom_ops/third_party/DeepGEMM"]
+	path = custom_ops/third_party/DeepGEMM
+	url = https://github.com/deepseek-ai/DeepGEMM.git
+	ignore = all
+[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)
 
-**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)
+
+**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/yaml/eb45-128k-wint4-a800-tp8.yaml

@@ -6,3 +6,4 @@ tensor_parallel_size: 8
 max_num_batched_tokens: 4096
 max_num_partial_prefills: 3
 max_long_partial_prefills: 3
+quantization: wint4

benchmarks/yaml/eb45-128k-wint4-tp1-plas.yaml

@@ -0,0 +1,6 @@
+tensor_parallel_size: 1
+max_model_len: 131072
+max_num_seqs: 32
+quantization: wint4
+max_num_batched_tokens: 8192
+plas_attention_config: '{"plas_encoder_top_k_left": 50, "plas_encoder_top_k_right": 60, "plas_decoder_top_k_left": 100, "plas_decoder_top_k_right": 120}'

benchmarks/yaml/eb45-128k-wint8-a800-tp8.yaml

@@ -6,3 +6,4 @@ tensor_parallel_size: 8
 max_num_batched_tokens: 4096
 max_num_partial_prefills: 3
 max_long_partial_prefills: 3
+quantization: wint8

benchmarks/yaml/eb45-32k-wint2-tp4.yaml

@@ -0,0 +1,5 @@
+max_model_len: 32768
+max_num_seqs: 256
+kv_cache_ratio: 0.75
+tensor_parallel_size: 4
+gpu_memory_utilization: 0.9

benchmarks/yaml/eb45-32k-wint4-tp4_decode.yaml

@@ -13,3 +13,4 @@ pd_comm_port: "2334"
 max_num_batched_tokens: 384
 max_num_partial_prefills: 3
 max_long_partial_prefills: 3
+quantization: wint4

benchmarks/yaml/eb45-32k-wint4-tp4_prefill.yaml

@@ -10,3 +10,4 @@ engine_worker_queue_port: 6677
 cache_transfer_protocol: "rdma,ipc"
 rdma_comm_ports: "7675,7676,7677,7678"
 pd_comm_port: "2333"
+quantization: wint4

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/eb45-vl-128k-wint4-h800-tp8.yaml

@@ -0,0 +1,11 @@
+enable_mm: True
+max_model_len: 131072
+max_num_seqs: 56
+gpu_memory_utilization: 0.8
+kv_cache_ratio: 0.8
+tensor_parallel_size: 8
+quantization: wint4
+limit_mm_per_prompt: '{"image": 100, "video": 100}'
+enable_chunked_prefill: True
+max_num_batched_tokens: 384
+reasoning_parser: ernie-45-vl

benchmarks/yaml/eb45-vl-32k-wint8-a800-tp8.yaml

@@ -1,7 +1,7 @@
 enable_mm: True
 max_model_len: 32768
 max_num_seqs: 36
-gpu_memory_utilization: 0.95
+gpu_memory_utilization: 0.9
 kv_cache_ratio: 0.8
 tensor_parallel_size: 8
 quantization: wint8

benchmarks/yaml/eb45-vl-32k-wint8-h800-tp8.yaml

@@ -1,7 +1,7 @@
 enable_mm: True
 max_model_len: 32768
 max_num_seqs: 36
-gpu_memory_utilization: 0.8
+gpu_memory_utilization: 0.85
 kv_cache_ratio: 0.8
 tensor_parallel_size: 8
 quantization: wint8

benchmarks/yaml/eb45-vl-lite-32k-bf16-a800-tp1.yaml

@@ -0,0 +1,9 @@
+enable_mm: True
+max_model_len: 32768
+max_num_seqs: 128
+gpu_memory_utilization: 0.9
+kv_cache_ratio: 0.71
+tensor_parallel_size: 1
+enable_chunked_prefill: True
+max_num_batched_tokens: 384
+reasoning_parser: ernie-45-vl

benchmarks/yaml/eb45-vl-lite-32k-wint4-a800-tp1.yaml

@@ -0,0 +1,10 @@
+enable_mm: True
+max_model_len: 32768
+max_num_seqs: 128
+gpu_memory_utilization: 0.9
+kv_cache_ratio: 0.71
+tensor_parallel_size: 1
+enable_chunked_prefill: True
+max_num_batched_tokens: 384
+quantization: wint4
+reasoning_parser: ernie-45-vl

benchmarks/yaml/eb45-vl-lite-32k-wint8-a800-tp1.yaml

@@ -0,0 +1,10 @@
+enable_mm: True
+max_model_len: 32768
+max_num_seqs: 128
+gpu_memory_utilization: 0.9
+kv_cache_ratio: 0.71
+tensor_parallel_size: 1
+enable_chunked_prefill: True
+max_num_batched_tokens: 384
+quantization: wint8
+reasoning_parser: ernie-45-vl

benchmarks/yaml/request_yaml/eb45-vl-128k.yaml

@@ -0,0 +1 @@
+max_tokens: 131071

benchmarks/yaml/request_yaml/eb45-vl-32k.yaml

@@ -0,0 +1 @@
+max_tokens: 12288

benchmarks/yaml/request_yaml/x1-128k.yaml

@@ -0,0 +1,8 @@
+top_p: 0.95
+temperature: 0.6
+metadata:
+  min_tokens: 1
+max_tokens: 131071
+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

benchmarks/yaml/x1-a3b-128k-wint8-h800-tp1.yaml

@@ -0,0 +1,6 @@
+tensor_parallel_size: 1
+max_model_len: 131072
+max_num_seqs: 32
+reasoning_parser: ernie_x1
+tool_call_parser: ernie_x1
+load_choices: "default_v1"

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
 
@@ -126,27 +119,33 @@ 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())"`
   if [ "$is_xpu" = "True" ]; then
-    cd xpu_ops/src
+    cd xpu_ops
     bash build.sh ${TMP_DIR_REAL_PATH}
-    cd ../..
+    cd ..
   elif [ "$FD_CPU_USE_BF16" == "true" ]; then
     if [ "$FD_BUILDING_ARCS" == "" ]; then
       FD_CPU_USE_BF16=True ${python} setup_ops.py install --install-lib ${OPS_TMP_DIR}
@@ -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/cpu_ops/set_value_by_flags.cc

@@ -14,7 +14,7 @@
 
 #include "paddle/extension.h"
 
-void set_value_by_flag_and_id(const bool *stop_flags,
+void set_value_by_flags_and_idx(const bool *stop_flags,
                               int64_t *pre_ids_all,
                               const int64_t *input_ids,
                               const int *seq_lens_encoder,
@@ -50,7 +50,7 @@ void SetValueByFlagsAndIdx(const paddle::Tensor &pre_ids_all,
     int length = pre_ids_all_shape[1];
     int length_input_ids = input_ids.shape()[1];
 
-    set_value_by_flag_and_id(stop_flags.data<bool>(),
+    set_value_by_flags_and_idx(stop_flags.data<bool>(),
                              const_cast<int64_t *>(pre_ids_all.data<int64_t>()),
                              input_ids.data<int64_t>(),
                              seq_lens_encoder.data<int>(),

custom_ops/cpu_ops/update_inputs.cc

@@ -46,7 +46,7 @@ void update_inputs_kernel(bool *not_need_stop,
     not_need_stop[0] = stop_sum < stop_nums[0];
 }
 
-void UpdateInputes(const paddle::Tensor &stop_flags,
+void UpdateInputs(const paddle::Tensor &stop_flags,
                    const paddle::Tensor &not_need_stop,
                    const paddle::Tensor &seq_lens_this_time,
                    const paddle::Tensor &seq_lens_encoder,
@@ -90,4 +90,4 @@ PD_BUILD_STATIC_OP(update_inputs_cpu)
                     {"seq_lens_encoder", "seq_lens_encoder_out"},
                     {"seq_lens_decoder", "seq_lens_decoder_out"},
                     {"input_ids", "input_ids_out"}})
-    .SetKernelFn(PD_KERNEL(UpdateInputes));
+    .SetKernelFn(PD_KERNEL(UpdateInputs));

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,
@@ -156,8 +140,8 @@ std::vector<paddle::Tensor> AppendAttentionKernel(
           key_cache,
           value_cache,
           attn_mask,
-          cache_k_dequant_scales,
-          cache_v_dequant_scales,
+          cache_quant_type_str == "block_wise_fp8" ? cache_k_quant_scales : cache_k_dequant_scales,
+          cache_quant_type_str == "block_wise_fp8" ? cache_v_quant_scales : cache_v_dequant_scales,
           cache_k_zp,
           cache_v_zp,
           out_linear_shifts,
@@ -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) {
@@ -286,11 +273,15 @@ std::vector<paddle::Tensor> AppendAttentionKernel(
             cache_v_zp,
             cache_quant_type_str,
             use_neox_rotary_style,
+            rope_3d,
             max_input_length,
             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 {
         SpeculateWriteCacheWithRoPEKernel<data_t, data_t>(
             meta_data,
@@ -309,11 +300,15 @@ std::vector<paddle::Tensor> AppendAttentionKernel(
             cache_v_zp,
             cache_quant_type_str,
             use_neox_rotary_style,
+            rope_3d,
             max_input_length,
             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 {
       if (qkv_out_scales) {
@@ -322,7 +317,6 @@ std::vector<paddle::Tensor> AppendAttentionKernel(
             qkv,  // [token_num, num_heads, head_dim]
             seq_lens_decoder,
             seq_lens_encoder,
-            batch_id_per_token,
             cu_seqlens_q,
             block_tables,
             rotary_embs,
@@ -339,14 +333,16 @@ 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,
             qkv_out,  // [token_num, num_heads, head_dim]
             seq_lens_decoder,
             seq_lens_encoder,
-            batch_id_per_token,
             cu_seqlens_q,
             block_tables,
             rotary_embs,
@@ -363,7 +359,10 @@ 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 +391,6 @@ std::vector<paddle::Tensor> AppendAttentionKernel(
       cudaStreamWaitEvent(main_stream, decoder_event);
     }
   }
-
-  return {fmha_out, qkv_out};
 }
 
 std::vector<paddle::Tensor> AppendAttention(
@@ -429,7 +426,11 @@ 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 +465,60 @@ 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 +540,7 @@ std::vector<paddle::Tensor> AppendAttention(
           decoder_num_blocks,
           set_max_lengths,
           max_len_kv,
+          fmha_out,
           rotary_embs,
           attn_mask,
           qkv_bias,
@@ -499,7 +553,11 @@ 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 +572,183 @@ 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 +761,9 @@ 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 +797,11 @@ 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 +825,7 @@ 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 +861,11 @@ 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 +884,148 @@ 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 +1059,15 @@ 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 +1081,71 @@ 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,
@@ -51,6 +52,7 @@ __global__ void multi_query_append_attention_kernel(
     const float quant_min_bound,
     const float in_scale,
     const uint32_t chunk_size,
+    const int num_blocks_x_cpu,
     T *__restrict__ tmp_workspace,  // split kv [token_num, num_chunks,
                                     // num_heads, head_dim]
     float *__restrict__ tmp_m,      // [token_num, num_chunks, num_heads]
@@ -73,6 +75,11 @@ __global__ void multi_query_append_attention_kernel(
 
   block_table_now = block_table + batch_id * max_block_num_per_seq;
 
+  //When cudagraph capture prefill, may launch more gridDim.x
+  if(btid >= static_cast<uint32_t>(num_blocks_x_cpu)){
+    return;
+  }
+
   const uint32_t q_len = seq_lens[batch_id];
   if (q_len <= 0) {
     return;
@@ -141,6 +148,7 @@ __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 * 2 : nullptr;
   smem_t qo_smem(smem);
 
   uint32_t q_smem_offset_r = smem_t::get_permuted_offset<num_vecs_per_head>(
@@ -179,7 +187,7 @@ __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 +253,16 @@ __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 +418,8 @@ __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,
@@ -414,12 +428,14 @@ __global__ void multi_query_append_attention_warp1_4_kernel(
     const float quant_min_bound,
     const float in_scale,
     const uint32_t chunk_size,
+    const int num_blocks_x_cpu,
     T *__restrict__ tmp_workspace,  // split kv [token_num, num_chunks,
                                     // num_heads, head_dim]
     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 int32_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");
@@ -436,6 +452,11 @@ __global__ void multi_query_append_attention_warp1_4_kernel(
   const uint32_t num_rows_per_block = num_frags_x * 16;
   const int *block_table_now = block_table + batch_id * max_block_num_per_seq;
 
+  //When cudagraph capture prefill, may launch more gridDim.x
+  if(btid >= static_cast<uint32_t>(num_blocks_x_cpu)){
+    return;
+  }
+
   const uint32_t q_len = seq_lens[batch_id];
   if (q_len <= 0) {
     return;
@@ -502,7 +523,7 @@ __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 * 2 : nullptr;
   smem_t qo_smem(smem);
 
   uint32_t q_smem_offset_r = smem_t::get_permuted_offset<num_vecs_per_head>(
@@ -540,10 +561,9 @@ __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 +631,15 @@ __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 +905,7 @@ 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,
@@ -890,6 +914,7 @@ void MultiQueryAppendAttention(
           quant_min_bound,
           in_scale,
           chunk_size,
+          num_blocks_x_cpu,
           nullptr,
           nullptr,
           nullptr,
@@ -939,6 +964,7 @@ 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,
@@ -947,6 +973,7 @@ void MultiQueryAppendAttention(
           quant_min_bound,
           in_scale,
           chunk_size,
+          num_blocks_x_cpu,
           reinterpret_cast<NV_TYPE *>(tmp_workspace->ptr()),
           static_cast<float *>(tmp_m->ptr()),
           static_cast<float *>(tmp_d->ptr()),
@@ -1061,12 +1088,18 @@ 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);
 
+    int32_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 +1137,9 @@ 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,
@@ -1112,11 +1148,13 @@ void MultiQueryAppendAttention(
           quant_min_bound,
           in_scale,
           chunk_size,
+          num_blocks_x_cpu,
           nullptr,
           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 +1199,8 @@ 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 +1210,9 @@ 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,
@@ -1180,11 +1221,13 @@ void MultiQueryAppendAttention(
           quant_min_bound,
           in_scale,
           chunk_size,
+          num_blocks_x_cpu,
           reinterpret_cast<NV_TYPE *>(tmp_workspace->ptr()),
           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 +1251,8 @@ 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 +1269,14 @@ 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 +1287,8 @@ 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,
@@ -56,6 +57,7 @@ __global__ void multi_query_append_attention_c4_kernel(
     const float quant_min_bound,
     const float in_scale,
     const uint32_t chunk_size,
+    const int num_blocks_x_cpu,
     T *__restrict__ tmp_workspace,  // split kv [token_num, num_chunks,
                                     // num_heads, head_dim]
     float *__restrict__ tmp_m,      // [token_num, num_chunks, num_heads]
@@ -84,6 +86,11 @@ __global__ void multi_query_append_attention_c4_kernel(
 
   block_table_now = block_table + batch_id * max_block_num_per_seq;
 
+  //When cudagraph capture prefill, may launch more gridDim.x
+  if(btid >= static_cast<uint32_t>(num_blocks_x_cpu)){
+    return;
+  }
+
   const uint32_t q_len = seq_lens[batch_id];
   if (q_len <= 0) {
     return;
@@ -172,6 +179,7 @@ __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 * 2 : nullptr;
   smem_t qo_smem(smem);
 
   uint32_t q_smem_offset_r = smem_t::get_permuted_offset<num_vecs_per_head>(
@@ -248,7 +256,7 @@ __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 +341,15 @@ __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 +516,8 @@ __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,
@@ -513,12 +526,14 @@ __global__ void multi_query_append_attention_c4_warp1_4_kernel(
     const float quant_min_bound,
     const float in_scale,
     const uint32_t chunk_size,
+    const int num_blocks_x_cpu,
     T *__restrict__ tmp_workspace,  // split kv [token_num, num_chunks,
                                     // num_heads, head_dim]
     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 int32_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>();
@@ -541,6 +556,11 @@ __global__ void multi_query_append_attention_c4_warp1_4_kernel(
   const uint32_t num_rows_per_block = num_frags_x * 16;
   const int *block_table_now = block_table + batch_id * max_block_num_per_seq;
 
+  //When cudagraph capture prefill, may launch more gridDim.x
+  if(btid >= static_cast<uint32_t>(num_blocks_x_cpu)){
+    return;
+  }
+
   const uint32_t q_len = seq_lens[batch_id];
   if (q_len <= 0) {
     return;
@@ -627,7 +647,7 @@ __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 * 2 : nullptr;
   smem_t qo_smem(smem);
 
   uint32_t q_smem_offset_r = smem_t::get_permuted_offset<num_vecs_per_head>(
@@ -703,10 +723,9 @@ __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 +807,15 @@ __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 +1110,7 @@ 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,
@@ -1096,6 +1119,7 @@ void MultiQueryAppendC4Attention(
           quant_min_bound,
           in_scale,
           chunk_size,
+          num_blocks_x_cpu,
           nullptr,
           nullptr,
           nullptr,
@@ -1151,6 +1175,7 @@ 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,
@@ -1159,6 +1184,7 @@ void MultiQueryAppendC4Attention(
           quant_min_bound,
           in_scale,
           chunk_size,
+          num_blocks_x_cpu,
           reinterpret_cast<NV_TYPE *>(tmp_workspace->ptr()),
           static_cast<float *>(tmp_m->ptr()),
           static_cast<float *>(tmp_d->ptr()),
@@ -1285,10 +1311,18 @@ 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);
+    int32_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 +1368,9 @@ 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,
@@ -1342,11 +1379,13 @@ void MultiQueryAppendC4Attention(
           quant_min_bound,
           in_scale,
           chunk_size,
+          num_blocks_x_cpu,
           nullptr,
           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 +1431,15 @@ 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 +1449,9 @@ 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,
@@ -1418,11 +1460,13 @@ void MultiQueryAppendC4Attention(
           quant_min_bound,
           in_scale,
           chunk_size,
+          num_blocks_x_cpu,
           reinterpret_cast<NV_TYPE *>(tmp_workspace->ptr()),
           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 +1489,8 @@ 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 +1507,14 @@ 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 +1525,8 @@ 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

@@ -32,14 +32,15 @@ template <typename T,
           typename OutT = T,
           bool ENABLE_PREFILL = true,
           bool is_scale_channel_wise = false,
-          bool IsFP8=false>
+          bool IsFP8 = false,
+          bool IsDynamicC8 = false>
 __global__ void multi_query_append_attention_c8_kernel(
     T *__restrict__ q,             // [token_num, (num_heads + 2* kv_num_head) * head_dim]
     CacheT *__restrict__ cache_k,  // [max_block_num, num_heads, block_size,
                                    // head_dim]
     CacheT *__restrict__ cache_v,
-    const T *__restrict__ cache_k_scale,  // [num_kv_heads]
-    const T *__restrict__ cache_v_scale,  // [num_kv_heads]
+    const T *__restrict__ cache_k_scale,  // [num_kv_heads] or [max_block_num, num_heads, block_size]
+    const T *__restrict__ cache_v_scale,  // [num_kv_heads] or [max_block_num, num_heads, block_size]
     const T *__restrict__ shift_bias,     // [q_num_heads * HEAD_DIM]
     const T *__restrict__ smooth_weight,  // [q_num_heads * HEAD_DIM]
     const int *__restrict__ seq_lens,
@@ -48,6 +49,7 @@ __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,
@@ -56,6 +58,7 @@ __global__ void multi_query_append_attention_c8_kernel(
     const float quant_min_bound,
     const float in_scale,
     const uint32_t chunk_size,
+    const int num_blocks_x_cpu,
     T *__restrict__ tmp_workspace,  // split kv [token_num, num_chunks,
                                     // num_heads, head_dim]
     float *__restrict__ tmp_m,      // [token_num, num_chunks, num_heads]
@@ -85,33 +88,40 @@ __global__ void multi_query_append_attention_c8_kernel(
 
   block_table_now = block_table + batch_id * max_block_num_per_seq;
 
+  //When cudagraph capture prefill, may launch more gridDim.x
+  if(btid >= static_cast<uint32_t>(num_blocks_x_cpu)){
+    return;
+  }
+
   const uint32_t q_len = seq_lens[batch_id];
   if (q_len <= 0) {
     return;
   }
 
-  T cache_k_scale_reg[num_frags_y * 4];
-  T cache_v_scale_reg[num_frags_y * 2];
-  if (is_scale_channel_wise) {
-    int scale_col_base = threadIdx.x % 4 * 2 + kv_head_idx * HEAD_DIM;
-    const T *cache_k_scale_cur_head = cache_k_scale + scale_col_base;
-    for (int i = 0; i < num_frags_y; ++i) {
-      const int scale_idx = i * 16;
-      cache_k_scale_reg[i * 4] = cache_k_scale_cur_head[scale_idx];
-      cache_k_scale_reg[i * 4 + 1] = cache_k_scale_cur_head[scale_idx + 1];
-      cache_k_scale_reg[i * 4 + 2] = cache_k_scale_cur_head[scale_idx + 8];
-      cache_k_scale_reg[i * 4 + 3] = cache_k_scale_cur_head[scale_idx + 9];
+  T cache_k_scale_reg[IsDynamicC8 ? num_frags_z * 2 : num_frags_y * 4];
+  T cache_v_scale_reg[IsDynamicC8 ? num_frags_z * 4 : num_frags_y * 2];
+  if constexpr (!IsDynamicC8) {
+    if constexpr (is_scale_channel_wise) {
+      int scale_col_base = threadIdx.x % 4 * 2 + kv_head_idx * HEAD_DIM;
+      const T *cache_k_scale_cur_head = cache_k_scale + scale_col_base;
+      for (int i = 0; i < num_frags_y; ++i) {
+        const int scale_idx = i * 16;
+        cache_k_scale_reg[i * 4] = cache_k_scale_cur_head[scale_idx];
+        cache_k_scale_reg[i * 4 + 1] = cache_k_scale_cur_head[scale_idx + 1];
+        cache_k_scale_reg[i * 4 + 2] = cache_k_scale_cur_head[scale_idx + 8];
+        cache_k_scale_reg[i * 4 + 3] = cache_k_scale_cur_head[scale_idx + 9];
+      }
+      scale_col_base = threadIdx.x / 4 + kv_head_idx * HEAD_DIM;
+      const T *cache_v_scale_cur_head = cache_v_scale + scale_col_base;
+      for (int i = 0; i < num_frags_y; ++i) {
+        const int scale_idx = i * 16;
+        cache_v_scale_reg[i * 2] = cache_v_scale_cur_head[scale_idx];
+        cache_v_scale_reg[i * 2 + 1] = cache_v_scale_cur_head[scale_idx + 8];
+      }
+    } else {
+      cache_k_scale_reg[0] = cache_k_scale[kv_head_idx];
+      cache_v_scale_reg[0] = cache_v_scale[kv_head_idx];
     }
-    scale_col_base = threadIdx.x / 4 + kv_head_idx * HEAD_DIM;
-    const T *cache_v_scale_cur_head = cache_v_scale + scale_col_base;
-    for (int i = 0; i < num_frags_y; ++i) {
-      const int scale_idx = i * 16;
-      cache_v_scale_reg[i * 2] = cache_v_scale_cur_head[scale_idx];
-      cache_v_scale_reg[i * 2 + 1] = cache_v_scale_cur_head[scale_idx + 8];
-    }
-  } else {
-    cache_k_scale_reg[0] = cache_k_scale[kv_head_idx];
-    cache_v_scale_reg[0] = cache_v_scale[kv_head_idx];
   }
 
   const uint32_t q_end =
@@ -179,6 +189,7 @@ __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 * 2 : nullptr;
   smem_t qo_smem(smem);
 
   uint32_t q_smem_offset_r = smem_t::get_permuted_offset<num_vecs_per_head>(
@@ -199,6 +210,13 @@ __global__ void multi_query_append_attention_c8_kernel(
   smem_t k_smem(smem + NUM_WARPS * num_frags_x * 16 * HEAD_DIM * sizeof(T)),
       v_smem(smem + NUM_WARPS * num_frags_x * 16 * HEAD_DIM * sizeof(T) +
              num_frags_z * 16 * HEAD_DIM * sizeof(CacheT));
+  T* k_smem_scale = nullptr;
+  T* v_smem_scale = nullptr;
+  if constexpr (IsDynamicC8) {
+    k_smem_scale = reinterpret_cast<T*>(smem + NUM_WARPS * num_frags_x * 16 * HEAD_DIM * sizeof(T) +
+                                         num_frags_z * 16 * HEAD_DIM * sizeof(CacheT) * 2);
+    v_smem_scale = k_smem_scale + num_frags_z * 16;
+  }
 
 
   const uint32_t num_iterations = div_up(
@@ -216,7 +234,7 @@ __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 =
@@ -280,10 +298,22 @@ __global__ void multi_query_append_attention_c8_kernel(
 
 #pragma unroll 1
   for (uint32_t iter = 0; iter < num_iterations; ++iter) {
+    if constexpr (IsDynamicC8) {
+      produce_k_dynamic_scale<BLOCK_SIZE, num_frags_z, NUM_WARP_Q, T>(
+        k_smem_scale,
+        cache_k_scale_reg,
+        block_table_now,
+        cache_k_scale,
+        kv_idx_base,
+        kv_num_heads,
+        kv_head_idx,
+        chunk_end
+      );
+    }
     wait_group<1>();
     __syncthreads();
     // s = qk
-    compute_qk_c8<num_frags_x, num_frags_y, num_frags_z, T, CacheT, is_scale_channel_wise, IsFP8>(
+    compute_qk_c8<num_frags_x, num_frags_y, num_frags_z, T, CacheT, is_scale_channel_wise, IsFP8, IsDynamicC8>(
         &qo_smem,
         &q_smem_offset_r,
         &k_smem,
@@ -300,12 +330,15 @@ __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
@@ -313,6 +346,7 @@ __global__ void multi_query_append_attention_c8_kernel(
         s_frag, o_frag, m_frag, d_frag);
     __syncthreads();
 
+    const int ori_kv_idx_base = kv_idx_base;
     kv_idx_base += num_frags_z * 16;
     produce_k_blockwise_c8<SharedMemFillMode::kNoFill,
                            NUM_WARPS,
@@ -331,6 +365,18 @@ __global__ void multi_query_append_attention_c8_kernel(
                                        chunk_end,
                                        const_k_offset);
     commit_group();
+    if constexpr (IsDynamicC8) {
+      produce_v_dynamic_scale<BLOCK_SIZE, num_frags_z, NUM_WARP_Q, T>(
+        v_smem_scale,
+        cache_v_scale_reg,
+        block_table_now,
+        cache_v_scale,
+        ori_kv_idx_base,
+        kv_num_heads,
+        kv_head_idx,
+        chunk_end
+      );
+    }
     wait_group<1>();
     __syncthreads();
 
@@ -341,7 +387,9 @@ __global__ void multi_query_append_attention_c8_kernel(
                      BLOCK_SIZE,
                      T,
                      CacheT,
-                     is_scale_channel_wise, IsFP8>(
+                     is_scale_channel_wise,
+                     IsFP8,
+                     IsDynamicC8>(
         &v_smem, &v_smem_offset_r, s_frag, o_frag, d_frag, cache_v_scale_reg);
     __syncthreads();
 
@@ -458,14 +506,15 @@ template <typename T,
           typename OutT = T,
           bool ENABLE_PREFILL = true,
           bool is_scale_channel_wise=false,
-          bool IsFP8=false>
+          bool IsFP8 = false,
+          bool IsDynamicC8 = false>
 __global__ void multi_query_append_attention_c8_warp1_4_kernel(
     T *__restrict__ q,             // [token_num, (num_heads + 2* kv_num_head) * head_dim]
     CacheT *__restrict__ cache_k,  // [max_block_num, num_heads, block_size,
                                    // head_dim]
     CacheT *__restrict__ cache_v,
-    const T *__restrict__ cache_k_scale,  // [num_kv_heads, head_dim]
-    const T *__restrict__ cache_v_scale,  // [num_kv_heads, head_dim]
+    const T *__restrict__ cache_k_scale,  // [num_kv_heads] or [max_block_num, num_heads, block_size]
+    const T *__restrict__ cache_v_scale,  // [num_kv_heads] or [max_block_num, num_heads, block_size]
     const T *__restrict__ shift_bias,     // [q_num_heads * HEAD_DIM]
     const T *__restrict__ smooth_weight,  // [q_num_heads * HEAD_DIM]
     const int *__restrict__ seq_lens,
@@ -474,6 +523,8 @@ __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,
@@ -482,12 +533,14 @@ __global__ void multi_query_append_attention_c8_warp1_4_kernel(
     const float quant_min_bound,
     const float in_scale,
     const uint32_t chunk_size,
+    const int num_blocks_x_cpu,
     T *__restrict__ tmp_workspace,  // split kv [token_num, num_chunks,
                                     // num_heads, head_dim]
     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 int32_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>();
@@ -510,32 +563,39 @@ __global__ void multi_query_append_attention_c8_warp1_4_kernel(
   const uint32_t num_rows_per_block = num_frags_x * 16;
   const int *block_table_now = block_table + batch_id * max_block_num_per_seq;
 
+  //When cudagraph capture prefill, may launch more gridDim.x
+  if(btid >= static_cast<uint32_t>(num_blocks_x_cpu)){
+    return;
+  }
+
   const uint32_t q_len = seq_lens[batch_id];
   if (q_len <= 0) {
     return;
   }
-  T cache_k_scale_reg[num_frags_y * 4];
-  T cache_v_scale_reg[num_frags_y * 2];
-  if (is_scale_channel_wise) {
-    int scale_col_base = threadIdx.x % 4 * 2 + kv_head_idx * HEAD_DIM;
-    const T *cache_k_scale_cur_head = cache_k_scale + scale_col_base;
-    for (int i = 0; i < num_frags_y; ++i) {
-      const int scale_idx = i * 16;
-      cache_k_scale_reg[i * 4] = cache_k_scale_cur_head[scale_idx];
-      cache_k_scale_reg[i * 4 + 1] = cache_k_scale_cur_head[scale_idx + 1];
-      cache_k_scale_reg[i * 4 + 2] = cache_k_scale_cur_head[scale_idx + 8];
-      cache_k_scale_reg[i * 4 + 3] = cache_k_scale_cur_head[scale_idx + 9];
+  T cache_k_scale_reg[IsDynamicC8 ? num_frags_z * 2 : num_frags_y * 4];
+  T cache_v_scale_reg[IsDynamicC8 ? num_frags_z * 4 : num_frags_y * 2];
+  if constexpr (!IsDynamicC8) {
+    if constexpr (is_scale_channel_wise) {
+      int scale_col_base = threadIdx.x % 4 * 2 + kv_head_idx * HEAD_DIM;
+      const T *cache_k_scale_cur_head = cache_k_scale + scale_col_base;
+      for (int i = 0; i < num_frags_y; ++i) {
+        const int scale_idx = i * 16;
+        cache_k_scale_reg[i * 4] = cache_k_scale_cur_head[scale_idx];
+        cache_k_scale_reg[i * 4 + 1] = cache_k_scale_cur_head[scale_idx + 1];
+        cache_k_scale_reg[i * 4 + 2] = cache_k_scale_cur_head[scale_idx + 8];
+        cache_k_scale_reg[i * 4 + 3] = cache_k_scale_cur_head[scale_idx + 9];
+      }
+      scale_col_base = threadIdx.x / 4 + kv_head_idx * HEAD_DIM;
+      const T *cache_v_scale_cur_head = cache_v_scale + scale_col_base;
+      for (int i = 0; i < num_frags_y; ++i) {
+        const int scale_idx = i * 16;
+        cache_v_scale_reg[i * 2] = cache_v_scale_cur_head[scale_idx];
+        cache_v_scale_reg[i * 2 + 1] = cache_v_scale_cur_head[scale_idx + 8];
+      }
+    } else {
+      cache_k_scale_reg[0] = cache_k_scale[kv_head_idx];
+      cache_v_scale_reg[0] = cache_v_scale[kv_head_idx];
     }
-    scale_col_base = threadIdx.x / 4 + kv_head_idx * HEAD_DIM;
-    const T *cache_v_scale_cur_head = cache_v_scale + scale_col_base;
-    for (int i = 0; i < num_frags_y; ++i) {
-      const int scale_idx = i * 16;
-      cache_v_scale_reg[i * 2] = cache_v_scale_cur_head[scale_idx];
-      cache_v_scale_reg[i * 2 + 1] = cache_v_scale_cur_head[scale_idx + 8];
-    }
-  } else {
-    cache_k_scale_reg[0] = cache_k_scale[kv_head_idx];
-    cache_v_scale_reg[0] = cache_v_scale[kv_head_idx];
   }
   const uint32_t q_end =
       min(q_len, div_up((tile_id + 1) * num_rows_per_block, GROUP_SIZE));
@@ -601,7 +661,7 @@ __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 * 2 : nullptr;
   smem_t qo_smem(smem);
 
   uint32_t q_smem_offset_r = smem_t::get_permuted_offset<num_vecs_per_head>(
@@ -626,6 +686,13 @@ __global__ void multi_query_append_attention_c8_warp1_4_kernel(
   smem_t k_smem(smem + num_frags_x * 16 * HEAD_DIM * sizeof(T)),
       v_smem(smem + num_frags_x * 16 * HEAD_DIM * sizeof(T) +
              NUM_WARP_KV * num_frags_z * 16 * HEAD_DIM * sizeof(CacheT));
+  T* k_smem_scale = nullptr;
+  T* v_smem_scale = nullptr;
+  if constexpr (IsDynamicC8) {
+    k_smem_scale = reinterpret_cast<T*>(smem + num_frags_x * 16 * HEAD_DIM * sizeof(T) +
+                                        NUM_WARP_KV * num_frags_z * 16 * HEAD_DIM * sizeof(CacheT) * 2);
+    v_smem_scale = k_smem_scale + NUM_WARP_KV * num_frags_z * 16;
+  }
 
   const uint32_t num_iterations = div_up(
       CAUSAL
@@ -642,7 +709,7 @@ __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 =
@@ -708,11 +775,23 @@ __global__ void multi_query_append_attention_c8_warp1_4_kernel(
   commit_group();
 #pragma unroll 1
   for (uint32_t iter = 0; iter < num_iterations; ++iter) {
+    if constexpr (IsDynamicC8) {
+      produce_k_dynamic_scale<BLOCK_SIZE, num_frags_z, NUM_WARP_Q, T>(
+        k_smem_scale,
+        cache_k_scale_reg,
+        block_table_now,
+        cache_k_scale,
+        kv_idx_base,
+        kv_num_heads,
+        kv_head_idx,
+        chunk_end
+      );
+    }
     wait_group<1>();
     __syncthreads();
 
     // s = qk
-    compute_qk_c8<num_frags_x, num_frags_y, num_frags_z, T, CacheT, is_scale_channel_wise, IsFP8>(
+    compute_qk_c8<num_frags_x, num_frags_y, num_frags_z, T, CacheT, is_scale_channel_wise, IsFP8, IsDynamicC8>(
         &qo_smem,
         &q_smem_offset_r,
         &k_smem,
@@ -728,12 +807,16 @@ __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
@@ -741,6 +824,7 @@ __global__ void multi_query_append_attention_c8_warp1_4_kernel(
         s_frag, o_frag, m_frag, d_frag);
     __syncthreads();
 
+    const uint32_t ori_kv_idx_base = kv_idx_base;
     kv_idx_base += NUM_WARP_KV * num_frags_z * 16;
     produce_k_blockwise_c8<SharedMemFillMode::kNoFill,
                            NUM_WARPS,
@@ -759,6 +843,18 @@ __global__ void multi_query_append_attention_c8_warp1_4_kernel(
                                        chunk_end,
                                        const_k_offset);
     commit_group();
+    if constexpr (IsDynamicC8) {
+      produce_v_dynamic_scale<BLOCK_SIZE, num_frags_z, NUM_WARP_Q, T>(
+        v_smem_scale,
+        cache_v_scale_reg,
+        block_table_now,
+        cache_v_scale,
+        ori_kv_idx_base,
+        kv_num_heads,
+        kv_head_idx,
+        chunk_end
+      );
+    }
     wait_group<1>();
     __syncthreads();
 
@@ -769,7 +865,9 @@ __global__ void multi_query_append_attention_c8_warp1_4_kernel(
                                   BLOCK_SIZE,
                                   T,
                                   CacheT,
-                                  is_scale_channel_wise, IsFP8>(
+                                  is_scale_channel_wise,
+                                  IsFP8,
+                                  IsDynamicC8>(
         &v_smem, &v_smem_offset_r, s_frag, o_frag, d_frag, cache_v_scale_reg);
     __syncthreads();
 
@@ -883,7 +981,8 @@ template <typename T,
           uint32_t NUM_WARP_Q,
           typename OutT = T,
           bool ENABLE_PREFILL = true,
-          bool IsFP8=false>
+          bool IsFP8 = false,
+          bool IsDynamicC8 = false>
 void MultiQueryAppendC8Attention(
     const AppendAttnMetaData &meta_data,
     const paddle::Tensor &qkv,
@@ -941,7 +1040,8 @@ void MultiQueryAppendC8Attention(
     constexpr uint32_t num_frags_z = BLOCK_SIZE / 16;
     constexpr uint32_t smem_size =
         num_warps * num_frags_x * 16 * HEAD_DIM * sizeof(T) +
-        num_frags_z * 16 * HEAD_DIM * sizeof(uint8_t) * 2;
+        num_frags_z * 16 * HEAD_DIM * sizeof(uint8_t) * 2 +
+        num_frags_z * 16 * sizeof(T) * 2;
     auto split_kv_kernel =
         multi_query_append_attention_c8_kernel<NV_TYPE,
                                                uint8_t,
@@ -958,7 +1058,9 @@ void MultiQueryAppendC8Attention(
                                                num_frags_y,
                                                OUT_NV_TYPE,
                                                ENABLE_PREFILL,
-                                               false, IsFP8>;
+                                               false,
+                                               IsFP8,
+                                               IsDynamicC8>;
     if (is_scale_channel_wise) {
       split_kv_kernel =
         multi_query_append_attention_c8_kernel<NV_TYPE,
@@ -976,7 +1078,9 @@ void MultiQueryAppendC8Attention(
                                                num_frags_y,
                                                OUT_NV_TYPE,
                                                ENABLE_PREFILL,
-                                               true, IsFP8>;
+                                               true,
+                                               IsFP8,
+                                               IsDynamicC8>;
     }
     if (smem_size >= 48 * 1024) {
       cudaFuncSetAttribute(split_kv_kernel,
@@ -1010,7 +1114,9 @@ void MultiQueryAppendC8Attention(
                                                  num_frags_y,
                                                  OUT_NV_TYPE,
                                                  ENABLE_PREFILL,
-                                                 false, IsFP8>;
+                                                 false,
+                                                 IsFP8,
+                                                 IsDynamicC8>;
       if (is_scale_channel_wise) {
         nosplit_kv_kernel =
           multi_query_append_attention_c8_kernel<NV_TYPE,
@@ -1028,7 +1134,9 @@ void MultiQueryAppendC8Attention(
                                                  num_frags_y,
                                                  OUT_NV_TYPE,
                                                  ENABLE_PREFILL,
-                                                 true, IsFP8>;
+                                                 true,
+                                                 IsFP8,
+                                                 IsDynamicC8>;
       }
       if (smem_size >= 48 * 1024) {
         cudaFuncSetAttribute(nosplit_kv_kernel,
@@ -1054,6 +1162,7 @@ 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,
@@ -1062,6 +1171,7 @@ void MultiQueryAppendC8Attention(
           quant_min_bound,
           in_scale,
           chunk_size,
+          num_blocks_x_cpu,
           nullptr,
           nullptr,
           nullptr,
@@ -1111,6 +1221,7 @@ 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,
@@ -1119,6 +1230,7 @@ void MultiQueryAppendC8Attention(
           quant_min_bound,
           in_scale,
           chunk_size,
+          num_blocks_x_cpu,
           reinterpret_cast<NV_TYPE *>(tmp_workspace->ptr()),
           static_cast<float *>(tmp_m->ptr()),
           static_cast<float *>(tmp_d->ptr()),
@@ -1204,7 +1316,8 @@ void MultiQueryAppendC8Attention(
     constexpr uint32_t num_frags_z = BLOCK_SIZE / 16 / NUM_WARP_KV * 2;
     constexpr uint32_t smem_size =
         num_frags_x * 16 * HEAD_DIM * sizeof(T) +
-        NUM_WARP_KV * num_frags_z * 16 * HEAD_DIM * sizeof(uint8_t) * 2;
+        NUM_WARP_KV * num_frags_z * 16 * HEAD_DIM * sizeof(uint8_t) * 2 +
+        NUM_WARP_KV * num_frags_z * 16 * sizeof(T) * 2;
     auto split_kv_kernel =
         multi_query_append_attention_c8_warp1_4_kernel<NV_TYPE,
                                                        uint8_t,
@@ -1221,7 +1334,9 @@ void MultiQueryAppendC8Attention(
                                                        num_frags_y,
                                                        OUT_NV_TYPE,
                                                        ENABLE_PREFILL,
-                                                       false, IsFP8>;
+                                                       false,
+                                                       IsFP8,
+                                                       IsDynamicC8>;
     if (is_scale_channel_wise) {
       split_kv_kernel =
         multi_query_append_attention_c8_warp1_4_kernel<NV_TYPE,
@@ -1239,7 +1354,9 @@ void MultiQueryAppendC8Attention(
                                                        num_frags_y,
                                                        OUT_NV_TYPE,
                                                        ENABLE_PREFILL,
-                                                       true, IsFP8>;
+                                                       true,
+                                                       IsFP8,
+                                                       IsDynamicC8>;
     }
     if (smem_size >= 48 * 1024) {
       cudaFuncSetAttribute(split_kv_kernel,
@@ -1254,10 +1371,17 @@ 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);
+    int32_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,
@@ -1274,7 +1398,9 @@ void MultiQueryAppendC8Attention(
                                                          num_frags_y,
                                                          OUT_NV_TYPE,
                                                          ENABLE_PREFILL,
-                                                         false, IsFP8>;
+                                                         false,
+                                                         IsFP8,
+                                                         IsDynamicC8>;
       if (is_scale_channel_wise) {
         nosplit_kv_kernel =
           multi_query_append_attention_c8_warp1_4_kernel<NV_TYPE,
@@ -1292,7 +1418,9 @@ void MultiQueryAppendC8Attention(
                                                          num_frags_y,
                                                          OUT_NV_TYPE,
                                                          ENABLE_PREFILL,
-                                                         true, IsFP8>;
+                                                         true,
+                                                         IsFP8,
+                                                         IsDynamicC8>;
       }
       if (smem_size >= 48 * 1024) {
         cudaFuncSetAttribute(nosplit_kv_kernel,
@@ -1318,6 +1446,9 @@ 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,
@@ -1326,11 +1457,13 @@ void MultiQueryAppendC8Attention(
           quant_min_bound,
           in_scale,
           chunk_size,
+          num_blocks_x_cpu,
           nullptr,
           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 +1510,8 @@ 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 +1521,9 @@ 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,
@@ -1396,11 +1532,13 @@ void MultiQueryAppendC8Attention(
           quant_min_bound,
           in_scale,
           chunk_size,
+          num_blocks_x_cpu,
           reinterpret_cast<NV_TYPE *>(tmp_workspace->ptr()),
           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 +1556,8 @@ 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 +1574,14 @@ 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 +1592,8 @@ 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,
@@ -1517,6 +1655,7 @@ void CascadeAppendAttentionC8Kernel(
     const bool causal,
     const bool is_decoder,
     const bool enable_prefill,
+    const std::string& cache_quant_type_str,
     cudaStream_t& stream,
     paddle::Tensor* out) {
   const auto token_num = meta_data.token_nums;
@@ -1525,6 +1664,7 @@ void CascadeAppendAttentionC8Kernel(
   const auto num_heads = meta_data.q_num_heads;
   const auto group_size = meta_data.q_num_heads / meta_data.kv_num_heads;
   const auto head_dim = meta_data.head_dims;
+  bool is_dynamic_cfp8 = cache_quant_type_str == "block_wise_fp8";
 
   DISPATCH_CAUSAL(
       causal,
@@ -1543,43 +1683,46 @@ void CascadeAppendAttentionC8Kernel(
                       BLOCK_SIZE,
                       {DISPATCH_BLOCKSHAPE_Q(
                           block_shape_q, BLOCK_SHAPE_Q, NUM_WARP_Q, {
-                            MultiQueryAppendC8Attention<T,
-                                                        GROUP_SIZE,
-                                                        HEAD_DIM,
-                                                        BLOCK_SIZE,
-                                                        CAUSAL,
-                                                        BLOCK_SHAPE_Q,
-                                                        NUM_WARP_Q,
-                                                        OutT,
-                                                        ENABLE_PREFILL, IsFP8>(
-                                meta_data,
-                                qkv,
-                                cache_k,
-                                cache_v,
-                                attn_mask,
-                                cache_k_scale.get(),
-                                cache_v_scale.get(),
-                                shift_bias,
-                                smooth_weight,
-                                seq_lens_q,
-                                seq_lens_kv,
-                                seq_lens_encoder,
-                                batch_id_per_token,
-                                cu_seqlens_q,
-                                block_table,
-                                batch_ids,
-                                tile_ids_per_batch,
-                                num_blocks,
-                                max_seq_len,
-                                max_dec_len,
-                                quant_max_bound,
-                                quant_min_bound,
-                                in_scale,
-                                max_partition_size,
-                                encoder_max_partition_size,
-                                speculate_max_draft_token_num,
-                                is_decoder,
-                                stream,
-                                out);
-                          })})})})})})
+                            DISPATCH_DyCfp8(is_dynamic_cfp8, IsDynamicC8, {
+                              MultiQueryAppendC8Attention<T,
+                                                          GROUP_SIZE,
+                                                          HEAD_DIM,
+                                                          BLOCK_SIZE,
+                                                          CAUSAL,
+                                                          BLOCK_SHAPE_Q,
+                                                          NUM_WARP_Q,
+                                                          OutT,
+                                                          ENABLE_PREFILL,
+                                                          IsFP8,
+                                                          IsDynamicC8>(
+                                  meta_data,
+                                  qkv,
+                                  cache_k,
+                                  cache_v,
+                                  attn_mask,
+                                  cache_k_scale.get(),
+                                  cache_v_scale.get(),
+                                  shift_bias,
+                                  smooth_weight,
+                                  seq_lens_q,
+                                  seq_lens_kv,
+                                  seq_lens_encoder,
+                                  batch_id_per_token,
+                                  cu_seqlens_q,
+                                  block_table,
+                                  batch_ids,
+                                  tile_ids_per_batch,
+                                  num_blocks,
+                                  max_seq_len,
+                                  max_dec_len,
+                                  quant_max_bound,
+                                  quant_min_bound,
+                                  in_scale,
+                                  max_partition_size,
+                                  encoder_max_partition_size,
+                                  speculate_max_draft_token_num,
+                                  is_decoder,
+                                  stream,
+                                  out);
+      })})})})})})})
 }

custom_ops/gpu_ops/append_attn/append_attention_func.cuh

@@ -384,6 +384,113 @@ __device__ __forceinline__ void produce_v_blockwise_c8(
   }
 }
 
+template<uint32_t block_size,
+         uint32_t num_frags_z,
+         uint32_t NUM_WARP_Q,
+         typename T>
+__device__ __forceinline__ void produce_k_dynamic_scale(
+  T* k_smem_scale,
+  T* cache_k_reg,
+  const int* block_table_now,
+  const T* cache_k_scale,
+  const uint32_t kv_idx,
+  const uint32_t kv_num_heads,
+  const uint32_t kv_head_idx,
+  const uint32_t chunk_end
+) {
+  const uint32_t tx = threadIdx.x, ty = threadIdx.y;
+  if constexpr (NUM_WARP_Q == 4) {
+    // 4 warps shared block_size
+    const uint32_t tid = ty * 32 + tx;
+    int block_id = __ldg(&block_table_now[kv_idx / block_size]);
+    if (block_id < 0) block_id = 0;
+    const T* cache_k_scale_now = cache_k_scale + block_id * kv_num_heads * block_size + kv_head_idx * block_size;
+    if (tid < block_size) {
+      k_smem_scale[tid] = cache_k_scale_now[tid];
+    }
+    __syncthreads();
+    const uint32_t row_id = tx / 4;
+    for (uint32_t fz = 0; fz < num_frags_z; fz++) {
+      cache_k_reg[fz * 2] = k_smem_scale[fz * 16 + row_id];
+      cache_k_reg[fz * 2 + 1] = k_smem_scale[fz * 16 + row_id + 8];
+    }
+  } else {
+    // 1 warp 32 tokens
+    const uint32_t kv_idx_now = kv_idx + block_size * ty / 2;
+    int block_id = __ldg(&block_table_now[kv_idx_now / block_size]);
+    if (block_id < 0) block_id = 0;
+    const T* cache_k_scale_now = cache_k_scale + block_id * kv_num_heads * block_size + kv_head_idx * block_size;
+    const int kv_idx_this_thread = kv_idx + ty * 32 + tx;
+    if (kv_idx_this_thread < chunk_end) {
+      k_smem_scale[ty * 32 + tx] = cache_k_scale_now[(ty  % 2) * 32 + tx];
+    } else {
+      k_smem_scale[ty * 32 + tx] = 0;
+    }
+    __syncwarp();
+    const uint32_t row_id = tx / 4;
+    for (uint32_t fz = 0; fz < num_frags_z; fz++) {
+      cache_k_reg[fz * 2] = k_smem_scale[ty * 32 + fz * 16 + row_id];
+      cache_k_reg[fz * 2 + 1] = k_smem_scale[ty * 32 + fz * 16 + row_id + 8];
+    }
+  }
+}
+
+template<uint32_t block_size,
+         uint32_t num_frags_z,
+         uint32_t NUM_WARP_Q,
+         typename T>
+__device__ __forceinline__ void produce_v_dynamic_scale(
+  T* v_smem_scale,
+  T* cache_v_reg,
+  const int* block_table_now,
+  const T* cache_v_scale,
+  const uint32_t kv_idx,
+  const uint32_t kv_num_heads,
+  const uint32_t kv_head_idx,
+  const uint32_t chunk_end
+) {
+  const uint32_t tx = threadIdx.x, ty = threadIdx.y;
+
+  if constexpr (NUM_WARP_Q == 4) {
+    // 4 warps shared block_size
+    const uint32_t tid = ty * 32 + tx;
+    int block_id = __ldg(&block_table_now[kv_idx / block_size]);
+    if (block_id < 0) block_id = 0;
+    const T* cache_v_scale_now = cache_v_scale + block_id * kv_num_heads * block_size + kv_head_idx * block_size;
+    if (tid < block_size) {
+      v_smem_scale[tid] = cache_v_scale_now[tid];
+    }
+    __syncthreads();
+    const uint32_t row_id = tx % 4 * 2;
+    for (uint32_t fz = 0; fz < num_frags_z; fz++) {
+      cache_v_reg[fz * 4] = v_smem_scale[fz * 16 + row_id];
+      cache_v_reg[fz * 4 + 1] = v_smem_scale[fz * 16 + row_id + 1];
+      cache_v_reg[fz * 4 + 2] = v_smem_scale[fz * 16 + row_id + 8];
+      cache_v_reg[fz * 4 + 3] = v_smem_scale[fz * 16 + row_id + 9];
+    }
+  } else {
+    // 1 warp 32 tokens
+    const uint32_t kv_idx_now = kv_idx + block_size * ty / 2;
+    int block_id = __ldg(&block_table_now[kv_idx_now / block_size]);
+    if (block_id < 0) block_id = 0;
+    const T* cache_v_scale_now = cache_v_scale + block_id * kv_num_heads * block_size + kv_head_idx * block_size;
+    const int kv_idx_this_thread = kv_idx + ty * 32 + tx;
+    if (kv_idx_this_thread < chunk_end) {
+      v_smem_scale[ty * 32 + tx] = cache_v_scale_now[(ty  % 2) * 32 + tx];
+    } else {
+      v_smem_scale[ty * 32 + tx] = 0;
+    }
+    __syncwarp();
+    const uint32_t row_id = tx % 4 * 2;
+    for (uint32_t fz = 0; fz < num_frags_z; fz++) {
+      cache_v_reg[fz * 4] = v_smem_scale[ty * 32 + fz * 16 + row_id];
+      cache_v_reg[fz * 4 + 1] = v_smem_scale[ty * 32 + fz * 16 + row_id + 1];
+      cache_v_reg[fz * 4 + 2] = v_smem_scale[ty * 32 + fz * 16 + row_id + 8];
+      cache_v_reg[fz * 4 + 3] = v_smem_scale[ty * 32 + fz * 16 + row_id + 9];
+    }
+  }
+}
+
 template <SharedMemFillMode fill_mode,
           uint32_t num_warps,
           uint32_t block_size,
@@ -816,7 +923,8 @@ template <uint32_t num_frags_x,
           typename T,
           typename CacheT,
           bool is_scale_channel_wise = false,
-          bool IsFP8=false>
+          bool IsFP8 = false,
+          bool IsDynamicC8 = false>
 __device__ __forceinline__ void compute_qk_c8(smem_t* q_smem,
                                               uint32_t* q_smem_offset_r,
                                               smem_t* k_smem,
@@ -860,20 +968,27 @@ __device__ __forceinline__ void compute_qk_c8(smem_t* q_smem,
         convert_c8<T,IsFP8>(b_frag_dq_T, b_frag[fy * 2]);
         convert_c8<T,IsFP8>(b_frag_dq_T + 4, b_frag[fy * 2 + 1]);
         // scale zp
-        if constexpr (is_scale_channel_wise) {
-          const int scale_col = (ky * 2 + fy) * 4;
-          b_frag_dq_T[0] *= cache_k_scale[scale_col];
-          b_frag_dq_T[1] *= cache_k_scale[scale_col + 1];
-          b_frag_dq_T[2] *= cache_k_scale[scale_col + 2];
-          b_frag_dq_T[3] *= cache_k_scale[scale_col + 3];
-          b_frag_dq_T[4] *= cache_k_scale[scale_col];
-          b_frag_dq_T[5] *= cache_k_scale[scale_col + 1];
-          b_frag_dq_T[6] *= cache_k_scale[scale_col + 2];
-          b_frag_dq_T[7] *= cache_k_scale[scale_col + 3];
+        if constexpr (!IsDynamicC8) {
+          if constexpr (is_scale_channel_wise) {
+            const int scale_col = (ky * 2 + fy) * 4;
+            b_frag_dq_T[0] *= cache_k_scale[scale_col];
+            b_frag_dq_T[1] *= cache_k_scale[scale_col + 1];
+            b_frag_dq_T[2] *= cache_k_scale[scale_col + 2];
+            b_frag_dq_T[3] *= cache_k_scale[scale_col + 3];
+            b_frag_dq_T[4] *= cache_k_scale[scale_col];
+            b_frag_dq_T[5] *= cache_k_scale[scale_col + 1];
+            b_frag_dq_T[6] *= cache_k_scale[scale_col + 2];
+            b_frag_dq_T[7] *= cache_k_scale[scale_col + 3];
+          } else {
+#pragma unroll
+            for (uint32_t b_i = 0; b_i < 8; ++b_i) {
+              b_frag_dq_T[b_i] *= cache_k_scale[0];
+            }
+          }
         } else {
 #pragma unroll
           for (uint32_t b_i = 0; b_i < 8; ++b_i) {
-            b_frag_dq_T[b_i] *= cache_k_scale[0];
+            b_frag_dq_T[b_i] *= cache_k_scale[fz * 2 + b_i / 4];
           }
         }
 #pragma unroll
@@ -905,12 +1020,15 @@ 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 int32_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 +1042,21 @@ __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 * 2 + 1] || kv_idx < mask_offset[q_idx * 2]) : 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 && attn_mask_len > 0 && q_idx < static_cast<uint32_t>(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 +1064,7 @@ __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) +
@@ -1078,7 +1208,9 @@ template <uint32_t num_frags_x,
           uint32_t block_size,
           typename T,
           typename CacheT,
-          bool is_scale_channel_wise = false, bool IsFP8=false>
+          bool is_scale_channel_wise = false,
+          bool IsFP8 = false,
+          bool IsDynamicC8 = false>
 __device__ __forceinline__ void compute_sfm_v_c8(
     smem_t* v_smem,
     uint32_t* v_smem_offset_r,
@@ -1120,16 +1252,28 @@ __device__ __forceinline__ void compute_sfm_v_c8(
         convert_c8<T,IsFP8>(b_frag_dq_T, b_frag[fz * 2]);
         convert_c8<T,IsFP8>(b_frag_dq_T + 4, b_frag[fz * 2 + 1]);
         // scale zp
-        if constexpr (is_scale_channel_wise) {
+        if constexpr (!IsDynamicC8) {
+          if constexpr (is_scale_channel_wise) {
 #pragma unroll
-          for (uint32_t b_i = 0; b_i < 8; ++b_i) {
-            b_frag_dq_T[b_i] *= cache_v_scale[b_i / 4 + fy * 2];
+            for (uint32_t b_i = 0; b_i < 8; ++b_i) {
+              b_frag_dq_T[b_i] *= cache_v_scale[b_i / 4 + fy * 2];
+            }
+          } else {
+#pragma unroll
+            for (uint32_t b_i = 0; b_i < 8; ++b_i) {
+              b_frag_dq_T[b_i] *= cache_v_scale[0];
+            }
           }
         } else {
-#pragma unroll
-          for (uint32_t b_i = 0; b_i < 8; ++b_i) {
-            b_frag_dq_T[b_i] *= cache_v_scale[0];
-          }
+          const int scale_col = (kz * 2 + fz) * 4;
+          b_frag_dq_T[0] *= cache_v_scale[scale_col];
+          b_frag_dq_T[1] *= cache_v_scale[scale_col + 1];
+          b_frag_dq_T[2] *= cache_v_scale[scale_col + 2];
+          b_frag_dq_T[3] *= cache_v_scale[scale_col + 3];
+          b_frag_dq_T[4] *= cache_v_scale[scale_col];
+          b_frag_dq_T[5] *= cache_v_scale[scale_col + 1];
+          b_frag_dq_T[6] *= cache_v_scale[scale_col + 2];
+          b_frag_dq_T[7] *= cache_v_scale[scale_col + 3];
         }
 #pragma unroll
         for (uint32_t fx = 0; fx < num_frags_x; ++fx) {  // m: num_frags_x * 16
@@ -1156,7 +1300,9 @@ template <uint32_t num_frags_x,
           uint32_t block_size,
           typename T,
           typename CacheT,
-          bool is_scale_channel_wise = false, bool IsFP8=false>
+          bool is_scale_channel_wise = false,
+          bool IsFP8 = false,
+          bool IsDynamicC8 = false>
 __device__ __forceinline__ void compute_sfm_v_c8_iter_sq_bvec(
     smem_t* v_smem,
     uint32_t* v_smem_offset_r,
@@ -1200,16 +1346,28 @@ __device__ __forceinline__ void compute_sfm_v_c8_iter_sq_bvec(
         convert_c8<T,IsFP8>(b_frag_dq_T, b_frag[fz * 2]);
         convert_c8<T,IsFP8>(b_frag_dq_T + 4, b_frag[fz * 2 + 1]);
         // scale zp
-        if constexpr (is_scale_channel_wise) {
+        if constexpr (!IsDynamicC8) {
+          if constexpr (is_scale_channel_wise) {
 #pragma unroll
-          for (uint32_t b_i = 0; b_i < 8; ++b_i) {
-            b_frag_dq_T[b_i] *= cache_v_scale[b_i / 4 + fy * 2];
+            for (uint32_t b_i = 0; b_i < 8; ++b_i) {
+              b_frag_dq_T[b_i] *= cache_v_scale[b_i / 4 + fy * 2];
+            }
+          } else {
+            #pragma unroll
+            for (uint32_t b_i = 0; b_i < 8; ++b_i) {
+              b_frag_dq_T[b_i] *= cache_v_scale[0];
+            }
           }
         } else {
-          #pragma unroll
-          for (uint32_t b_i = 0; b_i < 8; ++b_i) {
-            b_frag_dq_T[b_i] *= cache_v_scale[0];
-          }
+          const int scale_col = (kz * 2 + fz) * 4;
+          b_frag_dq_T[0] *= cache_v_scale[scale_col];
+          b_frag_dq_T[1] *= cache_v_scale[scale_col + 1];
+          b_frag_dq_T[2] *= cache_v_scale[scale_col + 2];
+          b_frag_dq_T[3] *= cache_v_scale[scale_col + 3];
+          b_frag_dq_T[4] *= cache_v_scale[scale_col];
+          b_frag_dq_T[5] *= cache_v_scale[scale_col + 1];
+          b_frag_dq_T[6] *= cache_v_scale[scale_col + 2];
+          b_frag_dq_T[7] *= cache_v_scale[scale_col + 3];
         }
 #pragma unroll
         for (uint32_t fx = 0; fx < num_frags_x; ++fx) {  // m: num_frags_x * 16

custom_ops/gpu_ops/append_attn/append_attention_kernel.h

@@ -103,6 +103,7 @@ void CascadeAppendAttentionC8Kernel(
     const bool causal,
     const bool is_decoder,
     const bool enable_prefill,
+    const std::string& cache_quant_type_str,
     cudaStream_t& stream,
     paddle::Tensor* out);
 
@@ -264,9 +265,10 @@ void CascadeAppendAttentionKernel(
                                                 causal,
                                                 is_decoder,
                                                 enable_prefill,
+                                                cache_quant_type_str,
                                                 stream,
                                                 out);
-    } else if (cache_quant_type_str == "cache_fp8") {
+    } else if (cache_quant_type_str == "cache_fp8" or cache_quant_type_str == "block_wise_fp8") {
         CascadeAppendAttentionC8Kernel<T, OutT, true>(meta_data,
                                                 qkv,
                                                 cache_k,
@@ -299,6 +301,7 @@ void CascadeAppendAttentionKernel(
                                                 causal,
                                                 is_decoder,
                                                 enable_prefill,
+                                                cache_quant_type_str,
                                                 stream,
                                                 out);
     } else if (cache_quant_type_str == "cache_int4_zp") {

custom_ops/gpu_ops/append_attn/decoder_write_cache_with_rope_kernel.cu

@@ -16,12 +16,11 @@
 #include "utils.cuh"
 
 template <typename T, typename QKV_TYPE>
-void append_decode_cache_rope(const QKV_TYPE* qkv,
+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,
@@ -38,6 +37,68 @@ void append_decode_cache_rope(const QKV_TYPE* qkv,
                               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,
+                                            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,
+                              T* value_cache,
+                              T* qkv_out,
+                              const int* block_tables,
+                              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 rotary_dim,
+                              const int block_size,
+                              const int bsz,
+                              const cudaStream_t& stream,
+                              const bool use_neox_style,
                               const bool rope_3d) {
   const uint32_t elem_nums =
       use_neox_style ? bsz * (num_heads + 2 * kv_num_heads) * dim_head / 2
@@ -57,7 +118,6 @@ void append_decode_cache_rope(const QKV_TYPE* qkv,
               value_cache,
               qkv_out,
               block_tables,
-              batch_id_per_token,
               cu_seqlens_q,
               seq_lens,
               seq_lens_encoder,
@@ -71,15 +131,37 @@ 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,
                                                 qkv_out,
                                                 block_tables,
-                                                batch_id_per_token,
                                                 cu_seqlens_q,
                                                 seq_lens,
                                                 seq_lens_encoder,
@@ -91,7 +173,9 @@ 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) {
@@ -102,7 +186,6 @@ void append_decode_cache_rope(const QKV_TYPE* qkv,
               value_cache,
               qkv_out,
               block_tables,
-              batch_id_per_token,
               cu_seqlens_q,
               seq_lens,
               seq_lens_encoder,
@@ -125,7 +208,6 @@ void append_decode_cache_rope(const QKV_TYPE* qkv,
                                                 value_cache,
                                                 qkv_out,
                                                 block_tables,
-                                                batch_id_per_token,
                                                 cu_seqlens_q,
                                                 seq_lens,
                                                 seq_lens_encoder,
@@ -149,7 +231,6 @@ void append_decode_cache_int8_rope(const QKV_TYPE* qkv,
                                    uint8_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,
@@ -182,7 +263,6 @@ void append_decode_cache_int8_rope(const QKV_TYPE* qkv,
               value_cache,
               qkv_out,
               block_tables,
-              batch_id_per_token,
               cu_seqlens_q,
               seq_lens,
               seq_lens_encoder,
@@ -198,7 +278,8 @@ 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>>>(
@@ -207,7 +288,6 @@ void append_decode_cache_int8_rope(const QKV_TYPE* qkv,
               value_cache,
               qkv_out,
               block_tables,
-              batch_id_per_token,
               cu_seqlens_q,
               seq_lens,
               seq_lens_encoder,
@@ -221,7 +301,8 @@ 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) {
@@ -232,7 +313,6 @@ void append_decode_cache_int8_rope(const QKV_TYPE* qkv,
               value_cache,
               qkv_out,
               block_tables,
-              batch_id_per_token,
               cu_seqlens_q,
               seq_lens,
               seq_lens_encoder,
@@ -248,7 +328,8 @@ 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>>>(
@@ -257,7 +338,6 @@ void append_decode_cache_int8_rope(const QKV_TYPE* qkv,
               value_cache,
               qkv_out,
               block_tables,
-              batch_id_per_token,
               cu_seqlens_q,
               seq_lens,
               seq_lens_encoder,
@@ -271,7 +351,8 @@ void append_decode_cache_int8_rope(const QKV_TYPE* qkv,
               block_size,
               127.0f,
               -127.0f,
-              kv_num_heads);
+              kv_num_heads,
+              rope_3d);
     }
   }
 }
@@ -282,7 +363,6 @@ void append_decode_cache_int4_rope(const QKV_TYPE* qkv,
                                    uint8_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,
@@ -317,7 +397,6 @@ void append_decode_cache_int4_rope(const QKV_TYPE* qkv,
               value_cache,
               qkv_out,
               block_tables,
-              batch_id_per_token,
               cu_seqlens_q,
               seq_lens,
               seq_lens_encoder,
@@ -335,7 +414,8 @@ 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>>>(
@@ -344,7 +424,6 @@ void append_decode_cache_int4_rope(const QKV_TYPE* qkv,
               value_cache,
               qkv_out,
               block_tables,
-              batch_id_per_token,
               cu_seqlens_q,
               seq_lens,
               seq_lens_encoder,
@@ -360,7 +439,8 @@ 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) {
@@ -371,7 +451,6 @@ void append_decode_cache_int4_rope(const QKV_TYPE* qkv,
               value_cache,
               qkv_out,
               block_tables,
-              batch_id_per_token,
               cu_seqlens_q,
               seq_lens,
               seq_lens_encoder,
@@ -389,7 +468,8 @@ 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>>>(
@@ -398,7 +478,6 @@ void append_decode_cache_int4_rope(const QKV_TYPE* qkv,
               value_cache,
               qkv_out,
               block_tables,
-              batch_id_per_token,
               cu_seqlens_q,
               seq_lens,
               seq_lens_encoder,
@@ -414,7 +493,8 @@ void append_decode_cache_int4_rope(const QKV_TYPE* qkv,
               block_size,
               7.0f,
               -8.0f,
-              kv_num_heads);
+              kv_num_heads,
+              rope_3d);
     }
   }
 }
@@ -424,7 +504,6 @@ void DecoderWriteCacheWithRoPEKernel(
     const paddle::Tensor& qkv,
     const paddle::Tensor& seq_lens,
     const paddle::Tensor& seq_lens_encoder,
-    const paddle::Tensor& batch_id_per_token,
     const paddle::Tensor& cu_seqlens_q,
     const paddle::Tensor& block_tables,
     const paddle::optional<paddle::Tensor>& rotary_embs,
@@ -441,7 +520,10 @@ 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_;
@@ -458,119 +540,126 @@ 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>(),
+          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 if (cache_quant_type_str == "block_wise_fp8") {
+      constexpr int num_warps = 4;
+      const int all_warps =
+          ((num_heads + 2 * kv_num_heads) + num_warps - 1) / num_warps * num_warps;
+      dim3 grids(bsz, all_warps / num_warps);
+      append_decode_cache_int8_rope_qk_norm_kernel<DataType_, 4, 0, 128, false, true>
+          <<<grids, num_warps * 32, 0, stream>>>(
+              reinterpret_cast<const DataType_*>(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>(),
+              cu_seqlens_q.data<int>(),
+              seq_lens.data<int>(),
+              seq_lens_encoder.data<int>(),
+              cos_emb,
+              sin_emb,
+              const_cast<DataType_*>(reinterpret_cast<const DataType_*>(cache_k_scale.get().data<T>())),
+              const_cast<DataType_*>(reinterpret_cast<const DataType_*>((cache_v_scale.get().data<T>()))),
+              q_norm_weight.get().data<float>(),
+              k_norm_weight.get().data<float>(),
+              max_seq_len,
+              max_blocks_per_seq,
+              num_heads,
+              block_size,
+              127.0f,
+              -127.0f,
+              kv_num_heads,
+              rope_3d,
+              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 just supports cache_quant_type none/block_wise_fp8");
     }
-  } 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>(),
+          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>(),
           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>(),
@@ -596,49 +685,144 @@ 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>(),
+            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>(),
+            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 == "block_wise_fp8") {
+      constexpr int num_warps = 4;
+      const int all_warps =
+          ((num_heads + 2 * kv_num_heads) + num_warps - 1) / num_warps * num_warps;
+      dim3 grids(bsz, all_warps / num_warps);
+      append_decode_cache_int8_rope_qk_norm_kernel<DataType_, 4, 0, 128, false, true>
+          <<<grids, num_warps * 32, 0, stream>>>(
+              reinterpret_cast<const DataType_*>(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>(),
+              cu_seqlens_q.data<int>(),
+              seq_lens.data<int>(),
+              seq_lens_encoder.data<int>(),
+              cos_emb,
+              sin_emb,
+              const_cast<DataType_*>(reinterpret_cast<const DataType_*>(cache_k_scale.get().data<T>())),
+              const_cast<DataType_*>(reinterpret_cast<const DataType_*>((cache_v_scale.get().data<T>()))),
+              nullptr,
+              nullptr,
+              max_seq_len,
+              max_blocks_per_seq,
+              num_heads,
+              block_size,
+              127.0f,
+              -127.0f,
+              kv_num_heads,
+              rope_3d,
+              rms_norm_eps);
+    } 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>(),
+          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]");
+    }
   }
 }
 
@@ -650,7 +834,6 @@ template void DecoderWriteCacheWithRoPEKernel<paddle::bfloat16, int>(
               // kv_num_heads, head_dim] if GQA)
     const paddle::Tensor& seq_lens,
     const paddle::Tensor& seq_lens_encoder,
-    const paddle::Tensor& batch_id_per_token,
     const paddle::Tensor& cu_seqlens_q,
     const paddle::Tensor& block_tables,
     const paddle::optional<paddle::Tensor>& rotary_embs,
@@ -667,7 +850,10 @@ 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>(
@@ -677,7 +863,6 @@ DecoderWriteCacheWithRoPEKernel<paddle::bfloat16, paddle::bfloat16>(
               // kv_num_heads, head_dim] if GQA)
     const paddle::Tensor& seq_lens,
     const paddle::Tensor& seq_lens_encoder,
-    const paddle::Tensor& batch_id_per_token,
     const paddle::Tensor& cu_seqlens_q,
     const paddle::Tensor& block_tables,
     const paddle::optional<paddle::Tensor>& rotary_embs,
@@ -694,7 +879,10 @@ 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,
@@ -703,7 +891,6 @@ template void DecoderWriteCacheWithRoPEKernel<paddle::float16, int>(
               // kv_num_heads, head_dim] if GQA)
     const paddle::Tensor& seq_lens,
     const paddle::Tensor& seq_lens_encoder,
-    const paddle::Tensor& batch_id_per_token,
     const paddle::Tensor& cu_seqlens_q,
     const paddle::Tensor& block_tables,
     const paddle::optional<paddle::Tensor>& rotary_embs,
@@ -720,7 +907,10 @@ 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,
@@ -729,7 +919,6 @@ template void DecoderWriteCacheWithRoPEKernel<paddle::float16, paddle::float16>(
               // kv_num_heads, head_dim] if GQA)
     const paddle::Tensor& seq_lens,
     const paddle::Tensor& seq_lens_encoder,
-    const paddle::Tensor& batch_id_per_token,
     const paddle::Tensor& cu_seqlens_q,
     const paddle::Tensor& block_tables,
     const paddle::optional<paddle::Tensor>& rotary_embs,
@@ -746,4 +935,7 @@ 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

@@ -23,7 +23,6 @@ void DecoderWriteCacheWithRoPEKernel(
               // kv_num_heads, head_dim] if GQA)
     const paddle::Tensor& seq_lens,
     const paddle::Tensor& seq_lens_encoder,
-    const paddle::Tensor& batch_id_per_token,
     const paddle::Tensor& cu_seqlens_q,
     const paddle::Tensor& block_tables,
     const paddle::optional<paddle::Tensor>& rotary_embs,
@@ -40,4 +39,6 @@ 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_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") {
@@ -134,7 +178,7 @@ void EncoderWriteCacheWithRopeKernel(
                                     stream,
                                     key_cache_out,
                                     value_cache_out);
-  } else if (cache_quant_type_str == "cache_int8" or cache_quant_type_str == "cache_fp8") {
+  } else if (cache_quant_type_str == "cache_int8" or cache_quant_type_str == "cache_fp8" or cache_quant_type_str == "block_wise_fp8") {
     DISPATCH_HEAD_DIM(
         head_dim, HEAD_DIM, {DISPATCH_BLOCK_SIZE(block_size, BLOCK_SIZE, {
           CascadeAppendWriteCacheKVC8QKV<T, HEAD_DIM, BLOCK_SIZE>(
@@ -154,7 +198,7 @@ void EncoderWriteCacheWithRopeKernel(
               num_blocks,
               max_seq_len,
               is_scale_channel_wise,
-              cache_quant_type_str == "cache_fp8",
+              cache_quant_type_str,
               stream,
               key_cache_out,
               value_cache_out);

custom_ops/gpu_ops/append_attn/get_block_shape_and_split_kv_block.cu

@@ -11,10 +11,11 @@
 // 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 "cute/tensor.hpp"
 #include "helper.h"
 #include "paddle/extension.h"
 #include "paddle/phi/core/memory/memcpy.h"
+#include "utils.cuh"
 
 template <int THREADBLOCK_SIZE>
 __global__ void
@@ -116,6 +117,93 @@ void GetMaxLen(const paddle::Tensor &seq_lens_tensor,
       max_len_tensor.data<int>(), batch_size);
 }
 
+template <uint32_t config_size>
+__global__ void search_chunk_size_for_mla(
+    const int *__restrict__ seq_lens_q,
+    const int *__restrict__ seq_lens_encoder,
+    const int *__restrict__ seq_lens_decoder,
+    int *__restrict__ num_blocks_x,
+    int *__restrict__ res_chunk_size,
+    const int bsz,
+    const int set_chunk_size,
+    const int block_size,
+    const int sm_cout) {
+  const uint32_t conf_id = threadIdx.x;
+  int gridx = 0;
+  if (set_chunk_size > 0 && conf_id == 0) {
+    for (uint32_t bid = 0; bid < bsz; bid++) {
+      int seq_len = seq_lens_q[bid];
+      int seq_len_encoder = seq_lens_encoder[bid];
+      int seq_len_decoder = seq_lens_decoder[bid] + seq_len;
+      if (seq_len == 0 || seq_len_encoder > 0) continue;
+
+      int loop_times;
+      loop_times = cute::ceil_div(seq_len_decoder, set_chunk_size);
+      gridx += loop_times;
+    }
+    *num_blocks_x = gridx;
+    *res_chunk_size = set_chunk_size;
+  } else if (conf_id < config_size) {
+    __shared__ int gridx_shared[config_size];
+    // chunk_size is a multiple of 64
+    const int chunk_size = block_size << conf_id;
+    for (uint32_t bid = 0; bid < bsz; bid++) {
+      int seq_len = seq_lens_q[bid];
+      int seq_len_encoder = seq_lens_encoder[bid];
+      int seq_len_decoder = seq_lens_decoder[bid] + seq_len;
+      if (seq_len == 0 || seq_len_encoder > 0) continue;
+
+      int loop_times;
+      loop_times = cute::ceil_div(seq_len_decoder, chunk_size);
+      gridx += loop_times;
+    }
+    gridx_shared[conf_id] = gridx;
+    __syncthreads();
+    if (threadIdx.x == 0) {
+      uint32_t res_id = 0;
+      uint32_t max_last_wave_block = 0;
+      for (uint32_t i = 1; i < config_size; i++) {
+          uint32_t last_wave_block = gridx_shared[i] % sm_cout;
+          if (last_wave_block >= max_last_wave_block) {
+              res_id = i;
+              max_last_wave_block = last_wave_block;
+          }
+      }
+      *num_blocks_x = gridx_shared[res_id];
+      *res_chunk_size = block_size << res_id;
+    }
+  }
+}
+
+__global__ void split_block_for_mla(const int *__restrict__ seq_lens_q,
+                                    const int *__restrict__ seq_lens_encoder,
+                                    const int *__restrict__ seq_lens_decoder,
+                                    int *__restrict__ batch_ids,
+                                    int *__restrict__ tile_ids_per_batch,
+                                    const int bsz,
+                                    const int chunk_size) {
+  if (threadIdx.x == 0) {
+    int index = 0;
+    for (uint32_t bid = 0; bid < bsz; bid++) {
+      int seq_len = seq_lens_q[bid];
+      int seq_len_encoder = seq_lens_encoder[bid];
+      int seq_len_decoder = seq_lens_decoder[bid] + seq_len;
+
+      if (seq_len == 0) continue;
+
+      int loop_times;
+      loop_times = cute::ceil_div(seq_len_decoder, chunk_size);
+      if (seq_len_encoder > 0) {
+          loop_times = 0;
+      }
+      for (uint32_t tile_id = 0; tile_id < loop_times; tile_id++) {
+          batch_ids[index] = bid;
+          tile_ids_per_batch[index++] = tile_id;
+      }
+    }
+  }
+}
+
 __global__ void split_q_block(const int *__restrict__ seq_lens_q,
                               const int *__restrict__ seq_lens_encoder,
                               int *__restrict__ batch_ids,
@@ -191,14 +279,23 @@ get_max_len_kv_ernel(int *max_seq_lens_out, const int *seq_lens_this_time,
   }
 }
 
-std::vector<paddle::Tensor> GetBlockShapeAndSplitKVBlock(
+void GetBlockShapeAndSplitKVBlock(
     const paddle::Tensor &seq_lens_encoder,
     const paddle::Tensor &seq_lens_decoder,
     const paddle::Tensor &seq_lens_this_time,
     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
+    paddle::Tensor &decoder_num_blocks_cpu,     // Inplace, Pinned Memory
+    paddle::Tensor &decoder_num_blocks_device,  // Inplace
+    paddle::Tensor &decoder_chunk_size_device,  // Inplace
+    paddle::Tensor &max_len_tensor_cpu,         // Inplace, CPU
+    paddle::Tensor &encoder_batch_ids,          // Inplace
+    paddle::Tensor &encoder_tile_ids_per_batch, // Inplace
+    paddle::Tensor &encoder_num_blocks_x_cpu,   // Inplace, CPU
+    paddle::Tensor &kv_batch_ids,               // Inplace
+    paddle::Tensor &kv_tile_ids_per_batch,      // Inplace
+    paddle::Tensor &kv_num_blocks_x_cpu,        // Inplace, CPU
+    paddle::Tensor &max_len_kv_cpu,             // Inplace, CPU
     const int encoder_block_shape_q,
     const int decoder_block_shape_q,
     const int group_size,
@@ -223,31 +320,120 @@ std::vector<paddle::Tensor> GetBlockShapeAndSplitKVBlock(
   int max_system_len = max_len_cpu_ptr[6];
   int max_just_dec_len_without_system = max_len_cpu_ptr[7];
 
-  paddle::Tensor encoder_batch_ids;
-  paddle::Tensor encoder_tile_ids_per_batch;
-  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 max_len_kv_cpu;            /*cpu*/
-
   auto max_len_kv =
       GetEmptyTensor({1}, paddle::DataType::INT32, seq_lens_decoder.place());
   get_max_len_kv_ernel<128><<<1, 128, 0, stream>>>(
       max_len_kv.data<int>(), seq_lens_this_time.data<int>(),
       seq_lens_decoder.data<int>(), bsz);
 
-  max_len_kv_cpu = max_len_kv.copy_to(paddle::CPUPlace(), false);
 
+  max_len_kv_cpu.copy_(max_len_kv, max_len_kv_cpu.place(), false);
+
+  // decoder
+  if (max_dec_len_this_time > 0) {
+    const bool mla_use_tensorcore = GetMlaUseTensorcore();
+    if (mla_use_tensorcore && group_size <= 64) {
+      const int set_chunk_size = get_mla_dec_chunk_size(bsz);
+
+      PADDLE_ENFORCE_GPU_SUCCESS(cudaMemsetAsync(
+          decoder_chunk_size_device.data<int>(), 64, sizeof(int32_t), stream));
+
+      PADDLE_ENFORCE_GPU_SUCCESS(cudaMemsetAsync(
+          decoder_num_blocks_device.data<int>(), 0, sizeof(int32_t), stream));
+
+      int device;
+      cudaGetDevice(&device);
+      int sm_cout;
+      cudaDeviceGetAttribute(&sm_cout, cudaDevAttrMultiProcessorCount, device);
+      constexpr int config_size =
+          12;  // search space for chunk size:[64, 128, 256, ... 131072]
+
+      search_chunk_size_for_mla<config_size>
+          <<<1, 32, 0, stream>>>(seq_lens_this_time.data<int>(),
+                                 seq_lens_encoder.data<int>(),
+                                 seq_lens_decoder.data<int>(),
+                                 decoder_num_blocks_device.data<int>(),
+                                 decoder_chunk_size_device.data<int>(),
+                                 bsz,
+                                 set_chunk_size,
+                                 block_size,
+                                 sm_cout);
+
+      decoder_num_blocks_cpu.copy_(
+          decoder_num_blocks_device, decoder_num_blocks_cpu.place(), false);
+      auto decoder_chunk_size_cpu =
+          decoder_chunk_size_device.copy_to(paddle::CPUPlace(), false);
+      const int chunk_size = decoder_chunk_size_cpu.data<int>()[0];
+
+      //  NOTE: (changwenbin) When using auto_chunk,
+      // decode_max_tile_size must take into account the maximum case, where * 1024 can cover 128K.
+      // const uint32_t decoder_batch_shape = seq_lens_decoder.dims()[0] * 1024;
+
+      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 * 1024 * 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));
+
+
+      split_block_for_mla<<<1, 32, 0, stream>>>(
+          seq_lens_this_time.data<int>(),
+          seq_lens_encoder.data<int>(),
+          seq_lens_decoder.data<int>(),
+          decoder_batch_ids.data<int>(),
+          decoder_tile_ids_per_batch.data<int>(),
+          bsz,
+          chunk_size);
+
+    } else {
+        // Note:(changwenbin)In order to adapt to cudagraph, the maximum value should be taken here
+        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 * 1024 * 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_device.data<int>(), 0, sizeof(int32_t), stream));
+
+        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_device.data<int>(),
+            bsz,
+            decoder_block_shape_q,
+            group_size);
+
+        decoder_num_blocks_cpu.copy_(
+            decoder_num_blocks_device, decoder_num_blocks_cpu.place(), false);
+        PADDLE_ENFORCE_GPU_SUCCESS(cudaMemsetAsync(
+            decoder_chunk_size_device.data<int>(), 64, sizeof(int32_t), stream));
+    }
+  } else {
+      PADDLE_ENFORCE_GPU_SUCCESS(cudaMemsetAsync(
+          decoder_chunk_size_device.data<int>(), 64, sizeof(int32_t), stream));
+      PADDLE_ENFORCE_GPU_SUCCESS(cudaMemsetAsync(
+          decoder_num_blocks_device.data<int>(), 0, sizeof(int32_t), stream));
+      decoder_num_blocks_cpu.copy_(
+          decoder_num_blocks_device, decoder_num_blocks_cpu.place(), false);
+  }
+
+  // encoder
   if (max_enc_len_this_time > 0) {
-    const uint32_t max_tile_size_per_bs_kv =
-        div_up(max_enc_dec_len_this_time, block_size);
-    kv_batch_ids =
-        GetEmptyTensor({bsz * max_tile_size_per_bs_kv}, paddle::DataType::INT32,
-                       seq_lens_encoder.place());
-    kv_tile_ids_per_batch =
-        GetEmptyTensor({bsz * max_tile_size_per_bs_kv}, paddle::DataType::INT32,
-                       seq_lens_encoder.place());
+    const uint32_t max_tile_size_per_bs_kv = div_up(max_enc_dec_len_this_time, block_size);
+    const uint32_t kv_batch_shape = bsz * max_tile_size_per_bs_kv;
+    PADDLE_ENFORCE_GPU_SUCCESS(cudaMemsetAsync(kv_batch_ids.data<int>(), 0, kv_batch_shape * sizeof(int32_t), stream));
+    PADDLE_ENFORCE_GPU_SUCCESS(cudaMemsetAsync(kv_tile_ids_per_batch.data<int>(), 0, kv_batch_shape * sizeof(int32_t), stream));
     auto kv_num_blocks_x =
         GetEmptyTensor({1}, paddle::DataType::INT32, seq_lens_encoder.place());
 
@@ -258,16 +444,12 @@ std::vector<paddle::Tensor> GetBlockShapeAndSplitKVBlock(
         kv_tile_ids_per_batch.data<int>(), kv_num_blocks_x.data<int>(), bsz,
         block_size, block_size);
 
-    kv_num_blocks_x_cpu = kv_num_blocks_x.copy_to(paddle::CPUPlace(), false);
-
-    const uint32_t encoder_max_tile_size_per_bs_q =
-        div_up((max_enc_dec_len_this_time * group_size), encoder_block_shape_q);
-    encoder_batch_ids =
-        GetEmptyTensor({bsz * encoder_max_tile_size_per_bs_q},
-                       paddle::DataType::INT32, seq_lens_encoder.place());
-    encoder_tile_ids_per_batch =
-        GetEmptyTensor({bsz * encoder_max_tile_size_per_bs_q},
-                       paddle::DataType::INT32, seq_lens_encoder.place());
+    kv_num_blocks_x_cpu.copy_(kv_num_blocks_x, kv_num_blocks_x_cpu.place(), false);
+    // Clear buffer
+    const uint32_t encoder_max_tile_size_per_bs_q = div_up((max_enc_dec_len_this_time * group_size), encoder_block_shape_q);
+    const uint32_t encoder_batch_shape = bsz * encoder_max_tile_size_per_bs_q;
+    PADDLE_ENFORCE_GPU_SUCCESS(cudaMemsetAsync(encoder_batch_ids.data<int>(), 0, encoder_batch_shape * sizeof(int32_t), stream));
+    PADDLE_ENFORCE_GPU_SUCCESS(cudaMemsetAsync(encoder_tile_ids_per_batch.data<int>(), 0, encoder_batch_shape * sizeof(int32_t), stream));
     auto encoder_num_blocks_x =
         GetEmptyTensor({1}, paddle::DataType::INT32, seq_lens_encoder.place());
     split_q_block<<<1, 32, 0, stream>>>(seq_lens_encoder.data<int>(), nullptr,
@@ -275,54 +457,9 @@ std::vector<paddle::Tensor> GetBlockShapeAndSplitKVBlock(
                                         encoder_tile_ids_per_batch.data<int>(),
                                         encoder_num_blocks_x.data<int>(), bsz,
                                         encoder_block_shape_q, group_size);
-    encoder_num_blocks_x_cpu =
-        encoder_num_blocks_x.copy_to(paddle::CPUPlace(), false);
-  } else {
-    encoder_batch_ids =
-        GetEmptyTensor({0}, paddle::DataType::INT32, seq_lens_encoder.place());
-    encoder_tile_ids_per_batch =
-        GetEmptyTensor({0}, paddle::DataType::INT32, seq_lens_encoder.place());
-    encoder_num_blocks_x_cpu =
-        GetEmptyTensor({0}, paddle::DataType::INT32, paddle::CPUPlace());
-    kv_batch_ids =
-        GetEmptyTensor({0}, paddle::DataType::INT32, seq_lens_encoder.place());
-    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());
+    encoder_num_blocks_x_cpu.copy_(encoder_num_blocks_x, encoder_num_blocks_x_cpu.place(), false);
   }
 
-  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);
-  }
-
-  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*/
-    max_len_kv_cpu,           /*cpu*/
-  };
 }
 
 PD_BUILD_STATIC_OP(get_block_shape_and_split_kv_block)
@@ -332,17 +469,20 @@ PD_BUILD_STATIC_OP(get_block_shape_and_split_kv_block)
       "seq_lens_this_time",
       "decoder_batch_ids",
       "decoder_tile_ids_per_batch",
-      "decoder_num_blocks_x_cpu",
-      "max_len_tensor_cpu"
+      "decoder_num_blocks_cpu",
+      "decoder_num_blocks_device",
+      "decoder_chunk_size_device",
+      "max_len_tensor_cpu",
+      "encoder_batch_ids",
+      "encoder_tile_ids_per_batch",
+      "encoder_num_blocks_x_cpu",
+      "kv_batch_ids",
+      "kv_tile_ids_per_batch",
+      "kv_num_blocks_x_cpu",
+      "max_len_kv_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",

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,
@@ -213,7 +217,7 @@ __global__ void append_cache_kv_c16(
 
   // load k_smem 64 rows 128 cols
   for (int fz = 0; fz < 4; fz++) { // 4 rows pre warp once, 16 rows all 4 warps once, need 4 iter
-    for (int fy = 0; fy < 2; fy++) { // 8 * 128b = 64 * bf16 noce, need 2 iter
+    for (int fy = 0; fy < 2; fy++) { // 8 * 128b = 64 * bf16 once, need 2 iter
       k_smem.load_128b_async<SharedMemFillMode::kNoFill>(
             k_smem_offset_w, cur_cache_k + k_read_idx, end_idx > 0);
       k_smem_offset_w =
@@ -231,7 +235,7 @@ __global__ void append_cache_kv_c16(
   // deal k_smem 64 rows 128 cols
   for (int fz = 0; fz < 1; fz++) { // 16 rows pre warp once, 64 rows all 4 warps once, need 1 iter
     uint32_t row_idx = wid * 16 + tid / 4;
-    for (int fy = 0; fy < 8; fy++) { // 2 * 128b = 16 * bf16 noce, need 8 iter
+    for (int fy = 0; fy < 8; fy++) { // 2 * 128b = 16 * bf16 once, need 8 iter
       uint32_t col_idx = fy * 16 + tid % 4 * 2;
       k_smem.ldmatrix_m8n8x4(k_smem_offset_r, kv_frag);
       // layout
@@ -274,7 +278,7 @@ __global__ void append_cache_kv_c16(
 
   // load v_smem 64 rows 128 cols
   for (int fz = 0; fz < 4; fz++) { // // 4 rows pre warp once, 16 rows all 4 warps once, need 4 iter
-    for (int fy = 0; fy < 2; fy++) { // 8 * 128b = 64 * bf16 noce, need 2 iter
+    for (int fy = 0; fy < 2; fy++) { // 8 * 128b = 64 * bf16 once, need 2 iter
       v_smem.load_128b_async<SharedMemFillMode::kNoFill>(
             v_smem_offset_w, cur_cache_v + v_read_idx, end_idx > 0);
       v_smem_offset_w =
@@ -292,7 +296,7 @@ __global__ void append_cache_kv_c16(
   // deal v_smem 64 rows 128 cols
   for (int fz = 0; fz < 1; fz++) { //  16 rows pre warp once, 64 rows all 4 warps once, need 1 iter
     uint32_t row_idx = wid * 16 + tid / 4;
-    for (int fy = 0; fy < 8; fy++) { // 2 * 128b = 16 * bf16 noce, need 8 iter
+    for (int fy = 0; fy < 8; fy++) { // 2 * 128b = 16 * bf16 once, need 8 iter
       uint32_t col_idx = fy * 16 + tid % 4 * 2;
       v_smem.ldmatrix_m8n8x4(v_smem_offset_r, kv_frag);
       // layout
@@ -396,7 +400,7 @@ __global__ void append_cache_kv_c8(
 
   // load v_smem 64 rows, 128 cols
   for (int fz = 0; fz < 4; fz++) { // 4 rows pre warp once, 16 rows all 4 warps once, need 4 iter
-    for (int fy = 0; fy < 1; fy++) { // 8 * 128b = 128 * uint8 noce, need 1 iter
+    for (int fy = 0; fy < 1; fy++) { // 8 * 128b = 128 * uint8 once, need 1 iter
       k_smem.load_128b_async<SharedMemFillMode::kNoFill>(
             k_smem_offset_w, cur_cache_k + k_read_idx, end_idx > 0);
       k_smem_offset_w =
@@ -414,7 +418,7 @@ __global__ void append_cache_kv_c8(
   // deal k_smem 64 rows, 128 cols
   for (int fz = 0; fz < 1; fz++) { // 16 rows pre warp once, 64 rows all 4 warps once, need 1 iter
     uint32_t row_idx = wid * 16 + tid / 4;
-    for (int fy = 0; fy < 4; fy++) { // 2 * 128b = 32 * uint8 noce, need 4 iter
+    for (int fy = 0; fy < 4; fy++) { // 2 * 128b = 32 * uint8 once, need 4 iter
       uint32_t col_idx = fy * 32 + tid % 4 * 2;
       k_smem.ldmatrix_m8n8x4(k_smem_offset_r, k_frag);
       // layout
@@ -462,7 +466,7 @@ __global__ void append_cache_kv_c8(
                           tid % 4 * num_elems_per_128b<CacheT>();
   // load v_smem 128 rows 64 cols
   for (int fy = 0; fy < 4; fy++) { // 8 rows pre warp once, 32 rows all 4 warps once, need 4 iter
-    for (int fz = 0; fz < 1; fz++) { // 4 * 128b = 64 * uint8 noce, need 1 iter
+    for (int fz = 0; fz < 1; fz++) { // 4 * 128b = 64 * uint8 once, need 1 iter
       v_smem.load_128b_async<SharedMemFillMode::kNoFill>(
               v_smem_offset_w, cur_cache_v + v_read_idx, end_idx > 0);
       v_smem_offset_w =
@@ -481,7 +485,7 @@ __global__ void append_cache_kv_c8(
   // deal v_smem 128 rows 64 cols
   for (int fy = 0; fy < 2; fy++) { // 16 rows pre warp once, 64 rows all 4 warps once, need 2 iter
     uint32_t dim_idx = fy * NUM_WARPS * 16 + wid * 16 + tid / 4;
-    for (int fz = 0; fz < 2; fz++) { // 2 * 128b = 32 * uint8 noce, need 2 iter
+    for (int fz = 0; fz < 2; fz++) { // 2 * 128b = 32 * uint8 once, need 2 iter
       uint32_t kv_idx = fz * 32 + tid % 4 * 2;
       v_smem.ldmatrix_m8n8x4(v_smem_offset_r, v_frag);
       // layout
@@ -610,7 +614,7 @@ __global__ void append_cache_kv_c4(
 
   // load k_smem 64 rows 128 cols
   for (int fz = 0; fz < 2; fz++) { // 4 rows pre warp once, 16 rows all 4 warps once, need 4 iter
-    for (int fy = 0; fy < 1; fy++) { // 4 * 128b = 128 * int4 noce, need 1 iter
+    for (int fy = 0; fy < 1; fy++) { // 4 * 128b = 128 * int4 once, need 1 iter
       k_smem.load_128b_async<SharedMemFillMode::kNoFill>(
             k_smem_offset_w, cur_cache_k + k_read_idx, end_idx > 0);
       k_smem_offset_w =
@@ -628,7 +632,7 @@ __global__ void append_cache_kv_c4(
   // deal k_smem 64 rows 128 cols
   for (int fz = 0; fz < 1; fz++) { // 16 rows pre warp once, 64 rows all 4 warps once, need 1 iter
     uint32_t row_idx = wid * 16 + tid / 4;
-    for (int fy = 0; fy < 2; fy++) { // 2 * 128b = 64 * int4 noce, need 2 iter
+    for (int fy = 0; fy < 2; fy++) { // 2 * 128b = 64 * int4 once, need 2 iter
       uint32_t col_idx = fy * 64 + tid % 4 * 2;
       k_smem.ldmatrix_m8n8x4(k_smem_offset_r, k_frag);
 
@@ -681,7 +685,7 @@ __global__ void append_cache_kv_c4(
                           tid % 2 * num_elems_per_128b<CacheT>();
   // load v_smem 128 rows 64 rows
   for (int fy = 0; fy < 2; fy++) { // 16 rows pre warp once, 64 rows all 4 warps once, need 2 iter
-    for (int fz = 0; fz < 1; fz++) { // 2 * 128b = 64 * int4 noce, need 1 iter
+    for (int fz = 0; fz < 1; fz++) { // 2 * 128b = 64 * int4 once, need 1 iter
       v_smem.load_128b_async<SharedMemFillMode::kNoFill>(
               v_smem_offset_w, cur_cache_v + v_read_idx, end_idx > 0);
       v_smem_offset_w =
@@ -700,7 +704,7 @@ __global__ void append_cache_kv_c4(
   // deal v_smem 128 rows 64 cols
   for (int fy = 0; fy < 2; fy++) { // 16 rows pre warp once, 64 rows all 4 warps once, need 2 iter
     uint32_t dim_idx = fy * NUM_WARPS * 16 + wid * 16 + tid / 4;
-    for (int fz = 0; fz < 1; fz++) { // 2 * 128b = 64 * int4 noce, need 1 iter
+    for (int fz = 0; fz < 1; fz++) { // 2 * 128b = 64 * int4 once, need 1 iter
       uint32_t kv_idx = fz * 64 + tid % 4 * 2;
       v_smem.ldmatrix_m8n8x4(v_smem_offset_r, v_frag);
       // layout
@@ -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;
@@ -953,8 +958,9 @@ 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) {
@@ -994,7 +1000,7 @@ std::vector<paddle::Tensor> GQARopeWriteCacheKernel(
       stream,
       const_cast<paddle::Tensor*>(&key_cache),
       const_cast<paddle::Tensor*>(&value_cache));
-  } else if (cache_quant_type == "cache_int8" || cache_quant_type == "cache_fp8") {
+  } else if (cache_quant_type == "cache_int8" || cache_quant_type == "cache_fp8" || cache_quant_type == "block_wise_fp8") {
     CascadeAppendWriteCacheKVC8QKV<data_t, 128, 64>(
         meta_data,
         *const_cast<paddle::Tensor*>(&key_cache),
@@ -1012,7 +1018,7 @@ std::vector<paddle::Tensor> GQARopeWriteCacheKernel(
         kv_num_blocks_data,
         max_seq_len,
         false, // is_scale_channel_wise
-        cache_quant_type == "cache_fp8", // is_fp8
+        cache_quant_type,
         stream,
         const_cast<paddle::Tensor*>(&key_cache),
         const_cast<paddle::Tensor*>(&value_cache));

custom_ops/gpu_ops/append_attn/speculate_write_cache_with_rope_impl.cuh

@@ -18,6 +18,168 @@
 #include "mma_tensor_op.cuh"
 #include "utils.cuh"
 
+template <typename T, int VecSize = 1, typename InT = T>
+__global__ void append_speculate_cache_T_rope_qk_norm_kernel(
+    const InT* __restrict__ qkv,  // [token_num, num_heads + 2 * gqa_group_size,
+                                  // head_size]
+    T* __restrict__ key_cache,    // [num_blocks, gqa_group_size, block_size,
+                                  // head_size // 2]
+    T* __restrict__ value_cache,  // [num_blocks, gqa_group_size, block_size,
+                                  // head_size // 2]
+    T* __restrict__ q_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_decoder,  // [bsz]
+    const float* __restrict__ cos_emb,
+    const float* __restrict__ sin_emb,
+    const float*
+        qkv_out_scales,   // [(num_heads + 2 * gqa_group_size) * head_size]
+    const T* qkv_biases,  // [num_head + 2 * gqa_group_size, dim_head]
+    const int max_seq_len,
+    const int max_blocks_per_seq,
+    const int num_heads,
+    const int output_inner_dim,
+    const int head_size,
+    const int block_size,
+    const int elem_cnt,
+    const int gqa_group_size,
+    const float* q_norm_weight,
+    const float* k_norm_weight,
+    const float rms_norm_eps,
+    const bool rope_3d) {
+  using LoadT = AlignedVector<T, VecSize>;
+  using LoadFloat = AlignedVector<float, VecSize>;
+  using LoadInT = AlignedVector<InT, VecSize>;
+  constexpr int HalfVecSize = VecSize / 2;
+  using LoadEmbT = AlignedVector<float, HalfVecSize>;
+  LoadInT src_vec;
+  LoadFloat scale_vec;
+  LoadT bias_vec;
+  LoadEmbT cos_emb_vec;
+  LoadEmbT sin_emb_vec;
+  LoadFloat tmp_vec;
+  LoadFloat q_norm_vec;
+  LoadFloat 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 * gqa_group_size) * head_size;
+  const int half_head_size = head_size / 2;
+  for (int global_hi = global_warp_idx; global_hi < all_head_dim; global_hi += all_warp_num) {
+    int64_t linear_index = global_hi * head_size + threadIdx.x * VecSize;
+    const int token_id = linear_index / hidden_size;
+    const int ori_bi = batch_id_per_token[token_id];
+    if (seq_lens_decoder[ori_bi] == 0) continue;
+    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];
+    const int write_seq_id =
+        seq_lens_decoder[ori_bi] + token_id - start_token_idx;
+    if (write_seq_id == 0) continue;
+
+    const int* block_table_now = block_tables + ori_bi * max_blocks_per_seq;
+    const int block_idx = block_table_now[write_seq_id / block_size];
+    if (block_idx < 0) {
+      printf(
+          "Fatal Error!!!, block idx %d when write_seq_id is %d\n some key var "
+          "%d %d %d %d\n",
+          block_idx,
+          write_seq_id,
+          ori_bi,
+          seq_lens_decoder[ori_bi],
+          token_id,
+          cu_seqlens_q[ori_bi]);
+    }
+    const int block_offset = write_seq_id % block_size;
+
+    const int write_q_idx =
+        token_id * output_inner_dim * head_size + hi * head_size + h_bias;
+
+    const int bias_idx = hi * head_size + h_bias;
+    Load<InT, VecSize>(&qkv[linear_index], &src_vec);
+    if (qkv_biases) {
+      Load<T, VecSize>(&qkv_biases[bias_idx], &bias_vec);
+    }
+    if (qkv_out_scales) {
+      Load<float, VecSize>(&qkv_out_scales[bias_idx], &scale_vec);
+    }
+    if (hi < num_heads + gqa_group_size) {
+      // q k rope
+      const int64_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++) {
+      // 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 (qkv_out_scales) {
+        input_left *= scale_vec[2 * i];
+        input_right *= scale_vec[2 * i + 1];
+      }
+      if (qkv_biases) {
+        input_left = input_left + static_cast<float>(bias_vec[2 * i]);
+        input_right = input_right + static_cast<float>(bias_vec[2 * i + 1]);
+      }
+      if (hi < num_heads + gqa_group_size) {
+        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 {
+        bias_vec[2 * i] = static_cast<T>(input_left);
+        bias_vec[2 * i + 1] = static_cast<T>(input_right);
+      }
+    }
+    if (hi < (num_heads + gqa_group_size)) {
+      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) {
+        Load<float, VecSize>(&q_norm_weight[threadIdx.x * VecSize], &q_norm_vec);
+        #pragma unroll
+        for (int i = 0; i < VecSize; i++) {
+          bias_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);
+        #pragma unroll
+        for (int i = 0; i < VecSize; i++) {
+          bias_vec[i] = static_cast<T>(tmp_vec[i] * row_inv_var * k_norm_vec[i]);
+        }
+      }
+    }
+    if (hi < num_heads) {
+      // write q
+      Store<T, VecSize>(bias_vec, &q_out[write_q_idx]);
+    } else {
+      //  write k/v
+      const int kv_head_idx = (hi - num_heads) % gqa_group_size;
+      const int tgt_idx = (block_idx * gqa_group_size * block_size * head_size +
+                           kv_head_idx * block_size * head_size +
+                           block_offset * head_size + h_bias);
+      // write
+      if (hi < num_heads + gqa_group_size) {
+        Store<T, VecSize>(bias_vec, &key_cache[tgt_idx]);
+      } else {
+        Store<T, VecSize>(bias_vec, &value_cache[tgt_idx]);
+      }
+    }
+  }
+}
+
 template <int VecSize = 4, int HeadDim = 128>
 __global__ void append_clear_cache_int8_block(
     uint8_t* __restrict__ key_cache,    // [num_blocks, gqa_group_size,
@@ -193,7 +355,8 @@ __global__ void append_speculate_cache_rope_kernel(
     const int head_size,
     const int block_size,
     const int elem_cnt,
-    const int gqa_group_size) {
+    const int gqa_group_size,
+    const bool rope_3d) {
   using LoadT = AlignedVector<T, VecSize>;
   using LoadFloat = AlignedVector<float, VecSize>;
   using LoadInT = AlignedVector<InT, VecSize>;
@@ -253,8 +416,9 @@ __global__ void append_speculate_cache_rope_kernel(
     if (hi < num_heads + gqa_group_size) {
       // q k rope
       const int64_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);
+      int64_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++) {
@@ -326,7 +490,8 @@ __global__ void append_speculate_cache_neox_rope_kernel(
     const int head_size,
     const int block_size,
     const int elem_cnt,
-    const int gqa_group_size) {
+    const int gqa_group_size,
+    const bool rope_3d) {
   using LoadT = AlignedVector<T, VecSize>;
   using LoadFloat = AlignedVector<float, VecSize>;
   using LoadInT = AlignedVector<InT, VecSize>;
@@ -390,8 +555,9 @@ __global__ void append_speculate_cache_neox_rope_kernel(
     if (hi < num_heads + gqa_group_size) {
       // q k rope
       const int64_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);
+      int64_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++) {
@@ -476,7 +642,8 @@ __global__ void append_speculate_cache_int8_rope_kernel(
     const int block_size,
     const float max_bound,
     const float min_bound,
-    const int gqa_group_size) {
+    const int gqa_group_size,
+    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;
@@ -522,8 +689,9 @@ __global__ void append_speculate_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);
       if (qkv_out_scales) {
         Load<float, VecSize>(&qkv_out_scales[bias_idx], &out_scale_vec);
       }
@@ -583,10 +751,11 @@ __global__ void append_speculate_cache_int8_rope_kernel(
     T scale;
     if (head_idx < num_heads + gqa_group_size) {
       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);
       scale = __ldg(&cache_k_scales[kv_head_idx]);
     } else {
       scale = __ldg(&cache_v_scales[kv_head_idx]);
@@ -708,7 +877,8 @@ __global__ void append_speculate_cache_int8_neox_rope_kernel(
     const int block_size,
     const float max_bound,
     const float min_bound,
-    const int gqa_group_size) {
+    const int gqa_group_size,
+    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;
@@ -757,8 +927,9 @@ __global__ void append_speculate_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);
       if (qkv_out_scales) {
         Load<float, VecSize>(&qkv_out_scales[bias_idx_left],
                              &left_out_scale_vec);
@@ -853,10 +1024,11 @@ __global__ void append_speculate_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++) {
@@ -1088,7 +1260,8 @@ __global__ void append_speculate_cache_int4_rope_kernel(
     const int block_size,
     const float max_bound,
     const float min_bound,
-    const int gqa_group_size) {
+    const int gqa_group_size,
+    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;
@@ -1145,8 +1318,9 @@ __global__ void append_speculate_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
@@ -1235,10 +1409,11 @@ __global__ void append_speculate_cache_int4_rope_kernel(
     //                              &out_scale_vec2);
     if (head_idx < num_heads + gqa_group_size) {
       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_scales[cache_idx], &scale_vec1);
       Load<T, HALF_K_VEC_SIZE>(&cache_k_scales[cache_idx + 8], &scale_vec2);
       Load<T, HALF_K_VEC_SIZE>(&cache_k_zero_points[cache_idx], &zp_vec1);
@@ -1431,7 +1606,8 @@ __global__ void append_speculate_cache_int4_neox_rope_kernel(
     const int block_size,
     const float max_bound,
     const float min_bound,
-    const int gqa_group_size) {
+    const int gqa_group_size,
+    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;
@@ -1581,10 +1757,11 @@ __global__ void append_speculate_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 : 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_scales[left_cache_idx],
                                  &left_scale_vec1);
         Load<T, HALF_K_VEC_SIZE>(&cache_k_scales[left_cache_idx + 8],

custom_ops/gpu_ops/append_attn/speculate_write_cache_with_rope_kernel.cu

@@ -15,6 +15,78 @@
 #include "speculate_write_cache_with_rope_kernel.h"
 #include "utils.cuh"
 
+template <typename T, typename QKV_TYPE>
+void append_speculate_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 int token_num,
+                              const cudaStream_t& stream,
+                              const bool use_neox_style,
+                              const float* q_norm_weight,
+                              const float* k_norm_weight,
+                              const float rms_norm_eps,
+                              const bool rope_3d) {
+  int output_inner_dim = num_heads + 2 * kv_num_heads;
+  const uint32_t elem_nums =
+      use_neox_style ? token_num * (num_heads + 2 * kv_num_heads) * dim_head / 2
+                     : token_num * (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);
+  if (use_neox_style) {
+      PD_THROW(
+          "append_speculate_cache_rope_qk_norm not support neox rope yet");
+  } else {
+    dim3 block_dim(kWarpSize, blocksize / kWarpSize, 1);
+    append_speculate_cache_T_rope_qk_norm_kernel<T, PackSize>
+      <<<grid_size, block_dim, 0, stream>>>(qkv,
+                                            key_cache,
+                                            value_cache,
+                                            qkv_out,
+                                            block_tables,
+                                            batch_id_per_token,
+                                            cu_seqlens_q,
+                                            seq_lens,
+                                            cos_emb,
+                                            sin_emb,
+                                            qkv_out_scales,
+                                            qkv_biases,
+                                            max_seq_len,
+                                            max_blocks_per_seq,
+                                            num_heads,
+                                            output_inner_dim,
+                                            dim_head,
+                                            block_size,
+                                            elem_nums,
+                                            kv_num_heads,
+                                            q_norm_weight,
+                                            k_norm_weight,
+                                            rms_norm_eps,
+                                            rope_3d);
+  }
+}
+
 // rope + write
 template <typename T, typename QKV_TYPE>
 void append_speculate_cache_rope(const QKV_TYPE* qkv,
@@ -39,7 +111,8 @@ void append_speculate_cache_rope(const QKV_TYPE* qkv,
                                  const int bsz,
                                  const int token_num,
                                  const cudaStream_t& stream,
-                                 const bool use_neox_style) {
+                                 const bool use_neox_style,
+                                 const bool rope_3d) {
   int output_inner_dim = num_heads + 2 * kv_num_heads;
 
   const uint32_t elem_nums =
@@ -73,7 +146,8 @@ void append_speculate_cache_rope(const QKV_TYPE* qkv,
             dim_head,
             block_size,
             elem_nums,
-            kv_num_heads);
+            kv_num_heads,
+            rope_3d);
   } else {
     append_speculate_cache_rope_kernel<T, PackSize>
         <<<grid_size, threads_per_block, 0, stream>>>(
@@ -96,7 +170,8 @@ void append_speculate_cache_rope(const QKV_TYPE* qkv,
             dim_head,
             block_size,
             elem_nums,
-            kv_num_heads);
+            kv_num_heads,
+            rope_3d);
   }
 }
 
@@ -125,7 +200,8 @@ void append_speculate_cache_int8_rope(const QKV_TYPE* qkv,
                                       const int bsz,
                                       const int token_num,
                                       const cudaStream_t& stream,
-                                      const bool use_neox_style) {
+                                      const bool use_neox_style,
+                                      const bool rope_3d) {
   constexpr int num_warps = 4;
   const int all_warps =
       ((num_heads + 2 * kv_num_heads) + num_warps - 1) / num_warps * num_warps;
@@ -167,7 +243,8 @@ void append_speculate_cache_int8_rope(const QKV_TYPE* qkv,
                                                block_size,
                                                127.0f,
                                                -127.0f,
-                                               kv_num_heads);
+                                               kv_num_heads,
+                                               rope_3d);
   } else {
     append_speculate_cache_int8_rope_kernel<T, 4, 0, 128, QKV_TYPE, IsFP8>
         <<<grids, num_warps * 32, 0, stream>>>(qkv,
@@ -191,7 +268,8 @@ void append_speculate_cache_int8_rope(const QKV_TYPE* qkv,
                                                block_size,
                                                127.0f,
                                                -127.0f,
-                                               kv_num_heads);
+                                               kv_num_heads,
+                                               rope_3d);
   }
 }
 
@@ -222,7 +300,8 @@ void append_speculate_cache_int4_rope(const QKV_TYPE* qkv,
                                       const int bsz,
                                       const int token_num,
                                       const cudaStream_t& stream,
-                                      const bool use_neox_style) {
+                                      const bool use_neox_style,
+                                      const bool rope_3d) {
   constexpr int num_warps = 4;
   const int all_warps =
       ((num_heads + 2 * kv_num_heads) + num_warps - 1) / num_warps * num_warps;
@@ -266,7 +345,8 @@ void append_speculate_cache_int4_rope(const QKV_TYPE* qkv,
                                                block_size,
                                                7.0f,
                                                -8.0f,
-                                               kv_num_heads);
+                                               kv_num_heads,
+                                               rope_3d);
   } else {
     append_speculate_cache_int4_rope_kernel<T, 4>
         <<<grids, num_warps * 32, 0, stream>>>(qkv,
@@ -292,7 +372,8 @@ void append_speculate_cache_int4_rope(const QKV_TYPE* qkv,
                                                block_size,
                                                7.0f,
                                                -8.0f,
-                                               kv_num_heads);
+                                               kv_num_heads,
+                                               rope_3d);
   }
 }
 template <typename T, typename QKV_TYPE>
@@ -313,11 +394,15 @@ void SpeculateWriteCacheWithRoPEKernel(
     const paddle::optional<paddle::Tensor>& cache_v_zp,
     const std::string& cache_quant_type_str,
     const bool use_neox_rotary_style,
+    const bool rope_3d,
     const int max_seq_len,
     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_;
@@ -342,142 +427,185 @@ void SpeculateWriteCacheWithRoPEKernel(
             ? rotary_embs.get().data<float>() + max_seq_len * dim_head
             : rotary_embs.get().data<float>() + max_seq_len * dim_head / 2;
   }
-  if (cache_quant_type_str == "none") {
-    append_speculate_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,
-        token_nums,
-        stream,
-        use_neox_rotary_style);
-  } else if (cache_quant_type_str == "cache_int8") {
-    append_speculate_cache_int8_rope(
-        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,
-        token_nums,
-        stream,
-        use_neox_rotary_style);
-  } else if (cache_quant_type_str == "cache_fp8") {
-    append_speculate_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,
-        token_nums,
-        stream,
-        use_neox_rotary_style);
-  } else if (cache_quant_type_str == "cache_int4_zp") {
-    append_speculate_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,
-        token_nums,
-        stream,
-        use_neox_rotary_style);
+  if (q_norm_weight && k_norm_weight) {
+    if (cache_quant_type_str == "none") {
+      append_speculate_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,
+          token_nums,
+          stream,
+          use_neox_rotary_style,
+          reinterpret_cast<const float*>(q_norm_weight.get().data<float>()),
+          reinterpret_cast<const float*>(k_norm_weight.get().data<float>()),
+          rms_norm_eps,
+          rope_3d);
+    } else {
+      PD_THROW(
+          "append_decode_cache_rope_qk_norm not support cachekv quant yet");
+    }
+
   } else {
-    PD_THROW(
-        "cache_quant_type_str should be one of [none, cache_int8, "
-        "cache_int4_zp]");
+    if (cache_quant_type_str == "none") {
+        append_speculate_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,
+            token_nums,
+            stream,
+            use_neox_rotary_style,
+            rope_3d);
+    } else if (cache_quant_type_str == "cache_int8") {
+        append_speculate_cache_int8_rope(
+            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,
+            token_nums,
+            stream,
+            use_neox_rotary_style,
+            rope_3d);
+    } else if (cache_quant_type_str == "cache_fp8") {
+        append_speculate_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,
+            token_nums,
+            stream,
+            use_neox_rotary_style,
+            rope_3d);
+    } else if (cache_quant_type_str == "cache_int4_zp") {
+        append_speculate_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,
+            token_nums,
+            stream,
+            use_neox_rotary_style,
+            rope_3d);
+    } else {
+        PD_THROW(
+            "cache_quant_type_str should be one of [none, cache_int8, "
+            "cache_int4_zp]");
+    }
   }
 }
 
@@ -500,11 +628,15 @@ template void SpeculateWriteCacheWithRoPEKernel<paddle::bfloat16, int>(
     const paddle::optional<paddle::Tensor>& cache_v_zp,
     const std::string& cache_quant_type_str,
     const bool use_neox_rotary_style,
+    const bool rope_3d,
     const int max_seq_len,
     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
 SpeculateWriteCacheWithRoPEKernel<paddle::bfloat16, paddle::bfloat16>(
@@ -526,11 +658,15 @@ SpeculateWriteCacheWithRoPEKernel<paddle::bfloat16, paddle::bfloat16>(
     const paddle::optional<paddle::Tensor>& cache_v_zp,
     const std::string& cache_quant_type_str,
     const bool use_neox_rotary_style,
+    const bool rope_3d,
     const int max_seq_len,
     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 SpeculateWriteCacheWithRoPEKernel<paddle::float16, int>(
     const AppendAttnMetaData& meta_data,
@@ -551,11 +687,15 @@ template void SpeculateWriteCacheWithRoPEKernel<paddle::float16, int>(
     const paddle::optional<paddle::Tensor>& cache_v_zp,
     const std::string& cache_quant_type_str,
     const bool use_neox_rotary_style,
+    const bool rope_3d,
     const int max_seq_len,
     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
@@ -578,8 +718,12 @@ SpeculateWriteCacheWithRoPEKernel<paddle::float16, paddle::float16>(
     const paddle::optional<paddle::Tensor>& cache_v_zp,
     const std::string& cache_quant_type_str,
     const bool use_neox_rotary_style,
+    const bool rope_3d,
     const int max_seq_len,
     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/speculate_write_cache_with_rope_kernel.h

@@ -35,8 +35,12 @@ void SpeculateWriteCacheWithRoPEKernel(
     const paddle::optional<paddle::Tensor>& cache_v_zp,
     const std::string& cache_quant_type_str,
     const bool use_neox_rotary_style,
+    const bool rope_3d,
     const int max_seq_len,
     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/append_attention_c8_bfloat16_bfloat16_kernel.cu

@@ -56,6 +56,7 @@ CascadeAppendAttentionC8Kernel<paddle::bfloat16, paddle::bfloat16, false>(
     const bool causal,
     const bool is_decoder,
     const bool enable_prefill,
+    const std::string& cache_quant_type_str,
     cudaStream_t& stream,
     paddle::Tensor* out);
 
@@ -103,5 +104,6 @@ CascadeAppendAttentionC8Kernel<paddle::bfloat16, paddle::bfloat16, true>(
     const bool causal,
     const bool is_decoder,
     const bool enable_prefill,
+    const std::string& cache_quant_type_str,
     cudaStream_t& stream,
     paddle::Tensor* out);

custom_ops/gpu_ops/append_attn/template_instantiation/append_attention_c8_bfloat16_fp8_kernel.cu

@@ -54,6 +54,7 @@ template void CascadeAppendAttentionC8Kernel<paddle::bfloat16, paddle::float8_e4
     const bool causal,
     const bool is_decoder,
     const bool enable_prefill,
+    const std::string& cache_quant_type_str,
     cudaStream_t& stream,
     paddle::Tensor* out);
 
@@ -98,5 +99,6 @@ template void CascadeAppendAttentionC8Kernel<paddle::bfloat16, paddle::float8_e4
     const bool causal,
     const bool is_decoder,
     const bool enable_prefill,
+    const std::string& cache_quant_type_str,
     cudaStream_t& stream,
     paddle::Tensor* out);

custom_ops/gpu_ops/append_attn/template_instantiation/append_attention_c8_bfloat16_int8_kernel.cu

@@ -54,6 +54,7 @@ template void CascadeAppendAttentionC8Kernel<paddle::bfloat16, int8_t, false>(
     const bool causal,
     const bool is_decoder,
     const bool enable_prefill,
+    const std::string& cache_quant_type_str,
     cudaStream_t& stream,
     paddle::Tensor* out);
 
@@ -100,5 +101,6 @@ template void CascadeAppendAttentionC8Kernel<paddle::bfloat16, int8_t, true>(
     const bool causal,
     const bool is_decoder,
     const bool enable_prefill,
+    const std::string& cache_quant_type_str,
     cudaStream_t& stream,
     paddle::Tensor* out);

custom_ops/gpu_ops/append_attn/template_instantiation/append_attention_c8_float16_float16_kernel.cu

@@ -54,6 +54,7 @@ template void CascadeAppendAttentionC8Kernel<paddle::float16, paddle::float16, f
     const bool causal,
     const bool is_decoder,
     const bool enable_prefill,
+    const std::string& cache_quant_type_str,
     cudaStream_t& stream,
     paddle::Tensor* out);
 
@@ -100,5 +101,6 @@ template void CascadeAppendAttentionC8Kernel<paddle::float16, paddle::float16, t
     const bool causal,
     const bool is_decoder,
     const bool enable_prefill,
+    const std::string& cache_quant_type_str,
     cudaStream_t& stream,
     paddle::Tensor* out);

custom_ops/gpu_ops/append_attn/template_instantiation/append_attention_c8_float16_fp8_kerne.cu

@@ -54,6 +54,7 @@ template void CascadeAppendAttentionC8Kernel<paddle::float16, paddle::float8_e4m
     const bool causal,
     const bool is_decoder,
     const bool enable_prefill,
+    const std::string& cache_quant_type_str,
     cudaStream_t& stream,
     paddle::Tensor* out);
 
@@ -99,5 +100,6 @@ template void CascadeAppendAttentionC8Kernel<paddle::float16, paddle::float8_e4m
     const bool causal,
     const bool is_decoder,
     const bool enable_prefill,
+    const std::string& cache_quant_type_str,
     cudaStream_t& stream,
     paddle::Tensor* out);

custom_ops/gpu_ops/append_attn/template_instantiation/append_attention_c8_float16_int8_kerne.cu

@@ -54,6 +54,7 @@ template void CascadeAppendAttentionC8Kernel<paddle::float16, int8_t, false>(
     const bool causal,
     const bool is_decoder,
     const bool enable_prefill,
+    const std::string& cache_quant_type_str,
     cudaStream_t& stream,
     paddle::Tensor* out);
 
@@ -99,5 +100,6 @@ template void CascadeAppendAttentionC8Kernel<paddle::float16, int8_t, true>(
     const bool causal,
     const bool is_decoder,
     const bool enable_prefill,
+    const std::string& cache_quant_type_str,
     cudaStream_t& stream,
     paddle::Tensor* 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__                                              \
@@ -437,6 +441,15 @@ __forceinline__ __host__ __device__ void vec_cast<nv_bfloat16, float>(
     PD_THROW("not support the group_size", group_size);      \
   }
 
+#define DISPATCH_DyCfp8(is_dynamic_cfp8, IsDynamicC8, ...) \
+  if (is_dynamic_cfp8) {                                   \
+    constexpr bool IsDynamicC8 = true;                     \
+    __VA_ARGS__                                            \
+  } else {                                                 \
+    constexpr bool IsDynamicC8 = false;                    \
+    __VA_ARGS__                                            \
+  }
+
 #define DISPATCH_MLA_GROUP_SIZE(group_size, GROUP_SIZE, ...) \
   if (group_size == 8) {                              \
     constexpr size_t GROUP_SIZE = 8;                         \
@@ -474,6 +487,9 @@ __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 +575,37 @@ 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

@@ -63,7 +63,7 @@ std::vector<paddle::Tensor> AppendAttention(
     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 &decoder_num_blocks_cpu,
     const paddle::Tensor &set_max_lengths, const paddle::Tensor &max_len_kv,
     const paddle::optional<paddle::Tensor> &rotary_embs,
     const paddle::optional<paddle::Tensor> &attn_mask,
@@ -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_cpu,
+    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 +254,9 @@ 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,
+    const int hadamard_block_size);
 
 paddle::Tensor MoeExpertFFNWint2Func(
     const paddle::Tensor& permute_input,
@@ -231,14 +299,23 @@ paddle::Tensor OpenShmAndGetMetaSignalFunc(const int rank, const int device_id,
 paddle::Tensor InitSignalLayerwiseFunc(const paddle::Tensor &kv_signal_metadata,
                                        const int layer_id);
 
-std::vector<paddle::Tensor> GetBlockShapeAndSplitKVBlock(
+void GetBlockShapeAndSplitKVBlock(
     const paddle::Tensor &seq_lens_encoder,
     const paddle::Tensor &seq_lens_decoder,
     const paddle::Tensor &seq_lens_this_time,
     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 &decoder_num_blocks_cpu,     // Inplace, Pinned Memory
+    paddle::Tensor &decoder_num_blocks_device,  // Inplace
+    paddle::Tensor &decoder_chunk_size_device,  // Inplace
     paddle::Tensor &max_len_tensor_cpu,         // Inplace, Pinned Memory
+    paddle::Tensor &encoder_batch_ids,          // Inplace
+    paddle::Tensor &encoder_tile_ids_per_batch, // Inplace
+    paddle::Tensor &encoder_num_blocks_x_cpu,   // Inplace, Pinned Memory
+    paddle::Tensor &kv_batch_ids,               // Inplace
+    paddle::Tensor &kv_tile_ids_per_batch,      // Inplace
+    paddle::Tensor &kv_num_blocks_x_cpu,        // Inplace, Pinned Memory
+    paddle::Tensor &max_len_kv_cpu,             // Inplace, Pinned Memory
     const int encoder_block_shape_q,
     const int decoder_block_shape_q,
     const int group_size,
@@ -311,9 +388,11 @@ void RecoverDecodeTask(const paddle::Tensor &stop_flags,
                    const paddle::Tensor &step_seq_lens_decoder,
                    const paddle::Tensor &block_tables,
                    const paddle::Tensor &is_block_step,
-                   const int block_size);
-
-
+                   const paddle::optional<paddle::Tensor> &draft_tokens,
+                   const paddle::optional<paddle::Tensor> &step_draft_tokens,
+                   const paddle::optional<paddle::Tensor> &step_seq_lens_this_time,
+                   const int block_size,
+                   const int max_draft_tokens);
 
 paddle::Tensor
 GroupSwigluWithMasked(const paddle::Tensor &fc1_out_tensor,
@@ -323,7 +402,7 @@ 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,
@@ -337,8 +416,8 @@ std::vector<paddle::Tensor> MoEDeepGEMMDePermute(
     const paddle::Tensor &topk_idx, const paddle::Tensor &topk_weights);
 
 void TextImageIndexOut(const paddle::Tensor &token_type_ids,
-                       const paddle::Tensor &text_input,
-                       const paddle::Tensor &image_input);
+                        paddle::Tensor &text_input,
+                        paddle::Tensor &image_input);
 
 void TextImageGatherScatter(paddle::Tensor &input, paddle::Tensor &text_input,
                             paddle::Tensor &image_input,
@@ -396,23 +475,18 @@ std::vector<paddle::Tensor> MultiHeadLatentAttention(
     const paddle::Tensor& query,
     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& cu_seqlens_q,
     const paddle::Tensor& batch_id_per_token,
     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& decoder_num_blocks_cpu,
-    const paddle::Tensor& max_enc_len_this_time,
+    const paddle::Tensor& decoder_num_blocks_device,
+    const paddle::Tensor& decoder_chunk_size_device,
     const paddle::Tensor& max_dec_len_this_time,
     const paddle::Tensor& max_len_kv,
     const paddle::optional<paddle::Tensor>& attn_mask,
@@ -526,7 +600,7 @@ 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);
@@ -609,7 +683,7 @@ 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,
@@ -640,6 +714,22 @@ void SpeculateSaveWithOutputMsgStatic(const paddle::Tensor& accept_tokens,
 void SpeculateClearAcceptNums(const paddle::Tensor& accept_num,
                               const paddle::Tensor& seq_lens_decoder);
 
+void SpeculateScheduleCache(const paddle::Tensor &draft_tokens,
+                            const paddle::Tensor &block_tables,
+                            const paddle::Tensor &stop_flags,
+                            const paddle::Tensor &seq_lens_this_time,
+                            const paddle::Tensor &seq_lens_decoder,
+                            const paddle::Tensor &step_seq_lens_decoder,
+                            const paddle::Tensor &step_draft_tokens,
+                            const paddle::Tensor &step_seq_lens_this_time,
+                            const paddle::Tensor &accept_num,
+                            const paddle::Tensor &accept_tokens,
+                            const paddle::Tensor &is_block_step,
+                            const paddle::Tensor &not_need_stop,
+                            const paddle::Tensor &stop_nums,
+                            const int block_size,
+                            const int max_draft_tokens);
+
 void NgramMatch(const paddle::Tensor &input_ids,
         const paddle::Tensor &input_ids_len,
         const paddle::Tensor &pre_ids,
@@ -654,6 +744,20 @@ 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,
@@ -669,9 +773,12 @@ void DraftModelPreprocess(const paddle::Tensor& draft_tokens,
                           const paddle::Tensor& seq_lens_decoder,
                           const paddle::Tensor& step_idx,
                           const paddle::Tensor& not_need_stop,
+                          const paddle::Tensor& is_block_step,
                           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,
@@ -680,7 +787,8 @@ void DraftModelPreprocess(const paddle::Tensor& draft_tokens,
                           const paddle::Tensor& base_model_draft_tokens,
                           const int max_draft_token,
                           const bool truncate_first_token,
-                          const bool splitwise_prefill);
+                          const bool splitwise_prefill,
+                          const bool kvcache_scheduler_v1);
 
 
 void DraftModelUpdate(const paddle::Tensor& inter_next_tokens,
@@ -761,6 +869,33 @@ 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"),
@@ -814,6 +949,7 @@ 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
@@ -845,7 +981,7 @@ 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
@@ -869,12 +1005,33 @@ PYBIND11_MODULE(fastdeploy_ops, m) {
 
   m.def("per_token_quant_padding", &PerTokenQuantPadding, py::arg("input"),
         py::arg("block_size"),
-        "per token per block quant and padding tranpose scale");
+        "per token per block quant and padding transpose scale");
 
   m.def("masked_per_token_quant", &MaskedPerTokenQuant, py::arg("input"),
         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
@@ -891,7 +1048,7 @@ PYBIND11_MODULE(fastdeploy_ops, m) {
   m.def("moe_expert_ffn", &MoeExpertFFNFunc, "moe export ffn function");
 
   /**
-   * moe/fused_moe/moe_ffn_wint2.cu
+   * moe/fused_moe/moe_expert_ffn_wint2.cu
    * moe_expert_ffn_wint2
    */
   m.def("moe_expert_ffn_wint2", &MoeExpertFFNWint2Func, "moe export ffn wint2 function");
@@ -1088,7 +1245,7 @@ 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");
 
@@ -1096,8 +1253,12 @@ PYBIND11_MODULE(fastdeploy_ops, m) {
 
   m.def("speculate_clear_accept_nums",&SpeculateClearAcceptNums, "speculate_clear_accept_nums function");
 
+  m.def("speculate_schedule_cache",&SpeculateScheduleCache, "SpeculateScheduleCache function");
+
   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");
@@ -1111,4 +1272,14 @@ 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();
@@ -163,3 +163,12 @@ 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/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/threadblock/wint2x_mma_base.h

@@ -89,11 +89,11 @@ public:
       GemmShape<Shape::kM / WarpGemm::kM, Shape::kN / WarpGemm::kN,
                 Shape::kK / WarpGemm::kK>;
 
-  /// Number of warp-level GEMM oeprations
+  /// Number of warp-level GEMM operations
   static int const kWarpGemmIterations =
       (WarpGemm::kK / Operator::Policy::MmaShape::kK);
 
-  /// Number of warp-level GEMM oeprations per load for B
+  /// Number of warp-level GEMM operations per load for B
   static constexpr int kWarpGemmIterationsPerLoadForB =
       Operator::IteratorB::InstructionShape::kRow / Operator::InstructionShape::kK;
   static_assert(!(kWarpGemmIterations % kWarpGemmIterationsPerLoadForB), "");

custom_ops/gpu_ops/cutlass_kernels/fp8_gemm_fused/dual_gemm/thread/left_gelu_and_mul.h

@@ -117,7 +117,7 @@ class LeftGELUAndMul {
     CUTLASS_HOST_DEVICE
     FragmentOutput operator()(FragmentAccumulator const &lhs,
                               FragmentAccumulator const &rhs) const {
-        // Convert source to interal compute numeric type
+        // Convert source to internal compute numeric type
         NumericArrayConverter<ElementCompute, ElementAccumulator, kCount, Round>
             accumulator_to_compute;
 

custom_ops/gpu_ops/cutlass_kernels/fp8_gemm_fused/dual_gemm/thread/left_silu_and_mul.h

@@ -117,7 +117,7 @@ class LeftSiLUAndMul {
     CUTLASS_HOST_DEVICE
     FragmentOutput operator()(FragmentAccumulator const &lhs,
                               FragmentAccumulator const &rhs) const {
-        // Convert source to interal compute numeric type
+        // Convert source to internal compute numeric type
         NumericArrayConverter<ElementCompute, ElementAccumulator, kCount, Round>
             accumulator_to_compute;
 

custom_ops/gpu_ops/cutlass_kernels/fp8_gemm_fused/dual_gemm/threadblock/dual_mma_base.h

@@ -92,7 +92,7 @@ class DualMmaBase {
                               Shape::kN / WarpGemm::kN,
                               Shape::kK / WarpGemm::kK>;
 
-  /// Number of warp-level GEMM oeprations
+  /// Number of warp-level GEMM operations
   static int const kWarpGemmIterations =
       (WarpGemm::kK / Operator0::Policy::MmaShape::kK);
 

custom_ops/gpu_ops/cutlass_kernels/w4a8_moe/cutlass_extensions/epilogue/threadblock/epilogue_per_row_per_col_scale_nf4.h

@@ -219,7 +219,7 @@ class EpilogueVisitorPerRowPerColNf4 {
       iterator_C_.clear_mask();
     }
     // NOTE(wangbojun) Currently, this kernel don't hanve implantention for
-    // adding elementwise beta, we keep this here for future useage beta_ =
+    // adding elementwise beta, we keep this here for future usage beta_ =
     // (params.elementwise.beta_ptr ? *params.elementwise.beta_ptr :
     // params.elementwise.beta); if (beta_ == ElementAccumulator()) {
     //     iterator_C_.clear_mask();

custom_ops/gpu_ops/cutlass_kernels/w4a8_moe/cutlass_extensions/epilogue/threadblock/epilogue_tensor_op_int32.h

@@ -176,7 +176,7 @@ struct Nf4DefaultIteratorsTensorOp<cutlass::bfloat16_t,
 ///
 /// Satisfies: ReadableTileIterator
 ///
-template <typename ThreadMap_  ///< Thread map (conept: OutputTileThreadMap)
+template <typename ThreadMap_  ///< Thread map (concept: OutputTileThreadMap)
           >
 class SharedLoadIteratorMixed<ThreadMap_, int32_t, 32, 16, 8, 8> {
  public:

custom_ops/gpu_ops/cutlass_kernels/w4a8_moe/cutlass_extensions/gemm/threadblock/default_mma_nf4_int8_interleaved.h

@@ -64,7 +64,7 @@ template <
     typename InstructionShape_,
     /// Number of stages used in the pipelined mainloop
     int Stages,
-    /// Operation perfomed by GEMM
+    /// Operation performed by GEMM
     typename Operator,
     /// Store the accumulators in row major or column major.  Row major is used
     /// when output layout is interleaved.

custom_ops/gpu_ops/cutlass_kernels/w4a8_moe/cutlass_extensions/gemm/threadblock/int8_mma_base.h

@@ -133,7 +133,7 @@ public:
     /// Shape describing the number of warps filling the CTA
     using WarpCount = GemmShape<Shape::kM / WarpGemm::kM, Shape::kN / WarpGemm::kN, Shape::kK / WarpGemm::kK>;
 
-    /// Number of warp-level GEMM oeprations
+    /// Number of warp-level GEMM operations
     static int const kWarpGemmIterations = (WarpGemm::kK / Operator::Policy::MmaShape::kK);
     static_assert(Operator::IteratorB::InstructionShape::kRow>=Operator::InstructionShape::kK,"");
     static constexpr int kNumKIterationsPerWarpBLoad =

custom_ops/gpu_ops/cutlass_kernels/w4a8_moe/cutlass_extensions/gemm/threadblock/int8_mma_multistage.h

@@ -509,7 +509,7 @@ public:
                     this->warp_tile_iterator_B_.load(warp_frag_B[(warp_tileB_k_load_offset + 1) % 2]);
                     ++this->warp_tile_iterator_B_;
                 }
-                // TOOD(wangbojun) lds_converter can be remove for int8 B input
+                // TODO(wangbojun) lds_converter can be remove for int8 B input
                 typename TransformBAfterLDS::result_type converted_frag_B =
                     lds_converter(warp_frag_B[warp_tileB_k_load_offset % 2]);
 

custom_ops/gpu_ops/cutlass_kernels/w4a8_moe/cutlass_extensions/gemm/threadblock/nf4_int8_mma_base.h

@@ -96,7 +96,7 @@ public:
     /// Shape describing the number of warps filling the CTA
     using WarpCount = GemmShape<Shape::kM / WarpGemm::kM, Shape::kN / WarpGemm::kN, Shape::kK / WarpGemm::kK>;
 
-    /// Number of warp-level GEMM oeprations
+    /// Number of warp-level GEMM operations
     static int const kWarpGemmIterations = (WarpGemm::kK / Operator::Policy::MmaShape::kK);
     static_assert(Operator::IteratorB::InstructionShape::kRow>=Operator::InstructionShape::kK,"");
     static constexpr int kNumKIterationsPerWarpBLoad =

custom_ops/gpu_ops/cutlass_kernels/w4a8_moe/cutlass_extensions/gemm/threadblock/nf4_int8_mma_multistage.h

@@ -646,7 +646,7 @@ public:
                     //             );
                     // }
                 }
-                // TOOD(wangbojun) lds_converter can be remove for int8 B input
+                // TODO(wangbojun) lds_converter can be remove for int8 B input
                 // int4
                 // typename TransformBAfterLDS::result_type converted_frag_B =
                 //     lds_converter(warp_frag_B[warp_tileB_k_load_offset % 2]);

custom_ops/gpu_ops/env.h

@@ -59,6 +59,15 @@ inline uint32_t get_cascade_attention_num_threads() {
 inline bool get_mla_use_tensorcore() {
     static const char* mla_use_tensorcore_env = std::getenv("FLAGS_mla_use_tensorcore");
     static const uint32_t mla_use_tensorcore =
-            mla_use_tensorcore_env == nullptr ? 1 : std::stoul(std::string(mla_use_tensorcore_env));
+            mla_use_tensorcore_env == nullptr ? 0 : std::stoul(std::string(mla_use_tensorcore_env));
     return mla_use_tensorcore != 0 ? true : false;
 }
+inline int get_mla_dec_chunk_size(int bsz) {
+    static const char* mla_dec_chunk_size_env =
+        std::getenv("FLAGS_mla_dec_chunk_size");
+    static const int mla_dec_chunk_size =
+        mla_dec_chunk_size_env == nullptr
+            ? -1
+            : std::stoi(std::string(mla_dec_chunk_size_env));
+    return bsz > 1 ? mla_dec_chunk_size : 64;
+}

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_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(
@@ -123,16 +126,14 @@ 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};
 }
 
 PD_BUILD_STATIC_OP(get_padding_offset)
-    .Inputs({"input_ids", "token_num", "cum_offsets", "seq_len"})
+    .Inputs({"input_ids", "cum_offsets", "token_num", "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

@@ -193,6 +193,12 @@ 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 +515,7 @@ 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 +548,7 @@ __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 +563,11 @@ inline int GetSMVersion() {
   return sm_version;
 
 }
+
+inline bool GetMlaUseTensorcore() {
+  static const bool flags_mla_use_tensorcore = get_mla_use_tensorcore();
+  static const bool enable_mla_tensorcore = GetSMVersion() >= 90 ? true : false;
+  const bool mla_use_tensorcore =
+      flags_mla_use_tensorcore && enable_mla_tensorcore;
+  return mla_use_tensorcore;
+}