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

@@ -44,7 +44,7 @@ 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}" \
@@ -143,7 +143,7 @@ jobs:
           -v "${CACHE_DIR}/ConfigDir:/root/.config" \
           -e TZ="Asia/Shanghai" \
           --gpus '"device='"${DEVICES}"'"' ${docker_image} /bin/bash -xc '
-          python -m pip install paddlepaddle-gpu==3.2.0 -i https://www.paddlepaddle.org.cn/packages/stable/cu126/
+          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
 
@@ -160,6 +160,7 @@ jobs:
           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 \

.github/workflows/_base_test.yml

@@ -44,7 +44,7 @@ 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}" \
@@ -143,7 +143,7 @@ jobs:
           -v "${CACHE_DIR}/ConfigDir:/root/.config" \
           -e TZ="Asia/Shanghai" \
           --gpus '"device='"${DEVICES}"'"' ${docker_image} /bin/bash -xc '
-          python -m pip install paddlepaddle-gpu==3.2.0 -i https://www.paddlepaddle.org.cn/packages/stable/cu126/
+          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
 
@@ -160,6 +160,7 @@ jobs:
           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 \

.github/workflows/_build_linux.yml

@@ -55,7 +55,7 @@ on:
 jobs:
   fd-build:
     runs-on: [self-hosted, GPU-Build]
-    timeout-minutes: 240
+    timeout-minutes: 360
     outputs:
       wheel_path: ${{ steps.set_output.outputs.wheel_path }}
     steps:
@@ -149,7 +149,7 @@ jobs:
             elif [[ "${PADDLEVERSION}" != "" ]];then
               python -m pip install paddlepaddle-gpu==${PADDLEVERSION} -i https://www.paddlepaddle.org.cn/packages/stable/cu126/
             else
-              python -m pip install paddlepaddle-gpu==3.2.0 -i https://www.paddlepaddle.org.cn/packages/stable/cu126/
+              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

.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/_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}" \
@@ -116,7 +117,6 @@ jobs:
             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 $(pwd):/workspace \
@@ -133,7 +133,7 @@ jobs:
           -v "${CACHE_DIR}/ConfigDir:/root/.config" \
           -e TZ="Asia/Shanghai" \
           --gpus '"device='"${DEVICES}"'"' ${docker_image} /bin/bash -xc '
-          python -m pip install paddlepaddle-gpu==3.2.0 -i https://www.paddlepaddle.org.cn/packages/stable/cu126/
+          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
 
@@ -147,6 +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 \

.github/workflows/_pre_ce_test.yml

@@ -46,7 +46,7 @@ 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}" \
@@ -142,7 +142,7 @@ jobs:
           --gpus "\"device=${DEVICES}\"" ${docker_image} /bin/bash -c '
           git config --global --add safe.directory /workspace/FastDeploy
           cd FastDeploy
-          python -m pip install paddlepaddle-gpu==3.2.0 -i https://www.paddlepaddle.org.cn/packages/stable/cu126/
+          python -m pip install --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

@@ -44,7 +44,7 @@ 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}" \
@@ -146,7 +146,7 @@ jobs:
           -v "${CACHE_DIR}/ConfigDir:/root/.config" \
           -e TZ="Asia/Shanghai" \
           --gpus '"device='"${DEVICES}"'"' ${docker_image} /bin/bash -xc '
-          python -m pip install paddlepaddle-gpu==3.2.0 -i https://www.paddlepaddle.org.cn/packages/stable/cu126/
+          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
 

.github/workflows/_unit_test_coverage.yml

@@ -41,7 +41,7 @@ jobs:
 
   run_tests_with_coverage:
     runs-on: [self-hosted, GPU-h1z1-2Cards]
-    timeout-minutes: 60
+    timeout-minutes: 90
     needs: check_cov_skip
     if: needs.check_cov_skip.outputs.can-skip != 'true'
     outputs:
@@ -60,7 +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}" \
@@ -168,13 +168,10 @@ jobs:
           git config --global --add safe.directory /workspace/FastDeploy
           cd FastDeploy
           git diff origin/${BASE_REF}..HEAD --unified=0 > diff.txt
-          python -m pip install paddlepaddle-gpu==3.2.0 -i https://www.paddlepaddle.org.cn/packages/stable/cu126/
+          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 coverage
-          python -m pip install diff-cover
-          python -m pip install pytest-cov
-          python -m pip install jsonschema aistudio_sdk==0.3.5
+          python -m pip install -r scripts/unittest_requirement.txt
           python -m pip install ${fd_wheel_url}
           rm -rf fastdeploy
           # coverage subprocess use

.github/workflows/ce_job.yml

@@ -9,7 +9,7 @@ on:
 permissions: read-all
 
 concurrency:
-  group: ${{ github.ref }}-${{ github.sha }}
+  group: CE-Job-${{ github.ref }}-${{ github.sha }}
   cancel-in-progress: true
 
 jobs:
@@ -199,13 +199,13 @@ jobs:
           ls
           python ${push_file} ${filename} ${target_path}
           target_path_stripped="${target_path#paddle-qa/}"
-          WHEEL_PATH=https://paddle-qa.bj.bcebos.com/${target_path_stripped}/${fd_wheel_name}
-          echo "commit wheel url is ${WHEEL_PATH}"
+          WHEEL_PATH=https://paddle-qa.bj.bcebos.com/${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/${target_path_stripped_latest}/${fd_wheel_name}
+          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:
@@ -224,9 +224,9 @@ jobs:
           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 }})
+          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 }}
@@ -238,11 +238,11 @@ jobs:
           ls
           python ${push_file} ${filename} ${target_path}
           target_path_stripped="${target_path#paddle-qa/}"
-          WHEEL_PATH=https://paddle-qa.bj.bcebos.com/${target_path_stripped}/${fd_wheel_name}
-          echo "commit wheel url is ${WHEEL_PATH}"
+          WHEEL_PATH=https://paddle-qa.bj.bcebos.com/${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/${target_path_stripped_latest}/${fd_wheel_name}
+          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/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/pr_build_and_test.yml

@@ -21,7 +21,7 @@ jobs:
     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: "89,90"
+      COMPILE_ARCH: "90"
       WITH_NIGHTLY_BUILD: "OFF"
       FD_VERSION: "0.0.0"
 

.github/workflows/publish_job.yml

@@ -13,7 +13,7 @@ on:
 permissions: read-all
 
 concurrency:
-  group: ${{ github.ref }}-${{ github.sha }}
+  group: Publish-Job-${{ github.ref }}-${{ github.sha }}
   cancel-in-progress: true
 
 
@@ -319,3 +319,13 @@ jobs:
       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"

.gitmodules

@@ -1,6 +1,7 @@
 [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

README.md

@@ -59,7 +59,7 @@ FastDeploy supports inference deployment on **NVIDIA GPUs**, **Kunlunxin XPUs**,
 - [Iluvatar GPU](./docs/get_started/installation/iluvatar_gpu.md)
 - [Enflame GCU](./docs/get_started/installation/Enflame_gcu.md)
 - [Hygon DCU](./docs/get_started/installation/hygon_dcu.md)
-- [MetaX GPU](./docs/get_started/installation/metax_gpu.md.md)
+- [MetaX GPU](./docs/get_started/installation/metax_gpu.md)
 
 **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!
 

README_CN.md

@@ -57,7 +57,7 @@ FastDeploy 支持在**英伟达（NVIDIA）GPU**、**昆仑芯（Kunlunxin）XPU
 - [天数 CoreX](./docs/zh/get_started/installation/iluvatar_gpu.md)
 - [燧原 S60](./docs/zh/get_started/installation/Enflame_gcu.md)
 - [海光 DCU](./docs/zh/get_started/installation/hygon_dcu.md)
-- [沐曦 GPU](./docs/zh/get_started/installation/metax_gpu.md.md)
+- [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！
 

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-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

@@ -2,7 +2,7 @@ top_p: 0.95
 temperature: 0.6
 metadata:
   min_tokens: 1
-max_tokens: 65535
+max_tokens: 131071
 repetition_penalty: 1.0
 frequency_penalty: 0
 presence_penalty: 0

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

@@ -143,9 +143,9 @@ function build_and_install_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}

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

@@ -140,8 +140,8 @@ void 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,
@@ -273,11 +273,15 @@ void 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,
@@ -296,11 +300,15 @@ void 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) {
@@ -309,7 +317,6 @@ void 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,
@@ -336,7 +343,6 @@ void AppendAttentionKernel(
             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,

custom_ops/gpu_ops/append_attn/append_attention_c16_impl.cuh

@@ -52,6 +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]
@@ -74,6 +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;
@@ -142,7 +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 : nullptr;
+  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>(
@@ -422,13 +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 uint32_t attn_mask_len = -1) {
+    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");
@@ -445,6 +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;
@@ -511,7 +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 : nullptr;
+  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>(
@@ -902,6 +914,7 @@ void MultiQueryAppendAttention(
           quant_min_bound,
           in_scale,
           chunk_size,
+          num_blocks_x_cpu,
           nullptr,
           nullptr,
           nullptr,
@@ -960,6 +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()),
@@ -1075,7 +1089,7 @@ void MultiQueryAppendAttention(
       chunk_size = static_cast<uint32_t>(encoder_max_partition_size);
     }
 
-    uint32_t attn_mask_len;
+    int32_t attn_mask_len;
     if (attn_mask) {
         attn_mask_len = attn_mask.get().shape()[1];
     } else {
@@ -1134,6 +1148,7 @@ void MultiQueryAppendAttention(
           quant_min_bound,
           in_scale,
           chunk_size,
+          num_blocks_x_cpu,
           nullptr,
           nullptr,
           nullptr,
@@ -1206,6 +1221,7 @@ 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()),

custom_ops/gpu_ops/append_attn/append_attention_c4_impl.cuh

@@ -57,6 +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]
@@ -85,6 +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;
@@ -173,7 +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 : nullptr;
+  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>(
@@ -520,13 +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 uint32_t attn_mask_len = -1) {
+    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>();
@@ -549,6 +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;
@@ -635,7 +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 : nullptr;
+  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>(
@@ -1107,6 +1119,7 @@ void MultiQueryAppendC4Attention(
           quant_min_bound,
           in_scale,
           chunk_size,
+          num_blocks_x_cpu,
           nullptr,
           nullptr,
           nullptr,
@@ -1171,6 +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()),
@@ -1299,7 +1313,7 @@ void MultiQueryAppendC4Attention(
     }
 
     const int num_chunks = div_up(max_seq_len, chunk_size);
-    uint32_t attn_mask_len;
+    int32_t attn_mask_len;
     if (attn_mask) {
         attn_mask_len = attn_mask.get().shape()[1];
     } else {
@@ -1365,6 +1379,7 @@ void MultiQueryAppendC4Attention(
           quant_min_bound,
           in_scale,
           chunk_size,
+          num_blocks_x_cpu,
           nullptr,
           nullptr,
           nullptr,
@@ -1445,6 +1460,7 @@ 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()),

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,
@@ -57,6 +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]
@@ -86,33 +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 =
@@ -180,7 +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 : nullptr;
+  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>(
@@ -201,6 +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(
@@ -282,10 +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,
@@ -318,6 +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,
@@ -336,6 +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();
 
@@ -346,7 +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();
 
@@ -463,14 +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,
@@ -489,13 +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 uint32_t attn_mask_len = -1) {
+    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>();
@@ -518,32 +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));
@@ -609,7 +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 : nullptr;
+  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>(
@@ -634,6 +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
@@ -716,11 +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,
@@ -753,6 +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,
@@ -771,6 +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();
 
@@ -781,7 +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();
 
@@ -895,7 +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,
@@ -953,7 +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,
@@ -970,7 +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,
@@ -988,7 +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,
@@ -1022,7 +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,
@@ -1040,7 +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,
@@ -1075,6 +1171,7 @@ void MultiQueryAppendC8Attention(
           quant_min_bound,
           in_scale,
           chunk_size,
+          num_blocks_x_cpu,
           nullptr,
           nullptr,
           nullptr,
@@ -1133,6 +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()),
@@ -1218,7 +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,
@@ -1235,7 +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,
@@ -1253,7 +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,
@@ -1269,7 +1372,7 @@ void MultiQueryAppendC8Attention(
     }
 
     const int num_chunks = div_up(max_seq_len, chunk_size);
-    uint32_t attn_mask_len;
+    int32_t attn_mask_len;
     if (attn_mask) {
         attn_mask_len = attn_mask.get().shape()[1];
     } else {
@@ -1295,7 +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,
@@ -1313,7 +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,
@@ -1350,6 +1457,7 @@ void MultiQueryAppendC8Attention(
           quant_min_bound,
           in_scale,
           chunk_size,
+          num_blocks_x_cpu,
           nullptr,
           nullptr,
           nullptr,
@@ -1424,6 +1532,7 @@ 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()),
@@ -1546,6 +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;
@@ -1554,6 +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,
@@ -1572,43 +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
@@ -911,7 +1026,7 @@ __device__ __forceinline__ void mask_s(const bool* attn_mask,
                                        const uint32_t qo_len,
                                        const uint32_t kv_len,
                                        const uint32_t chunk_end,
-                                       const uint32_t attn_mask_len,
+                                       const int32_t attn_mask_len,
                                        float (*s_frag)[num_frags_z][8],
                                        const int *mask_offset = nullptr) {
   const uint32_t tx = threadIdx.x;
@@ -929,13 +1044,13 @@ __device__ __forceinline__ void mask_s(const bool* attn_mask,
                                   8 * (reg_id / 4) + reg_id % 2;
           bool out_of_boundary;
           if (mask_offset) {
-            out_of_boundary = q_idx < qo_len ? (kv_idx > mask_offset[q_idx]) : true;
+            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 && q_idx < attn_mask_len) {
+            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;
@@ -1093,7 +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,
@@ -1135,16 +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
@@ -1171,7 +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,
@@ -1215,16 +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_impl.cuh

@@ -18,6 +18,53 @@
 #include "mma_tensor_op.cuh"
 #include "utils.cuh"
 
+
+// Note(ZKK)
+// This function is very easy!
+// just make HeadDim data to be new HeadDim data!
+
+template <typename T, int VecSize=8, int HEAD_DIM=128, int NUM_THREADS=32>
+__device__ __forceinline__ void apply_rope(
+  const T* input,
+  const float* cos_emb,
+  const float* sin_emb,
+  T* output,
+  const int thread_id) {
+
+  using LoadT = AlignedVector<T, VecSize>;
+  using LoadBiasT = AlignedVector<T, VecSize>;
+  using LoadOutScaleT = AlignedVector<float, VecSize>;
+  constexpr int HalfVecSize = VecSize / 2;
+  using LoadEmbT = AlignedVector<float, HalfVecSize>;
+
+  LoadT src_vec;
+  LoadBiasT out_vec;
+  LoadEmbT cos_emb_vec;
+  LoadEmbT sin_emb_vec;
+
+#pragma unroll
+    for (uint32_t head_bias = thread_id * VecSize; head_bias < HEAD_DIM; head_bias +=  NUM_THREADS * VecSize) {
+      Load<T, VecSize>(&input[head_bias], &src_vec);
+      const uint32_t emb_idx = head_bias / 2;
+      Load<float, HalfVecSize>(&cos_emb[emb_idx], &cos_emb_vec);
+      Load<float, HalfVecSize>(&sin_emb[emb_idx], &sin_emb_vec);
+#pragma unroll
+      for (int i = 0; i < HalfVecSize; i++) {
+
+        float input_left = static_cast<float>(src_vec[2 * i]);
+        float input_right = static_cast<float>(src_vec[2 * i + 1]);
+
+        const float cos_tmp = cos_emb_vec[i];
+        const float sin_tmp = sin_emb_vec[i];
+        out_vec[2 * i] =
+            static_cast<T>(input_left * cos_tmp - input_right * sin_tmp);
+        out_vec[2 * i + 1] =
+            static_cast<T>(input_right * cos_tmp + input_left * sin_tmp);
+      }
+      Store<T, VecSize>(out_vec, &output[head_bias]);
+    }
+}
+
 template <typename T, int VecSize = 1>
 __global__ void append_decode_cache_T_rope_qk_norm_kernel(
     const T* __restrict__ quant_qkv,  // [bsz, num_heads + 2 * kv_num_heads,
@@ -28,7 +75,7 @@ __global__ void append_decode_cache_T_rope_qk_norm_kernel(
                                   // head_size // 2]
     T* __restrict__ qkv_out,
     const int* __restrict__ block_tables,     // [bsz, max_blocks_per_seq]
-    const int* __restrict__ batch_id_per_token,  // [num_tokens]
+
     const int* __restrict__ cu_seqlens_q,
     const int* __restrict__ seq_lens,          // [bsz]
     const int* __restrict__ seq_lens_encoder,  // [bsz]
@@ -120,7 +167,6 @@ __global__ void append_decode_cache_T_rope_qk_norm_kernel(
       float row_variance =
           max(warp_m2 / head_size, 0.0f);
       float row_inv_var = Rsqrt(row_variance + rms_norm_eps);
-
       if (hi < num_heads) { // q
         Load<float, VecSize>(&q_norm_weight[threadIdx.x * VecSize], &q_norm_vec);
         #pragma unroll
@@ -129,6 +175,7 @@ __global__ void append_decode_cache_T_rope_qk_norm_kernel(
         }
       } else { // k
         Load<float, VecSize>(&k_norm_weight[threadIdx.x * VecSize], &k_norm_vec);
+        #pragma unroll
         for (int i = 0; i < VecSize; i++) {
           out_vec[i] = static_cast<T>(tmp_vec[i] * row_inv_var * k_norm_vec[i]);
         }
@@ -164,7 +211,7 @@ __global__ void append_decode_cache_T_rope_kernel(
                                   // head_size // 2]
     T* __restrict__ qkv_out,
     const int* __restrict__ block_tables,     // [bsz, max_blocks_per_seq]
-    const int* __restrict__ batch_id_per_token,  // [num_tokens]
+
     const int* __restrict__ cu_seqlens_q,
     const int* __restrict__ seq_lens,          // [bsz]
     const int* __restrict__ seq_lens_encoder,  // [bsz]
@@ -270,7 +317,7 @@ __global__ void append_decode_cache_T_rope_kernel(
                                   // head_size // 2]
     T* __restrict__ qkv_out,
     const int* __restrict__ block_tables,     // [bsz, max_blocks_per_seq]
-    const int* __restrict__ batch_id_per_token,  // [num_tokens]
+
     const int* __restrict__ cu_seqlens_q,
     const int* __restrict__ seq_lens,          // [bsz]
     const int* __restrict__ seq_lens_encoder,  // [bsz]
@@ -435,10 +482,10 @@ __global__ void append_decode_cache_T_neox_partial_rope_kernel(
     if (hi < num_heads && h_bias >= half_rotary_dim){
       continue;
     }
-    if (seq_lens_encoder[ori_bi] > 0) continue;
+    const int start_token_idx = cu_seqlens_q[ori_bi];
+    if (seq_lens_encoder[ori_bi] > 0) return;
     const int write_seq_id = seq_lens[ori_bi];
     if (write_seq_id == 0) continue;
-    const int start_token_idx = cu_seqlens_q[ori_bi];
 
     const int* block_table_now = nullptr;
 
@@ -527,7 +574,6 @@ __global__ void append_decode_cache_T_neox_rope_kernel(
                                   // head_size // 2]
     T* __restrict__ qkv_out,
     const int* __restrict__ block_tables,     // [bsz, max_blocks_per_seq]
-    const int* __restrict__ batch_id_per_token,  // [num_tokens]
     const int* __restrict__ cu_seqlens_q,
     const int* __restrict__ seq_lens,          // [bsz]
     const int* __restrict__ seq_lens_encoder,  // [bsz]
@@ -641,7 +687,6 @@ __global__ void append_decode_cache_T_neox_rope_kernel(
                                   // head_size // 2]
     T* __restrict__ qkv_out,
     const int* __restrict__ block_tables,     // [bsz, max_blocks_per_seq]
-    const int* __restrict__ batch_id_per_token,  // [num_tokens]
     const int* __restrict__ cu_seqlens_q,
     const int* __restrict__ seq_lens,          // [bsz]
     const int* __restrict__ seq_lens_encoder,  // [bsz]
@@ -765,6 +810,293 @@ __global__ void append_decode_cache_T_neox_rope_kernel(
   }
 }
 
+template <typename T, int VecSize = 4, int RoundType = 0, int HeadDim = 128, bool is_scale_channel_wise=false, bool IsFP8=true>
+__global__ void append_decode_cache_int8_rope_qk_norm_kernel(
+    const T* __restrict__ quant_qkv,    // [bsz, num_heads + 2 * kv_num_heads,
+                                        // head_size]
+    uint8_t* __restrict__ key_cache,    // [num_blocks, kv_num_heads,
+                                        // block_size, head_size // 2]
+    uint8_t* __restrict__ value_cache,  // [num_blocks, kv_num_heads,
+                                        // block_size, head_size // 2]
+    T* __restrict__ qkv_out,
+    const int* __restrict__ block_tables,     // [bsz, max_blocks_per_seq]
+    const int* __restrict__ cu_seqlens_q,
+    const int* __restrict__ seq_lens,          // [bsz]
+    const int* __restrict__ seq_lens_encoder,  // [bsz]
+    const float* __restrict__ cos_emb,
+    const float* __restrict__ sin_emb,
+    T* __restrict__ cache_k_scale,
+    T* __restrict__ cache_v_scale,
+    const float* q_norm_weight,
+    const float* k_norm_weight,
+    const int max_seq_len,
+    const int max_blocks_per_seq,
+    const int num_heads,
+    const int block_size,
+    const float max_bound,
+    const float min_bound,
+    const int kv_num_heads,
+    const bool rope_3d,
+    const float rms_norm_eps) {
+  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;
+  const int tid = threadIdx.x;
+  const int wid = tid / 32;
+  const int lane_id = tid % 32;
+  const int bid = blockIdx.x, head_idx = blockIdx.y * NUM_WARPS + wid;
+  int q_head_idx, k_head_idx, v_idx;
+  const int64_t hidden_size = (num_heads + 2 * kv_num_heads) * HeadDim;
+  constexpr int half_head_size = HeadDim / 2;
+  const int start_token_idx = cu_seqlens_q[bid];
+  if (seq_lens_encoder[bid] > 0) return;
+  const int write_seq_id = seq_lens[bid];
+  if (write_seq_id == 0) return;
+  const int* block_table_now = nullptr;
+
+  block_table_now = block_tables + bid * max_blocks_per_seq;
+  const int block_idx = __ldg(&block_table_now[write_seq_id / block_size]);
+  const int block_offset = write_seq_id % block_size;
+
+  int cache_offset;
+  if (head_idx < num_heads) {
+    cache_offset = 0;
+  } else if (head_idx < num_heads + 2 * kv_num_heads) {
+    cache_offset = block_idx * kv_num_heads * block_size + (head_idx - num_heads) % kv_num_heads * block_size + block_offset;
+  }
+  T *cache_k_scale_now = cache_k_scale + cache_offset;
+  T *cache_v_scale_now = cache_v_scale + cache_offset;
+
+  float thread_m2 = 0.0f;
+  float warp_m2 = 0.0f;
+
+  if (head_idx < num_heads) {
+    // q
+    using LoadT = AlignedVector<T, VecSize>;
+    using LoadBiasT = AlignedVector<T, VecSize>;
+    using LoadOutScaleT = AlignedVector<float, VecSize>;
+    constexpr int HalfVecSize = VecSize / 2;
+    using LoadEmbT = AlignedVector<float, HalfVecSize>;
+
+    LoadT src_vec;
+    LoadBiasT out_vec;
+    LoadEmbT cos_emb_vec;
+    LoadEmbT sin_emb_vec;
+    const T* qkv_now = quant_qkv + start_token_idx * hidden_size;
+    T* qkv_out_now = qkv_out + start_token_idx * hidden_size;
+#pragma unroll
+    for (uint32_t head_bias = lane_id * VecSize; head_bias < HeadDim;
+         head_bias += 32 * VecSize) {
+      const int bias_idx = head_idx * HeadDim + head_bias;
+      Load<T, VecSize>(&qkv_now[bias_idx], &src_vec);
+      // q rope
+      const uint32_t emb_idx = write_seq_id * half_head_size + head_bias / 2;
+      const 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
+        float input_left = static_cast<float>(src_vec[2 * i]);
+        float input_right = static_cast<float>(src_vec[2 * i + 1]);
+
+        const float cos_tmp = cos_emb_vec[i];
+        const float sin_tmp = sin_emb_vec[i];
+        float tmp1 = input_left * cos_tmp - input_right * sin_tmp;
+        float tmp2 = input_right * cos_tmp + input_left * sin_tmp;
+        thread_m2 += tmp1 * tmp1 + tmp2 * tmp2;
+        out_vec[2 * i] =
+            static_cast<T>(tmp1);
+        out_vec[2 * i + 1] =
+            static_cast<T>(tmp2);
+      }
+      // qk norm
+      if (q_norm_weight) {
+        WelfordWarpAllReduce<float, 32>(thread_m2, &warp_m2);
+        float row_variance =
+            max(warp_m2 / HeadDim, 0.0f);
+        float row_inv_var = Rsqrt(row_variance + rms_norm_eps);
+        LoadOutScaleT q_norm_vec;
+        Load<float, VecSize>(&q_norm_weight[lane_id * VecSize], &q_norm_vec);
+        #pragma unroll
+        for (int i = 0; i < VecSize; i++) {
+          out_vec[i] = static_cast<T>(static_cast<float>(out_vec[i]) * row_inv_var * q_norm_vec[i]);
+        }
+      }
+      Store<T, VecSize>(out_vec, &qkv_out_now[bias_idx]);
+    }
+  } else if (head_idx < num_heads + 2 * kv_num_heads) {
+    // k
+    constexpr int KV_VEC_SIZE = 16 / sizeof(uint8_t);  // 16
+    using LoadPadKVT = AlignedVector<uint8_t, KV_VEC_SIZE>;
+    const uint32_t kv_head_idx = (head_idx - num_heads) % kv_num_heads;
+    if (block_offset == 0) {
+      // pad zero for this kv_head_idx for this block
+      LoadPadKVT pad_cache_vec;
+      *(reinterpret_cast<uint4*>(pad_cache_vec.val)) = make_uint4(0, 0, 0, 0);
+      if (head_idx < num_heads + kv_num_heads) {
+        constexpr int num_vecs_per_head_dim = HeadDim / KV_VEC_SIZE;
+        constexpr int num_token_each_time = 32 / num_vecs_per_head_dim;
+        const uint32_t tgt_idx =
+            (block_idx * kv_num_heads + kv_head_idx) * block_size * HeadDim +
+            lane_id % num_vecs_per_head_dim * KV_VEC_SIZE;
+        for (int block_i = lane_id / num_vecs_per_head_dim;
+             block_i < block_size;
+             block_i += num_token_each_time) {
+          Store<uint8_t, KV_VEC_SIZE>(pad_cache_vec,
+                                      &key_cache[tgt_idx + block_i * HeadDim]);
+        }
+      } else {
+        const int num_vecs_per_head_dim = block_size / KV_VEC_SIZE;
+        const int num_token_each_time = 32 / num_vecs_per_head_dim;
+        const uint32_t tgt_idx =
+            (block_idx * kv_num_heads + kv_head_idx) * HeadDim * block_size +
+            lane_id % num_vecs_per_head_dim * KV_VEC_SIZE;
+        for (int block_i = lane_id / num_vecs_per_head_dim; block_i < HeadDim;
+             block_i += num_token_each_time) {
+          Store<uint8_t, KV_VEC_SIZE>(
+              pad_cache_vec, &value_cache[tgt_idx + block_i * block_size]);
+        }
+      }
+      __syncwarp();
+    }
+
+    constexpr int K_VEC_SIZE = 4;
+    constexpr int HALF_K_VEC_SIZE = 2;
+    using LoadKVResT = AlignedVector<uint8_t, K_VEC_SIZE>;
+    using LoadKVT = AlignedVector<uint8_t, HALF_K_VEC_SIZE>;
+    using LoadT = AlignedVector<T, HALF_K_VEC_SIZE>;
+    using LoadBiasT = AlignedVector<T, HALF_K_VEC_SIZE>;
+    using LoadOutScaleT = AlignedVector<float, HALF_K_VEC_SIZE>;
+    using LoadEmbT = AlignedVector<float, 1>;
+    LoadKVResT cache_vec;
+    LoadT src_vec1, src_vec2;
+    LoadBiasT out_vec1, out_vec2;
+    LoadEmbT cos_emb_vec1, cos_emb_vec2;
+    LoadEmbT sin_emb_vec1, sin_emb_vec2;
+
+    const T* qkv_now = quant_qkv + start_token_idx * hidden_size;
+    const int head_bias = lane_id / 4 * 16 + lane_id % 4 * 2;
+    const int bias_idx = head_idx * HeadDim + head_bias;
+    Load<T, HALF_K_VEC_SIZE>(&qkv_now[bias_idx], &src_vec1);
+    Load<T, HALF_K_VEC_SIZE>(&qkv_now[bias_idx + 8], &src_vec2);
+    T scale = T(1.0f);
+    const int k_head_idx = head_idx - num_heads;
+    const int v_head_idx = head_idx - num_heads - kv_num_heads;
+    if (head_idx < num_heads + kv_num_heads) {
+      const uint32_t emb_idx = write_seq_id * half_head_size + head_bias / 2;
+      const 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);
+    }
+
+    float input_left = static_cast<float>(src_vec1[0]);
+    float input_right = static_cast<float>(src_vec1[1]);
+    if (head_idx < num_heads + kv_num_heads) {
+      float cos_tmp = cos_emb_vec1[0];
+      float sin_tmp = sin_emb_vec1[0];
+      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;
+      out_vec1[0] =
+          static_cast<T>(tmp1);
+      out_vec1[1] =
+          static_cast<T>(tmp2);
+    } else {
+      out_vec1[0] = src_vec1[0];
+      out_vec1[1] = src_vec1[1];
+    }
+
+    // rope
+    input_left = static_cast<float>(src_vec2[0]);
+    input_right = static_cast<float>(src_vec2[1]);
+    if (head_idx < num_heads + kv_num_heads) {
+      float cos_tmp = cos_emb_vec2[0];
+      float sin_tmp = sin_emb_vec2[0];
+      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;
+      out_vec2[0] = static_cast<T>(tmp1);
+      out_vec2[1] = static_cast<T>(tmp2);
+    } else {
+      out_vec2[0] = src_vec2[0];
+      out_vec2[1] = src_vec2[1];
+    }
+    if (k_norm_weight) {
+      if (head_idx < num_heads + kv_num_heads) {
+        LoadOutScaleT k_norm_vec1, k_norm_vec2;
+        Load<float, HALF_K_VEC_SIZE>(&k_norm_weight[head_bias], &k_norm_vec1);
+        Load<float, HALF_K_VEC_SIZE>(&k_norm_weight[head_bias + 8], &k_norm_vec2);
+        // qk norm
+        WelfordWarpAllReduce<float, 32>(thread_m2, &warp_m2);
+        float row_variance =
+            max(warp_m2 / HeadDim, 0.0f);
+        float row_inv_var = Rsqrt(row_variance + rms_norm_eps);
+
+        for (int i = 0; i < HALF_K_VEC_SIZE; i++) {
+          out_vec1[i] = static_cast<T>(static_cast<float>(out_vec1[i]) * row_inv_var * k_norm_vec1[i]);
+          out_vec2[i] = static_cast<T>(static_cast<float>(out_vec2[i]) * row_inv_var * k_norm_vec2[i]);
+        }
+      }
+    }
+    // reduce max, 1 head per warp
+    T local_max = -INFINITY;
+#pragma unroll
+    for (int i = 0; i < HALF_K_VEC_SIZE; i++) {
+      local_max = __hmax(local_max, __habs(out_vec1[i]));
+      local_max = __hmax(local_max, __habs(out_vec2[i]));
+    }
+#pragma unroll
+    for (int m_offset = 16; m_offset > 0; m_offset /= 2) {
+      local_max = __hmax(local_max, __shfl_xor_sync(0xffffffff, local_max, m_offset));
+    }
+
+    scale = __hdiv(448, local_max);
+
+    if (lane_id == 0) {
+      if (head_idx < num_heads + kv_num_heads) {
+        cache_k_scale_now[0] = __hdiv(1, scale);
+      } else {
+        cache_v_scale_now[0] = __hdiv(1, scale);
+      }
+    }
+
+#pragma unroll
+    for (uint32_t i = 0; i < HALF_K_VEC_SIZE; i++) {
+      cache_vec[i] = QuantToC8<T,true, IsFP8, RoundType>(scale, out_vec1[i], max_bound, min_bound);
+      cache_vec[i + HALF_K_VEC_SIZE] = QuantToC8<T,true, IsFP8, RoundType>(scale, out_vec2[i], max_bound, min_bound);
+    }
+    if (head_idx < num_heads + kv_num_heads) {
+      const int start_block_16 =
+          block_offset / 16 * 16 + block_offset % 8 + lane_id / 4 % 2 * 8;
+      const uint32_t tgt_cache_idx =
+          block_idx * kv_num_heads * block_size * HeadDim +
+          kv_head_idx * block_size * HeadDim + start_block_16 * HeadDim +
+          lane_id / 4 / 2 * 32 + (block_offset % 16) / 8 * 16 + lane_id % 4 * 4;
+      Store<uint8_t, K_VEC_SIZE>(cache_vec, &key_cache[tgt_cache_idx]);
+    } else {
+      const uint32_t base_tgt_cache_idx =
+          block_idx * kv_num_heads * HeadDim * block_size +
+          kv_head_idx * HeadDim * block_size +
+          (lane_id / 4 * 16 + lane_id % 4 * 2) * block_size +
+          block_offset / 16 % 2 * 8 * block_size + block_offset / 16 / 2 * 32;
+      const uint32_t tgt_cache_idx1 = base_tgt_cache_idx +
+                                      block_offset % 8 / 2 * 4     // per 4
+                                      + block_offset % 16 / 8 * 2  // per 2
+                                      + block_offset % 2;          // per 1
+      const uint32_t tgt_cache_idx2 = tgt_cache_idx1 + block_size;
+      const uint32_t tgt_cache_idx3 = tgt_cache_idx1 + 16;
+      const uint32_t tgt_cache_idx4 = tgt_cache_idx3 + block_size;
+      value_cache[tgt_cache_idx1] = cache_vec[0];
+      value_cache[tgt_cache_idx2] = cache_vec[1];
+      value_cache[tgt_cache_idx3] = cache_vec[2];
+      value_cache[tgt_cache_idx4] = cache_vec[3];
+    }
+  }
+}
+
 template <typename T, int VecSize = 4, int RoundType = 0, int HeadDim = 128, bool is_scale_channel_wise=false, bool IsFP8=false>
 __global__ void append_decode_cache_int8_rope_kernel(
     const T* __restrict__ quant_qkv,    // [bsz, num_heads + 2 * kv_num_heads,
@@ -775,7 +1107,6 @@ __global__ void append_decode_cache_int8_rope_kernel(
                                         // block_size, head_size // 2]
     T* __restrict__ qkv_out,
     const int* __restrict__ block_tables,     // [bsz, max_blocks_per_seq]
-    const int* __restrict__ batch_id_per_token,  // [num_tokens]
     const int* __restrict__ cu_seqlens_q,
     const int* __restrict__ seq_lens,          // [bsz]
     const int* __restrict__ seq_lens_encoder,  // [bsz]
@@ -813,44 +1144,18 @@ __global__ void append_decode_cache_int8_rope_kernel(
 
   if (head_idx < num_heads) {
     // q
-    using LoadT = AlignedVector<T, VecSize>;
-    using LoadBiasT = AlignedVector<T, VecSize>;
-    using LoadOutScaleT = AlignedVector<float, VecSize>;
-    constexpr int HalfVecSize = VecSize / 2;
-    using LoadEmbT = AlignedVector<float, HalfVecSize>;
+    const T* qkv_now = quant_qkv + start_token_idx * hidden_size + head_idx * HeadDim;
+    T* qkv_out_now = qkv_out + start_token_idx * hidden_size + head_idx * HeadDim;
 
-    LoadT src_vec;
-    LoadBiasT out_vec;
-    LoadEmbT cos_emb_vec;
-    LoadEmbT sin_emb_vec;
-    const T* qkv_now = quant_qkv + start_token_idx * hidden_size;
-    T* qkv_out_now = qkv_out + start_token_idx * hidden_size;
-#pragma unroll
-    for (uint32_t head_bias = lane_id * VecSize; head_bias < HeadDim;
-         head_bias += 32 * VecSize) {
-      const int bias_idx = head_idx * HeadDim + head_bias;
-      Load<T, VecSize>(&qkv_now[bias_idx], &src_vec);
+    uint32_t emb_offset = write_seq_id * half_head_size;
+    emb_offset += rope_3d ? bid * max_seq_len * HeadDim : 0;
+    apply_rope<T, VecSize, HeadDim, 32>(
+      qkv_now,
+      cos_emb + emb_offset,
+      sin_emb + emb_offset,
+      qkv_out_now,
+      lane_id);
 
-      // q rope
-      const uint32_t emb_idx = write_seq_id * half_head_size + head_bias / 2;
-      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
-        float input_left = static_cast<float>(src_vec[2 * i]);
-        float input_right = static_cast<float>(src_vec[2 * i + 1]);
-
-        const float cos_tmp = cos_emb_vec[i];
-        const float sin_tmp = sin_emb_vec[i];
-        out_vec[2 * i] =
-            static_cast<T>(input_left * cos_tmp - input_right * sin_tmp);
-        out_vec[2 * i + 1] =
-            static_cast<T>(input_right * cos_tmp + input_left * sin_tmp);
-      }
-      Store<T, VecSize>(out_vec, &qkv_out_now[bias_idx]);
-    }
   } else if (head_idx < num_heads + 2 * kv_num_heads) {
     // k
     constexpr int KV_VEC_SIZE = 16 / sizeof(uint8_t);  // 16
@@ -1025,7 +1330,6 @@ __global__ void append_decode_cache_int8_rope_kernel(
                                         // block_size, head_size // 2]
     T* __restrict__ qkv_out,
     const int* __restrict__ block_tables,     // [bsz, max_blocks_per_seq]
-    const int* __restrict__ batch_id_per_token,  // [num_tokens]
     const int* __restrict__ cu_seqlens_q,
     const int* __restrict__ seq_lens,          // [bsz]
     const int* __restrict__ seq_lens_encoder,  // [bsz]
@@ -1330,7 +1634,6 @@ __global__ void append_decode_cache_int8_neox_rope_kernel(
                                         // block_size, head_size // 2]
     T* __restrict__ qkv_out,
     const int* __restrict__ block_tables,     // [bsz, max_blocks_per_seq]
-    const int* __restrict__ batch_id_per_token,  // [num_tokens]
     const int* __restrict__ cu_seqlens_q,
     const int* __restrict__ seq_lens,          // [bsz]
     const int* __restrict__ seq_lens_encoder,  // [bsz]
@@ -1632,7 +1935,7 @@ __global__ void append_decode_cache_int8_neox_rope_kernel(
                                         // block_size, head_size // 2]
     T* __restrict__ qkv_out,
     const int* __restrict__ block_tables,     // [bsz, max_blocks_per_seq]
-    const int* __restrict__ batch_id_per_token,  // [num_tokens]
+
     const int* __restrict__ cu_seqlens_q,
     const int* __restrict__ seq_lens,          // [bsz]
     const int* __restrict__ seq_lens_encoder,  // [bsz]
@@ -2029,7 +2332,7 @@ __global__ void append_decode_cache_int4_rope_kernel(
                                         // block_size, head_size // 2]
     T* __restrict__ qkv_out,
     const int* __restrict__ block_tables,     // [bsz, max_blocks_per_seq]
-    const int* __restrict__ batch_id_per_token,  // [num_tokens]
+
     const int* __restrict__ cu_seqlens_q,
     const int* __restrict__ seq_lens,          // [bsz]
     const int* __restrict__ seq_lens_encoder,  // [bsz]
@@ -2070,44 +2373,18 @@ __global__ void append_decode_cache_int4_rope_kernel(
 
   if (head_idx < num_heads) {
     // q
-    using LoadT = AlignedVector<T, VecSize>;
-    using LoadBiasT = AlignedVector<T, VecSize>;
-    using LoadOutScaleT = AlignedVector<float, VecSize>;
-    constexpr int HalfVecSize = VecSize / 2;
-    using LoadEmbT = AlignedVector<float, HalfVecSize>;
+    const T* qkv_now = quant_qkv + start_token_idx * hidden_size + head_idx * HeadDim;
+    T* qkv_out_now = qkv_out + start_token_idx * hidden_size + head_idx * HeadDim;
 
-    LoadT src_vec;
-    LoadBiasT out_vec;
-    LoadEmbT cos_emb_vec;
-    LoadEmbT sin_emb_vec;
-    const T* qkv_now = quant_qkv + start_token_idx * hidden_size;
-    T* qkv_out_now = qkv_out + start_token_idx * hidden_size;
-#pragma unroll
-    for (uint32_t head_bias = lane_id * VecSize; head_bias < HeadDim;
-         head_bias += 32 * VecSize) {
-      const int bias_idx = head_idx * HeadDim + head_bias;
-      Load<T, VecSize>(&qkv_now[bias_idx], &src_vec);
+    uint32_t emb_offset = write_seq_id * half_head_size;
+    emb_offset += rope_3d ? bid * max_seq_len * HeadDim : 0;
+    apply_rope<T, VecSize, HeadDim, 32>(
+      qkv_now,
+      cos_emb + emb_offset,
+      sin_emb + emb_offset,
+      qkv_out_now,
+      lane_id);
 
-      // q rope
-      const uint32_t emb_idx = write_seq_id * half_head_size + head_bias / 2;
-      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
-        float input_left = static_cast<float>(src_vec[2 * i]);
-        float input_right = static_cast<float>(src_vec[2 * i + 1]);
-
-        const float cos_tmp = cos_emb_vec[i];
-        const float sin_tmp = sin_emb_vec[i];
-        out_vec[2 * i] =
-            static_cast<T>(input_left * cos_tmp - input_right * sin_tmp);
-        out_vec[2 * i + 1] =
-            static_cast<T>(input_right * cos_tmp + input_left * sin_tmp);
-      }
-      Store<T, VecSize>(out_vec, &qkv_out_now[bias_idx]);
-    }
   } else if (head_idx < num_heads + 2 * kv_num_heads) {
     // k
     constexpr int KV_VEC_SIZE = 16 / sizeof(uint8_t);  // 16
@@ -2327,7 +2604,7 @@ __global__ void append_decode_cache_int4_rope_kernel(
                                         // block_size, head_size // 2]
     T* __restrict__ qkv_out,
     const int* __restrict__ block_tables,     // [bsz, max_blocks_per_seq]
-    const int* __restrict__ batch_id_per_token,  // [num_tokens]
+
     const int* __restrict__ cu_seqlens_q,
     const int* __restrict__ seq_lens,          // [bsz]
     const int* __restrict__ seq_lens_encoder,  // [bsz]
@@ -2658,7 +2935,7 @@ __global__ void append_decode_cache_int4_neox_rope_kernel(
                                         // block_size, head_size // 2]
     T* __restrict__ qkv_out,
     const int* __restrict__ block_tables,     // [bsz, max_blocks_per_seq]
-    const int* __restrict__ batch_id_per_token,  // [num_tokens]
+
     const int* __restrict__ cu_seqlens_q,
     const int* __restrict__ seq_lens,          // [bsz]
     const int* __restrict__ seq_lens_encoder,  // [bsz]
@@ -3031,7 +3308,7 @@ __global__ void append_decode_cache_int4_neox_rope_kernel(
                                         // block_size, head_size // 2]
     T* __restrict__ qkv_out,
     const int* __restrict__ block_tables,     // [bsz, max_blocks_per_seq]
-    const int* __restrict__ batch_id_per_token,  // [num_tokens]
+
     const int* __restrict__ cu_seqlens_q,
     const int* __restrict__ seq_lens,          // [bsz]
     const int* __restrict__ seq_lens_encoder,  // [bsz]

custom_ops/gpu_ops/append_attn/decoder_write_cache_with_rope_kernel.cu

@@ -21,7 +21,6 @@ void append_decode_cache_rope_qk_norm(const QKV_TYPE* qkv,
                               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,
@@ -59,7 +58,6 @@ void append_decode_cache_rope_qk_norm(const QKV_TYPE* qkv,
                                             value_cache,
                                             qkv_out,
                                             block_tables,
-                                            batch_id_per_token,
                                             cu_seqlens_q,
                                             seq_lens,
                                             seq_lens_encoder,
@@ -84,7 +82,6 @@ void append_decode_cache_rope(const QKV_TYPE* qkv,
                               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,
@@ -121,7 +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,
@@ -166,7 +162,6 @@ 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,
@@ -191,7 +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,
@@ -214,7 +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,
@@ -238,7 +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,
@@ -271,7 +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,
@@ -297,7 +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,
@@ -323,7 +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,
@@ -349,7 +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,
@@ -375,7 +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,
@@ -410,7 +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,
@@ -438,7 +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,
@@ -466,7 +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,
@@ -494,7 +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,
@@ -521,7 +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,
@@ -582,7 +564,6 @@ void DecoderWriteCacheWithRoPEKernel(
           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>(),
@@ -605,9 +586,39 @@ void DecoderWriteCacheWithRoPEKernel(
           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 {
       PD_THROW(
-          "append_decode_cache_rope_qk_norm not support cachekv quant yet");
+          "append_decode_cache_rope_qk_norm just supports cache_quant_type none/block_wise_fp8");
     }
   } else {
     if (cache_quant_type_str == "none") {
@@ -617,7 +628,6 @@ void DecoderWriteCacheWithRoPEKernel(
           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>(),
@@ -650,7 +660,6 @@ void DecoderWriteCacheWithRoPEKernel(
           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>(),
@@ -683,7 +692,6 @@ void DecoderWriteCacheWithRoPEKernel(
             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>(),
@@ -717,7 +725,6 @@ void DecoderWriteCacheWithRoPEKernel(
             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>(),
@@ -743,6 +750,36 @@ void DecoderWriteCacheWithRoPEKernel(
             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),
@@ -750,7 +787,6 @@ void DecoderWriteCacheWithRoPEKernel(
           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>(),
@@ -798,7 +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,
@@ -828,7 +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,
@@ -857,7 +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,
@@ -886,7 +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,

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,

custom_ops/gpu_ops/append_attn/encoder_write_cache_with_rope_impl.cuh

@@ -449,8 +449,8 @@ __global__ void GQAVariableLengthRotaryQKNormKernel(
   const int half_lastdim = last_dim / 2;
   const int offset = (q_num_head + kv_num_head) * last_dim;
   const int all_head_num = elem_cnt / last_dim;
-  for (int gloabl_hi = global_warp_idx; gloabl_hi < all_head_num; gloabl_hi += all_warp_num) {
-    int64_t linear_index = gloabl_hi * last_dim + threadIdx.x * VecSize;
+  for (int global_hi = global_warp_idx; global_hi < all_head_num; global_hi += all_warp_num) {
+    int64_t linear_index = global_hi * last_dim + threadIdx.x * VecSize;
     const int token_idx = linear_index / offset;
     const int ori_bi = batch_id_per_token[token_idx];
     if (seq_lens[ori_bi] == 0) continue;
@@ -1300,6 +1300,411 @@ __global__ void append_write_cache_kv_c8_qkv(
   }
 }
 
+template <typename T,
+          uint32_t num_frags_y,
+          uint32_t num_frags_z,
+          uint32_t HEAD_DIM,
+          uint32_t BLOCK_SIZE,
+          uint32_t NUM_WARPS,
+          bool is_need_kv_quant,
+          bool IsFP8 = true>
+__global__ void append_write_cache_kv_c8_qkv_dynamic(
+    uint8_t *__restrict__ cache_k,
+    uint8_t *__restrict__ cache_v,
+    const T *__restrict__ qkv_input,
+    T *__restrict__ cache_k_scales, // [block_num, num_heads, block_size]
+    T *__restrict__ cache_v_scales, // [block_num, num_heads, block_size]
+    const int *__restrict__ batch_ids,
+    const int *__restrict__ tile_ids,
+    const int *__restrict__ seq_lens_this_time,
+    const int *__restrict__ seq_lens_decoder,
+    const int *__restrict__ batch_id_per_token,
+    const int *__restrict__ cu_seqlens_q,
+    const int *__restrict__ block_tables,
+    const int max_seq_len,
+    const int max_blocks_per_seq,
+    const int num_heads,
+    const int kv_num_heads) {
+  constexpr uint32_t num_vecs_per_head = HEAD_DIM / num_elems_per_128b<T>();
+  constexpr uint32_t pad_len = BLOCK_SIZE;
+  const uint32_t btid = blockIdx.x, kv_head_idx = blockIdx.z;
+  const T cache_k_scale = cache_k_scales[kv_head_idx];
+  const T cache_v_scale = cache_v_scales[kv_head_idx];
+  const uint32_t tid = threadIdx.x, wid = threadIdx.y;
+  const uint32_t batch_id = batch_ids[btid];
+  const uint32_t tile_id = tile_ids[btid];
+  const uint32_t seq_len_this_time = seq_lens_this_time[batch_id];
+  if (seq_len_this_time <= 0) {
+    return;
+  }
+  const int *block_table_now = nullptr;
+
+  block_table_now = block_tables + batch_id * max_blocks_per_seq;
+
+  const uint32_t num_rows_per_block =
+      NUM_WARPS * num_frags_z * 16;  // BLOCK_SIZE
+  const uint32_t start_len = seq_lens_decoder[batch_id];
+  const uint32_t bf_pad_len = start_len % pad_len;
+  const uint32_t start_len_pad = start_len - bf_pad_len;
+  const uint32_t end_len = start_len + seq_len_this_time;
+
+  const uint32_t tile_start = start_len_pad + tile_id * num_rows_per_block;
+  int block_id = __ldg(&block_table_now[tile_start / BLOCK_SIZE]);
+  uint32_t chunk_start = tile_start + wid * num_frags_z * 16 + tid / 8;
+
+  const uint32_t start_token_idx = cu_seqlens_q[batch_id];
+  const uint32_t kv_batch_stride = (num_heads + 2 * kv_num_heads) * HEAD_DIM;
+  const uint32_t kv_h_stride = HEAD_DIM;
+  __shared__ T k_smem_ori[num_rows_per_block * HEAD_DIM];
+  __shared__ T v_smem_ori[num_rows_per_block * HEAD_DIM];
+  __shared__ T v_scale_smem[BLOCK_SIZE];
+  if (tile_start >= start_len) {
+    constexpr int KV_VEC_SIZE = 16 / sizeof(uint8_t);  // 16
+    using LoadPadKVT = AlignedVector<uint8_t, KV_VEC_SIZE>;
+    // pad zero for this kv_head_idx for this block
+    LoadPadKVT pad_cache_vec;
+    *(reinterpret_cast<uint4*>(pad_cache_vec.val)) = make_uint4(0, 0, 0, 0);
+    // reset k
+    constexpr int num_vecs_per_head_k = HEAD_DIM / KV_VEC_SIZE;
+    constexpr int num_token_each_time_k = 32 / num_vecs_per_head_k;
+    uint32_t tgt_idx =
+        (block_id * kv_num_heads + kv_head_idx) * BLOCK_SIZE * HEAD_DIM +
+        tid % num_vecs_per_head_k * KV_VEC_SIZE;
+    for (int block_i = tid / num_vecs_per_head_k;
+          block_i < BLOCK_SIZE;
+          block_i += num_token_each_time_k) {
+      Store<uint8_t, KV_VEC_SIZE>(pad_cache_vec,
+                                  &cache_k[tgt_idx + block_i * HEAD_DIM]);
+    }
+
+    // reset v
+    const int num_vecs_per_head_v = BLOCK_SIZE / KV_VEC_SIZE;
+    const int num_token_each_time_v = 32 / num_vecs_per_head_v;
+    tgt_idx =
+        (block_id * kv_num_heads + kv_head_idx) * HEAD_DIM * BLOCK_SIZE +
+        tid % num_vecs_per_head_v * KV_VEC_SIZE;
+    for (int block_i = tid / num_vecs_per_head_v; block_i < HEAD_DIM;
+          block_i += num_token_each_time_v) {
+      Store<uint8_t, KV_VEC_SIZE>(
+          pad_cache_vec, &cache_v[tgt_idx + block_i * BLOCK_SIZE]);
+    }
+  }
+  smem_t k_smem(k_smem_ori);
+  smem_t v_smem(v_smem_ori);
+
+  uint32_t kv_smem_offset_w = smem_t::get_permuted_offset<num_vecs_per_head>(
+      wid * num_frags_z * 16 + tid / 8, tid % 8);  // 4 * 8 per warp
+
+  /*
+   0 | 1
+   2 | 3
+  */
+  uint32_t k_smem_offset_r = smem_t::get_permuted_offset<num_vecs_per_head>(
+      wid * num_frags_z * 16 + 8 * (tid / 16) + tid % 8, (tid % 16) / 8);
+
+  constexpr uint32_t num_frags_v = num_frags_y / NUM_WARPS;
+  /*
+   0 | 2
+   1 | 3
+  */
+  uint32_t v_smem_offset_r = smem_t::get_permuted_offset<num_vecs_per_head>(
+      tid % 16, wid * num_frags_v * 2 + tid / 16);
+
+  // load kv gmem to smem
+  const uint32_t real_start_token_idx = start_token_idx - bf_pad_len +
+                                        tile_id * num_rows_per_block +
+                                        wid * num_frags_z * 16 + tid / 8;
+  uint32_t k_read_idx = real_start_token_idx * kv_batch_stride +
+                        (num_heads + kv_head_idx) * kv_h_stride +
+                        tid % 8 * num_elems_per_128b<T>();
+  uint32_t v_read_idx = real_start_token_idx * kv_batch_stride +
+                        (num_heads + kv_num_heads + kv_head_idx) * kv_h_stride +
+                        tid % 8 * num_elems_per_128b<T>();
+#pragma unroll
+  for (uint32_t fz = 0; fz < num_frags_z; ++fz) {
+#pragma unroll
+    for (uint32_t j = 0; j < 4; ++j) {
+#pragma unroll
+      for (uint32_t fy = 0; fy < num_frags_y / 4;
+           ++fy) {  // (num_frags_y * 16) / (8 *  num_elems_per_128b<T>())
+        if (chunk_start >= start_len && chunk_start < end_len) {
+          k_smem.load_128b_async<SharedMemFillMode::kNoFill>(
+              kv_smem_offset_w, qkv_input + k_read_idx, chunk_start < end_len);
+          v_smem.load_128b_async<SharedMemFillMode::kNoFill>(
+              kv_smem_offset_w, qkv_input + v_read_idx, chunk_start < end_len);
+        }
+        kv_smem_offset_w =
+            k_smem.advance_offset_by_column<8>(kv_smem_offset_w, fy);
+        k_read_idx += 8 * num_elems_per_128b<T>();
+        v_read_idx += 8 * num_elems_per_128b<T>();
+      }
+      kv_smem_offset_w =
+          k_smem.advance_offset_by_row<4, num_vecs_per_head>(kv_smem_offset_w) -
+          2 * num_frags_y;
+      chunk_start += 4;
+      k_read_idx +=
+          4 * kv_batch_stride - 2 * num_frags_y * num_elems_per_128b<T>();
+      v_read_idx +=
+          4 * kv_batch_stride - 2 * num_frags_y * num_elems_per_128b<T>();
+    }
+  }
+  commit_group();
+  wait_group<0>();
+  __syncthreads();
+
+  // reduce scale
+  // 16 rows per warp
+  uint32_t kv_reduce_frag[4];
+  T *kv_reduce_frag_T = reinterpret_cast<T*>(kv_reduce_frag);
+
+   T k_local_max_value[num_frags_z * 2];
+   T v_local_max_value[num_frags_z * 2];
+#pragma unroll
+  for (int i = 0; i < num_frags_z * 2; i++) {
+    k_local_max_value[i] = -INFINITY;
+  }
+#pragma unroll
+  for (int i = 0; i < num_frags_z * 2; i++) {
+    v_local_max_value[i] = -INFINITY;
+  }
+  const int num_kv_heads = gridDim.z;
+  const int scale_offset = block_id * num_kv_heads * BLOCK_SIZE + kv_head_idx * BLOCK_SIZE;
+  T *cache_k_scale_now = cache_k_scales + scale_offset;
+  T *cache_v_scale_now = cache_v_scales + scale_offset;
+  // k scale
+#pragma unroll
+  for (uint32_t fz = 0; fz < num_frags_z; ++fz) {
+#pragma unroll
+    for (uint32_t fy = 0; fy < num_frags_y; ++fy) {
+      // reduce per thread, 4 threads each row
+      k_smem.ldmatrix_m8n8x4(k_smem_offset_r, kv_reduce_frag);
+#pragma unroll
+      for (int i = 0; i < 4; i++) {
+        k_local_max_value[fz * 2] = __hmax(__habs(kv_reduce_frag_T[i]), k_local_max_value[fz * 2]);
+      }
+#pragma unroll
+      for (int i = 0; i < 4; i++) {
+        k_local_max_value[fz * 2 + 1] = __hmax(__habs(kv_reduce_frag_T[i + 4]), k_local_max_value[fz * 2 + 1]);
+      }
+      k_smem_offset_r = k_smem.advance_offset_by_column<2>(k_smem_offset_r, fy);
+    }
+    // reduce per row
+    for (int i = 0; i < 2; i++) {
+      T local_max_value = __habs(k_local_max_value[fz * 2 + i]);
+      local_max_value = __hmax(local_max_value, __shfl_xor_sync(0xffffffff, local_max_value, 2));
+      local_max_value = __hmax(local_max_value, __shfl_xor_sync(0xffffffff, local_max_value, 1));
+      // used for quant
+      k_local_max_value[fz * 2 + i] = __hdiv(448, local_max_value);
+    }
+    // store
+    if (tid % 4 == 0) {
+      const int offset_now = wid * num_frags_z * 16 + tid / 4;
+      // used for dequant
+      if (tile_start + offset_now >= start_len) {
+        if (tile_start + offset_now < end_len) {
+          cache_k_scale_now[offset_now] = __hdiv(1, k_local_max_value[fz * 2]);
+        } else {
+          cache_k_scale_now[offset_now] = 0;
+        }
+      }
+      if (tile_start + offset_now + 8 >= start_len) {
+        if (tile_start + offset_now + 8 < end_len) {
+          cache_k_scale_now[offset_now + 8] = __hdiv(1, k_local_max_value[fz * 2 + 1]);
+        } else {
+          cache_k_scale_now[offset_now + 8] = 0;
+        }
+      }
+    }
+    __syncthreads();
+    k_smem_offset_r -= 2 * num_frags_y; // num_frags_z = 1
+  }
+  // v scale
+  #pragma unroll
+  for (uint32_t fz = 0; fz < num_frags_z; ++fz) {
+#pragma unroll
+    for (uint32_t fy = 0; fy < num_frags_y; ++fy) {
+      // reduce per thread, 4 threads each row
+      v_smem.ldmatrix_m8n8x4(k_smem_offset_r, kv_reduce_frag);
+#pragma unroll
+      for (int i = 0; i < 4; i++) {
+        v_local_max_value[fz * 2] = __hmax(__habs(kv_reduce_frag_T[i]), v_local_max_value[fz * 2]);
+      }
+#pragma unroll
+      for (int i = 0; i < 4; i++) {
+        v_local_max_value[fz * 2 + 1] = __hmax(__habs(kv_reduce_frag_T[i + 4]), v_local_max_value[fz * 2 + 1]);
+      }
+      k_smem_offset_r = v_smem.advance_offset_by_column<2>(k_smem_offset_r, fy);
+    }
+    // reduce per row
+    for (int i = 0; i < 2; i++) {
+      T local_max_value = __habs(v_local_max_value[fz * 2 + i]);
+      local_max_value = __hmax(local_max_value, __shfl_xor_sync(0xffffffff, local_max_value, 2));
+      local_max_value = __hmax(local_max_value, __shfl_xor_sync(0xffffffff, local_max_value, 1));
+      v_local_max_value[fz * 2 + i] = __hdiv(448, local_max_value);
+    }
+    // store
+    if (tid % 4 == 0) {
+      const int offset_now = wid * num_frags_z * 16 + tid / 4;
+      // used for dequant
+      if (tile_start + offset_now >= start_len) {
+        if (tile_start + offset_now < end_len) {
+          cache_v_scale_now[offset_now] = __hdiv(1, v_local_max_value[fz * 2]);
+          v_scale_smem[offset_now] = v_local_max_value[fz * 2];
+        } else {
+          cache_v_scale_now[offset_now] = 0;
+          v_scale_smem[offset_now] = 0;
+        }
+      }
+      if (tile_start + offset_now + 8 >= start_len) {
+        if (tile_start + offset_now + 8 < end_len) {
+          cache_v_scale_now[offset_now + 8] = __hdiv(1, v_local_max_value[fz * 2 + 1]);
+          v_scale_smem[offset_now + 8] = v_local_max_value[fz * 2 + 1];
+        } else {
+          cache_v_scale_now[offset_now + 8] = 0;
+          v_scale_smem[offset_now + 8] = 0;
+        }
+      }
+    }
+    __syncthreads();
+    k_smem_offset_r -= 2 * num_frags_y; // num_frags_z = 1
+  }
+  __syncthreads();
+
+  // mask, quant, store
+  using LoadKVT = AlignedVector<uint8_t, 4>;
+  LoadKVT cache_vec1;
+  LoadKVT cache_vec2;
+
+  uint32_t chunk_start_k = tile_start + wid * num_frags_z * 16 + tid / 4;
+  uint32_t kv_frag[4];
+  const uint32_t write_n_stride = kv_num_heads * BLOCK_SIZE * HEAD_DIM;
+  const uint32_t write_h_stride = BLOCK_SIZE * HEAD_DIM;
+  const uint32_t write_b_stride = HEAD_DIM;
+  const uint32_t write_d_stride = BLOCK_SIZE;
+  uint32_t k_write_idx = block_id * write_n_stride +
+                         kv_head_idx * write_h_stride +
+                         (wid * num_frags_z * 16 + tid / 4) * write_b_stride +
+                         tid % 4 * 4;  // 4 * int8 = 8 * int4 = 32bit
+#pragma unroll
+  for (uint32_t fz = 0; fz < num_frags_z; ++fz) {
+    uint32_t k_write_idx_now_z = k_write_idx + fz * 16 * write_b_stride;
+#pragma unroll
+    for (uint32_t fy = 0; fy < num_frags_y; ++fy) {
+      uint32_t k_write_idx_now = k_write_idx_now_z +
+                                 fy % 2 * 8 * write_b_stride +
+                                 fy / 2 * 32;  // + fy % 2 * 16;
+      // load
+      k_smem.ldmatrix_m8n8x4(k_smem_offset_r, kv_frag);
+      // quant
+      T *k_frag_T = reinterpret_cast<T *>(kv_frag);
+      if (bf_pad_len != 0) {
+        Load<uint8_t, 4>(cache_k + k_write_idx_now, &cache_vec1);
+        Load<uint8_t, 4>(cache_k + k_write_idx_now + 16, &cache_vec2);
+      }
+#pragma unroll
+      for (uint32_t v_id = 0; v_id < 8; ++v_id) {
+        uint8_t uint_quant_value;
+        if (chunk_start_k + (v_id / 4) * 8 >= start_len &&
+            chunk_start_k + (v_id / 4) * 8 < end_len) {
+          uint_quant_value = QuantToC8<T, is_need_kv_quant, IsFP8>(k_local_max_value[fz * 2 + v_id / 4], k_frag_T[v_id], 127.0f, -127.0f);
+        } else {
+          uint_quant_value = 0;
+        }
+        if (bf_pad_len != 0) {
+          if (v_id < 4) {
+            cache_vec1[v_id] |= uint_quant_value;
+          } else {
+            cache_vec2[v_id % 4] |= uint_quant_value;
+          }
+        } else {
+          if (v_id < 4) {
+            cache_vec1[v_id] = uint_quant_value;
+          } else {
+            cache_vec2[v_id - 4] = uint_quant_value;
+          }
+        }
+      }
+      // store
+      Store<uint8_t, 4>(cache_vec1, cache_k + k_write_idx_now);
+      Store<uint8_t, 4>(cache_vec2, cache_k + k_write_idx_now + 16);
+      k_smem_offset_r = k_smem.advance_offset_by_column<2>(k_smem_offset_r, fy);
+    }
+    k_smem_offset_r =
+        k_smem.advance_offset_by_row<16, num_vecs_per_head>(k_smem_offset_r) -
+        2 * num_frags_y;
+    chunk_start_k += 16;
+  }
+
+  uint32_t chunk_start_v = tile_start + tid % 4 * 2;
+  uint32_t v_write_idx = block_id * write_n_stride +
+                         kv_head_idx * write_h_stride +
+                         (wid * num_frags_v * 16 + tid / 4) * write_d_stride +
+                         tid % 4 * 4;  // 4 * int8 = 8 * int4 = 32bit
+  const uint32_t num_frags_z_v = num_frags_z * NUM_WARPS;
+  T v_scales[num_frags_z_v * 4];
+  for (int v_i = 0; v_i < num_frags_z_v; v_i++) {
+    const int offset = v_i * 16;
+    const int t_offset = tid % 4 * 2;
+    v_scales[v_i * 4] = v_scale_smem[offset + t_offset];
+    v_scales[v_i * 4 + 1] = v_scale_smem[offset + t_offset + 1];
+    v_scales[v_i * 4 + 2] = v_scale_smem[offset + t_offset + 8];
+    v_scales[v_i * 4 + 3] = v_scale_smem[offset + t_offset + 9];
+  }
+
+#pragma unroll
+  for (uint32_t fy = 0; fy < num_frags_v; ++fy) {
+    uint32_t v_write_idx_now_v = v_write_idx + fy * 16 * write_d_stride;
+#pragma unroll
+    for (uint32_t fz = 0; fz < num_frags_z_v; ++fz) {
+      uint32_t v_write_idx_now = v_write_idx_now_v +
+                                 fz % 2 * 8 * write_d_stride +
+                                 fz / 2 * 32;  // + fz % 2 * 16;
+      // load
+      v_smem.ldmatrix_m8n8x4_trans(v_smem_offset_r, kv_frag);
+      // quant
+      T *v_frag_T = reinterpret_cast<T *>(kv_frag);
+      if (bf_pad_len != 0) {
+        Load<uint8_t, 4>(cache_v + v_write_idx_now, &cache_vec1);
+        Load<uint8_t, 4>(cache_v + v_write_idx_now + 16, &cache_vec2);
+      }
+#pragma unroll
+      for (uint32_t v_id = 0; v_id < 8; ++v_id) {
+        uint8_t uint_quant_value;
+        if (chunk_start_v + v_id % 2 + (v_id % 4) / 2 * 8 >= start_len &&
+            chunk_start_v + v_id % 2 + (v_id % 4) / 2 * 8 < end_len) {
+          uint_quant_value = QuantToC8<T, is_need_kv_quant, IsFP8>(v_scales[fz * 4 + v_id % 4], v_frag_T[v_id], 127.0f, -127.0f);
+          // store now
+        } else {
+          uint_quant_value = 0;
+        }
+        if (bf_pad_len != 0) {
+          if (v_id < 4) {
+            cache_vec1[v_id] |= uint_quant_value;
+          } else {
+            cache_vec2[v_id % 4] |= uint_quant_value;
+          }
+        } else {
+          if (v_id < 4) {
+            cache_vec1[v_id] = uint_quant_value;
+          } else {
+            cache_vec2[v_id % 4] = uint_quant_value;
+          }
+        }
+      }
+      // store
+      Store<uint8_t, 4>(cache_vec1, cache_v + v_write_idx_now);
+      Store<uint8_t, 4>(cache_vec2, cache_v + v_write_idx_now + 16);
+      chunk_start_v += 16;
+      v_smem_offset_r =
+          k_smem.advance_offset_by_row<16, num_vecs_per_head>(v_smem_offset_r);
+    }
+    v_smem_offset_r = k_smem.advance_offset_by_column<2>(
+                          v_smem_offset_r, wid * num_frags_v + fy) -
+                      16 * num_frags_z_v * num_vecs_per_head;
+    chunk_start_v -= 16 * num_frags_z_v;
+  }
+}
+
 // Write Cache KV in Append
 template <typename T,
           uint32_t num_frags_y,
@@ -2107,10 +2512,11 @@ void CascadeAppendWriteCacheKVC8QKV(
     int num_blocks_x_cpu,
     int max_seq_len,
     bool is_scale_channel_wise,
-    const bool is_fp8,
+    const std::string& cache_quant_type,
     cudaStream_t &stream,
     paddle::Tensor *cache_k_out,
     paddle::Tensor *cache_v_out) {
+  using NV_TYPE = typename cascade_attn_type_traits<T>::type;
   auto max_blocks_per_seq = meta_data.max_blocks_per_seq;
   auto num_tokens = meta_data.token_nums;
   auto num_heads = meta_data.q_num_heads;
@@ -2128,49 +2534,77 @@ void CascadeAppendWriteCacheKVC8QKV(
   dim3 blocks(32, num_warps);
 
   const uint32_t smem_size = (BLOCK_SIZE * HEAD_DIM) * sizeof(T) * 2;
-  auto kernel_fn = append_write_cache_kv_c8_qkv<T,
-                                                num_frags_y,
-                                                num_frags_z,
-                                                HEAD_DIM,
-                                                BLOCK_SIZE,
-                                                num_warps,
-                                                true, false>;
-  if (is_fp8) {
-    kernel_fn = append_write_cache_kv_c8_qkv<T,
-                                                num_frags_y,
-                                                num_frags_z,
-                                                HEAD_DIM,
-                                                BLOCK_SIZE,
-                                                num_warps,
-                                                true, true>;
+  if (cache_quant_type != "block_wise_fp8") {
+    auto kernel_fn = append_write_cache_kv_c8_qkv<T,
+                                                  num_frags_y,
+                                                  num_frags_z,
+                                                  HEAD_DIM,
+                                                  BLOCK_SIZE,
+                                                  num_warps,
+                                                  true, false>;
+    if (cache_quant_type == "cache_fp8") {
+      kernel_fn = append_write_cache_kv_c8_qkv<T,
+                                              num_frags_y,
+                                              num_frags_z,
+                                              HEAD_DIM,
+                                              BLOCK_SIZE,
+                                              num_warps,
+                                              true, true>;
+    }
+    if (is_scale_channel_wise) {
+      kernel_fn = append_write_cache_kv_c8_qkv<T,
+                                              num_frags_y,
+                                              num_frags_z,
+                                              HEAD_DIM,
+                                              BLOCK_SIZE,
+                                              num_warps,
+                                              false>;
+    }
+    cudaFuncSetAttribute(
+        kernel_fn, cudaFuncAttributeMaxDynamicSharedMemorySize, smem_size);
+    kernel_fn<<<grids, blocks, 0, stream>>>(cache_k_out->data<uint8_t>(),
+                                            cache_v_out->data<uint8_t>(),
+                                            qkv.data<T>(),
+                                            cache_k_scale.data<T>(),
+                                            cache_v_scale.data<T>(),
+                                            batch_ids.data<int>(),
+                                            tile_ids_per_batch.data<int>(),
+                                            seq_lens_this_time.data<int>(),
+                                            seq_lens_decoder.data<int>(),
+                                            batch_id_per_token.data<int>(),
+                                            cu_seqlens_q.data<int>(),
+                                            block_table.data<int>(),
+                                            max_seq_len,
+                                            max_blocks_per_seq,
+                                            num_heads,
+                                            kv_num_heads);
+  } else {
+    auto kernel_fn = append_write_cache_kv_c8_qkv_dynamic<NV_TYPE,
+                                                          num_frags_y,
+                                                          num_frags_z,
+                                                          HEAD_DIM,
+                                                          BLOCK_SIZE,
+                                                          num_warps,
+                                                          true, true>;
+    cudaFuncSetAttribute(
+        kernel_fn, cudaFuncAttributeMaxDynamicSharedMemorySize, smem_size);
+    kernel_fn<<<grids, blocks, 0, stream>>>(cache_k_out->data<uint8_t>(),
+                                            cache_v_out->data<uint8_t>(),
+                                            reinterpret_cast<const NV_TYPE*>(qkv.data<T>()),
+                                            const_cast<NV_TYPE*>(reinterpret_cast<const NV_TYPE*>(cache_k_scale.data<T>())),
+                                            const_cast<NV_TYPE*>(reinterpret_cast<const NV_TYPE*>(cache_v_scale.data<T>())),
+                                            batch_ids.data<int>(),
+                                            tile_ids_per_batch.data<int>(),
+                                            seq_lens_this_time.data<int>(),
+                                            seq_lens_decoder.data<int>(),
+                                            batch_id_per_token.data<int>(),
+                                            cu_seqlens_q.data<int>(),
+                                            block_table.data<int>(),
+                                            max_seq_len,
+                                            max_blocks_per_seq,
+                                            num_heads,
+                                            kv_num_heads);
   }
-  if (is_scale_channel_wise) {
-    kernel_fn = append_write_cache_kv_c8_qkv<T,
-                                             num_frags_y,
-                                             num_frags_z,
-                                             HEAD_DIM,
-                                             BLOCK_SIZE,
-                                             num_warps,
-                                             false>;
-  }
-  cudaFuncSetAttribute(
-      kernel_fn, cudaFuncAttributeMaxDynamicSharedMemorySize, smem_size);
-  kernel_fn<<<grids, blocks, 0, stream>>>(cache_k_out->data<uint8_t>(),
-                                          cache_v_out->data<uint8_t>(),
-                                          qkv.data<T>(),
-                                          cache_k_scale.data<T>(),
-                                          cache_v_scale.data<T>(),
-                                          batch_ids.data<int>(),
-                                          tile_ids_per_batch.data<int>(),
-                                          seq_lens_this_time.data<int>(),
-                                          seq_lens_decoder.data<int>(),
-                                          batch_id_per_token.data<int>(),
-                                          cu_seqlens_q.data<int>(),
-                                          block_table.data<int>(),
-                                          max_seq_len,
-                                          max_blocks_per_seq,
-                                          num_heads,
-                                          kv_num_heads);
 }
 
 template <typename T, uint32_t HEAD_DIM, uint32_t BLOCK_SIZE>

custom_ops/gpu_ops/append_attn/encoder_write_cache_with_rope_kernel.h

@@ -178,7 +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>(
@@ -198,7 +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, paddle::CPUPlace());
+    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

@@ -217,7 +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 =
@@ -235,7 +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
@@ -278,7 +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 =
@@ -296,7 +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
@@ -400,7 +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 =
@@ -418,7 +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
@@ -466,7 +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 =
@@ -485,7 +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
@@ -614,7 +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 =
@@ -632,7 +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);
 
@@ -685,7 +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 =
@@ -704,7 +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
@@ -1000,7 +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),
@@ -1018,7 +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/utils.cuh

@@ -441,6 +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;                         \

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,
@@ -105,7 +105,7 @@ void AppendAttentionWithOutput(
     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,
     paddle::Tensor &fmha_out,
     const paddle::optional<paddle::Tensor> &rotary_embs,
@@ -255,7 +255,8 @@ paddle::Tensor MoeExpertFFNFunc(
     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 int estimate_total_token_nums);
+    const int estimate_total_token_nums,
+    const int hadamard_block_size);
 
 paddle::Tensor MoeExpertFFNWint2Func(
     const paddle::Tensor& permute_input,
@@ -298,14 +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,
@@ -378,9 +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,
@@ -404,8 +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,
@@ -463,23 +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,
@@ -707,6 +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,
@@ -750,6 +773,7 @@ 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,
@@ -763,7 +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,
@@ -980,7 +1005,7 @@ 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"),
@@ -1023,7 +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");
@@ -1228,6 +1253,8 @@ 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");

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/get_padding_offset.cu

@@ -132,7 +132,7 @@ std::vector<paddle::DataType> GetPaddingOffsetInferDtype(
 }
 
 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",
               "batch_id_per_token",
               "cu_seqlens_q",

custom_ops/gpu_ops/helper.h

@@ -563,3 +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;
+}

custom_ops/gpu_ops/int8_gemm_with_cutlass/epilogue_tensor_op_int32.h

@@ -171,7 +171,7 @@ struct DefaultIteratorsTensorOp<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/mla_attn/batch_mla_with_paged_kv_cache.cu

@@ -70,7 +70,6 @@ void BatchMLAWithPagedKVCacheKernel(
     const paddle::optional<paddle::Tensor>& smooth_weight,  // [num_kv_heads, head_dim]
     const paddle::Tensor& seq_lens_this_time,
     const paddle::Tensor& seq_lens_decoder,
-    const paddle::Tensor& seq_lens_encoder,
     const paddle::Tensor& cu_seqlens_q,
     const paddle::Tensor& batch_id_per_token,
     const paddle::Tensor& block_tables,
@@ -78,9 +77,8 @@ void BatchMLAWithPagedKVCacheKernel(
     const paddle::Tensor& tile_ids_per_batch,
     const paddle::Tensor& num_blocks_x_device,
     const std::string& cache_quant_type_str,
-    const int num_blocks_x,
+    const paddle::Tensor& decoder_chunk_size_device,
     const int max_seq_len,
-    const int max_dec_len,
     const float softmax_scale,
     const float quant_max_bound,
     const float quant_min_bound,
@@ -97,14 +95,12 @@ void BatchMLAWithPagedKVCacheKernel(
   const auto q_head_num = meta_data.q_num_heads;
   const auto max_block_num_per_seq = meta_data.max_blocks_per_seq;
   const auto max_block_num = bsz * max_block_num_per_seq;
-  const uint32_t chunk_size = get_max_partition_size(bsz);
-
 
   int q_head_dim = meta_data.head_dims;
   int k_head_dim = meta_data.head_dims;
   int v_head_dim = meta_data.head_dims_v;
   // int num_chunks = max_dec_len / chunk_size;
-  int num_chunks = div_up(max_dec_len, chunk_size);
+  int num_chunks = div_up(max_seq_len, 64);
 
   auto *allocator = paddle::GetAllocator(q.place());
   phi::Allocator::AllocationPtr O_tmp, m_tmp, d_tmp;
@@ -127,14 +123,14 @@ void BatchMLAWithPagedKVCacheKernel(
   params.d = reinterpret_cast<float*>(d_tmp->ptr());
   params.block_tables = const_cast<int*>(block_tables.data<int>());
   params.seq_lens_this_time = const_cast<int*>(seq_lens_this_time.data<int>());
-  params.seq_lens_encoder = const_cast<int*>(seq_lens_encoder.data<int>());
   params.seq_lens_decoder = const_cast<int*>(seq_lens_decoder.data<int>());
   params.cumsum_q_seqlens = const_cast<int*>(cu_seqlens_q.data<int>());
   params.batch_id_per_token = const_cast<int*>(batch_id_per_token.data<int>());
   params.batch_ids = const_cast<int*>(batch_ids.data<int>());
   params.tile_ids_per_batch = const_cast<int*>(tile_ids_per_batch.data<int>());
   params.num_blocks_x = const_cast<int*>(num_blocks_x_device.data<int>());
-  params.num_blocks_x_int = num_blocks_x;
+  params.chunk_size_device =
+      const_cast<int*>(decoder_chunk_size_device.data<int>());
   params.q_stride_bsz = q_head_num * q_head_dim;
   params.q_stride_head_num = q_head_dim;
   params.kv_stride_block_num = block_size * k_head_dim;
@@ -151,7 +147,6 @@ void BatchMLAWithPagedKVCacheKernel(
   params.block_size = block_size;
   params.max_draft_token_num = draft_token_num;
   params.sm_scale = softmax_scale;
-  params.chunk_size = chunk_size;
   params.chunk_num = num_chunks;
 
   if (q_head_dim == 576) {
@@ -176,7 +171,6 @@ template void BatchMLAWithPagedKVCacheKernel<paddle::bfloat16>(
     const paddle::optional<paddle::Tensor>& smooth_weight,  // [num_kv_heads, head_dim]
     const paddle::Tensor& seq_lens_this_time,
     const paddle::Tensor& seq_lens_decoder,
-    const paddle::Tensor& seq_lens_encoder,
     const paddle::Tensor& cu_seqlens_q,
     const paddle::Tensor& batch_id_per_token,
     const paddle::Tensor& block_tables,
@@ -184,9 +178,8 @@ template void BatchMLAWithPagedKVCacheKernel<paddle::bfloat16>(
     const paddle::Tensor& tile_ids_per_batch,
     const paddle::Tensor& num_blocks_x_device,
     const std::string& cache_quant_type_str,
-    const int num_blocks_x,
+    const paddle::Tensor& decoder_chunk_size_device,
     const int max_seq_len,
-    const int max_dec_len,
     const float softmax_scale,
     const float quant_max_bound,
     const float quant_min_bound,
@@ -210,7 +203,6 @@ template void BatchMLAWithPagedKVCacheKernel<paddle::float16>(
     const paddle::optional<paddle::Tensor>& smooth_weight,  // [num_kv_heads, head_dim]
     const paddle::Tensor& seq_lens_this_time,
     const paddle::Tensor& seq_lens_decoder,
-    const paddle::Tensor& seq_lens_encoder,
     const paddle::Tensor& cu_seqlens_q,
     const paddle::Tensor& batch_id_per_token,
     const paddle::Tensor& block_tables,
@@ -218,9 +210,8 @@ template void BatchMLAWithPagedKVCacheKernel<paddle::float16>(
     const paddle::Tensor& tile_ids_per_batch,
     const paddle::Tensor& num_blocks_x_device,
     const std::string& cache_quant_type_str,
-    const int num_blocks_x,
+    const paddle::Tensor& decoder_chunk_size_device,
     const int max_seq_len,
-    const int max_dec_len,
     const float softmax_scale,
     const float quant_max_bound,
     const float quant_min_bound,

custom_ops/gpu_ops/mla_attn/batch_mla_with_paged_kv_cache.h

@@ -47,7 +47,6 @@ void BatchMLAWithPagedKVCacheKernel(
     const paddle::optional<paddle::Tensor>& smooth_weight,  // [num_kv_heads, head_dim]
     const paddle::Tensor& seq_lens_this_time,
     const paddle::Tensor& seq_lens_decoder,
-    const paddle::Tensor& seq_lens_encoder,
     const paddle::Tensor& cu_seqlens_q,
     const paddle::Tensor& batch_id_per_token,
     const paddle::Tensor& block_tables,
@@ -55,9 +54,8 @@ void BatchMLAWithPagedKVCacheKernel(
     const paddle::Tensor& tile_ids_per_batch,
     const paddle::Tensor& num_blocks_x_device,
     const std::string& cache_quant_type_str,
-    const int num_blocks_x,
+    const paddle::Tensor& decoder_chunk_size_device,
     const int max_seq_len,
-    const int max_dec_len,
     const float softmax_scale,
     const float quant_max_bound,
     const float quant_min_bound,

custom_ops/gpu_ops/mla_attn/mainloop_load.cuh

@@ -128,12 +128,13 @@ struct CollectiveMainloop {
     DTypeMD const* d_ptr;
     IdType const* kv_block_tables;
     IdType const* seq_lens_this_time;
-    IdType const* seq_lens_encoder;
+    // IdType const* seq_lens_encoder;
     IdType const* seq_lens_decoder;
     IdType const* cumsum_q_seqlens;
     IdType const* batch_ids;
     IdType const* tile_ids_per_batch;
     IdType const* num_blocks_x;
+    IdType const* chunk_size_device;
     float sm_scale;
     int bsz;
     int max_block_num;
@@ -144,7 +145,7 @@ struct CollectiveMainloop {
     int kv_stride_block_size;
     int o_stride_bsz;
     int o_stride_head_num;
-    int chunk_size;
+    // int chunk_size;
     int chunk_num;
     int max_draft_token_num;
   };
@@ -160,12 +161,13 @@ struct CollectiveMainloop {
     DTypeMD* d_ptr;
     IdType* kv_block_tables;
     IdType* seq_lens_this_time;
-    IdType* seq_lens_encoder;
+    // IdType* seq_lens_encoder;
     IdType* seq_lens_decoder;
     IdType* cumsum_q_seqlens;
     IdType* batch_ids;
     IdType* tile_ids_per_batch;
     IdType* num_blocks_x;
+    IdType* chunk_size_device;
     float sm_scale;
     int bsz;
     int max_block_num;
@@ -176,7 +178,7 @@ struct CollectiveMainloop {
     int kv_stride_block_size;
     int o_stride_bsz;
     int o_stride_head_num;
-    int chunk_size;
+    // int chunk_size;
     int chunk_num;
     int max_draft_token_num;
     TMA_KV tma_load_KV;
@@ -198,12 +200,13 @@ struct CollectiveMainloop {
             const_cast<DTypeMD*>(args.d_ptr),
             const_cast<IdType*>(args.kv_block_tables),
             const_cast<IdType*>(args.seq_lens_this_time),
-            const_cast<IdType*>(args.seq_lens_encoder),
+            // const_cast<IdType*>(args.seq_lens_encoder),
             const_cast<IdType*>(args.seq_lens_decoder),
             const_cast<IdType*>(args.cumsum_q_seqlens),
             const_cast<IdType*>(args.batch_ids),
             const_cast<IdType*>(args.tile_ids_per_batch),
             const_cast<IdType*>(args.num_blocks_x),
+            const_cast<IdType*>(args.chunk_size_device),
             args.sm_scale,
             args.bsz,
             args.max_block_num,
@@ -214,7 +217,7 @@ struct CollectiveMainloop {
             args.kv_stride_block_size,
             args.o_stride_bsz,
             args.o_stride_head_num,
-            args.chunk_size,
+            // args.chunk_size,
             args.chunk_num,
             args.max_draft_token_num,
             tma_load_KV
@@ -281,9 +284,9 @@ struct CollectiveMainloop {
     auto gmem_thr_copy_kv = gmem_tiled_copy_kv.get_slice(thread_idx);
 
     static constexpr int BLOCK_SHAPE_KV = get<1>(TileShape_QKD{});
-    const int start_len = tile_idx * mainloop_params.chunk_size;
+    const int start_len = tile_idx * mainloop_params.chunk_size_device[0];
     const int start_tile_idx = start_len / BLOCK_SHAPE_KV;
-    const int end_tile_idx = cute::ceil_div(min(start_len + mainloop_params.chunk_size, kv_len), BLOCK_SHAPE_KV) - 1;
+    const int end_tile_idx = cute::ceil_div(min(start_len + mainloop_params.chunk_size_device[0], kv_len), BLOCK_SHAPE_KV) - 1;
 
     auto kv_block_tables = make_tensor(make_gmem_ptr(mainloop_params.kv_block_tables), make_layout(make_shape(mainloop_params.bsz, mainloop_params.max_block_num_per_seq), make_stride(mainloop_params.max_block_num_per_seq, 1)));
 
@@ -322,9 +325,9 @@ struct CollectiveMainloop {
                       group_modes<0, 2>(sK), group_modes<0, 2>(gKV));
 
     static constexpr int BLOCK_SHAPE_KV = get<1>(TileShape_QKD{});
-    const int start_len = tile_idx * mainloop_params.chunk_size;
+    const int start_len = tile_idx * mainloop_params.chunk_size_device[0];
     const int start_tile_idx = start_len / BLOCK_SHAPE_KV;
-    const int end_tile_idx = cute::ceil_div(min(start_len + mainloop_params.chunk_size, kv_len), BLOCK_SHAPE_KV) - 1;
+    const int end_tile_idx = cute::ceil_div(min(start_len + mainloop_params.chunk_size_device[0], kv_len), BLOCK_SHAPE_KV) - 1;
 
     auto kv_block_tables = make_tensor(make_gmem_ptr(mainloop_params.kv_block_tables), make_layout(make_shape(mainloop_params.bsz, mainloop_params.max_block_num_per_seq), make_stride(mainloop_params.max_block_num_per_seq, 1)));
 

custom_ops/gpu_ops/mla_attn/mainloop_mma.cuh

@@ -57,7 +57,7 @@ CUTLASS_DEVICE void mma_f16(const Params& mainloop_params,
   using SmemLayoutVtOneStage = typename Ktraits::SmemLayoutVtOneStage;
   static_assert(is_rmem<FrgTensorO>::value, "O tensor must be rmem resident.");
 
-  const int chunk_num_this_seq = cute::ceil_div(kv_len, mainloop_params.chunk_size);
+  const int chunk_num_this_seq = cute::ceil_div(kv_len, mainloop_params.chunk_size_device[0]);
 
   static constexpr int BLOCK_SHAPE_Q = get<0>(TileShape_QKD{});
   static constexpr int BLOCK_SHAPE_KV = get<1>(TileShape_QKD{});
@@ -84,9 +84,9 @@ CUTLASS_DEVICE void mma_f16(const Params& mainloop_params,
   Tensor tOrV2 = threadMmaPVSS.partition_fragment_B(sVt_s2);
   Tensor tOrP_CS2 = threadMmaPVSS.partition_fragment_A(sPSS);
 
-  const int start_len = tile_idx * mainloop_params.chunk_size;
+  const int start_len = tile_idx * mainloop_params.chunk_size_device[0];
   const int start_tile_idx = start_len / BLOCK_SHAPE_KV;
-  const int end_tile_idx =cute::ceil_div(min(start_len + mainloop_params.chunk_size, kv_len), BLOCK_SHAPE_KV) - 1;
+  const int end_tile_idx =cute::ceil_div(min(start_len + mainloop_params.chunk_size_device[0], kv_len), BLOCK_SHAPE_KV) - 1;
   int kv_tile_idx = end_tile_idx;
 
   auto consumer_wait = [](auto& pipeline, auto& smem_pipe_read) {
@@ -263,7 +263,7 @@ CUTLASS_DEVICE void mma_f16_two_stages(const Params& mainloop_params,
   using SmemLayoutVtOneStage = typename Ktraits::SmemLayoutVtOneStage;
   static_assert(is_rmem<FrgTensorO>::value, "O tensor must be rmem resident.");
 
-  const int chunk_num_this_seq = cute::ceil_div(kv_len, mainloop_params.chunk_size);
+  const int chunk_num_this_seq = cute::ceil_div(kv_len, mainloop_params.chunk_size_device[0]);
 
   static constexpr int BLOCK_SHAPE_Q = get<0>(TileShape_QKD{});
   static constexpr int BLOCK_SHAPE_KV = get<1>(TileShape_QKD{});
@@ -295,9 +295,9 @@ CUTLASS_DEVICE void mma_f16_two_stages(const Params& mainloop_params,
   Tensor tOrV4 = threadMmaPVSS.partition_fragment_B(sVt_s4);
   Tensor tOrP_CS2 = threadMmaPVSS.partition_fragment_A(sPSS);
 
-  const int start_len = tile_idx * mainloop_params.chunk_size;
+  const int start_len = tile_idx * mainloop_params.chunk_size_device[0];
   const int start_tile_idx = start_len / BLOCK_SHAPE_KV;
-  const int end_tile_idx = cute::ceil_div(min(start_len + mainloop_params.chunk_size, kv_len), BLOCK_SHAPE_KV) - 1;
+  const int end_tile_idx = cute::ceil_div(min(start_len + mainloop_params.chunk_size_device[0], kv_len), BLOCK_SHAPE_KV) - 1;
   int kv_tile_idx = end_tile_idx;
 
   auto consumer_wait = [](auto& pipeline, auto& smem_pipe_read) {

custom_ops/gpu_ops/mla_attn/mla_hopper.cuh

@@ -62,13 +62,12 @@ struct Params {
     alignas(16) DTypeQ *Q; // [token_num, head_num, dim_head]
     alignas(16) DTypeKV *KV; // [max_block_num, block_size, dim_head]
     alignas(16) DTypeO *O; // [token_num, head_num, dim_head]
-    alignas(16) DTypeO *O_tmp; // [num_chunks, bsz, head_num, dim_head]
-    alignas(16) float *m; // [num_chunks, bsz * max_draft_token_num * head_num]
-    alignas(16) float *d; // [num_chunks, bsz * max_draft_token_num * head_num]
+    alignas(16) DTypeO *O_tmp; // [max_num_chunks, bsz, head_num, dim_head]
+    alignas(16) float *m; // [max_num_chunks, bsz * max_draft_token_num * head_num]
+    alignas(16) float *d; // [max_num_chunks, bsz * max_draft_token_num * head_num]
 
     alignas(16) IdType *block_tables;
     alignas(16) IdType *seq_lens_this_time;
-    alignas(16) IdType *seq_lens_encoder;
     alignas(16) IdType *seq_lens_decoder;
     alignas(16) IdType *cumsum_q_seqlens;
     alignas(16) IdType *batch_id_per_token;
@@ -76,7 +75,7 @@ struct Params {
     alignas(16) IdType *batch_ids;
     alignas(16) IdType *tile_ids_per_batch;
     alignas(16) IdType *num_blocks_x;
-
+    alignas(16) IdType *chunk_size_device;
 
     uint32_t q_stride_bsz;
     uint32_t q_stride_head_num;
@@ -96,9 +95,7 @@ struct Params {
     int vo_head_dim;
     int block_size;
     int max_draft_token_num;
-    int chunk_size;
     int chunk_num;
-    int num_blocks_x_int;
 
     float sm_scale;
 };
@@ -118,7 +115,7 @@ struct Params {
     return cudaErrorNotSupported;                            \
   }
 
-template <typename CollectiveMainloop, typename CollectiveEpilogue, typename Ktraits, bool CAUSAL, int SM_COUNT = 132, bool USE_REG_EALLOC=false, bool USE_FIXED_BLOCK=false>
+template <typename CollectiveMainloop, typename CollectiveEpilogue, typename Ktraits, bool CAUSAL, int SM_COUNT = 132, bool USE_REG_EALLOC=false, bool USE_FIXED_BLOCK=true>
 __global__ void __launch_bounds__(Ktraits::NUM_WARPS * cutlass::NumThreadsPerWarp, 1)
 MLAWithKVCacheKernel(CUTE_GRID_CONSTANT
                      typename CollectiveMainloop::Params const mainloop_params,
@@ -137,6 +134,7 @@ MLAWithKVCacheKernel(CUTE_GRID_CONSTANT
   static constexpr int BLOCK_SHAPE_Q = Ktraits::BLOCK_SHAPE_Q;
   static constexpr int BLOCK_SHAPE_KV = Ktraits::BLOCK_SHAPE_KV;
   const int num_blocks_x = mainloop_params.num_blocks_x[0];
+  const int chunk_size = mainloop_params.chunk_size_device[0];
 
   static constexpr bool use_tma_load_kv = CollectiveMainloop::USE_TMA_LOAD_KV;
 
@@ -205,58 +203,10 @@ MLAWithKVCacheKernel(CUTE_GRID_CONSTANT
 
     PipelineStateQ smem_pipe_write_q = cutlass::make_producer_start_state<MainloopPipelineQ>();
     PipelineState smem_pipe_write_kv = cutlass::make_producer_start_state<MainloopPipeline>();
-    if constexpr(USE_FIXED_BLOCK) {
-      for (int i = blockIdx.x; i < num_blocks_x; i += SM_COUNT) {
-        const int bid = mainloop_params.batch_ids[i];
-        const int tile_id = mainloop_params.tile_ids_per_batch[i];
-        const int seq_len_now = mainloop_params.seq_lens_this_time[bid];
-        const int seq_len_encoder_now = mainloop_params.seq_lens_encoder[bid];
-        const int seq_len_decoder_now = mainloop_params.seq_lens_decoder[bid] + seq_len_now;
-        const int start_token_idx = mainloop_params.cumsum_q_seqlens[bid];
-        cutlass::arch::NamedBarrier::sync(Ktraits::NUM_THREADS,
-                                          /*id=*/static_cast<int>(NamedBarriers::kWG0WG1WG2Sync));
-
-        // load Q
-        collective_mainloop.load_q(
-            mainloop_params,
-            pipeline_q,
-            smem_pipe_write_q,
-            shared_storage,
-            threadIdx.x,
-            bid);
-
-        if constexpr (!use_tma_load_kv) {
-          // load kv
-          collective_mainloop.load_kv(
-              mainloop_params,
-              pipeline_kv,
-              smem_pipe_write_kv,
-              shared_storage,
-              bid,
-              seq_len_decoder_now,
-              tile_id
-          );
-        } else {
-          if (warp_idx_in_warpgroup == 0) {
-            // load kv tma
-            collective_mainloop.load_kv_tma(
-                mainloop_params,
-                pipeline_kv,
-                smem_pipe_write_kv,
-                shared_storage,
-                bid,
-                seq_len_decoder_now,
-                tile_id
-            );
-          }
-        }
-      }
-    } else {
-      const int block_id = blockIdx.x;
-      const int bid = mainloop_params.batch_ids[block_id];
-      const int tile_id = mainloop_params.tile_ids_per_batch[block_id];
+    for (int i = blockIdx.x; i < num_blocks_x; i += SM_COUNT) {
+      const int bid = mainloop_params.batch_ids[i];
+      const int tile_id = mainloop_params.tile_ids_per_batch[i];
       const int seq_len_now = mainloop_params.seq_lens_this_time[bid];
-      const int seq_len_encoder_now = mainloop_params.seq_lens_encoder[bid];
       const int seq_len_decoder_now = mainloop_params.seq_lens_decoder[bid] + seq_len_now;
       const int start_token_idx = mainloop_params.cumsum_q_seqlens[bid];
       cutlass::arch::NamedBarrier::sync(Ktraits::NUM_THREADS,
@@ -309,76 +259,12 @@ MLAWithKVCacheKernel(CUTE_GRID_CONSTANT
     Tensor tOrO = partition_fragment_C(tiled_mma_pv, select<0, 1>(TileShape_PDV{}));
 
     auto attention_updater = OnlineSoftmax<2 * size<1>(tOrO), /*WITH_SCALE=*/true>(mainloop_params.sm_scale);
-    if constexpr(USE_FIXED_BLOCK) {
-      for (int i = blockIdx.x; i < num_blocks_x; i += SM_COUNT) {
-        clear(tOrO);
-        clear(attention_updater.scores_scale);
-        const int bid = mainloop_params.batch_ids[i];
-        const int tile_id = mainloop_params.tile_ids_per_batch[i];
-        const int seq_len_now = mainloop_params.seq_lens_this_time[bid];
-        const int seq_len_encoder_now = mainloop_params.seq_lens_encoder[bid];
-        const int seq_len_decoder_now = mainloop_params.seq_lens_decoder[bid] + seq_len_now;
-        const int start_token_idx = mainloop_params.cumsum_q_seqlens[bid];
-        cutlass::arch::NamedBarrier::sync(Ktraits::NUM_THREADS,
-                                          /*id=*/static_cast<int>(NamedBarriers::kWG0WG1WG2Sync));
-
-        if constexpr (BLOCK_SHAPE_KV == 64) {
-          mma_f16<Ktraits, CAUSAL>(
-            mainloop_params,
-            pipeline_q,
-            smem_pipe_read_q,
-            pipeline_kv,
-            smem_pipe_read_kv,
-            tOrO,
-            attention_updater,
-            threadIdx.x - NUM_COPY_THREADS,
-            bid,
-            seq_len_decoder_now,
-            seq_len_now,
-            tile_id,
-            shared_storage);
-        } else if (BLOCK_SHAPE_KV == 32) {
-          mma_f16_two_stages<Ktraits, CAUSAL>(
-            mainloop_params,
-            pipeline_q,
-            smem_pipe_read_q,
-            pipeline_kv,
-            smem_pipe_read_kv,
-            tOrO,
-            attention_updater,
-            threadIdx.x - NUM_COPY_THREADS,
-            bid,
-            seq_len_decoder_now,
-            seq_len_now,
-            tile_id,
-            shared_storage);
-        }
-
-        collective_epilogue.store(
-            epilogue_params,
-            tOrO,
-            attention_updater.get_lse(),
-            shared_storage,
-            tiled_mma_pv,
-            threadIdx.x - NUM_COPY_THREADS,
-            bid,
-            mainloop_params.bsz,
-            seq_len_now,
-            start_token_idx,
-            tile_id,
-            seq_len_decoder_now,
-            mainloop_params.chunk_size,
-            mainloop_params.max_draft_token_num,
-            mainloop_params.o_stride_bsz);
-      }
-    } else {
-      const int block_id = blockIdx.x;
+    for (int i = blockIdx.x; i < num_blocks_x; i += SM_COUNT) {
       clear(tOrO);
       clear(attention_updater.scores_scale);
-      const int bid = mainloop_params.batch_ids[block_id];
-      const int tile_id = mainloop_params.tile_ids_per_batch[block_id];
+      const int bid = mainloop_params.batch_ids[i];
+      const int tile_id = mainloop_params.tile_ids_per_batch[i];
       const int seq_len_now = mainloop_params.seq_lens_this_time[bid];
-      const int seq_len_encoder_now = mainloop_params.seq_lens_encoder[bid];
       const int seq_len_decoder_now = mainloop_params.seq_lens_decoder[bid] + seq_len_now;
       const int start_token_idx = mainloop_params.cumsum_q_seqlens[bid];
       cutlass::arch::NamedBarrier::sync(Ktraits::NUM_THREADS,
@@ -429,15 +315,15 @@ MLAWithKVCacheKernel(CUTE_GRID_CONSTANT
           start_token_idx,
           tile_id,
           seq_len_decoder_now,
-          mainloop_params.chunk_size,
+          chunk_size,
           mainloop_params.max_draft_token_num,
           mainloop_params.o_stride_bsz);
-    }
+      }
   }
 }
 
 
-template <typename KernelTraits, bool CAUSAL, typename Params, bool USE_REG_EALLOC=false, bool USE_FIXED_BLOCK=false>
+template <typename KernelTraits, bool CAUSAL, typename Params, bool USE_REG_EALLOC=false, bool USE_FIXED_BLOCK=true>
 cudaError_t BatchMLAWithPagedKVCacheKernelTraitsDispatched(Params& params,
                                                            cudaStream_t stream) {
   using DTypeQ = typename KernelTraits::DTypeQ;
@@ -460,12 +346,12 @@ cudaError_t BatchMLAWithPagedKVCacheKernelTraitsDispatched(Params& params,
       params.d,
       params.block_tables,
       params.seq_lens_this_time,
-      params.seq_lens_encoder,
       params.seq_lens_decoder,
       params.cumsum_q_seqlens,
       params.batch_ids,
       params.tile_ids_per_batch,
       params.num_blocks_x,
+      params.chunk_size_device,
       params.sm_scale,
       params.bsz,
       params.max_block_num,
@@ -476,7 +362,6 @@ cudaError_t BatchMLAWithPagedKVCacheKernelTraitsDispatched(Params& params,
       params.kv_stride_block_size,
       params.o_stride_bsz,
       params.o_stride_head_num,
-      params.chunk_size,
       params.chunk_num,
       params.max_draft_token_num
   });
@@ -500,13 +385,9 @@ cudaError_t BatchMLAWithPagedKVCacheKernelTraitsDispatched(Params& params,
   cudaOccupancyMaxActiveBlocksPerMultiprocessor(
       &act_blocks_per_sm, kernel, KernelTraits::NUM_WARPS * 32, smem_size);
 
-  int gridx;
-  if constexpr(USE_FIXED_BLOCK) {
-    gridx = multiprocessor_count;
-  } else {
-    gridx = params.num_blocks_x_int;
-  }
-  dim3 grid_dims = {gridx, 1, 1};
+  // NOTE: (changwenbin) Here the grid size is fixed so that MLA can be captured
+  // by the graph.
+  dim3 grid_dims = {multiprocessor_count, 1, 1};
   static constexpr int ctaSize = KernelTraits::NUM_WARPS * 32;
   dim3 block_dims(ctaSize, 1, 1);
   kernel<<<grid_dims, block_dims, smem_size, stream>>>(
@@ -517,37 +398,38 @@ cudaError_t BatchMLAWithPagedKVCacheKernelTraitsDispatched(Params& params,
     constexpr int merge_block_size = 256;
     constexpr int blockx = KernelTraits::HEAD_DIM_VO / vec_size;
     constexpr int blocky = (merge_block_size + blockx - 1) / blockx;
-    dim3 grids_merge(min(multiprocessor_count, params.token_num), params.q_num_head); // 128k is too large
+    dim3 grids_merge(multiprocessor_count, params.q_num_head); // 128k is too large
     dim3 blocks_merge(blockx, blocky);
-    merge_multi_chunks_kernel<NV_TYPE, vec_size, blocky, KernelTraits::HEAD_DIM_VO><<<grids_merge, blocks_merge, 0, stream>>>(
-      reinterpret_cast<NV_TYPE*>(params.O_tmp),
-      params.m,
-      params.d,
-      params.seq_lens_this_time,
-      params.seq_lens_decoder,
-      params.seq_lens_encoder,
-      params.cumsum_q_seqlens,
-      params.batch_id_per_token,
-      reinterpret_cast<NV_TYPE*>(params.O),
-      params.chunk_num,
-      params.q_num_head,
-      params.chunk_size,
-      params.vo_head_dim,
-      params.token_num,
-      params.bsz,
-      params.max_draft_token_num
-    );
+    merge_multi_chunks_kernel<NV_TYPE,
+                              vec_size,
+                              blocky,
+                              KernelTraits::HEAD_DIM_VO>
+        <<<grids_merge, blocks_merge, 0, stream>>>(
+            reinterpret_cast<NV_TYPE *>(params.O_tmp),
+            params.m,
+            params.d,
+            params.seq_lens_this_time,
+            params.seq_lens_decoder,
+            params.cumsum_q_seqlens,
+            params.batch_id_per_token,
+            params.chunk_size_device,
+            reinterpret_cast<NV_TYPE *>(params.O),
+            params.q_num_head,
+            params.vo_head_dim,
+            params.token_num,
+            params.bsz,
+            params.max_draft_token_num);
   }
   return cudaSuccess;
 }
 
-template <uint32_t HEAD_DIM_QK, uint32_t HEAD_DIM_VO, typename NV_TYPE, typename Params, bool USE_REG_EALLOC=false, bool USE_FIXED_BLOCK=false>
+template <uint32_t HEAD_DIM_QK, uint32_t HEAD_DIM_VO, typename NV_TYPE, typename Params, bool USE_REG_EALLOC=false, bool USE_FIXED_BLOCK=true>
 cudaError_t BatchMLAWithPagedKVCacheDispatched(Params& params, cudaStream_t stream) {
   constexpr bool CAUSAL = true;
   if constexpr (HEAD_DIM_QK == 576) {
     DISPATCH_GROUP_SIZE(params.q_num_head, GROUP_SIZE,
       BatchMLAWithPagedKVCacheKernelTraitsDispatched<
-          AttentionKernelTraits</*USE_TMA_LOAD_KV=*/false,
+          AttentionKernelTraits</*USE_TMA_LOAD_KV=*/true,
                                 HEAD_DIM_QK,
                                 HEAD_DIM_VO,
                                 GROUP_SIZE,

custom_ops/gpu_ops/mla_attn/utils.cuh

@@ -249,18 +249,16 @@ struct prefill_softmax_state_t {
 };
 
 template <typename T, int vec_size, uint32_t bdy, uint32_t HEAD_DIM>
-__global__ void merge_multi_chunks_kernel(const T * __restrict__ multi_out, // [num_chunks, bsz, max_draft_token, num_heads, head_dim]
-                                          const float * __restrict__ multi_m, // [num_chunks, bsz, max_draft_token, num_heads]
-                                          const float * __restrict__ multi_d, // [num_chunks, bsz, max_draft_token, num_heads]
+__global__ void merge_multi_chunks_kernel(const T * __restrict__ multi_out, // [max_num_chunks, bsz, max_draft_token, num_heads, head_dim]
+                                          const float * __restrict__ multi_m, // [max_num_chunks, bsz, max_draft_token, num_heads]
+                                          const float * __restrict__ multi_d, // [max_num_chunks, bsz, max_draft_token, num_heads]
                                           const int * __restrict__ seq_lens_this_time,
                                           const int * __restrict__ seq_lens_decoder,
-                                          const int * __restrict__ seq_lens_encoder,
                                           const int *__restrict__ cu_seqlens_q,
                                           const int * __restrict__ batch_id_per_token,
+                                          const int * __restrict__ chunk_size_device,
                                           T * __restrict__ out, // [token_num, num_heads, head_dim]
-                                          const int num_chunks,
                                           const int num_heads,
-                                          const int chunk_size,
                                           const int head_dim,
                                           const int token_num,
                                           const int bsz,
@@ -271,13 +269,15 @@ __global__ void merge_multi_chunks_kernel(const T * __restrict__ multi_out, // [
   __shared__ float md_smem[bdy * 2];
   for (int qid = blockIdx.x; qid < token_num; qid += gridDim.x) {
     const uint32_t bid = batch_id_per_token[qid];
+    // NOTE : (changwenbin) Batch_id_per_token is initialized to [:]=-1, Marking meaningless batch IDs.
+    if (bid == -1) continue;
     const int seq_len_q = seq_lens_this_time[bid];
     if (seq_len_q == 0) continue;
     const uint32_t local_seq_id = qid - cu_seqlens_q[bid];
     int seq_len_kv = seq_lens_decoder[bid];
     if (seq_len_kv == 0) continue;
     seq_len_kv += seq_len_q;
-    const int num_chunks_this_seq = cute::ceil_div(seq_len_kv, chunk_size);
+    const int num_chunks_this_seq = cute::ceil_div(seq_len_kv, chunk_size_device[0]);
     if (num_chunks_this_seq <= 1) {
       // not need merge
       continue;

custom_ops/gpu_ops/moba_attn/moba_decoder_attn/moba_decoder_attn.cu

@@ -383,7 +383,7 @@ __global__ __launch_bounds__(Kernel_traits::kNThreads) void moba_decoder_attenti
 
 
 template<typename Kernel_traits, typename ParamType>
-inline __device__ float caluate_logit_scale(const int partition_num, const int pack_max_partition_num, ParamType &params, char * shared_mem, const int seq_len, const int *qk_gate_topk_idx_ptr) {
+inline __device__ float calculate_logit_scale(const int partition_num, const int pack_max_partition_num, ParamType &params, char * shared_mem, const int seq_len, const int *qk_gate_topk_idx_ptr) {
     constexpr int32_t kNFloatPacksize = 16 / sizeof(float);
     constexpr int32_t kNReduceThreads = Kernel_traits::kNReduceThreads;
     const int32_t bi = blockIdx.z;
@@ -524,7 +524,7 @@ __global__ void __launch_bounds__(Kernel_traits::kNReduceThreads) moba_decoder_a
     const int kv_head_idx = head_idx / Kernel_traits::kGqaGroupSize;
     const int * qk_gate_topk_idx_ptr = params.qk_gate_topk_idx_ptr + (bi * params.kv_head_num + kv_head_idx) * Kernel_traits::kMaxN;
 
-    float inv_global_exp_sum = caluate_logit_scale<Kernel_traits>(partition_num, pack_max_partition_num, params, shared_mem, seq_len, qk_gate_topk_idx_ptr);
+    float inv_global_exp_sum = calculate_logit_scale<Kernel_traits>(partition_num, pack_max_partition_num, params, shared_mem, seq_len, qk_gate_topk_idx_ptr);
 
 
     using T_vec = Vec<cuteType, kNReducePacksize>;

custom_ops/gpu_ops/moba_attn/moba_encoder_attn/moba_encoder_write_cache.cu

@@ -40,7 +40,7 @@ __global__ void write_encoder_cachekv_c16(
 
     if (seq_len == 0) return;
 
-    const int ramian_tokens = seq_len - block_idx;
+    const int remain_tokens = seq_len - block_idx;
 
     const int32_t *block_table_now = block_tables + bidb * max_blocks_per_seq;
     const uint32_t physical_block_number = block_table_now[blockIdx.x + seq_len_decoder[bidb] / kBlockSize];
@@ -51,7 +51,7 @@ __global__ void write_encoder_cachekv_c16(
 
         #pragma unroll
         for (int i = row_idx; i < kBlockSize; i += 128 / (kHeadDim / kPackSize)) {
-            if (i < ramian_tokens) {
+            if (i < remain_tokens) {
                 *reinterpret_cast<float4*>(cache + i * kHeadDim) = *reinterpret_cast<const float4*>(k_input + base_load_idx + i * kv_head_num * kHeadDim);
             }
         }
@@ -62,7 +62,7 @@ __global__ void write_encoder_cachekv_c16(
 
         #pragma unroll
         for (int i = row_idx; i < kBlockSize; i += 128 / (kHeadDim / kPackSize)) {
-            if (i < ramian_tokens) {
+            if (i < remain_tokens) {
                 *reinterpret_cast<float4*>(cache + i * kHeadDim) = *reinterpret_cast<const float4*>(v_input + base_load_idx + i * kv_head_num * kHeadDim);
             }
         }

custom_ops/gpu_ops/moba_attn/moba_process/moba_get_kv_from_cache.cu

@@ -50,14 +50,14 @@ __global__ void get_kv_from_cache_c16_kernel(
     const int physical_block_number = block_tables[bidb * max_blocks_per_seq + block_idx];
 
 
-    const int ramian_tokens = seq_len - base_token_idx;
+    const int remain_tokens = seq_len - base_token_idx;
 
     if (bidh < kv_head_num) {
         const int cache_offset = physical_block_number * kv_head_num * kBlockSize * kHeadDim + bidh * kBlockSize * kHeadDim + col_idx;
         const int base_store_idx = (base_token_idx + cu_seq_k[bidb]) * kv_head_num * kHeadDim + bidh * kHeadDim + col_idx;
         #pragma unroll
         for (int i = row_idx; i < kBlockSize; i += 128 / (kHeadDim / kPackSize)) {
-            if (i < ramian_tokens) {
+            if (i < remain_tokens) {
                 *reinterpret_cast<float4*>(k_input + base_store_idx + i * kv_head_num * kHeadDim) = *reinterpret_cast<const float4*>(cache_k + cache_offset + i * kHeadDim);
             }
         }
@@ -67,7 +67,7 @@ __global__ void get_kv_from_cache_c16_kernel(
         const int base_store_idx = (base_token_idx + cu_seq_k[bidb]) * kv_head_num * kHeadDim + bidh * kHeadDim + col_idx;
         #pragma unroll
         for (int i = row_idx; i < kBlockSize; i += 128 / (kHeadDim / kPackSize)) {
-            if (i < ramian_tokens) {
+            if (i < remain_tokens) {
                 *reinterpret_cast<float4*>(v_input + base_store_idx + i * kv_head_num * kHeadDim) = *reinterpret_cast<const float4*>(cache_v + cache_offset + i * kHeadDim);
             }
         }

custom_ops/gpu_ops/moe/ep_moe_expert_dispatch.cu

@@ -448,137 +448,71 @@ void EPMoeDispatchKernel(const paddle::Tensor& input,
   auto place = input.place();
   const int gridx = min(132 * 8, num_rows);
   if (moe_quant_type == "w4a8") {
-    if (num_experts_per_rank == 8) {
-      permute_x_kernel<data_t, int8_t, 8><<<gridx, 512, 0, stream>>>(
-        input.data<data_t>(),
-        topk_ids.data<int64_t>(),
-        topk_weights.data<float>(),
-        token_nums_per_expert.data<int>(),
-        up_gate_proj_in_scale ? up_gate_proj_in_scale.get().data<float>() : nullptr,
-        moe_topk,
-        num_rows,
-        token_nums_this_rank,
-        hidden_size,
-        permute_input->data<int8_t>(),
-        permute_indices_per_token->data<int>(),
-        dst_weights->data<float>(),
-        dst_indices->data<int>(),
-        cumsum_idx_gpu->data<int>(),
-        token_nums_per_expert_cumsum->data<int64_t>(),
-        expert_idx_per_token->data<int64_t>(),
-        127.0,
-        -127.0
-      );
-    } else if (num_experts_per_rank == 16) {
-      permute_x_kernel<data_t, int8_t, 16><<<gridx, 512, 0, stream>>>(
-        input.data<data_t>(),
-        topk_ids.data<int64_t>(),
-        topk_weights.data<float>(),
-        token_nums_per_expert.data<int>(),
-        up_gate_proj_in_scale ? up_gate_proj_in_scale.get().data<float>() : nullptr,
-        moe_topk,
-        num_rows,
-        token_nums_this_rank,
-        hidden_size,
-        permute_input->data<int8_t>(),
-        permute_indices_per_token->data<int>(),
-        dst_weights->data<float>(),
-        dst_indices->data<int>(),
-        cumsum_idx_gpu->data<int>(),
-        token_nums_per_expert_cumsum->data<int64_t>(),
-        expert_idx_per_token->data<int64_t>(),
-        127.0,
-        -127.0
-      );
-    }
+    DISPATCH_NUM_EXPERTS_PER_RANK(num_experts_per_rank, NUM_EXPERTS_PER_RANK,
+    permute_x_kernel<data_t, int8_t, NUM_EXPERTS_PER_RANK><<<gridx, 512, 0, stream>>>(
+      input.data<data_t>(),
+      topk_ids.data<int64_t>(),
+      topk_weights.data<float>(),
+      token_nums_per_expert.data<int>(),
+      up_gate_proj_in_scale ? up_gate_proj_in_scale.get().data<float>() : nullptr,
+      moe_topk,
+      num_rows,
+      token_nums_this_rank,
+      hidden_size,
+      permute_input->data<int8_t>(),
+      permute_indices_per_token->data<int>(),
+      dst_weights->data<float>(),
+      dst_indices->data<int>(),
+      cumsum_idx_gpu->data<int>(),
+      token_nums_per_expert_cumsum->data<int64_t>(),
+      expert_idx_per_token->data<int64_t>(),
+      127.0,
+      -127.0
+    );)
   } else if (moe_quant_type == "w4afp8") {
-    if (num_experts_per_rank == 8) {
-      permute_x_kernel<data_t, data_t_fp8, 8, 512><<<gridx, 512, 0, stream>>>(
-        input.data<data_t>(),
-        topk_ids.data<int64_t>(),
-        topk_weights.data<float>(),
-        token_nums_per_expert.data<int>(),
-        up_gate_proj_in_scale ? up_gate_proj_in_scale.get().data<float>() : nullptr,
-        moe_topk,
-        num_rows,
-        token_nums_this_rank,
-        hidden_size,
-        permute_input->data<data_t_fp8>(),
-        permute_indices_per_token->data<int>(),
-        dst_weights->data<float>(),
-        dst_indices->data<int>(),
-        cumsum_idx_gpu->data<int>(),
-        token_nums_per_expert_cumsum->data<int64_t>(),
-        expert_idx_per_token->data<int64_t>(),
-        448.0f,
-        -448.0f
-      );
-    } else if (num_experts_per_rank == 16) {
-      permute_x_kernel<data_t, data_t_fp8, 16, 512><<<gridx, 512, 0, stream>>>(
-        input.data<data_t>(),
-        topk_ids.data<int64_t>(),
-        topk_weights.data<float>(),
-        token_nums_per_expert.data<int>(),
-        up_gate_proj_in_scale ? up_gate_proj_in_scale.get().data<float>() : nullptr,
-        moe_topk,
-        num_rows,
-        token_nums_this_rank,
-        hidden_size,
-        permute_input->data<data_t_fp8>(),
-        permute_indices_per_token->data<int>(),
-        dst_weights->data<float>(),
-        dst_indices->data<int>(),
-        cumsum_idx_gpu->data<int>(),
-        token_nums_per_expert_cumsum->data<int64_t>(),
-        expert_idx_per_token->data<int64_t>(),
-        448.0f,
-        -448.0f
-      );
-    }
+    DISPATCH_NUM_EXPERTS_PER_RANK(num_experts_per_rank, NUM_EXPERTS_PER_RANK,
+    permute_x_kernel<data_t, data_t_fp8, NUM_EXPERTS_PER_RANK, 512><<<gridx, 512, 0, stream>>>(
+      input.data<data_t>(),
+      topk_ids.data<int64_t>(),
+      topk_weights.data<float>(),
+      token_nums_per_expert.data<int>(),
+      up_gate_proj_in_scale ? up_gate_proj_in_scale.get().data<float>() : nullptr,
+      moe_topk,
+      num_rows,
+      token_nums_this_rank,
+      hidden_size,
+      permute_input->data<data_t_fp8>(),
+      permute_indices_per_token->data<int>(),
+      dst_weights->data<float>(),
+      dst_indices->data<int>(),
+      cumsum_idx_gpu->data<int>(),
+      token_nums_per_expert_cumsum->data<int64_t>(),
+      expert_idx_per_token->data<int64_t>(),
+      448.0f,
+      -448.0f
+    );)
   } else {
-    if (num_experts_per_rank == 8) {
-      permute_x_kernel<data_t, data_t, 8><<<gridx, 512, 0, stream>>>(
-        input.data<data_t>(),
-        topk_ids.data<int64_t>(),
-        topk_weights.data<float>(),
-        token_nums_per_expert.data<int>(),
-        up_gate_proj_in_scale ? up_gate_proj_in_scale.get().data<float>() : nullptr,
-        moe_topk,
-        num_rows,
-        token_nums_this_rank,
-        hidden_size,
-        permute_input->data<data_t>(),
-        permute_indices_per_token->data<int>(),
-        dst_weights->data<float>(),
-        dst_indices->data<int>(),
-        cumsum_idx_gpu->data<int>(),
-        token_nums_per_expert_cumsum->data<int64_t>(),
-        expert_idx_per_token->data<int64_t>(),
-        127.0,
-        -127.0
-      );
-    } else if (num_experts_per_rank == 16) {
-      permute_x_kernel<data_t, data_t, 16><<<gridx, 512, 0, stream>>>(
-        input.data<data_t>(),
-        topk_ids.data<int64_t>(),
-        topk_weights.data<float>(),
-        token_nums_per_expert.data<int>(),
-        up_gate_proj_in_scale ? up_gate_proj_in_scale.get().data<float>() : nullptr,
-        moe_topk,
-        num_rows,
-        token_nums_this_rank,
-        hidden_size,
-        permute_input->data<data_t>(),
-        permute_indices_per_token->data<int>(),
-        dst_weights->data<float>(),
-        dst_indices->data<int>(),
-        cumsum_idx_gpu->data<int>(),
-        token_nums_per_expert_cumsum->data<int64_t>(),
-        expert_idx_per_token->data<int64_t>(),
-        127.0,
-        -127.0
-      );
-    }
+    DISPATCH_NUM_EXPERTS_PER_RANK(num_experts_per_rank, NUM_EXPERTS_PER_RANK,
+    permute_x_kernel<data_t, data_t, NUM_EXPERTS_PER_RANK><<<gridx, 512, 0, stream>>>(
+      input.data<data_t>(),
+      topk_ids.data<int64_t>(),
+      topk_weights.data<float>(),
+      token_nums_per_expert.data<int>(),
+      up_gate_proj_in_scale ? up_gate_proj_in_scale.get().data<float>() : nullptr,
+      moe_topk,
+      num_rows,
+      token_nums_this_rank,
+      hidden_size,
+      permute_input->data<data_t>(),
+      permute_indices_per_token->data<int>(),
+      dst_weights->data<float>(),
+      dst_indices->data<int>(),
+      cumsum_idx_gpu->data<int>(),
+      token_nums_per_expert_cumsum->data<int64_t>(),
+      expert_idx_per_token->data<int64_t>(),
+      127.0,
+      -127.0
+    );)
   }
 }
 

custom_ops/gpu_ops/moe/fast_hardamard_kernel.cu

@@ -872,16 +872,14 @@ void MoeFastHardamardWrapper(const T *x_data,
                           const int64_t dim,
                           const int num_max_tokens_per_expert,
                           bool used_in_ep_low_latency,
+                          const int hadamard_block_size,
                           OutT* out,
                           cudaStream_t &stream) {
   bool FLAGS_hardamard_use_diagonal_block_matrix = true;
 
-  static const char* FLAGS_hardamard_moe_block_size = std::getenv("FLAGS_hardamard_moe_block_size");
-  static const int32_t hardamard_moe_block_size = FLAGS_hardamard_moe_block_size != nullptr ?
-    stoi(std::string(FLAGS_hardamard_moe_block_size)) : 512;
   constexpr int kThreads = 128;
   if (FLAGS_hardamard_use_diagonal_block_matrix) {
-    const int VecSize = hardamard_moe_block_size / kThreads; // 128 / 128 = 1
+    const int VecSize = hadamard_block_size / kThreads;
     const int logN = int(ceil(std::log2(kThreads * VecSize)));
     constexpr int kNChunks = 1;
     DISPATCH_SP_VS(VecSize, VEC_SIZE, {
@@ -991,6 +989,7 @@ template void MoeFastHardamardWrapper<phi::dtype::float16, phi::dtype::float16>(
   const int64_t dim,
   const int num_max_tokens_per_expert,
   bool used_in_ep_low_latency,
+  const int hadamard_block_size,
   phi::dtype::float16 *out,
   cudaStream_t &stream
 );
@@ -1009,6 +1008,7 @@ template void MoeFastHardamardWrapper<phi::dtype::float16, int8_t>(
   const int64_t dim,
   const int num_max_tokens_per_expert,
   bool used_in_ep_low_latency,
+  const int hadamard_block_size,
   int8_t *out,
   cudaStream_t &stream
 );
@@ -1027,6 +1027,7 @@ template void MoeFastHardamardWrapper<phi::dtype::bfloat16, phi::dtype::bfloat16
   const int64_t dim,
   const int num_max_tokens_per_expert,
   bool used_in_ep_low_latency,
+  const int hadamard_block_size,
   phi::dtype::bfloat16 *out,
   cudaStream_t &stream
 );
@@ -1045,6 +1046,7 @@ template void MoeFastHardamardWrapper<phi::dtype::bfloat16, int8_t>(
   const int64_t dim,
   const int num_max_tokens_per_expert,
   bool used_in_ep_low_latency,
+  const int hadamard_block_size,
   int8_t *out,
   cudaStream_t &stream
 );

custom_ops/gpu_ops/moe/fast_hardamard_kernel.h

@@ -32,5 +32,6 @@ void MoeFastHardamardWrapper(const T *x_data,
                             const int64_t dim,
                             const int num_max_tokens_per_expert,
                             bool used_in_ep_low_latency,
+                            const int hadamard_block_size,
                             OutT* out,
                             cudaStream_t &stream);

custom_ops/gpu_ops/moe/fused_moe_helper.h

@@ -236,7 +236,7 @@ public:
                                                num_experts, k, stream);
     }
 
-    topk_gating_softmax_kernelLauncher<float, int>::run(
+    topk_gating_softmax_kernelLauncher<float, int>(
         gating_output, nullptr, expert_scales_float, softmax_out_,
         expert_for_source_row, source_rows_, softmax_max_prob, num_rows,
         num_experts, k, group_moe, stream);
@@ -248,7 +248,7 @@ public:
                 permuted_experts_, source_rows_, permuted_rows_, k * num_rows,
                 false, stream);
 
-    initialize_moe_routing_kernelLauncher<T>::run(
+    initialize_moe_routing_kernelLauncher(
         input_activations, permuted_data_, permuted_rows_, nullptr, nullptr,
         expanded_source_row_to_expanded_dest_row, num_rows, num_rows,
         hidden_size, k, stream);
@@ -335,14 +335,14 @@ public:
             num_experts, down_proj_quant_args, stream);
       }
 
-      finalize_moe_routing_kernelLauncher<T>::run(
+      finalize_moe_routing_kernelLauncher(
           fc2_result, output_, fc2_expert_biases,
           reinterpret_cast<float *>(expert_scales_float),
           expanded_source_row_to_expanded_dest_row, expert_for_source_row,
           num_rows, hidden_size, k, static_cast<int>(1), norm_topk_prob,
           routed_scaling_factor, stream);
     } else {
-      finalize_moe_routing_kernelLauncher<T>::run(
+      finalize_moe_routing_kernelLauncher(
           // fc2_result,
           fc1_out, output_,
           fc1_expert_biases, // fc2_expert_biases,

custom_ops/gpu_ops/moe/fused_moe_op.h

@@ -1139,9 +1139,7 @@ void topk_gating_softmax_launcher_helper(const T* input,
 }
 
 template <typename T, typename IdxT = int>
-struct topk_gating_softmax_kernelLauncher{
-
-static void run(const T* input,
+void topk_gating_softmax_kernelLauncher(const T* input,
                                         const T* gating_correction_bias,
                                         T* output,
                                         T* softmax,
@@ -1221,7 +1219,6 @@ static void run(const T* input,
     }
   }
 }
-};
 
 // ========================== Permutation things
 // =======================================
@@ -1316,9 +1313,7 @@ __global__ void initialize_moe_routing_kernel(
 }
 
 template <typename T, typename OutT = T>
-struct initialize_moe_routing_kernelLauncher{
-
-static void run(
+void initialize_moe_routing_kernelLauncher(
     const T* unpermuted_input,
     OutT* permuted_output,
     const int* expanded_dest_row_to_expanded_source_row,
@@ -1361,7 +1356,6 @@ static void run(
             num_rows * k);
   }
 }
-};
 
 // ============================== Infer GEMM sizes
 // =================================
@@ -1472,8 +1466,7 @@ __global__ void finalize_moe_routing_kernel(
 }
 
 template <typename T>
-struct finalize_moe_routing_kernelLauncher{
-static void run(
+void finalize_moe_routing_kernelLauncher(
     const T* expanded_permuted_rows,
     T* reduced_unpermuted_output,
     const T* bias,
@@ -1505,5 +1498,4 @@ static void run(
             routed_scaling_factor,
             num_rows);
 }
-};
 }  // namespace phi

custom_ops/gpu_ops/moe/moe_dispatch.cu

@@ -36,6 +36,9 @@ void MoeDispatchKernel(
     paddle::Tensor *topk_idx, paddle::Tensor *expert_idx_per_token) {
   using namespace phi;
 
+  if (num_rows == 0){
+    return;
+  }
   typedef PDTraits<T> traits_;
   typedef typename traits_::DataType DataType_;
   typedef typename traits_::data_t data_t;
@@ -80,7 +83,7 @@ void MoeDispatchKernel(
   if (group_moe) {
     paddle::Tensor softmax_max_prob_tensor =
         GetEmptyTensor({num_rows, moe_topk}, paddle::DataType::FLOAT32, place);
-    // (TODO: check fill sucess ?)
+    // (TODO: check fill success ?)
     paddle::experimental::fill(softmax_max_prob_tensor, 0.f);
     softmax_max_prob = softmax_max_prob_tensor.data<float>();
   }
@@ -100,7 +103,7 @@ void MoeDispatchKernel(
     softmax_out_ = nullptr;
   }
 
-  topk_gating_softmax_kernelLauncher<float, int>::run(
+  topk_gating_softmax_kernelLauncher(
       gating_output.data<float>(),
       gating_correction_bias ? gating_correction_bias.get().data<float>()
                              : nullptr,
@@ -114,13 +117,13 @@ void MoeDispatchKernel(
 
   if (w4a8_in_scale) {
     if (permute_input->dtype() == paddle::DataType::INT8) {
-      initialize_moe_routing_kernelLauncher<data_t, int8_t>::run(
+      initialize_moe_routing_kernelLauncher(
         input.data<data_t>(), permute_input->data<int8_t>(), permuted_rows_,
         expert_idx_per_token->data<int32_t>(), w4a8_in_scale->data<float>(),
         permute_indices_per_token->data<int32_t>(), num_rows, num_rows,
         hidden_size, moe_topk, stream);
     } else if (permute_input->dtype() == paddle::DataType::FLOAT8_E4M3FN) {
-      initialize_moe_routing_kernelLauncher<data_t, float8_e4m3fn>::run(
+      initialize_moe_routing_kernelLauncher(
         input.data<data_t>(), permute_input->data<float8_e4m3fn>(),
         permuted_rows_, expert_idx_per_token->data<int32_t>(),
         w4a8_in_scale->data<float>(),
@@ -128,7 +131,7 @@ void MoeDispatchKernel(
         hidden_size, moe_topk, stream);
     }
   } else {
-    initialize_moe_routing_kernelLauncher<data_t>::run(
+    initialize_moe_routing_kernelLauncher(
         input.data<data_t>(), permute_input->data<data_t>(), permuted_rows_,
         expert_idx_per_token->data<int32_t>(), nullptr,
         permute_indices_per_token->data<int32_t>(), num_rows, num_rows,
@@ -185,6 +188,15 @@ std::vector<paddle::Tensor> MoeExpertDispatch(
   auto expert_idx_per_token =
       GetEmptyTensor({num_rows * moe_topk}, paddle::DataType::INT32, place);
 
+  if (token_rows == 0){
+    return {permute_input,
+            tokens_expert_prefix_sum,
+            permute_indices_per_token,
+            topk_weight,
+            topk_idx,
+            expert_idx_per_token};
+  }
+
   switch (input_type) {
   case paddle::DataType::BFLOAT16:
     MoeDispatchKernel<paddle::DataType::BFLOAT16>(

custom_ops/gpu_ops/moe/moe_ffn.cu

@@ -35,7 +35,8 @@ void MoeFFNKernel(const paddle::Tensor& permute_input,
                   const std::string& quant_method,
                   paddle::Tensor ffn_out,
                   bool used_in_ep_low_latency,
-                  const int estimate_total_token_nums) {
+                  const int estimate_total_token_nums,
+                  const int hadamard_block_size) {
     using namespace phi;
     typedef PDTraits<T> traits_;
     typedef typename traits_::DataType DataType_;
@@ -291,6 +292,7 @@ void MoeFFNKernel(const paddle::Tensor& permute_input,
             inter_size / 2,
             num_max_tokens_per_expert,
             used_in_ep_low_latency,
+            hadamard_block_size,
             reinterpret_cast<int8_t *>(int8_act_out->ptr()),
             stream
         );
@@ -340,6 +342,7 @@ void MoeFFNKernel(const paddle::Tensor& permute_input,
             inter_size / 2,
             num_max_tokens_per_expert,
             used_in_ep_low_latency,
+            hadamard_block_size,
             act_out_tensor.data<data_t>(),
             stream
         );
@@ -403,13 +406,15 @@ paddle::Tensor MoeExpertFFNFunc(
     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 int estimate_total_token_nums) {
+    const int estimate_total_token_nums, const int hadamard_block_size) {
 
 const auto t_type = (quant_method == "w4a8") ? up_gate_proj_scale.get().dtype() :
                     (quant_method == "w4afp8") ? paddle::DataType::BFLOAT16 :
                     permute_input.dtype();
     auto ffn_out = paddle::empty_like(permute_input, t_type);
-
+    if(permute_input.numel() == 0){
+        return ffn_out;
+    }
     switch (t_type) {
         case paddle::DataType::BFLOAT16:
             MoeFFNKernel<paddle::DataType::BFLOAT16>(permute_input,
@@ -424,7 +429,8 @@ const auto t_type = (quant_method == "w4a8") ? up_gate_proj_scale.get().dtype()
                                                      quant_method,
                                                      ffn_out,
                                                      used_in_ep_low_latency,
-                                                     estimate_total_token_nums);
+                                                     estimate_total_token_nums,
+                                                     hadamard_block_size);
             break;
         case paddle::DataType::FLOAT16:
             MoeFFNKernel<paddle::DataType::FLOAT16>(permute_input,
@@ -439,7 +445,8 @@ const auto t_type = (quant_method == "w4a8") ? up_gate_proj_scale.get().dtype()
                                                     quant_method,
                                                     ffn_out,
                                                     used_in_ep_low_latency,
-                                                    estimate_total_token_nums);
+                                                    estimate_total_token_nums,
+                                                    hadamard_block_size);
             break;
         default:
             PD_THROW("Unsupported data type for MoeExpertFFN");
@@ -458,7 +465,8 @@ std::vector<paddle::Tensor> MoeExpertFFN(
     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 int estimate_total_token_nums) {
+    const int estimate_total_token_nums,
+    const int hadamard_block_size) {
     return {MoeExpertFFNFunc(permute_input,
                              tokens_expert_prefix_sum,
                              up_gate_proj_weight,
@@ -470,7 +478,8 @@ std::vector<paddle::Tensor> MoeExpertFFN(
                              expert_idx_per_token,
                              quant_method,
                              used_in_ep_low_latency,
-                             estimate_total_token_nums)};
+                             estimate_total_token_nums,
+                             hadamard_block_size)};
 }
 
 std::vector<std::vector<int64_t>> MoeExpertFFNInferShape(
@@ -485,7 +494,8 @@ std::vector<std::vector<int64_t>> MoeExpertFFNInferShape(
     const paddle::optional<std::vector<int64_t>>& expert_idx_per_token_shape,
     const std::string& quant_method,
     const bool used_in_ep_low_latency,
-    const int estimate_total_token_nums) {
+    const int estimate_total_token_nums,
+    const int hadamard_block_size) {
     return {permute_input_shape};
 }
 
@@ -499,7 +509,7 @@ std::vector<paddle::DataType> MoeExpertFFNInferDtype(
     const paddle::optional<paddle::DataType> &down_proj_scale_dtype,
     const paddle::optional<paddle::DataType> &down_proj_in_scale_dtype,
     const std::string &quant_method, const bool used_in_ep_low_latency,
-    const int estimate_total_token_nums) {
+    const int estimate_total_token_nums, const int hadamard_block_size) {
   if (quant_method == "w4a8" || quant_method == "w4afp8") {
     return {up_gate_proj_scale_dtype.get()};
   } else {
@@ -555,6 +565,8 @@ std::vector<paddle::DataType> MoeExpertFFNInferDtype(
  *                 Options: "none", "weight_only_int4", "weight_only_int8", "w4a8"
  *   - used_in_ep_low_latency: Whether running in low latency mode
  *                            Affects activation function implementation
+ *   - estimate_total_token_nums: estimate total token numbers
+ *   - hadamard_block_size: hadamard block size for w4a8/w4afp8 quantization
  *
  * Note:
  * - w4a8 mode requires additional workspace memory allocation
@@ -571,7 +583,7 @@ PD_BUILD_STATIC_OP(moe_expert_ffn)
              paddle::Optional("down_proj_in_scale"),
              paddle::Optional("expert_idx_per_token")})
     .Outputs({"output_tensor"})
-    .Attrs({"quant_method:std::string", "used_in_ep_low_latency:bool", "estimate_total_token_nums:int"})
+    .Attrs({"quant_method:std::string", "used_in_ep_low_latency:bool", "estimate_total_token_nums:int", "hadamard_block_size:int"})
     .SetKernelFn(PD_KERNEL(MoeExpertFFN))
     .SetInferShapeFn(PD_INFER_SHAPE(MoeExpertFFNInferShape))
     .SetInferDtypeFn(PD_INFER_DTYPE(MoeExpertFFNInferDtype));

custom_ops/gpu_ops/moe/moe_reduce.cu

@@ -36,7 +36,7 @@ void MoeReduceKernel(const paddle::Tensor &ffn_out,
   typedef typename traits_::data_t data_t;
   auto stream = ffn_out.stream();
 
-  finalize_moe_routing_kernelLauncher<data_t>::run(
+  finalize_moe_routing_kernelLauncher(
       ffn_out.data<data_t>(), output->data<data_t>(),
       down_proj_bias ? down_proj_bias->data<data_t>() : nullptr,
       top_k_weight.data<float>(), permute_indices_per_token.data<int32_t>(),
@@ -59,6 +59,10 @@ paddle::Tensor MoeExpertReduceFunc(
 
   auto output = GetEmptyTensor({num_rows, hidden_size}, input_type, place);
 
+  if(num_rows == 0){
+    return output;
+  }
+
   switch (input_type) {
   case paddle::DataType::BFLOAT16:
     MoeReduceKernel<paddle::DataType::BFLOAT16>(

custom_ops/gpu_ops/multi_head_latent_attention.cu

@@ -22,23 +22,18 @@ std::vector<paddle::Tensor> MultiHeadLatentAttentionKernel(
     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_chunk_size_device,
     const paddle::Tensor& max_dec_len_this_time,
     const paddle::Tensor& max_len_kv,
     const paddle::optional<paddle::Tensor>& attn_mask,
@@ -64,9 +59,12 @@ std::vector<paddle::Tensor> MultiHeadLatentAttentionKernel(
   typedef PDTraits<D> traits_;
   typedef typename traits_::data_t data_t;
 
-  int decoder_num_blocks_data = decoder_num_blocks_cpu.data<int>()[0];
+  // NOTE: (changwenbin) In cuda graph, it will be fixed in the capture stage
+  // int decoder_num_blocks_data = decoder_num_blocks_cpu.data<int>()[0];
   int max_dec_len_this_time_data = max_dec_len_this_time.data<int>()[0];
   int max_len_kv_data = max_len_kv.data<int>()[0];
+  // int chunk_size = decoder_chunk_size_cpu.data<int>()[0];
+  //
 
   const bool mla_use_tensorcore = get_mla_use_tensorcore();
   auto sm_version = GetSMVersion();
@@ -96,7 +94,6 @@ std::vector<paddle::Tensor> MultiHeadLatentAttentionKernel(
                                              out_linear_smooths,
                                              seq_lens_this_time,
                                              seq_lens_decoder,
-                                             seq_lens_encoder,
                                              cu_seqlens_q,
                                              batch_id_per_token,
                                              block_tables,
@@ -104,9 +101,8 @@ std::vector<paddle::Tensor> MultiHeadLatentAttentionKernel(
                                              decoder_tile_ids_per_batch,
                                              decoder_num_blocks,
                                              cache_quant_type_str,
-                                             decoder_num_blocks_data,
+                                             decoder_chunk_size_device,
                                              max_input_length,
-                                             max_len_kv_data,
                                              softmax_scale,
                                              quant_max_bound,
                                              quant_min_bound,
@@ -145,23 +141,18 @@ 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_chunk_size_device,
     const paddle::Tensor& max_dec_len_this_time,
     const paddle::Tensor& max_len_kv,
     const paddle::optional<paddle::Tensor>& attn_mask,
@@ -208,23 +199,18 @@ std::vector<paddle::Tensor> MultiHeadLatentAttention(
           query,
           key_cache,
           value_cache,
-          seq_lens_encoder,
           seq_lens_decoder,
           seq_lens_this_time,
           cu_seqlens_q,
           batch_id_per_token,
           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,
-          decoder_num_blocks_cpu,
-          max_enc_len_this_time,
+          decoder_chunk_size_device,
           max_dec_len_this_time,
           max_len_kv,
           attn_mask,
@@ -254,23 +240,18 @@ std::vector<paddle::Tensor> MultiHeadLatentAttention(
           query,
           key_cache,
           value_cache,
-          seq_lens_encoder,
           seq_lens_decoder,
           seq_lens_this_time,
           cu_seqlens_q,
           batch_id_per_token,
           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,
-          decoder_num_blocks_cpu,
-          max_enc_len_this_time,
+          decoder_chunk_size_device,
           max_dec_len_this_time,
           max_len_kv,
           attn_mask,
@@ -307,23 +288,18 @@ std::vector<std::vector<int64_t>> MultiHeadLatentAttentionInferShape(
     const std::vector<int64_t>& query_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>& cu_seqlens_q_shape,
     const std::vector<int64_t>& batch_id_per_token_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>& decoder_num_blocks_cpu_shape,
-    const std::vector<int64_t>& max_enc_len_this_time_shape,
+    const std::vector<int64_t>& decoder_chunk_size_device_shape,
     const std::vector<int64_t>& max_dec_len_this_time_shape,
     const std::vector<int64_t>& max_len_kv_shape,
     const paddle::optional<std::vector<int64_t>>& attn_mask_shape,
@@ -361,23 +337,18 @@ std::vector<paddle::DataType> MultiHeadLatentAttentionInferDtype(
     const paddle::DataType& query_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& cu_seqlens_q_dtype,
     const paddle::DataType& batch_id_per_token_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& decoder_num_blocks_cpu_dtype,
-    const paddle::DataType& max_enc_len_this_time_dtype,
+    const paddle::DataType& decoder_chunk_size_device_dtype,
     const paddle::DataType& max_dec_len_this_time_dtype,
     const paddle::DataType& max_len_kv_dtype,
     const paddle::optional<paddle::DataType>& attn_mask_dtype,
@@ -415,23 +386,18 @@ PD_BUILD_STATIC_OP(multi_head_latent_attention)
     .Inputs({"query",
              "key_cache",
              "value_cache",
-             "seq_lens_encoder",
              "seq_lens_decoder",
              "seq_lens_this_time",
              "cu_seqlens_q",
              "batch_id_per_token",
              "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",
-             "decoder_num_blocks_cpu",
-             "max_enc_len_this_time",
+             "decoder_chunk_size_device",
              "max_dec_len_this_time",
              "max_len_kv",
              paddle::Optional("attn_mask"),

custom_ops/gpu_ops/recover_decode_task.cu

@@ -15,31 +15,72 @@
 #include "helper.h"
 
 __global__ void recover_decode_task(bool *stop_flags,
-                                   int *seq_lens_this_time,
-                                   int *seq_lens_encoder,
-                                   int *seq_lens_decoder,
-                                   int *step_seq_lens_decoder,
-                                   int *block_tables,
-                                   bool *is_block_step,
-                                   const int bsz,
-                                   const int block_num_per_seq,
-                                   const int block_size) {
+                                    int *seq_lens_this_time,
+                                    int *seq_lens_encoder,
+                                    int *seq_lens_decoder,
+                                    int *step_seq_lens_decoder,
+                                    int *block_tables,
+                                    bool *is_block_step,
+                                    const int bsz,
+                                    const int block_num_per_seq,
+                                    const int block_size) {
     int thread_idx = threadIdx.x;
     if (thread_idx < bsz) {
         if(is_block_step[thread_idx] == true) {
             int *block_table_now = block_tables + thread_idx * block_num_per_seq;
             if (block_table_now[step_seq_lens_decoder[thread_idx] / block_size] != -1) {
-                    // can be recovered for decoding
-                    is_block_step[thread_idx] = false;
-                    seq_lens_this_time[thread_idx]= 1;
-                    stop_flags[thread_idx] = false;
-                    seq_lens_encoder[thread_idx] = 0;
-                    seq_lens_decoder[thread_idx] = step_seq_lens_decoder[thread_idx];
-                }
+                // can be recovered for decoding
+                is_block_step[thread_idx] = false;
+                seq_lens_this_time[thread_idx]= 1;
+                stop_flags[thread_idx] = false;
+                seq_lens_encoder[thread_idx] = 0;
+                seq_lens_decoder[thread_idx] = step_seq_lens_decoder[thread_idx];
+
+            }
         }
     }
 }
 
+__global__ void recover_spec_decode_task(bool *stop_flags,
+                                        int *seq_lens_this_time,
+                                        int *seq_lens_encoder,
+                                        int *seq_lens_decoder,
+                                        int *step_seq_lens_decoder,
+                                        int *block_tables,
+                                        bool *is_block_step,
+                                        int64_t *draft_tokens,
+                                        const int64_t *step_draft_tokens,
+                                        const int *step_seq_lens_this_time,
+                                        const int bsz,
+                                        const int block_num_per_seq,
+                                        const int block_size,
+                                        const int draft_tokens_len,
+                                        const int num_extra_tokens) {
+    int thread_idx = threadIdx.x;
+    if (thread_idx < bsz) {
+        if(is_block_step[thread_idx] == true) {
+            int *block_table_now = block_tables + thread_idx * block_num_per_seq;
+            int max_possible_block_idx = (step_seq_lens_decoder[thread_idx] + num_extra_tokens) / block_size;
+            max_possible_block_idx = min(max_possible_block_idx, block_num_per_seq);
+            if (block_table_now[max_possible_block_idx] != -1) {
+                // can be recovered for decoding
+                int64_t *draft_tokens_now = draft_tokens + thread_idx * draft_tokens_len;
+                const int64_t *step_draft_tokens_now = step_draft_tokens + thread_idx * draft_tokens_len;
+                is_block_step[thread_idx] = false;
+                seq_lens_this_time[thread_idx] = step_seq_lens_this_time[thread_idx];
+                stop_flags[thread_idx] = false;
+                seq_lens_encoder[thread_idx] = 0;
+                seq_lens_decoder[thread_idx] = step_seq_lens_decoder[thread_idx];
+                for (int i = 0; i < seq_lens_this_time[thread_idx]; i++) {
+                    draft_tokens_now[i] = step_draft_tokens_now[i];
+                }
+
+            }
+        }
+    }
+}
+
+
 void RecoverDecodeTask(const paddle::Tensor &stop_flags,
                    const paddle::Tensor &seq_lens_this_time,
                    const paddle::Tensor &seq_lens_encoder,
@@ -47,7 +88,11 @@ 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) {
 #ifdef PADDLE_WITH_CUSTOM_DEVICE
     auto dev_ctx = static_cast<const phi::CustomContext*>(paddle::experimental::DeviceContextPool::Instance().Get(seq_lens_this_time.place()));
     auto cu_stream = dev_ctx->stream();
@@ -56,17 +101,38 @@ void RecoverDecodeTask(const paddle::Tensor &stop_flags,
 #endif
     const int bsz = seq_lens_this_time.shape()[0];
     const int block_num_per_seq = block_tables.shape()[1];
-    recover_decode_task<<<1, 1024, 0, cu_stream>>>(
-        const_cast<bool *>(stop_flags.data<bool>()),
-        const_cast<int *>(seq_lens_this_time.data<int>()),
-        const_cast<int *>(seq_lens_encoder.data<int>()),
-        const_cast<int *>(seq_lens_decoder.data<int>()),
-        const_cast<int *>(step_seq_lens_decoder.data<int>()),
-        const_cast<int *>(block_tables.data<int>()),
-        const_cast<bool *>(is_block_step.data<bool>()),
-        bsz,
-        block_num_per_seq,
-        block_size);
+    if (draft_tokens) {
+        const int draft_tokens_len = draft_tokens.get_ptr()->shape()[1];
+        recover_spec_decode_task<<<1, 1024, 0, cu_stream>>>(
+            const_cast<bool *>(stop_flags.data<bool>()),
+            const_cast<int *>(seq_lens_this_time.data<int>()),
+            const_cast<int *>(seq_lens_encoder.data<int>()),
+            const_cast<int *>(seq_lens_decoder.data<int>()),
+            const_cast<int *>(step_seq_lens_decoder.data<int>()),
+            const_cast<int *>(block_tables.data<int>()),
+            const_cast<bool *>(is_block_step.data<bool>()),
+            const_cast<int64_t *>(draft_tokens.get_ptr()->data<int64_t>()),
+            step_draft_tokens.get_ptr()->data<int64_t>(),
+            step_seq_lens_this_time.get_ptr()->data<int>(),
+            bsz,
+            block_num_per_seq,
+            block_size,
+            draft_tokens_len,
+            max_draft_tokens * 2 + 1);
+
+    } else {
+        recover_decode_task<<<1, 1024, 0, cu_stream>>>(
+            const_cast<bool *>(stop_flags.data<bool>()),
+            const_cast<int *>(seq_lens_this_time.data<int>()),
+            const_cast<int *>(seq_lens_encoder.data<int>()),
+            const_cast<int *>(seq_lens_decoder.data<int>()),
+            const_cast<int *>(step_seq_lens_decoder.data<int>()),
+            const_cast<int *>(block_tables.data<int>()),
+            const_cast<bool *>(is_block_step.data<bool>()),
+            bsz,
+            block_num_per_seq,
+            block_size);
+    }
 }
 
 PD_BUILD_STATIC_OP(recover_decode_task)
@@ -76,8 +142,11 @@ PD_BUILD_STATIC_OP(recover_decode_task)
              "seq_lens_decoder",
              "step_seq_lens_decoder",
              "block_tables",
-             "is_block_step"})
-    .Attrs({"block_size: int"})
+             "is_block_step",
+             paddle::Optional("draft_tokens"),
+             paddle::Optional("step_draft_tokens"),
+             paddle::Optional("step_seq_lens_this_time")})
+    .Attrs({"block_size: int", "max_draft_tokens: int"})
     .Outputs({"seq_lens_this_time_out",
               "seq_lens_encoder_out",
               "seq_lens_decoder_out",

custom_ops/gpu_ops/save_output_msg_with_topk.cc

@@ -75,7 +75,7 @@ void SaveOutMmsgTopK(const paddle::Tensor& x,
         std::string inference_msg_id_env_str(inference_msg_id_env_p);
         inference_msg_id_from_env = std::stoi(inference_msg_id_env_str);
         if (inference_msg_id_from_env == 2) {
-            // 2 and -2 is perserve for no-output indication.
+            // 2 and -2 is preserve for no-output indication.
             throw std::runtime_error(
                 " INFERENCE_MSG_ID cannot be 2, please use other number.");
         }

custom_ops/gpu_ops/save_with_output_msg.cc

@@ -45,7 +45,7 @@ void save_kernel(const paddle::Tensor& x,
         std::string inference_msg_id_env_str(inference_msg_id_env_p);
         inference_msg_id_from_env = std::stoi(inference_msg_id_env_str);
         if (inference_msg_id_from_env == 2) {
-            // 2 and -2 is perserve for no-output indication.
+            // 2 and -2 is preserve for no-output indication.
             throw std::runtime_error(
                 " INFERENCE_MSG_ID cannot be 2, please use other number.");
         }

custom_ops/gpu_ops/set_value_by_flags_and_idx.cu

@@ -34,7 +34,7 @@ __global__ void set_value_by_flag_and_id(const bool *stop_flags,
         const int64_t *input_ids_now = input_ids + tid * length_input_ids;
         const int seq_len_dec = seq_lens_decoder[tid];
         const int seq_len_enc = seq_lens_encoder[tid];
-        if (seq_len_dec == 0 && seq_len_enc == 0) return;  // stoped
+        if (seq_len_dec == 0 && seq_len_enc == 0) return;  // stopped
         if (step_idx[tid] >= 0) {
             if (seq_len_enc > 0) { // encoder, get last token accord to seq_lens_encoder
                 pre_ids_all_now[step_idx[tid]] = input_ids_now[seq_len_enc - 1];

custom_ops/gpu_ops/speculate_decoding/draft_model/draft_model_preprocess.cu

@@ -15,7 +15,48 @@
 #include "helper.h"
 #include "paddle/extension.h"
 
-template <int THREADBLOCK_SIZE, bool TRCUNCATE_FIRST_TOKEN>
+
+#define DISPATCH_BLOCKSIZE(BLOCK_SIZE, ...) \
+  do {                                     \
+    constexpr int BlockSize = BLOCK_SIZE;  \
+    __VA_ARGS__;                           \
+  } while (0)
+
+#define DISPATCH_TRUNCATE_FIRST_TOKEN(truncate_first_token, TRUNCATE_FIRST_TOKEN, ...) \
+  do {                                                                                 \
+    if (truncate_first_token) {                                                        \
+      constexpr bool TRUNCATE_FIRST_TOKEN = true;                                     \
+      __VA_ARGS__;                                                                     \
+    } else {                                                                           \
+      constexpr bool TRUNCATE_FIRST_TOKEN = false;                                    \
+      __VA_ARGS__;                                                                     \
+    }                                                                                  \
+  } while (0)
+
+#define DISPATCH_KVCACHE_SCHEDULER(kvcache_scheduler_v1, KVCACHE_SCHEDULER_V1, ...) \
+  do {                                                                              \
+    if (kvcache_scheduler_v1) {                                                     \
+      constexpr bool KVCACHE_SCHEDULER_V1 = true;                                   \
+      __VA_ARGS__;                                                                  \
+    } else {                                                                        \
+      constexpr bool KVCACHE_SCHEDULER_V1 = false;                                  \
+      __VA_ARGS__;                                                                  \
+    }                                                                               \
+  } while (0)
+
+#define DISPATCH_SPLITWISE_PREFILL(splitwise_prefill, SPLITWISE_PREFILL, ...) \
+  do {                                                                        \
+    if (splitwise_prefill) {                                                  \
+      constexpr bool SPLITWISE_PREFILL = true;                                \
+      __VA_ARGS__;                                                            \
+    } else {                                                                  \
+      constexpr bool SPLITWISE_PREFILL = false;                               \
+      __VA_ARGS__;                                                            \
+    }                                                                         \
+  } while (0)
+
+
+template <int THREADBLOCK_SIZE, bool TRUNCATE_FIRST_TOKEN, bool KVCACHE_SCHEDULER_V1>
 __global__ void process_splitwise_prefill(
     int64_t* draft_tokens,
     int64_t* input_ids,
@@ -25,6 +66,7 @@ __global__ void process_splitwise_prefill(
     int* seq_lens_decoder,
     int64_t* step_idx,
     bool* not_need_stop,
+    bool* is_block_step,
     bool* batch_drop,
     int64_t* pre_ids,
     const int64_t* accept_tokens,
@@ -58,7 +100,7 @@ __global__ void process_splitwise_prefill(
       stop_flags[tid] = false;
       int64_t base_model_first_token = accept_tokens_now[0];
       int position = seq_len_encoder;
-      if (TRCUNCATE_FIRST_TOKEN) {
+      if (TRUNCATE_FIRST_TOKEN) {
         input_ids_now[position - 1] = base_model_first_token;
         seq_lens_this_time[tid] = seq_len_encoder;
       } else {
@@ -84,7 +126,7 @@ __global__ void process_splitwise_prefill(
 
 
 
-template <int THREADBLOCK_SIZE, bool TRCUNCATE_FIRST_TOKEN>
+template <int THREADBLOCK_SIZE, bool TRUNCATE_FIRST_TOKEN, bool KVCACHE_SCHEDULER_V1>
 __global__ void draft_model_preprocess_kernel(
     int64_t* draft_tokens,
     int64_t* input_ids,
@@ -94,6 +136,7 @@ __global__ void draft_model_preprocess_kernel(
     int* seq_lens_decoder,
     int64_t* step_idx,
     bool* not_need_stop,
+    bool* is_block_step,
     bool* batch_drop,
     int64_t* pre_ids,
     const int64_t* accept_tokens,
@@ -134,14 +177,26 @@ __global__ void draft_model_preprocess_kernel(
       base_model_draft_tokens_now[i] = -1;
     }
 
-    if (base_model_stop_flags[tid] && base_model_is_block_step[tid]) {
-      batch_drop[tid] = true;
-      stop_flags[tid] = true;
+    // 1. process block_step situation
+    //    -- In v0 mode, block_step will drop mtp query.
+    //    -- In v1 mode, block_step will continue to infer.
+    if constexpr(KVCACHE_SCHEDULER_V1) {
+      if (base_model_stop_flags[tid] && base_model_is_block_step[tid]) {
+        stop_flags[tid] = true;
+        is_block_step[tid] = true;
+        // Need to continue infer
+      }
+    } else {
+      if (base_model_stop_flags[tid] && base_model_is_block_step[tid]) {
+        batch_drop[tid] = true;
+        stop_flags[tid] = true;
+      }
     }
 
+    // 2. process normal query, not in any special case.
     if (!(base_model_stop_flags[tid] || batch_drop[tid])) {
       not_stop_flag = 1;
-      // 1. first token
+      // prefill generation
       if (seq_lens_encoder[tid] > 0) {
         // Can be extended to first few tokens
         int seq_len_encoder = seq_lens_encoder[tid];
@@ -149,14 +204,20 @@ __global__ void draft_model_preprocess_kernel(
         int64_t base_model_first_token = accept_tokens_now[0];
         pre_ids_now[0] = base_model_first_token;
         int position = seq_len_encoder;
-        if (TRCUNCATE_FIRST_TOKEN) {
+        if (TRUNCATE_FIRST_TOKEN) {
           input_ids_now[position - 1] = base_model_first_token;
           seq_lens_this_time[tid] = seq_len_encoder;
         } else {
           input_ids_now[position] = base_model_first_token;
           seq_lens_this_time[tid] = seq_len_encoder + 1;
         }
-      } else {
+      } else {  // decode generation
+        if constexpr (KVCACHE_SCHEDULER_V1) {
+        // 3. try to recover mtp infer in V1 mode
+          if (!base_model_is_block_step[tid] && is_block_step[tid]) {
+            is_block_step[tid] = false;
+          }
+        }
         if (stop_flags[tid]) {
           stop_flags[tid] = false;
           // TODO: check
@@ -189,99 +250,8 @@ __global__ void draft_model_preprocess_kernel(
   }
 }
 
-template <bool TRCUNCATE_FIRST_TOKEN>
-void DispatchRunner(
-    const cudaStream_t& stream,
-    int64_t* draft_tokens,
-    int64_t* input_ids,
-    bool* stop_flags,
-    int* seq_lens_this_time,
-    int* seq_lens_encoder,
-    int* seq_lens_decoder,
-    int64_t* step_idx,
-    bool* not_need_stop,
-    bool* batch_drop,
-    int64_t* pre_ids,
-    const int64_t* accept_tokens,
-    const int* accept_num,
-    const int* base_model_seq_lens_this_time,
-    const int* base_model_seq_lens_encoder,
-    const int* base_model_seq_lens_decoder,
-    const int64_t* base_model_step_idx,
-    const bool* base_model_stop_flags,
-    const bool* base_model_is_block_step,
-    int64_t* base_model_draft_tokens,
-    const int bsz,
-    const int num_model_step,
-    const int accept_tokens_len,
-    const int draft_tokens_len,
-    const int input_ids_len,
-    const int base_model_draft_tokens_len,
-    const int pre_ids_len,
-    const bool splitwise_prefill) {
-  constexpr int BlockSize = 512;
-  if (splitwise_prefill) {
-    process_splitwise_prefill<BlockSize, TRCUNCATE_FIRST_TOKEN>
-          <<<1, BlockSize, 0, stream>>>(
-              draft_tokens,
-              input_ids,
-              stop_flags,
-              seq_lens_this_time,
-              seq_lens_encoder,
-              seq_lens_decoder,
-              step_idx,
-              not_need_stop,
-              batch_drop,
-              pre_ids,
-              accept_tokens,
-              accept_num,
-              base_model_seq_lens_this_time,
-              base_model_seq_lens_encoder,
-              base_model_seq_lens_decoder,
-              base_model_step_idx,
-              base_model_stop_flags,
-              base_model_is_block_step,
-              base_model_draft_tokens,
-              bsz,
-              num_model_step,
-              accept_tokens_len,
-              draft_tokens_len,
-              input_ids_len,
-              base_model_draft_tokens_len,
-              pre_ids_len);
-  } else {
-      draft_model_preprocess_kernel<BlockSize, TRCUNCATE_FIRST_TOKEN>
-          <<<1, BlockSize, 0, stream>>>(
-              draft_tokens,
-              input_ids,
-              stop_flags,
-              seq_lens_this_time,
-              seq_lens_encoder,
-              seq_lens_decoder,
-              step_idx,
-              not_need_stop,
-              batch_drop,
-              pre_ids,
-              accept_tokens,
-              accept_num,
-              base_model_seq_lens_this_time,
-              base_model_seq_lens_encoder,
-              base_model_seq_lens_decoder,
-              base_model_step_idx,
-              base_model_stop_flags,
-              base_model_is_block_step,
-              base_model_draft_tokens,
-              bsz,
-              num_model_step,
-              accept_tokens_len,
-              draft_tokens_len,
-              input_ids_len,
-              base_model_draft_tokens_len,
-              pre_ids_len);
-  }
-}
 
-void DispatchTokenMode(
+void DispatchRunner(
     const cudaStream_t &stream,
     int64_t* draft_tokens,
     int64_t* input_ids,
@@ -291,6 +261,7 @@ void DispatchTokenMode(
     int* seq_lens_decoder,
     int64_t* step_idx,
     bool* not_need_stop,
+    bool* is_block_step,
     bool* batch_drop,
     int64_t* pre_ids,
     const int64_t* accept_tokens,
@@ -310,75 +281,79 @@ void DispatchTokenMode(
     const int base_model_draft_tokens_len,
     const int pre_ids_len,
     const bool truncate_first_token,
-    const bool splitwise_prefill) {
-  if (truncate_first_token) {
-    DispatchRunner<true>(
-              stream,
-              draft_tokens,
-              input_ids,
-              stop_flags,
-              seq_lens_this_time,
-              seq_lens_encoder,
-              seq_lens_decoder,
-              step_idx,
-              not_need_stop,
-              batch_drop,
-              pre_ids,
-              accept_tokens,
-              accept_num,
-              base_model_seq_lens_this_time,
-              base_model_seq_lens_encoder,
-              base_model_seq_lens_decoder,
-              base_model_step_idx,
-              base_model_stop_flags,
-              base_model_is_block_step,
-              base_model_draft_tokens,
-              bsz,
-              num_model_step,
-              accept_tokens_len,
-              draft_tokens_len,
-              input_ids_len,
-              base_model_draft_tokens_len,
-              pre_ids_len,
-              splitwise_prefill
-    );
-  } else {
-    DispatchRunner<false>(
-              stream,
-              draft_tokens,
-              input_ids,
-              stop_flags,
-              seq_lens_this_time,
-              seq_lens_encoder,
-              seq_lens_decoder,
-              step_idx,
-              not_need_stop,
-              batch_drop,
-              pre_ids,
-              accept_tokens,
-              accept_num,
-              base_model_seq_lens_this_time,
-              base_model_seq_lens_encoder,
-              base_model_seq_lens_decoder,
-              base_model_step_idx,
-              base_model_stop_flags,
-              base_model_is_block_step,
-              base_model_draft_tokens,
-              bsz,
-              num_model_step,
-              accept_tokens_len,
-              draft_tokens_len,
-              input_ids_len,
-              base_model_draft_tokens_len,
-              pre_ids_len,
-              splitwise_prefill
-    );
-  }
+    const bool splitwise_prefill,
+    const bool kvcache_scheduler_v1) {
+  DISPATCH_BLOCKSIZE(512, {
+    DISPATCH_TRUNCATE_FIRST_TOKEN(truncate_first_token, TRUNCATE_FIRST_TOKEN, {
+      DISPATCH_KVCACHE_SCHEDULER(kvcache_scheduler_v1, KVCACHE_SCHEDULER_V1, {
+        DISPATCH_SPLITWISE_PREFILL(splitwise_prefill, SPLITWISE_PREFILL, {
+          if constexpr (SPLITWISE_PREFILL) {
+            process_splitwise_prefill<BlockSize, TRUNCATE_FIRST_TOKEN, KVCACHE_SCHEDULER_V1>
+                <<<1, BlockSize, 0, stream>>>(
+                    draft_tokens,
+                    input_ids,
+                    stop_flags,
+                    seq_lens_this_time,
+                    seq_lens_encoder,
+                    seq_lens_decoder,
+                    step_idx,
+                    not_need_stop,
+                    is_block_step,
+                    batch_drop,
+                    pre_ids,
+                    accept_tokens,
+                    accept_num,
+                    base_model_seq_lens_this_time,
+                    base_model_seq_lens_encoder,
+                    base_model_seq_lens_decoder,
+                    base_model_step_idx,
+                    base_model_stop_flags,
+                    base_model_is_block_step,
+                    base_model_draft_tokens,
+                    bsz,
+                    num_model_step,
+                    accept_tokens_len,
+                    draft_tokens_len,
+                    input_ids_len,
+                    base_model_draft_tokens_len,
+                    pre_ids_len);
+          } else {
+            draft_model_preprocess_kernel<BlockSize, TRUNCATE_FIRST_TOKEN, KVCACHE_SCHEDULER_V1>
+                <<<1, BlockSize, 0, stream>>>(
+                    draft_tokens,
+                    input_ids,
+                    stop_flags,
+                    seq_lens_this_time,
+                    seq_lens_encoder,
+                    seq_lens_decoder,
+                    step_idx,
+                    not_need_stop,
+                    is_block_step,
+                    batch_drop,
+                    pre_ids,
+                    accept_tokens,
+                    accept_num,
+                    base_model_seq_lens_this_time,
+                    base_model_seq_lens_encoder,
+                    base_model_seq_lens_decoder,
+                    base_model_step_idx,
+                    base_model_stop_flags,
+                    base_model_is_block_step,
+                    base_model_draft_tokens,
+                    bsz,
+                    num_model_step,
+                    accept_tokens_len,
+                    draft_tokens_len,
+                    input_ids_len,
+                    base_model_draft_tokens_len,
+                    pre_ids_len);
+          }
+        });
+      });
+    });
+  });
 }
 
-
-
-
 void DraftModelPreprocess(const paddle::Tensor& draft_tokens,
                           const paddle::Tensor& input_ids,
                           const paddle::Tensor& stop_flags,
@@ -387,6 +362,7 @@ 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,
@@ -400,7 +376,8 @@ void DraftModelPreprocess(const paddle::Tensor& draft_tokens,
                           const paddle::Tensor& base_model_draft_tokens,
                           const int num_model_step,
                           const bool truncate_first_token,
-                          const bool splitwise_prefill) {
+                          const bool splitwise_prefill,
+                          const bool kvcache_scheduler_v1) {
   int real_bsz = seq_lens_this_time.shape()[0];
   int accept_tokens_len = accept_tokens.shape()[1];
   int input_ids_len = input_ids.shape()[1];
@@ -412,36 +389,38 @@ void DraftModelPreprocess(const paddle::Tensor& draft_tokens,
   auto not_need_stop_gpu =
       not_need_stop.copy_to(seq_lens_this_time.place(), false);
 
-  DispatchTokenMode(
-              cu_stream,
-              const_cast<int64_t*>(draft_tokens.data<int64_t>()),
-              const_cast<int64_t*>(input_ids.data<int64_t>()),
-              const_cast<bool*>(stop_flags.data<bool>()),
-              const_cast<int*>(seq_lens_this_time.data<int>()),
-              const_cast<int*>(seq_lens_encoder.data<int>()),
-              const_cast<int*>(seq_lens_decoder.data<int>()),
-              const_cast<int64_t*>(step_idx.data<int64_t>()),
-              const_cast<bool*>(not_need_stop_gpu.data<bool>()),
-              const_cast<bool*>(batch_drop.data<bool>()),
-              const_cast<int64_t*>(pre_ids.data<int64_t>()),
-              accept_tokens.data<int64_t>(),
-              accept_num.data<int>(),
-              base_model_seq_lens_this_time.data<int>(),
-              base_model_seq_lens_encoder.data<int>(),
-              base_model_seq_lens_decoder.data<int>(),
-              base_model_step_idx.data<int64_t>(),
-              base_model_stop_flags.data<bool>(),
-              base_model_is_block_step.data<bool>(),
-              const_cast<int64_t*>(base_model_draft_tokens.data<int64_t>()),
-              real_bsz,
-              num_model_step,
-              accept_tokens_len,
-              draft_tokens_len,
-              input_ids_len,
-              base_model_draft_tokens_len,
-              pre_ids_len,
-              truncate_first_token,
-              splitwise_prefill);
+  DispatchRunner(
+      cu_stream,
+      const_cast<int64_t*>(draft_tokens.data<int64_t>()),
+      const_cast<int64_t*>(input_ids.data<int64_t>()),
+      const_cast<bool*>(stop_flags.data<bool>()),
+      const_cast<int*>(seq_lens_this_time.data<int>()),
+      const_cast<int*>(seq_lens_encoder.data<int>()),
+      const_cast<int*>(seq_lens_decoder.data<int>()),
+      const_cast<int64_t*>(step_idx.data<int64_t>()),
+      const_cast<bool*>(not_need_stop_gpu.data<bool>()),
+      const_cast<bool*>(is_block_step.data<bool>()),
+      const_cast<bool*>(batch_drop.data<bool>()),
+      const_cast<int64_t*>(pre_ids.data<int64_t>()),
+      accept_tokens.data<int64_t>(),
+      accept_num.data<int>(),
+      base_model_seq_lens_this_time.data<int>(),
+      base_model_seq_lens_encoder.data<int>(),
+      base_model_seq_lens_decoder.data<int>(),
+      base_model_step_idx.data<int64_t>(),
+      base_model_stop_flags.data<bool>(),
+      base_model_is_block_step.data<bool>(),
+      const_cast<int64_t*>(base_model_draft_tokens.data<int64_t>()),
+      real_bsz,
+      num_model_step,
+      accept_tokens_len,
+      draft_tokens_len,
+      input_ids_len,
+      base_model_draft_tokens_len,
+      pre_ids_len,
+      truncate_first_token,
+      splitwise_prefill,
+      kvcache_scheduler_v1);
 
   auto not_need_stop_cpu =
       not_need_stop_gpu.copy_to(not_need_stop.place(), false);
@@ -459,6 +438,7 @@ PD_BUILD_STATIC_OP(draft_model_preprocess)
              "seq_lens_decoder",
              "step_idx",
              "not_need_stop",
+             "is_block_step",
              "batch_drop",
              "pre_ids",
              "accept_tokens",
@@ -480,7 +460,7 @@ PD_BUILD_STATIC_OP(draft_model_preprocess)
               "not_need_stop_out",
               "batch_drop_out",
               "pre_ids_out"})
-    .Attrs({"num_model_step: int", "truncate_first_token: bool", "splitwise_prefill: bool"})
+    .Attrs({"num_model_step: int", "truncate_first_token: bool", "splitwise_prefill: bool", "kvcache_scheduler_v1: bool"})
     .SetInplaceMap({{"draft_tokens", "draft_tokens_out"},
                     {"input_ids", "input_ids_out"},
                     {"stop_flags", "stop_flags_out"},

custom_ops/gpu_ops/speculate_decoding/draft_model/eagle_get_hidden_states.cu

@@ -63,7 +63,7 @@ __global__ void ComputeOrderKernel(
           position_map[in_offset++] = out_offset++;
         }
         in_offset += cur_base_model_seq_lens_this_time - accept_num;
-// (liuzichang): Temperary Reserved for debug
+// (liuzichang): Temporary Reserved for debug
 //         if (accept_num <= actual_draft_token_num) /*Accept partial draft tokens*/ {
 // #ifdef DEBUG_EAGLE_KERNEL
 //         printf("batch %d: accept_num <= actual_draft_token_num \n", i);