Merge branch 'main' into qiming/enale_eplb

Author: mayuyuace

.github/workflows/ut.yaml

@@ -64,7 +64,7 @@ jobs:
     timeout-minutes: 40
     container:                      
       image: localhost:5000/xpu-kernel-ci-image:latest
-      options: --device /dev/dri -v /dev/dri/by-path:/dev/dri/by-path --privileged
+      options: --device /dev/dri -v /dev/dri/by-path:/dev/dri/by-path --privileged -v ccache:/root/.ccache -e CCACHE_DIR=/root/.ccache
     steps:
       - name: Checkout
         uses: actions/checkout@8e8c483db84b4bee98b60c0593521ed34d9990e8 # v6.0.1
@@ -74,8 +74,15 @@ jobs:
       - name: build & install wheel
         run: |
           git config --global --add safe.directory "${GITHUB_WORKSPACE}"
+          export CCACHE_DIR=/root/.ccache
+          export CCACHE_BASEDIR="${GITHUB_WORKSPACE}"
+          export CCACHE_NOHASHDIR=1
+          export CCACHE_COMPILERCHECK=content
+          ccache -s || true
+          ccache -p || true
           uv pip install -r requirements.txt
-          MAX_JOB=128 uv pip install -e . -v
+          MAX_JOBS=128 uv pip install --no-build-isolation -e . -v
+          ccache -s || true
 
       - name: test 
         run: |
@@ -98,7 +105,7 @@ jobs:
     timeout-minutes: 40
     container: 
       image: localhost:5000/xpu-kernel-ci-image:latest
-      options: --device /dev/dri -v /dev/dri/by-path:/dev/dri/by-path --privileged 
+      options: --device /dev/dri -v /dev/dri/by-path:/dev/dri/by-path --privileged -v ccache:/root/.ccache -e CCACHE_DIR=/root/.ccache
     steps:
       - name: Checkout
         uses: actions/checkout@8e8c483db84b4bee98b60c0593521ed34d9990e8 # v6.0.1
@@ -108,8 +115,15 @@ jobs:
       - name: build & install wheel
         run: |
           git config --global --add safe.directory "${GITHUB_WORKSPACE}"
+          export CCACHE_DIR=/root/.ccache
+          export CCACHE_BASEDIR="${GITHUB_WORKSPACE}"
+          export CCACHE_NOHASHDIR=1
+          export CCACHE_COMPILERCHECK=content
+          ccache -s || true
+          ccache -p || true
           uv pip install -r requirements.txt
-          MAX_JOBS=128 uv pip install -e . -v
+          MAX_JOBS=128 uv pip install --no-build-isolation -e . -v
+          ccache -s || true
 
       - name: test 
         run: |

Dockerfile.xpu

@@ -14,6 +14,7 @@ RUN apt clean && apt-get update -y && \
     curl \
     ffmpeg \
     git \
+    ccache \
     libsndfile1 \
     libsm6 \
     libxext6 \

tests/test_fp8_quant.py

@@ -218,6 +218,15 @@ def seed_everything(seed):
 
 # override pytest parameters when enable mini pytest
 MINI_PYTEST_PARAMS = {
+    "test_per_block_fp8_quant": {
+        "num_tokens_block_quant": [1],
+        "hidden_size_block_quant": [64],
+        "group_size": [32],
+    },
+    "test_per_block_mxfp8_quant": {
+        "num_tokens_block_quant": [1],
+        "hidden_size_block_quant": [64],
+    },
     "default": {
         "num_tokens": [1],
         "hidden_size": [1],