Split kv decode

.github/workflows/pr-test-xpu.yml

@@ -67,6 +67,8 @@ jobs:
               python3 bench_merge_states_v2.py 2>&1 | tee merge_states.py.log \
               python3 bench_swiglu_alpha_limit.py 2>&1 | tee swiglu_alpha_limit.py.log \
               python3 bench_fused_qk_norm_rope.py 2>&1 | tee fused_qk_norm_rope.py.log \
+              python3 bench_per_token_group_quant_8bit.py 2>&1 | tee per_token_group_quant_8bit.py.log \
+              python3 bench_per_token_group_quant_mxfp4.py 2>&1 | tee per_token_group_quant_mxfp4.py.log \
             "
 
       - name: Copy logs from container

CMakeLists.txt

@@ -42,7 +42,7 @@ set(CUTLASS_ENABLE_HEADERS_ONLY ON CACHE BOOL "Enable headers only mode in cutla
 FetchContent_Declare(
     repo-cutlass-sycl
     GIT_REPOSITORY https://github.com/intel/sycl-tla.git
-    GIT_TAG        482b40e8bed0e9204311d1569c876b4573dfb952
+    GIT_TAG        64584484b4279b1b4184b508af445698a4a1b603
     GIT_SHALLOW    OFF
 )
 FetchContent_MakeAvailable(repo-cutlass-sycl)

Dockerfile.xpu_kernel

@@ -22,10 +22,10 @@ ARG SG_LANG_KERNEL_BRANCH=main
 # Install the latest UMD driver for SYCL-TLA
 RUN apt-get install -y software-properties-common && \
     add-apt-repository -y ppa:kobuk-team/intel-graphics && \
+    apt-get update  && \
     apt-get install -y libze-intel-gpu1 libze1 intel-metrics-discovery intel-opencl-icd clinfo intel-gsc && \
     apt-get install -y intel-media-va-driver-non-free libmfx-gen1 libvpl2 libvpl-tools libva-glx2 va-driver-all vainfo && \
-    apt-get install -y libze-dev intel-ocloc && \
-    apt-get update
+    apt-get install -y libze-dev intel-ocloc
 
 # Install Miniforge & PyTorch/Triton
 RUN curl -fsSL -v -o miniforge.sh -O https://github.com/conda-forge/miniforge/releases/download/25.1.1-0/Miniforge3-Linux-x86_64.sh && \
@@ -66,3 +66,4 @@ RUN --mount=type=secret,id=github_token \
 # Set the default shell to bash
 SHELL ["bash", "-c"]
 CMD ["bash", "-c", "source /root/.bashrc && exec bash"]
+USER root