add fatrelu

.github/workflows/ut.yaml

@@ -5,12 +5,23 @@ on:
     branches: [main]
   push:
     branches: [main]
+  schedule:
+    # Nightly full-scope run at 02:00 UTC
+    - cron: '0 2 * * *'
+  workflow_dispatch:
+    inputs:
+      test_scope:
+        description: 'Test scope: full, ci, mini, ondemand:llama, ondemand:deepseek'
+        required: false
+        default: 'ci'
 
 permissions: 
   contents: read
 
 env:
   REGISTRY: localhost:5000
+  # Nightly (schedule) → full scope; manual dispatch → user choice; PR/push → ci scope
+  XPU_KERNEL_TEST_SCOPE: ${{ github.event_name == 'schedule' && 'full' || github.event.inputs.test_scope || 'ci' }}
 
 jobs:
 
@@ -87,8 +98,9 @@ jobs:
 
       - name: test 
         run: |
-          ZE_AFFINITY_MASK=0,1 SKIP_HANG_KERNEL=1 SKIP_ACC_ERROR_KERNEL=1 pytest -v -s tests/
-          VLLM_XPU_FORCE_XE_DEFAULT_KERNEL=1 ZE_AFFINITY_MASK=0,1 pytest -v -s tests/fused_moe/test_grouped_gemm.py::test_grouped_gemm
+          echo "Running tests with XPU_KERNEL_TEST_SCOPE=${{ env.XPU_KERNEL_TEST_SCOPE }}"
+          XPU_KERNEL_TEST_SCOPE=${{ env.XPU_KERNEL_TEST_SCOPE }} ZE_AFFINITY_MASK=0,1 SKIP_HANG_KERNEL=1 SKIP_ACC_ERROR_KERNEL=1 pytest -v -s tests/
+          VLLM_XPU_FORCE_XE_DEFAULT_KERNEL=1 XPU_KERNEL_TEST_SCOPE=${{ env.XPU_KERNEL_TEST_SCOPE }} ZE_AFFINITY_MASK=0,1 pytest -v -s tests/fused_moe/test_grouped_gemm.py::test_grouped_gemm
 
   clean-repo-pvc:
     runs-on: self-hosted-pvc
@@ -129,11 +141,12 @@ jobs:
 
       - name: test 
         run: |
+          echo "Running tests with XPU_KERNEL_TEST_SCOPE=${{ env.XPU_KERNEL_TEST_SCOPE }}"
           # tests/test_moe_align_block_size.py, tests/test_moe_lora_align_sum.py takes much time than expected. ignore it for now.
-          ZE_AFFINITY_MASK=0,1 pytest -v -s tests/ --ignore=tests/test_lora_ops.py --ignore=tests/test_fp8_quant.py --ignore=tests/test_moe_align_block_size.py --ignore=tests/test_moe_lora_align_sum.py --ignore=tests/test_cache.py::test_swap_blocks --ignore=tests/test_topk_per_row.py --ignore=tests/test_lora_ops.py
+          XPU_KERNEL_TEST_SCOPE=${{ env.XPU_KERNEL_TEST_SCOPE }} ZE_AFFINITY_MASK=0,1 pytest -v -s tests/ --ignore=tests/test_lora_ops.py --ignore=tests/test_fp8_quant.py --ignore=tests/test_moe_align_block_size.py --ignore=tests/test_moe_lora_align_sum.py --ignore=tests/test_cache.py::test_swap_blocks --ignore=tests/test_topk_per_row.py --ignore=tests/test_lora_ops.py
           # fixme: Running lora UT separately to avoid OOM when running together with other tests.
-          ZE_AFFINITY_MASK=0,1 pytest -v -s tests/test_lora_ops.py
-          VLLM_XPU_FORCE_XE_DEFAULT_KERNEL=1 ZE_AFFINITY_MASK=0,1 pytest -v -s tests/fused_moe/test_grouped_gemm.py::test_grouped_gemm
+          XPU_KERNEL_TEST_SCOPE=${{ env.XPU_KERNEL_TEST_SCOPE }} ZE_AFFINITY_MASK=0,1 pytest -v -s tests/test_lora_ops.py
+          VLLM_XPU_FORCE_XE_DEFAULT_KERNEL=1 XPU_KERNEL_TEST_SCOPE=${{ env.XPU_KERNEL_TEST_SCOPE }} ZE_AFFINITY_MASK=0,1 pytest -v -s tests/fused_moe/test_grouped_gemm.py::test_grouped_gemm
 
   clean-repo-bmg:
     runs-on: self-hosted-bmg

.pre-commit-config.yaml

@@ -6,11 +6,6 @@ default_stages:
   - manual # Run in CI
 exclude: 'vllm/third_party/.*'
 repos:
-- repo: https://github.com/google/yapf
-  rev: v0.43.0
-  hooks:
-  - id: yapf
-    args: [--in-place, --verbose]
 - repo: https://github.com/astral-sh/ruff-pre-commit
   rev: v0.11.7
   hooks:

CMakeLists.txt

@@ -50,14 +50,35 @@ set(BUILD_SYCL_TLA_KERNELS
     ON
     CACHE BOOL "Build SYCL-TLA based kernels for XPU")
 # ARCHITECTURE OPTIONS
-set(VLLM_XPU_ENABLE_XE2 ON)
-set(VLLM_XPU_ENABLE_XE_DEFAULT ON)
-
-set(BASIC_KERNELS_ENABLED ON)
-set(FA2_KERNELS_ENABLED ON)
-set(GND_KERNELS_ENABLED ON)
-set(MOE_KERNELS_ENABLED ON)
-set(XPU_SPECIFIC_KERNELS_ENABLED ON)
+option(VLLM_XPU_ENABLE_XE2 "Enable XE2 architecture kernels" ON)
+option(VLLM_XPU_ENABLE_XE_DEFAULT "Enable XE Default architecture kernels" ON)
+
+# KERNEL OPTIONS — each controls whether the corresponding Python extension is
+# built. Override from the command line with -D<NAME>=OFF or via the
+# identically-named environment variable (read by setup.py).
+option(BASIC_KERNELS_ENABLED "Build basic kernels (_C extension)" ON)
+option(FA2_KERNELS_ENABLED
+       "Build Flash Attention 2 kernels (_vllm_fa2_C extension)" ON)
+option(MOE_KERNELS_ENABLED
+       "Build MoE kernels (_moe_C extension + grouped_gemm TLA)" ON)
+option(GDN_KERNELS_ENABLED "Build GDN attention kernels (gdn_attn TLA)" ON)
+option(XPU_SPECIFIC_KERNELS_ENABLED
+       "Build XPU-specific kernels (_xpu_C extension)" ON)
+option(XPUMEM_ALLOCATOR_ENABLED "Build xpumem_allocator extension" ON)
+
+message(STATUS "")
+message(STATUS "Kernel build configuration:")
+message(STATUS "  BUILD_SYCL_TLA_KERNELS       = ${BUILD_SYCL_TLA_KERNELS}")
+message(STATUS "  VLLM_XPU_ENABLE_XE2          = ${VLLM_XPU_ENABLE_XE2}")
+message(STATUS "  VLLM_XPU_ENABLE_XE_DEFAULT   = ${VLLM_XPU_ENABLE_XE_DEFAULT}")
+message(STATUS "  BASIC_KERNELS_ENABLED        = ${BASIC_KERNELS_ENABLED}")
+message(STATUS "  FA2_KERNELS_ENABLED          = ${FA2_KERNELS_ENABLED}")
+message(STATUS "  MOE_KERNELS_ENABLED          = ${MOE_KERNELS_ENABLED}")
+message(STATUS "  GDN_KERNELS_ENABLED          = ${GDN_KERNELS_ENABLED}")
+message(
+  STATUS "  XPU_SPECIFIC_KERNELS_ENABLED = ${XPU_SPECIFIC_KERNELS_ENABLED}")
+message(STATUS "  XPUMEM_ALLOCATOR_ENABLED     = ${XPUMEM_ALLOCATOR_ENABLED}")
+message(STATUS "")
 
 #
 # Try to find python package with an executable that exactly matches
@@ -313,23 +334,40 @@ if(BUILD_SYCL_TLA_KERNELS)
   # extensions shared library
   set(SYCL_TLA_COMPILE_OPTIONS "")
   if(VLLM_XPU_ENABLE_XE_DEFAULT)
-    add_subdirectory(csrc/xpu/grouped_gemm/xe_default)
-    list(APPEND GROUPED_GEMM_LIB_NAME "grouped_gemm_xe_default")
+    if(MOE_KERNELS_ENABLED)
+      add_subdirectory(csrc/xpu/grouped_gemm/xe_default)
+      list(APPEND GROUPED_GEMM_LIB_NAME "grouped_gemm_xe_default")
+    endif()
     list(APPEND SYCL_TLA_COMPILE_OPTIONS -DVLLM_XPU_ENABLE_XE_DEFAULT)
   endif()
   if(VLLM_XPU_ENABLE_XE2)
-    add_subdirectory(csrc/xpu/grouped_gemm/xe_2)
-    add_subdirectory(csrc/xpu/attn/xe_2)
-    add_subdirectory(csrc/xpu/gdn_attn/xe_2)
-    list(APPEND GROUPED_GEMM_LIB_NAME "grouped_gemm_xe_2")
-    list(APPEND ATTN_KERNEL_LIB_NAME "attn_kernels_xe_2")
-    list(APPEND GDN_ATTN_LIB_NAME "gdn_attn_kernels_xe_2")
+    if(MOE_KERNELS_ENABLED)
+      add_subdirectory(csrc/xpu/grouped_gemm/xe_2)
+      list(APPEND GROUPED_GEMM_LIB_NAME "grouped_gemm_xe_2")
+    endif()
+    if(FA2_KERNELS_ENABLED)
+      add_subdirectory(csrc/xpu/attn/xe_2)
+      list(APPEND ATTN_KERNEL_LIB_NAME "attn_kernels_xe_2")
+    endif()
+    if(GDN_KERNELS_ENABLED)
+      add_subdirectory(csrc/xpu/gdn_attn/xe_2)
+      list(APPEND GDN_ATTN_LIB_NAME "gdn_attn_kernels_xe_2")
+    endif()
     list(APPEND SYCL_TLA_COMPILE_OPTIONS -DVLLM_XPU_ENABLE_XE2)
   endif()
   list(APPEND VLLM_GPU_COMPILE_FLAGS ${SYCL_TLA_COMPILE_OPTIONS})
 
 endif()
 
+# Feature compile defines — these guard op registrations and interface code so
+# that disabled features don't pull in unbuilt TLA library symbols.
+if(MOE_KERNELS_ENABLED)
+  list(APPEND VLLM_GPU_COMPILE_FLAGS -DVLLM_MOE_ENABLED)
+endif()
+if(GDN_KERNELS_ENABLED)
+  list(APPEND VLLM_GPU_COMPILE_FLAGS -DVLLM_GDN_ENABLED)
+endif()
+
 # define vLLM XPU cmake variables
 
 set(VLLM_XPU_INCLUDE_DIR "")
@@ -352,9 +390,6 @@ list(
 #
 # xpumem_allocator extension - XPU memory allocator with Python callbacks
 #
-set(XPUMEM_ALLOCATOR_ENABLED
-    ON
-    CACHE BOOL "Build xpumem_allocator extension")
 if(XPUMEM_ALLOCATOR_ENABLED)
   message(STATUS "Enabling xpumem_allocator extension.")
 
@@ -404,8 +439,10 @@ if(BASIC_KERNELS_ENABLED)
   set(VLLM_EXT_SRC
       "csrc/cache.cpp"
       "csrc/layernorm.cpp"
+      "csrc/layernorm_quant.cpp"
       "csrc/activation.cpp"
       "csrc/pos_encoding_kernels.cpp"
+      "csrc/fused_qknorm_rope.cpp"
       "csrc/torch_bindings.cpp"
       "csrc/quantization/fp8/fp8_quant.cpp"
       "csrc/quantization/fp4/mxfp4_quant.cpp"
@@ -441,6 +478,14 @@ endif()
 #
 # flash attention _C extension
 #
+if(FA2_KERNELS_ENABLED AND NOT ATTN_KERNEL_LIB_NAME)
+  message(
+    WARNING
+      "FA2_KERNELS_ENABLED is ON but no attention kernel libraries are available. "
+      "The _vllm_fa2_C extension will be built without kernel libraries and may "
+      "fail at runtime. Enable BUILD_SYCL_TLA_KERNELS and VLLM_XPU_ENABLE_XE2 "
+      "for full FA2 functionality.")
+endif()
 if(FA2_KERNELS_ENABLED)
   message(STATUS "Enabling fa2 extension.")
   file(GLOB FA2_GEN_SRCS "csrc/flash_attn/*.cpp"
@@ -482,9 +527,14 @@ if(XPU_SPECIFIC_KERNELS_ENABLED)
       "csrc/xpu/sampler/topk_topp_sampler.cpp"
       "csrc/xpu/sycl/deepseek_scaling_rope.cpp"
       "csrc/xpu/rand/exponential.cpp"
-      "csrc/xpu/grouped_gemm/grouped_gemm_interface.cpp"
-      "csrc/xpu/utils.cpp"
-      "csrc/xpu/gdn_attn/gdn_attn_interface.cpp")
+      "csrc/xpu/utils.cpp")
+  if(MOE_KERNELS_ENABLED)
+    list(APPEND VLLM_EXT_XPU_SRC
+         "csrc/xpu/grouped_gemm/grouped_gemm_interface.cpp")
+  endif()
+  if(GDN_KERNELS_ENABLED)
+    list(APPEND VLLM_EXT_XPU_SRC "csrc/xpu/gdn_attn/gdn_attn_interface.cpp")
+  endif()
   include_directories("/usr/include")
   # TODO: check if we need this flags list(APPEND VLLM_GPU_FLAGS
   # "-gline-tables-only")