EngineX-Mthreads/enginex-mthreads-vllm
4
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity

Sync from v0.13

Browse Source
This commit is contained in:
xiezhongtao
2026-01-19 10:38:50 +08:00
parent b2ef04d792
commit 5aef6c175a
3714 changed files with 854317 additions and 89342 deletions
Download Patch File Download Diff File Expand all files Collapse all files

133
cmake/external_projects/flashmla.cmake Normal file
View File

@@ -0,0 +1,133 @@
include(FetchContent)
# If FLASH_MLA_SRC_DIR is set, flash-mla is installed from that directory
# instead of downloading.
# It can be set as an environment variable or passed as a cmake argument.
# The environment variable takes precedence.
if (DEFINED ENV{FLASH_MLA_SRC_DIR})
set(FLASH_MLA_SRC_DIR $ENV{FLASH_MLA_SRC_DIR})
endif()
if(FLASH_MLA_SRC_DIR)
FetchContent_Declare(
flashmla
SOURCE_DIR ${FLASH_MLA_SRC_DIR}
CONFIGURE_COMMAND ""
BUILD_COMMAND ""
)
else()
FetchContent_Declare(
flashmla
GIT_REPOSITORY https://github.com/vllm-project/FlashMLA
GIT_TAG 46d64a8ebef03fa50b4ae74937276a5c940e3f95
GIT_PROGRESS TRUE
CONFIGURE_COMMAND ""
BUILD_COMMAND ""
)
endif()
FetchContent_MakeAvailable(flashmla)
message(STATUS "FlashMLA is available at ${flashmla_SOURCE_DIR}")
# The FlashMLA kernels only work on hopper and require CUDA 12.3 or later.
# Only build FlashMLA kernels if we are building for something compatible with
# sm90a
set(SUPPORT_ARCHS)
if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.3)
list(APPEND SUPPORT_ARCHS 9.0a)
endif()
if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.8)
list(APPEND SUPPORT_ARCHS 10.0a)
endif()
cuda_archs_loose_intersection(FLASH_MLA_ARCHS "${SUPPORT_ARCHS}" "${CUDA_ARCHS}")
if(FLASH_MLA_ARCHS)
set(VLLM_FLASHMLA_GPU_FLAGS ${VLLM_GPU_FLAGS})
list(APPEND VLLM_FLASHMLA_GPU_FLAGS "--expt-relaxed-constexpr" "--expt-extended-lambda" "--use_fast_math")
set(FlashMLA_SOURCES
${flashmla_SOURCE_DIR}/csrc/torch_api.cpp
${flashmla_SOURCE_DIR}/csrc/pybind.cpp
${flashmla_SOURCE_DIR}/csrc/smxx/get_mla_metadata.cu
${flashmla_SOURCE_DIR}/csrc/smxx/mla_combine.cu
${flashmla_SOURCE_DIR}/csrc/sm90/decode/dense/splitkv_mla.cu
${flashmla_SOURCE_DIR}/csrc/sm90/decode/sparse_fp8/splitkv_mla.cu
${flashmla_SOURCE_DIR}/csrc/sm90/prefill/sparse/fwd.cu
${flashmla_SOURCE_DIR}/csrc/sm100/decode/sparse_fp8/splitkv_mla.cu
${flashmla_SOURCE_DIR}/csrc/sm100/prefill/dense/fmha_cutlass_fwd_sm100.cu
${flashmla_SOURCE_DIR}/csrc/sm100/prefill/dense/fmha_cutlass_bwd_sm100.cu
${flashmla_SOURCE_DIR}/csrc/sm100/prefill/sparse/fwd.cu
)
set(FlashMLA_Extension_SOURCES
${flashmla_SOURCE_DIR}/csrc/extension/torch_api.cpp
${flashmla_SOURCE_DIR}/csrc/extension/sm90/dense_fp8/pybind.cpp
${flashmla_SOURCE_DIR}/csrc/extension/sm90/dense_fp8/flash_fwd_mla_fp8_sm90.cu
${flashmla_SOURCE_DIR}/csrc/extension/sm90/dense_fp8/flash_fwd_mla_metadata.cu
)
set(FlashMLA_INCLUDES
${flashmla_SOURCE_DIR}/csrc
${flashmla_SOURCE_DIR}/csrc/sm90
${flashmla_SOURCE_DIR}/csrc/cutlass/include
${flashmla_SOURCE_DIR}/csrc/cutlass/tools/util/include
)
set(FlashMLA_Extension_INCLUDES
${flashmla_SOURCE_DIR}/csrc
${flashmla_SOURCE_DIR}/csrc/sm90
${flashmla_SOURCE_DIR}/csrc/extension/sm90/dense_fp8/
${flashmla_SOURCE_DIR}/csrc/cutlass/include
${flashmla_SOURCE_DIR}/csrc/cutlass/tools/util/include
)
set_gencode_flags_for_srcs(
SRCS "${FlashMLA_SOURCES}"
CUDA_ARCHS "${FLASH_MLA_ARCHS}")
set_gencode_flags_for_srcs(
SRCS "${FlashMLA_Extension_SOURCES}"
CUDA_ARCHS "${FLASH_MLA_ARCHS}")
define_extension_target(
_flashmla_C
DESTINATION vllm
LANGUAGE ${VLLM_GPU_LANG}
SOURCES ${FlashMLA_SOURCES}
COMPILE_FLAGS ${VLLM_GPU_FLAGS}
ARCHITECTURES ${VLLM_GPU_ARCHES}
INCLUDE_DIRECTORIES ${FlashMLA_INCLUDES}
USE_SABI 3
WITH_SOABI)
# Keep Stable ABI for the module, but *not* for CUDA/C++ files.
# This prevents Py_LIMITED_API from affecting nvcc and C++ compiles.
target_compile_options(_flashmla_C PRIVATE
$<$<COMPILE_LANGUAGE:CUDA>:-UPy_LIMITED_API>
$<$<COMPILE_LANGUAGE:CXX>:-UPy_LIMITED_API>)
define_extension_target(
_flashmla_extension_C
DESTINATION vllm
LANGUAGE ${VLLM_GPU_LANG}
SOURCES ${FlashMLA_Extension_SOURCES}
COMPILE_FLAGS ${VLLM_FLASHMLA_GPU_FLAGS}
ARCHITECTURES ${VLLM_GPU_ARCHES}
INCLUDE_DIRECTORIES ${FlashMLA_Extension_INCLUDES}
USE_SABI 3
WITH_SOABI)
# Keep Stable ABI for the module, but *not* for CUDA/C++ files.
# This prevents Py_LIMITED_API from affecting nvcc and C++ compiles.
target_compile_options(_flashmla_extension_C PRIVATE
$<$<COMPILE_LANGUAGE:CUDA>:-UPy_LIMITED_API>
$<$<COMPILE_LANGUAGE:CXX>:-UPy_LIMITED_API>)
else()
# Create empty targets for setup.py when not targeting sm90a systems
add_custom_target(_flashmla_C)
add_custom_target(_flashmla_extension_C)
endif()

97
cmake/external_projects/qutlass.cmake Normal file
View File

@@ -0,0 +1,97 @@
include(FetchContent)
set(CUTLASS_INCLUDE_DIR "${CUTLASS_INCLUDE_DIR}" CACHE PATH "Path to CUTLASS include/ directory")
if(DEFINED ENV{QUTLASS_SRC_DIR})
set(QUTLASS_SRC_DIR $ENV{QUTLASS_SRC_DIR})
endif()
if(QUTLASS_SRC_DIR)
FetchContent_Declare(
qutlass
SOURCE_DIR ${QUTLASS_SRC_DIR}
CONFIGURE_COMMAND ""
BUILD_COMMAND ""
)
else()
FetchContent_Declare(
qutlass
GIT_REPOSITORY https://github.com/IST-DASLab/qutlass.git
GIT_TAG 830d2c4537c7396e14a02a46fbddd18b5d107c65
GIT_PROGRESS TRUE
CONFIGURE_COMMAND ""
BUILD_COMMAND ""
)
endif()
FetchContent_Populate(qutlass)
if(NOT qutlass_SOURCE_DIR)
message(FATAL_ERROR "[QUTLASS] source directory could not be resolved.")
endif()
message(STATUS "[QUTLASS] QuTLASS is available at ${qutlass_SOURCE_DIR}")
cuda_archs_loose_intersection(QUTLASS_ARCHS "12.0a;10.0a" "${CUDA_ARCHS}")
if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.8 AND QUTLASS_ARCHS)
if(QUTLASS_ARCHS MATCHES "10\\.0a")
set(QUTLASS_TARGET_CC 100)
elseif(QUTLASS_ARCHS MATCHES "12\\.0a")
set(QUTLASS_TARGET_CC 120)
else()
message(FATAL_ERROR "[QUTLASS] internal error parsing CUDA_ARCHS='${QUTLASS_ARCHS}'.")
endif()
set(QUTLASS_SOURCES
${qutlass_SOURCE_DIR}/qutlass/csrc/bindings.cpp
${qutlass_SOURCE_DIR}/qutlass/csrc/gemm.cu
${qutlass_SOURCE_DIR}/qutlass/csrc/gemm_ada.cu
${qutlass_SOURCE_DIR}/qutlass/csrc/fused_quantize_mx.cu
${qutlass_SOURCE_DIR}/qutlass/csrc/fused_quantize_nv.cu
${qutlass_SOURCE_DIR}/qutlass/csrc/fused_quantize_mx_sm100.cu
${qutlass_SOURCE_DIR}/qutlass/csrc/fused_quantize_nv_sm100.cu
)
set(QUTLASS_INCLUDES
${qutlass_SOURCE_DIR}
${qutlass_SOURCE_DIR}/qutlass
${qutlass_SOURCE_DIR}/qutlass/csrc/include
${qutlass_SOURCE_DIR}/qutlass/csrc/include/cutlass_extensions
)
if(CUTLASS_INCLUDE_DIR AND EXISTS "${CUTLASS_INCLUDE_DIR}/cutlass/cutlass.h")
list(APPEND QUTLASS_INCLUDES "${CUTLASS_INCLUDE_DIR}")
elseif(EXISTS "${qutlass_SOURCE_DIR}/qutlass/third_party/cutlass/include/cutlass/cutlass.h")
list(APPEND QUTLASS_INCLUDES "${qutlass_SOURCE_DIR}/qutlass/third_party/cutlass/include")
message(STATUS "[QUTLASS] Using QuTLASS vendored CUTLASS headers (no vLLM CUTLASS detected).")
else()
message(FATAL_ERROR "[QUTLASS] CUTLASS headers not found. "
"Set -DCUTLASS_INCLUDE_DIR=/path/to/cutlass/include")
endif()
set_gencode_flags_for_srcs(
SRCS "${QUTLASS_SOURCES}"
CUDA_ARCHS "${QUTLASS_ARCHS}"
)
target_sources(_C PRIVATE ${QUTLASS_SOURCES})
target_include_directories(_C PRIVATE ${QUTLASS_INCLUDES})
target_compile_definitions(_C PRIVATE
QUTLASS_DISABLE_PYBIND=1
TARGET_CUDA_ARCH=${QUTLASS_TARGET_CC}
)
set_property(SOURCE ${QUTLASS_SOURCES} APPEND PROPERTY COMPILE_OPTIONS
$<$<COMPILE_LANGUAGE:CUDA>:--expt-relaxed-constexpr --use_fast_math -O3>
)
else()
if("${CMAKE_CUDA_COMPILER_VERSION}" VERSION_LESS "12.8")
message(STATUS
"[QUTLASS] Skipping build: CUDA 12.8 or newer is required (found ${CMAKE_CUDA_COMPILER_VERSION}).")
else()
message(STATUS
"[QUTLASS] Skipping build: no supported arch (12.0a / 10.0a) found in "
"CUDA_ARCHS='${CUDA_ARCHS}'.")
endif()
endif()

53
cmake/external_projects/triton_kernels.cmake Normal file
View File

@@ -0,0 +1,53 @@
# Install OpenAI triton_kernels from https://github.com/triton-lang/triton/tree/main/python/triton_kernels
set(DEFAULT_TRITON_KERNELS_TAG "v3.5.0")
# Set TRITON_KERNELS_SRC_DIR for use with local development with vLLM. We expect TRITON_KERNELS_SRC_DIR to
# be directly set to the triton_kernels python directory.
if (DEFINED ENV{TRITON_KERNELS_SRC_DIR})
message(STATUS "[triton_kernels] Fetch from $ENV{TRITON_KERNELS_SRC_DIR}")
FetchContent_Declare(
triton_kernels
SOURCE_DIR $ENV{TRITON_KERNELS_SRC_DIR}
)
else()
set(TRITON_GIT "https://github.com/triton-lang/triton.git")
message (STATUS "[triton_kernels] Fetch from ${TRITON_GIT}:${DEFAULT_TRITON_KERNELS_TAG}")
FetchContent_Declare(
triton_kernels
# TODO (varun) : Fetch just the triton_kernels directory from Triton
GIT_REPOSITORY https://github.com/triton-lang/triton.git
GIT_TAG ${DEFAULT_TRITON_KERNELS_TAG}
GIT_PROGRESS TRUE
SOURCE_SUBDIR python/triton_kernels/triton_kernels
)
endif()
# Fetch content
FetchContent_MakeAvailable(triton_kernels)
if (NOT triton_kernels_SOURCE_DIR)
message (FATAL_ERROR "[triton_kernels] Cannot resolve triton_kernels_SOURCE_DIR")
endif()
if (DEFINED ENV{TRITON_KERNELS_SRC_DIR})
set(TRITON_KERNELS_PYTHON_DIR "${triton_kernels_SOURCE_DIR}/")
else()
set(TRITON_KERNELS_PYTHON_DIR "${triton_kernels_SOURCE_DIR}/python/triton_kernels/triton_kernels/")
endif()
message (STATUS "[triton_kernels] triton_kernels is available at ${TRITON_KERNELS_PYTHON_DIR}")
add_custom_target(triton_kernels)
# Ensure the vllm/third_party directory exists before installation
install(CODE "file(MAKE_DIRECTORY \"\${CMAKE_INSTALL_PREFIX}/vllm/third_party/triton_kernels\")")
## Copy .py files to install directory.
install(DIRECTORY
${TRITON_KERNELS_PYTHON_DIR}
DESTINATION
vllm/third_party/triton_kernels/
COMPONENT triton_kernels
FILES_MATCHING PATTERN "*.py")

83
cmake/external_projects/vllm_flash_attn.cmake Normal file
View File

@@ -0,0 +1,83 @@
# vLLM flash attention requires VLLM_GPU_ARCHES to contain the set of target
# arches in the CMake syntax (75-real, 89-virtual, etc), since we clear the
# arches in the CUDA case (and instead set the gencodes on a per file basis)
# we need to manually set VLLM_GPU_ARCHES here.
if(VLLM_GPU_LANG STREQUAL "CUDA")
foreach(_ARCH ${CUDA_ARCHS})
string(REPLACE "." "" _ARCH "${_ARCH}")
list(APPEND VLLM_GPU_ARCHES "${_ARCH}-real")
endforeach()
endif()
#
# Build vLLM flash attention from source
#
# IMPORTANT: This has to be the last thing we do, because vllm-flash-attn uses the same macros/functions as vLLM.
# Because functions all belong to the global scope, vllm-flash-attn's functions overwrite vLLMs.
# They should be identical but if they aren't, this is a massive footgun.
#
# The vllm-flash-attn install rules are nested under vllm to make sure the library gets installed in the correct place.
# To only install vllm-flash-attn, use --component _vllm_fa2_C (for FA2) or --component _vllm_fa3_C (for FA3).
# If no component is specified, vllm-flash-attn is still installed.
# If VLLM_FLASH_ATTN_SRC_DIR is set, vllm-flash-attn is installed from that directory instead of downloading.
# This is to enable local development of vllm-flash-attn within vLLM.
# It can be set as an environment variable or passed as a cmake argument.
# The environment variable takes precedence.
if (DEFINED ENV{VLLM_FLASH_ATTN_SRC_DIR})
set(VLLM_FLASH_ATTN_SRC_DIR $ENV{VLLM_FLASH_ATTN_SRC_DIR})
endif()
if(VLLM_FLASH_ATTN_SRC_DIR)
FetchContent_Declare(
vllm-flash-attn SOURCE_DIR
${VLLM_FLASH_ATTN_SRC_DIR}
BINARY_DIR ${CMAKE_BINARY_DIR}/vllm-flash-attn
)
else()
FetchContent_Declare(
vllm-flash-attn
GIT_REPOSITORY https://github.com/vllm-project/flash-attention.git
GIT_TAG 86f8f157cf82aa2342743752b97788922dd7de43
GIT_PROGRESS TRUE
# Don't share the vllm-flash-attn build between build types
BINARY_DIR ${CMAKE_BINARY_DIR}/vllm-flash-attn
)
endif()
# Ensure the vllm/vllm_flash_attn directory exists before installation
install(CODE "file(MAKE_DIRECTORY \"\${CMAKE_INSTALL_PREFIX}/vllm/vllm_flash_attn\")" ALL_COMPONENTS)
# Make sure vllm-flash-attn install rules are nested under vllm/
# This is here to support installing all components under the same prefix with cmake --install.
# setup.py installs every component separately but uses the same prefix for all.
# ALL_COMPONENTS is used to avoid duplication for FA2 and FA3,
# and these statements don't hurt when installing neither component.
install(CODE "set(CMAKE_INSTALL_LOCAL_ONLY FALSE)" ALL_COMPONENTS)
install(CODE "set(OLD_CMAKE_INSTALL_PREFIX \"\${CMAKE_INSTALL_PREFIX}\")" ALL_COMPONENTS)
install(CODE "set(CMAKE_INSTALL_PREFIX \"\${CMAKE_INSTALL_PREFIX}/vllm/\")" ALL_COMPONENTS)
# Fetch the vllm-flash-attn library
FetchContent_MakeAvailable(vllm-flash-attn)
message(STATUS "vllm-flash-attn is available at ${vllm-flash-attn_SOURCE_DIR}")
# Restore the install prefix
install(CODE "set(CMAKE_INSTALL_PREFIX \"\${OLD_CMAKE_INSTALL_PREFIX}\")" ALL_COMPONENTS)
install(CODE "set(CMAKE_INSTALL_LOCAL_ONLY TRUE)" ALL_COMPONENTS)
# Copy over the vllm-flash-attn python files (duplicated for fa2 and fa3, in
# case only one is built, in the case both are built redundant work is done)
install(
DIRECTORY ${vllm-flash-attn_SOURCE_DIR}/vllm_flash_attn/
DESTINATION vllm/vllm_flash_attn
COMPONENT _vllm_fa2_C
FILES_MATCHING PATTERN "*.py"
)
install(
DIRECTORY ${vllm-flash-attn_SOURCE_DIR}/vllm_flash_attn/
DESTINATION vllm/vllm_flash_attn
COMPONENT _vllm_fa3_C
FILES_MATCHING PATTERN "*.py"
)