CMakeLists.txt 49 KB
Newer Older
1
cmake_minimum_required(VERSION 3.26)
bnellnm's avatar
bnellnm committed
2

3
4
5
6
7
8
9
10
11
12
13
# When building directly using CMake, make sure you run the install step
# (it places the .so files in the correct location).
#
# Example:
# mkdir build && cd build
# cmake -G Ninja -DVLLM_PYTHON_EXECUTABLE=`which python3` -DCMAKE_INSTALL_PREFIX=.. ..
# cmake --build . --target install
#
# If you want to only build one target, make sure to install it manually:
# cmake --build . --target _C
# cmake --install . --component _C
bnellnm's avatar
bnellnm committed
14
15
project(vllm_extensions LANGUAGES CXX)

16
17
18
19
set(CMAKE_CXX_STANDARD 17)
set(CMAKE_CXX_STANDARD_REQUIRED ON)


20
21
# CUDA by default, can be overridden by using -DVLLM_TARGET_DEVICE=... (used by setup.py)
set(VLLM_TARGET_DEVICE "cuda" CACHE STRING "Target device backend for vLLM")
bnellnm's avatar
bnellnm committed
22
message(STATUS "Build type: ${CMAKE_BUILD_TYPE}")
23
message(STATUS "Target device: ${VLLM_TARGET_DEVICE}")
bnellnm's avatar
bnellnm committed
24
25
26

include(${CMAKE_CURRENT_LIST_DIR}/cmake/utils.cmake)

27
28
29
# Suppress potential warnings about unused manually-specified variables
set(ignoreMe "${VLLM_PYTHON_PATH}")

30
31
32
# Prevent installation of dependencies (cutlass) by default.
install(CODE "set(CMAKE_INSTALL_LOCAL_ONLY TRUE)" ALL_COMPONENTS)

bnellnm's avatar
bnellnm committed
33
34
35
36
#
# Supported python versions.  These versions will be searched in order, the
# first match will be selected.  These should be kept in sync with setup.py.
#
37
set(PYTHON_SUPPORTED_VERSIONS "3.10" "3.11" "3.12" "3.13")
bnellnm's avatar
bnellnm committed
38
39

# Supported AMD GPU architectures.
40
set(HIP_SUPPORTED_ARCHS "gfx906;gfx908;gfx90a;gfx942;gfx950;gfx1030;gfx1100;gfx1101;gfx1150;gfx1151;gfx1152;gfx1153;gfx1200;gfx1201")
bnellnm's avatar
bnellnm committed
41

42
43
44
45
46
47
48
# ROCm installation prefix. Default to /opt/rocm but allow override via
# -DROCM_PATH=/your/rocm/path when invoking cmake.
if(NOT DEFINED ROCM_PATH)
  set(ROCM_PATH "/opt/rocm" CACHE PATH "ROCm installation prefix")
else()
  set(ROCM_PATH ${ROCM_PATH} CACHE PATH "ROCm installation prefix" FORCE)
endif()
bnellnm's avatar
bnellnm committed
49
50
51
52
53
54
55
56
#
# Supported/expected torch versions for CUDA/ROCm.
#
# Currently, having an incorrect pytorch version results in a warning
# rather than an error.
#
# Note: the CUDA torch version is derived from pyproject.toml and various
# requirements.txt files and should be kept consistent.  The ROCm torch
57
# versions are derived from docker/Dockerfile.rocm
bnellnm's avatar
bnellnm committed
58
#
59
60
set(TORCH_SUPPORTED_VERSION_CUDA "2.10.0")
set(TORCH_SUPPORTED_VERSION_ROCM "2.10.0")
bnellnm's avatar
bnellnm committed
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78

#
# Try to find python package with an executable that exactly matches
# `VLLM_PYTHON_EXECUTABLE` and is one of the supported versions.
#
if (VLLM_PYTHON_EXECUTABLE)
  find_python_from_executable(${VLLM_PYTHON_EXECUTABLE} "${PYTHON_SUPPORTED_VERSIONS}")
else()
  message(FATAL_ERROR
    "Please set VLLM_PYTHON_EXECUTABLE to the path of the desired python version"
    " before running cmake configure.")
endif()

#
# Update cmake's `CMAKE_PREFIX_PATH` with torch location.
#
append_cmake_prefix_path("torch" "torch.utils.cmake_prefix_path")

79
80
# Ensure the 'nvcc' command is in the PATH
find_program(NVCC_EXECUTABLE nvcc)
81
if (CUDA_FOUND AND NOT NVCC_EXECUTABLE)
82
83
84
    message(FATAL_ERROR "nvcc not found")
endif()

bnellnm's avatar
bnellnm committed
85
86
87
88
89
90
91
92
#
# Import torch cmake configuration.
# Torch also imports CUDA (and partially HIP) languages with some customizations,
# so there is no need to do this explicitly with check_language/enable_language,
# etc.
#
find_package(Torch REQUIRED)

93
94
95
# Supported NVIDIA architectures.
# This check must happen after find_package(Torch) because that's when CMAKE_CUDA_COMPILER_VERSION gets defined
if(DEFINED CMAKE_CUDA_COMPILER_VERSION AND
Johnny's avatar
Johnny committed
96
   CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL 13.0)
97
  set(CUDA_SUPPORTED_ARCHS "7.5;8.0;8.6;8.7;8.9;9.0;10.0;11.0;12.0;12.1")
Johnny's avatar
Johnny committed
98
elseif(DEFINED CMAKE_CUDA_COMPILER_VERSION AND
99
   CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL 12.8)
100
  set(CUDA_SUPPORTED_ARCHS "7.0;7.2;7.5;8.0;8.6;8.7;8.9;9.0;10.0;10.1;12.0;12.1")
101
102
103
104
else()
  set(CUDA_SUPPORTED_ARCHS "7.0;7.2;7.5;8.0;8.6;8.7;8.9;9.0")
endif()

105
106
107
108
109
110
111
112
#
# Forward the non-CUDA device extensions to external CMake scripts.
#
if (NOT VLLM_TARGET_DEVICE STREQUAL "cuda" AND
    NOT VLLM_TARGET_DEVICE STREQUAL "rocm")
    if (VLLM_TARGET_DEVICE STREQUAL "cpu")
        include(${CMAKE_CURRENT_LIST_DIR}/cmake/cpu_extension.cmake)
    else()
113
        return()
114
115
116
117
    endif()
    return()
endif()

bnellnm's avatar
bnellnm committed
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
#
# Set up GPU language and check the torch version and warn if it isn't
# what is expected.
#
if (NOT HIP_FOUND AND CUDA_FOUND)
  set(VLLM_GPU_LANG "CUDA")

  if (NOT Torch_VERSION VERSION_EQUAL ${TORCH_SUPPORTED_VERSION_CUDA})
    message(WARNING "Pytorch version ${TORCH_SUPPORTED_VERSION_CUDA} "
      "expected for CUDA build, saw ${Torch_VERSION} instead.")
  endif()
elseif(HIP_FOUND)
  set(VLLM_GPU_LANG "HIP")

  # Importing torch recognizes and sets up some HIP/ROCm configuration but does
  # not let cmake recognize .hip files. In order to get cmake to understand the
  # .hip extension automatically, HIP must be enabled explicitly.
  enable_language(HIP)

137
138
  # ROCm 5.X and 6.X
  if (ROCM_VERSION_DEV_MAJOR GREATER_EQUAL 5 AND
139
      Torch_VERSION VERSION_LESS ${TORCH_SUPPORTED_VERSION_ROCM})
140
    message(WARNING "Pytorch version >= ${TORCH_SUPPORTED_VERSION_ROCM} "
141
      "expected for ROCm build, saw ${Torch_VERSION} instead.")
bnellnm's avatar
bnellnm committed
142
143
144
145
146
  endif()
else()
  message(FATAL_ERROR "Can't find CUDA or HIP installation.")
endif()

147
148

if(VLLM_GPU_LANG STREQUAL "CUDA")
149
  #
150
  # For cuda we want to be able to control which architectures we compile for on
151
  # a per-file basis in order to cut down on compile time. So here we extract
152
  # the set of architectures we want to compile for and remove the from the
153
154
  # CMAKE_CUDA_FLAGS so that they are not applied globally.
  #
155
156
157
  clear_cuda_arches(CUDA_ARCH_FLAGS)
  extract_unique_cuda_archs_ascending(CUDA_ARCHS "${CUDA_ARCH_FLAGS}")
  message(STATUS "CUDA target architectures: ${CUDA_ARCHS}")
158
159
  # Filter the target architectures by the supported supported archs
  # since for some files we will build for all CUDA_ARCHS.
160
  cuda_archs_loose_intersection(CUDA_ARCHS
161
162
163
164
165
166
167
168
169
170
171
    "${CUDA_SUPPORTED_ARCHS}" "${CUDA_ARCHS}")
  message(STATUS "CUDA supported target architectures: ${CUDA_ARCHS}")
else()
  #
  # For other GPU targets override the GPU architectures detected by cmake/torch
  # and filter them by the supported versions for the current language.
  # The final set of arches is stored in `VLLM_GPU_ARCHES`.
  #
  override_gpu_arches(VLLM_GPU_ARCHES
    ${VLLM_GPU_LANG}
    "${${VLLM_GPU_LANG}_SUPPORTED_ARCHS}")
172
173
endif()

bnellnm's avatar
bnellnm committed
174
175
176
177
178
179
180
181
182
183
184
185
186
187
#
# Query torch for additional GPU compilation flags for the given
# `VLLM_GPU_LANG`.
# The final set of arches is stored in `VLLM_GPU_FLAGS`.
#
get_torch_gpu_compiler_flags(VLLM_GPU_FLAGS ${VLLM_GPU_LANG})

#
# Set nvcc parallelism.
#
if(NVCC_THREADS AND VLLM_GPU_LANG STREQUAL "CUDA")
  list(APPEND VLLM_GPU_FLAGS "--threads=${NVCC_THREADS}")
endif()

Johnny's avatar
Johnny committed
188
189
190
191
192
193
194
195
196
#
# Set compression mode for CUDA >=13.x.
#
if(VLLM_GPU_LANG STREQUAL "CUDA" AND
   DEFINED CMAKE_CUDA_COMPILER_VERSION AND
   CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL 13.0)
  list(APPEND VLLM_GPU_FLAGS "--compress-mode=size")
endif()

Aidyn-A's avatar
Aidyn-A committed
197
198
199
200
201
202
203
204
205
206
#
# Set CUDA include flags for CXX compiler.
#
if(VLLM_GPU_LANG STREQUAL "CUDA")
  set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -I${CUDA_TOOLKIT_ROOT_DIR}/include")
  if(CUDA_VERSION VERSION_GREATER_EQUAL 13.0)
    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -I${CUDA_TOOLKIT_ROOT_DIR}/include/cccl")
  endif()
endif()

207
208
#
# Use FetchContent for C++ dependencies that are compiled as part of vLLM's build process.
209
210
211
# setup.py will override FETCHCONTENT_BASE_DIR to play nicely with sccache.
# Each dependency that produces build artifacts should override its BINARY_DIR to avoid
# conflicts between build types. It should instead be set to ${CMAKE_BINARY_DIR}/<dependency>.
212
#
213
include(FetchContent)
214
file(MAKE_DIRECTORY ${FETCHCONTENT_BASE_DIR}) # Ensure the directory exists
215
message(STATUS "FetchContent base directory: ${FETCHCONTENT_BASE_DIR}")
216

217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
if(VLLM_GPU_LANG STREQUAL "HIP")
  #
  # Overriding the default -O set up by cmake, adding ggdb3 for the most verbose devug info
  #
  set(CMAKE_${VLLM_GPU_LANG}_FLAGS_DEBUG "${CMAKE_${VLLM_GPU_LANG}_FLAGS_DEBUG} -O0 -ggdb3")
  set(CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS_DEBUG} -O0 -ggdb3")

  #
  # Certain HIP functions are marked as [[nodiscard]], yet vllm ignores the result which generates
  # a lot of warnings that always mask real issues. Suppressing until this is properly addressed.
  #
  set(CMAKE_${VLLM_GPU_LANG}_FLAGS "${CMAKE_${VLLM_GPU_LANG}_FLAGS} -Wno-unused-result")
  set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-unused-result")
endif()

bnellnm's avatar
bnellnm committed
232
#
233
# Define other extension targets
bnellnm's avatar
bnellnm committed
234
235
#

236
237
238
239
240
241
242
243
244
245
246
#
# cumem_allocator extension
#

set(VLLM_CUMEM_EXT_SRC
  "csrc/cumem_allocator.cpp")

set_gencode_flags_for_srcs(
  SRCS "${VLLM_CUMEM_EXT_SRC}"
  CUDA_ARCHS "${CUDA_ARCHS}")

247
if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
248
  message(STATUS "Enabling cumem allocator extension.")
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
  if(VLLM_GPU_LANG STREQUAL "CUDA")
    # link against cuda driver library
    list(APPEND CUMEM_LIBS CUDA::cuda_driver)
  else()
    # link against rocm driver library. Prefer an absolute path to
    # libamdhip64.so inside ${ROCM_PATH}/lib if available, otherwise fall
    # back to linking by name "amdhip64".
    find_library(AMDHIP64_LIB
      NAMES amdhip64 libamdhip64.so
      PATHS ${ROCM_PATH}/lib
      NO_DEFAULT_PATH)
    if(AMDHIP64_LIB)
      message(STATUS "Found libamdhip64 at ${AMDHIP64_LIB}")
      list(APPEND CUMEM_LIBS ${AMDHIP64_LIB})
    else()
      message(WARNING "libamdhip64 not found in ${ROCM_PATH}/lib; falling back to linking 'amdhip64' by name")
      list(APPEND CUMEM_LIBS amdhip64)
    endif()
  endif()
268
  define_extension_target(
269
270
271
272
273
274
275
276
277
    cumem_allocator
    DESTINATION vllm
    LANGUAGE CXX
    SOURCES ${VLLM_CUMEM_EXT_SRC}
    LIBRARIES ${CUMEM_LIBS}
    USE_SABI 3.8
    WITH_SOABI)
endif()

bnellnm's avatar
bnellnm committed
278
279
280
281
282
#
# _C extension
#

set(VLLM_EXT_SRC
283
  "csrc/mamba/mamba_ssm/selective_scan_fwd.cu"
bnellnm's avatar
bnellnm committed
284
  "csrc/cache_kernels.cu"
285
  "csrc/cache_kernels_fused.cu"
286
287
  "csrc/attention/paged_attention_v1.cu"
  "csrc/attention/paged_attention_v2.cu"
288
  "csrc/attention/merge_attn_states.cu"
289
  "csrc/attention/vertical_slash_index.cu"
bnellnm's avatar
bnellnm committed
290
291
292
  "csrc/pos_encoding_kernels.cu"
  "csrc/activation_kernels.cu"
  "csrc/layernorm_kernels.cu"
293
  "csrc/fused_qknorm_rope_kernel.cu"
294
  "csrc/layernorm_quant_kernels.cu"
295
  "csrc/sampler.cu"
296
  "csrc/topk.cu"
297
  "csrc/cuda_view.cu"
bnellnm's avatar
bnellnm committed
298
  "csrc/quantization/gptq/q_gemm.cu"
299
300
  "csrc/quantization/w8a8/int8/scaled_quant.cu"
  "csrc/quantization/w8a8/fp8/common.cu"
301
  "csrc/quantization/fused_kernels/fused_layernorm_dynamic_per_token_quant.cu"
302
  "csrc/quantization/gguf/gguf_kernel.cu"
303
  "csrc/quantization/activation_kernels.cu"
bnellnm's avatar
bnellnm committed
304
  "csrc/cuda_utils_kernels.cu"
305
  "csrc/custom_all_reduce.cu"
306
  "csrc/torch_bindings.cpp")
bnellnm's avatar
bnellnm committed
307
308

if(VLLM_GPU_LANG STREQUAL "CUDA")
309
  SET(CUTLASS_ENABLE_HEADERS_ONLY ON CACHE BOOL "Enable only the header library")
310

311
  # Set CUTLASS_REVISION. Used for FetchContent. Also fixes some bogus messages when building.
312
  set(CUTLASS_REVISION "v4.4.2")
313

314
315
316
317
318
319
320
321
322
323
324
325
326
  # Use the specified CUTLASS source directory for compilation if VLLM_CUTLASS_SRC_DIR is provided
  if (DEFINED ENV{VLLM_CUTLASS_SRC_DIR})
    set(VLLM_CUTLASS_SRC_DIR $ENV{VLLM_CUTLASS_SRC_DIR})
  endif()

  if(VLLM_CUTLASS_SRC_DIR)
    if(NOT IS_ABSOLUTE VLLM_CUTLASS_SRC_DIR)
      get_filename_component(VLLM_CUTLASS_SRC_DIR "${VLLM_CUTLASS_SRC_DIR}" ABSOLUTE)
    endif()
    message(STATUS "The VLLM_CUTLASS_SRC_DIR is set, using ${VLLM_CUTLASS_SRC_DIR} for compilation")
    FetchContent_Declare(cutlass SOURCE_DIR ${VLLM_CUTLASS_SRC_DIR})
  else()
    FetchContent_Declare(
327
        cutlass
328
        GIT_REPOSITORY https://github.com/nvidia/cutlass.git
329
        # Please keep this in sync with CUTLASS_REVISION line above.
330
        GIT_TAG ${CUTLASS_REVISION}
331
        GIT_PROGRESS TRUE
332
333
334
335

        # Speed up CUTLASS download by retrieving only the specified GIT_TAG instead of the history.
        # Important: If GIT_SHALLOW is enabled then GIT_TAG works only with branch names and tags.
        # So if the GIT_TAG above is updated to a commit hash, GIT_SHALLOW must be set to FALSE
336
        GIT_SHALLOW TRUE
337
338
    )
  endif()
339
340
  FetchContent_MakeAvailable(cutlass)

bnellnm's avatar
bnellnm committed
341
342
  list(APPEND VLLM_EXT_SRC
    "csrc/quantization/awq/gemm_kernels.cu"
343
    "csrc/quantization/fp4/nvfp4_quant_entry.cu"
344
    "csrc/quantization/fp4/nvfp4_scaled_mm_entry.cu"
345
    "csrc/cutlass_extensions/common.cpp")
346
347
348
349
350
351
352
353

  set_gencode_flags_for_srcs(
    SRCS "${VLLM_EXT_SRC}"
    CUDA_ARCHS "${CUDA_ARCHS}")

  # Only build Marlin kernels if we are building for at least some compatible archs.
  # Keep building Marlin for 9.0 as there are some group sizes and shapes that
  # are not supported by Machete yet.
354
355
356

  # marlin arches for fp16 output
  cuda_archs_loose_intersection(MARLIN_ARCHS "8.0+PTX" "${CUDA_ARCHS}")
357
358
  # marlin has limited support for turing
  cuda_archs_loose_intersection(MARLIN_SM75_ARCHS "7.5" "${CUDA_ARCHS}")
359
360
361
362
363
364
  # marlin arches for bf16 output (we need 9.0 for bf16 atomicAdd PTX)
  cuda_archs_loose_intersection(MARLIN_BF16_ARCHS "8.0+PTX;9.0+PTX" "${CUDA_ARCHS}")
  # marlin arches for fp8 input
  # - sm80 doesn't support fp8 computation
  # - sm90 and sm100 don't support QMMA.16832.F32.E4M3.E4M3 SAAS instruction
  # so we only enable fp8 computation for SM89 (e.g. RTX 40x0)  and 12.0 (e.g. RTX 50x0)
Johnny's avatar
Johnny committed
365
  cuda_archs_loose_intersection(MARLIN_FP8_ARCHS "8.9;12.0;12.1" "${CUDA_ARCHS}")
366
367
  # marlin arches for other files
  cuda_archs_loose_intersection(MARLIN_OTHER_ARCHS "7.5;8.0+PTX" "${CUDA_ARCHS}")
368

369
  if (MARLIN_OTHER_ARCHS)
370
371
372
373
374
375

    #
    # For the Marlin kernels we automatically generate sources for various
    # preselected input type pairs and schedules.
    # Generate sources:
    set(MARLIN_GEN_SCRIPT
376
      ${CMAKE_CURRENT_SOURCE_DIR}/csrc/quantization/marlin/generate_kernels.py)
377
    file(MD5 ${MARLIN_GEN_SCRIPT} MARLIN_GEN_SCRIPT_HASH)
378
379
    list(JOIN CUDA_ARCHS "," CUDA_ARCHS_STR)
    set(MARLIN_GEN_SCRIPT_HASH_AND_ARCH "${MARLIN_GEN_SCRIPT_HASH}(ARCH:${CUDA_ARCHS_STR})")
380

381
382
    message(STATUS "Marlin generation script hash: ${MARLIN_GEN_SCRIPT_HASH_AND_ARCH}")
    message(STATUS "Last run Marlin generate script hash: $CACHE{MARLIN_GEN_SCRIPT_HASH_AND_ARCH}")
383

384
385
    if (NOT DEFINED CACHE{MARLIN_GEN_SCRIPT_HASH_AND_ARCH}
        OR NOT $CACHE{MARLIN_GEN_SCRIPT_HASH_AND_ARCH} STREQUAL ${MARLIN_GEN_SCRIPT_HASH_AND_ARCH})
386
387
      execute_process(
        COMMAND ${CMAKE_COMMAND} -E env
388
        PYTHONPATH=$ENV{PYTHONPATH}
389
          ${Python_EXECUTABLE} ${MARLIN_GEN_SCRIPT} ${CUDA_ARCHS_STR}
390
391
392
393
394
395
396
397
398
399
400
401
        RESULT_VARIABLE marlin_generation_result
        OUTPUT_VARIABLE marlin_generation_result
        OUTPUT_FILE ${CMAKE_CURRENT_BINARY_DIR}/marlin_generation.log
        ERROR_FILE ${CMAKE_CURRENT_BINARY_DIR}/marlin_generation.log
      )

      if (NOT marlin_generation_result EQUAL 0)
        message(FATAL_ERROR "Marlin generation failed."
                            " Result: \"${marlin_generation_result}\""
                            "\nCheck the log for details: "
                            "${CMAKE_CURRENT_BINARY_DIR}/marlin_generation.log")
      else()
402
403
        set(MARLIN_GEN_SCRIPT_HASH_AND_ARCH ${MARLIN_GEN_SCRIPT_HASH_AND_ARCH}
            CACHE STRING "Last run Marlin generate script hash and arch" FORCE)
404
405
406
407
408
409
        message(STATUS "Marlin generation completed successfully.")
      endif()
    else()
      message(STATUS "Marlin generation script has not changed, skipping generation.")
    endif()

410
    if (MARLIN_ARCHS)
411
      file(GLOB MARLIN_TEMPLATE_KERNEL_SRC "csrc/quantization/marlin/sm80_kernel_*_float16.cu")
412
413
414
415
416
417
418
419
420
      set_gencode_flags_for_srcs(
        SRCS "${MARLIN_TEMPLATE_KERNEL_SRC}"
        CUDA_ARCHS "${MARLIN_ARCHS}")
      if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8)
        set_source_files_properties(${MARLIN_TEMPLATE_KERNEL_SRC}
          PROPERTIES COMPILE_FLAGS "-static-global-template-stub=false")
      endif()
      list(APPEND VLLM_EXT_SRC ${MARLIN_TEMPLATE_KERNEL_SRC})

421
      file(GLOB MARLIN_TEMPLATE_BF16_KERNEL_SRC "csrc/quantization/marlin/sm80_kernel_*_bfloat16.cu")
422
423
424
425
426
427
428
429
      set_gencode_flags_for_srcs(
        SRCS "${MARLIN_TEMPLATE_BF16_KERNEL_SRC}"
        CUDA_ARCHS "${MARLIN_BF16_ARCHS}")
      if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8)
        set_source_files_properties(${MARLIN_TEMPLATE_BF16_KERNEL_SRC}
          PROPERTIES COMPILE_FLAGS "-static-global-template-stub=false")
      endif()
      list(APPEND VLLM_EXT_SRC ${MARLIN_TEMPLATE_BF16_KERNEL_SRC})
430
    endif()
431

432
    if (MARLIN_SM75_ARCHS)
433
      file(GLOB MARLIN_TEMPLATE_SM75_KERNEL_SRC "csrc/quantization/marlin/sm75_kernel_*.cu")
434
435
436
437
438
439
440
441
      set_gencode_flags_for_srcs(
        SRCS "${MARLIN_TEMPLATE_SM75_KERNEL_SRC}"
        CUDA_ARCHS "${MARLIN_SM75_ARCHS}")
      if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8)
        set_source_files_properties(${MARLIN_TEMPLATE_SM75_KERNEL_SRC}
          PROPERTIES COMPILE_FLAGS "-static-global-template-stub=false")
      endif()
      list(APPEND VLLM_EXT_SRC ${MARLIN_TEMPLATE_SM75_KERNEL_SRC})
442
443
    endif()

444
    if (MARLIN_FP8_ARCHS)
445
      file(GLOB MARLIN_TEMPLATE_FP8_KERNEL_SRC "csrc/quantization/marlin/sm89_kernel_*.cu")
446
447
448
449
450
451
452
453
454
455
      set_gencode_flags_for_srcs(
        SRCS "${MARLIN_TEMPLATE_FP8_KERNEL_SRC}"
        CUDA_ARCHS "${MARLIN_FP8_ARCHS}")
      if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8)
        set_source_files_properties(${MARLIN_TEMPLATE_FP8_KERNEL_SRC}
          PROPERTIES COMPILE_FLAGS "-static-global-template-stub=false")
      endif()
      list(APPEND VLLM_EXT_SRC ${MARLIN_TEMPLATE_FP8_KERNEL_SRC})
    endif()

456
    set(MARLIN_SRCS
457
458
459
460
       "csrc/quantization/marlin/marlin.cu"
       "csrc/quantization/marlin/marlin_int4_fp8_preprocess.cu"
       "csrc/quantization/marlin/gptq_marlin_repack.cu"
       "csrc/quantization/marlin/awq_marlin_repack.cu")
461
462
    set_gencode_flags_for_srcs(
      SRCS "${MARLIN_SRCS}"
463
      CUDA_ARCHS "${MARLIN_OTHER_ARCHS}")
464
    if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8)
465
      set_source_files_properties(${MARLIN_SRCS}
466
467
        PROPERTIES COMPILE_FLAGS "-static-global-template-stub=false")
    endif()
468
    list(APPEND VLLM_EXT_SRC "${MARLIN_SRCS}")
469

470
    message(STATUS "Building Marlin kernels for archs: ${MARLIN_OTHER_ARCHS}")
471
472
  else()
    message(STATUS "Not building Marlin kernels as no compatible archs found"
473
                   " in CUDA target architectures")
474
475
  endif()

476
477
  # Only build AllSpark kernels if we are building for at least some compatible archs.
  cuda_archs_loose_intersection(ALLSPARK_ARCHS "8.0;8.6;8.7;8.9" "${CUDA_ARCHS}")
478
  if (ALLSPARK_ARCHS)
479
480
481
482
483
484
485
486
487
488
    set(ALLSPARK_SRCS
       "csrc/quantization/gptq_allspark/allspark_repack.cu"
       "csrc/quantization/gptq_allspark/allspark_qgemm_w8a16.cu")
    set_gencode_flags_for_srcs(
      SRCS "${ALLSPARK_SRCS}"
      CUDA_ARCHS "${ALLSPARK_ARCHS}")
    list(APPEND VLLM_EXT_SRC "${ALLSPARK_SRCS}")
    message(STATUS "Building AllSpark kernels for archs: ${ALLSPARK_ARCHS}")
  else()
    message(STATUS "Not building AllSpark kernels as no compatible archs found"
489
                   " in CUDA target architectures")
490
491
  endif()

Johnny's avatar
Johnny committed
492
  # The nvfp4_scaled_mm_sm120 kernels for Blackwell SM12x require
493
  # CUDA 12.8 or later
Johnny's avatar
Johnny committed
494
495
496
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
    cuda_archs_loose_intersection(FP4_ARCHS "12.0f" "${CUDA_ARCHS}")
  else()
Johnny's avatar
Johnny committed
497
    cuda_archs_loose_intersection(FP4_ARCHS "12.0a;12.1a" "${CUDA_ARCHS}")
Johnny's avatar
Johnny committed
498
  endif()
499
500
501
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND FP4_ARCHS)
    set(SRCS
      "csrc/quantization/fp4/nvfp4_quant_kernels.cu"
502
      "csrc/quantization/fp4/activation_nvfp4_quant_fusion_kernels.cu"
503
504
505
      "csrc/quantization/fp4/nvfp4_experts_quant.cu"
      "csrc/quantization/fp4/nvfp4_scaled_mm_sm120_kernels.cu"
      "csrc/quantization/fp4/nvfp4_blockwise_moe_kernel.cu")
506
507
508
509
510
    set_gencode_flags_for_srcs(
      SRCS "${SRCS}"
      CUDA_ARCHS "${FP4_ARCHS}")
    list(APPEND VLLM_EXT_SRC "${SRCS}")
    list(APPEND VLLM_GPU_FLAGS "-DENABLE_NVFP4_SM120=1")
511
    list(APPEND VLLM_GPU_FLAGS "-DENABLE_CUTLASS_MOE_SM120=1")
512
513
514
515
516
517
518
    message(STATUS "Building NVFP4 for archs: ${FP4_ARCHS}")
  else()
    message(STATUS "Not building NVFP4 as no compatible archs were found.")
    # clear FP4_ARCHS
    set(FP4_ARCHS)
  endif()

519
  # FP4 Archs and flags
Johnny's avatar
Johnny committed
520
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
521
    cuda_archs_loose_intersection(FP4_ARCHS "10.0f;11.0f" "${CUDA_ARCHS}")
Johnny's avatar
Johnny committed
522
  else()
523
    cuda_archs_loose_intersection(FP4_ARCHS "10.0a;10.1a;10.3a" "${CUDA_ARCHS}")
Johnny's avatar
Johnny committed
524
  endif()
525
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND FP4_ARCHS)
526
    set(SRCS
527
      "csrc/quantization/fp4/nvfp4_quant_kernels.cu"
528
      "csrc/quantization/fp4/activation_nvfp4_quant_fusion_kernels.cu"
529
530
531
      "csrc/quantization/fp4/nvfp4_experts_quant.cu"
      "csrc/quantization/fp4/nvfp4_scaled_mm_kernels.cu"
      "csrc/quantization/fp4/nvfp4_blockwise_moe_kernel.cu")
532
533
534
535
    set_gencode_flags_for_srcs(
      SRCS "${SRCS}"
      CUDA_ARCHS "${FP4_ARCHS}")
    list(APPEND VLLM_EXT_SRC "${SRCS}")
536
    list(APPEND VLLM_GPU_FLAGS "-DENABLE_NVFP4_SM100=1")
537
    list(APPEND VLLM_GPU_FLAGS "-DENABLE_CUTLASS_MOE_SM100=1")
538
539
540
541
542
543
    message(STATUS "Building NVFP4 for archs: ${FP4_ARCHS}")
  else()
    message(STATUS "Not building NVFP4 as no compatible archs were found.")
    # clear FP4_ARCHS
    set(FP4_ARCHS)
  endif()
544

545
  # CUTLASS MLA Archs and flags
Johnny's avatar
Johnny committed
546
547
548
549
550
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
    cuda_archs_loose_intersection(MLA_ARCHS "10.0f;11.0f;12.0f" "${CUDA_ARCHS}")
  else()
    cuda_archs_loose_intersection(MLA_ARCHS "10.0a;10.1a;10.3a;12.0a;12.1a" "${CUDA_ARCHS}")
  endif()
551
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND MLA_ARCHS)
552
    set(SRCS
553
      "csrc/attention/mla/sm100_cutlass_mla_kernel.cu")
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
    set_gencode_flags_for_srcs(
      SRCS "${SRCS}"
      CUDA_ARCHS "${MLA_ARCHS}")
    list(APPEND VLLM_EXT_SRC "${SRCS}")
    list(APPEND VLLM_GPU_FLAGS "-DENABLE_CUTLASS_MLA=1")
    # Add MLA-specific include directories only to MLA source files
    set_source_files_properties(${SRCS}
      PROPERTIES INCLUDE_DIRECTORIES "${CUTLASS_DIR}/examples/77_blackwell_fmha;${CUTLASS_DIR}/examples/common")
    message(STATUS "Building CUTLASS MLA for archs: ${MLA_ARCHS}")
  else()
    message(STATUS "Not building CUTLASS MLA as no compatible archs were found.")
    # clear MLA_ARCHS
    set(MLA_ARCHS)
  endif()

569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
  # Expert-specialization MXFP8 blockscaled grouped kernels (SM100+).
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
    cuda_archs_loose_intersection(ES_MXFP8_GROUPED_MM_ARCHS "10.0f;11.0f" "${CUDA_ARCHS}")
  else()
    cuda_archs_loose_intersection(ES_MXFP8_GROUPED_MM_ARCHS "10.0a;10.1a;10.3a" "${CUDA_ARCHS}")
  endif()
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND ES_MXFP8_GROUPED_MM_ARCHS)
    set(SRCS
      "csrc/moe/mxfp8_moe/cutlass_mxfp8_grouped_mm.cu"
      "csrc/moe/mxfp8_moe/mxfp8_experts_quant.cu")
    set_gencode_flags_for_srcs(
      SRCS "${SRCS}"
      CUDA_ARCHS "${ES_MXFP8_GROUPED_MM_ARCHS}")
    list(APPEND VLLM_EXT_SRC "${SRCS}")
    list(APPEND VLLM_GPU_FLAGS "-DENABLE_ES_MXFP8_GROUPED_MM_SM100=1")
    message(STATUS "Building ES MXFP8 grouped kernels for archs: ${ES_MXFP8_GROUPED_MM_ARCHS}")
  else()
    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8
        AND ES_MXFP8_GROUPED_MM_ARCHS)
      message(STATUS "Not building ES MXFP8 grouped kernels as CUDA Compiler version is "
                     "not >= 12.8.")
    else()
      message(STATUS "Not building ES MXFP8 grouped kernels as no compatible archs found "
                     "in CUDA target architectures.")
    endif()
  endif()

596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
  # DeepSeek V3 fused A GEMM kernel (requires SM 9.0+, Hopper and later)
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
    cuda_archs_loose_intersection(DSV3_FUSED_A_GEMM_ARCHS "9.0a;10.0f;11.0f" "${CUDA_ARCHS}")
  else()
    cuda_archs_loose_intersection(DSV3_FUSED_A_GEMM_ARCHS "9.0a;10.0a;10.1a;10.3a" "${CUDA_ARCHS}")
  endif()
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.0 AND DSV3_FUSED_A_GEMM_ARCHS)
    set(DSV3_FUSED_A_GEMM_SRC "csrc/dsv3_fused_a_gemm.cu")
    set_gencode_flags_for_srcs(
      SRCS "${DSV3_FUSED_A_GEMM_SRC}"
      CUDA_ARCHS "${DSV3_FUSED_A_GEMM_ARCHS}")
    list(APPEND VLLM_EXT_SRC ${DSV3_FUSED_A_GEMM_SRC})
    message(STATUS "Building dsv3_fused_a_gemm for archs: ${DSV3_FUSED_A_GEMM_ARCHS}")
  else()
    message(STATUS "Not building dsv3_fused_a_gemm as no compatible archs found "
                   "in CUDA target architectures.")
  endif()

614
  #
615
  # Machete kernels
616

617
  # The machete kernels only work on hopper and require CUDA 12.0 or later.
618
619
  # Only build Machete kernels if we are building for something compatible with sm90a
  cuda_archs_loose_intersection(MACHETE_ARCHS "9.0a" "${CUDA_ARCHS}")
620
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.0 AND MACHETE_ARCHS)
621
    #
622
    # For the Machete kernels we automatically generate sources for various
623
624
    # preselected input type pairs and schedules.
    # Generate sources:
625
    set(MACHETE_GEN_SCRIPT
626
627
628
629
630
631
632
633
634
      ${CMAKE_CURRENT_SOURCE_DIR}/csrc/quantization/machete/generate.py)
    file(MD5 ${MACHETE_GEN_SCRIPT} MACHETE_GEN_SCRIPT_HASH)

    message(STATUS "Machete generation script hash: ${MACHETE_GEN_SCRIPT_HASH}")
    message(STATUS "Last run machete generate script hash: $CACHE{MACHETE_GEN_SCRIPT_HASH}")

    if (NOT DEFINED CACHE{MACHETE_GEN_SCRIPT_HASH}
        OR NOT $CACHE{MACHETE_GEN_SCRIPT_HASH} STREQUAL ${MACHETE_GEN_SCRIPT_HASH})
      execute_process(
635
        COMMAND ${CMAKE_COMMAND} -E env
636
        PYTHONPATH=${CMAKE_CURRENT_SOURCE_DIR}/csrc/cutlass_extensions/:${CUTLASS_DIR}/python/:${VLLM_PYTHON_PATH}:$ENV{PYTHONPATH}
637
638
639
640
641
642
643
644
645
          ${Python_EXECUTABLE} ${MACHETE_GEN_SCRIPT}
        RESULT_VARIABLE machete_generation_result
        OUTPUT_VARIABLE machete_generation_output
        OUTPUT_FILE ${CMAKE_CURRENT_BINARY_DIR}/machete_generation.log
        ERROR_FILE ${CMAKE_CURRENT_BINARY_DIR}/machete_generation.log
      )

      if (NOT machete_generation_result EQUAL 0)
        message(FATAL_ERROR "Machete generation failed."
646
                            " Result: \"${machete_generation_result}\""
647
648
649
                            "\nCheck the log for details: "
                            "${CMAKE_CURRENT_BINARY_DIR}/machete_generation.log")
      else()
650
        set(MACHETE_GEN_SCRIPT_HASH ${MACHETE_GEN_SCRIPT_HASH}
651
652
653
            CACHE STRING "Last run machete generate script hash" FORCE)
        message(STATUS "Machete generation completed successfully.")
      endif()
654
    else()
655
      message(STATUS "Machete generation script has not changed, skipping generation.")
656
    endif()
657

658
659
660
    # Add machete generated sources
    file(GLOB MACHETE_GEN_SOURCES "csrc/quantization/machete/generated/*.cu")
    list(APPEND VLLM_EXT_SRC ${MACHETE_GEN_SOURCES})
661

662
663
664
665
666
667
668
669
670
671
    # forward compatible
    set_gencode_flags_for_srcs(
      SRCS "${MACHETE_GEN_SOURCES}"
      CUDA_ARCHS "${MACHETE_ARCHS}")

    list(APPEND VLLM_EXT_SRC
      csrc/quantization/machete/machete_pytorch.cu)

    message(STATUS "Building Machete kernels for archs: ${MACHETE_ARCHS}")
  else()
672
    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.0
673
674
675
676
677
678
679
680
681
        AND MACHETE_ARCHS)
      message(STATUS "Not building Machete kernels as CUDA Compiler version is "
                     "not >= 12.0, we recommend upgrading to CUDA 12.0 or "
                     "later if you intend on running w4a16 quantized models on "
                     "Hopper.")
    else()
      message(STATUS "Not building Machete kernels as no compatible archs "
                     "found in CUDA target architectures")
    endif()
682
  endif()
683
684
685
686
687

  # Only build W4A8 kernels if we are building for something compatible with sm90a
  cuda_archs_loose_intersection(W4A8_ARCHS "9.0a" "${CUDA_ARCHS}")
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.0 AND W4A8_ARCHS)
    set(SRCS
688
689
690
691
       "csrc/quantization/cutlass_w4a8/w4a8_mm_entry.cu"
       "csrc/quantization/cutlass_w4a8/w4a8_grouped_mm_entry.cu"
       "csrc/quantization/cutlass_w4a8/w4a8_utils.cu"
       )
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712

    set_gencode_flags_for_srcs(
      SRCS "${SRCS}"
      CUDA_ARCHS "${W4A8_ARCHS}")

    list(APPEND VLLM_EXT_SRC "${SRCS}")

    message(STATUS "Building W4A8 kernels for archs: ${W4A8_ARCHS}")
  else()
    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.0
        AND W4A8_ARCHS)
      message(STATUS "Not building W4A8 kernels as CUDA Compiler version is "
                     "not >= 12.0, we recommend upgrading to CUDA 12.0 or "
                     "later if you intend on running w4a16 quantized models on "
                     "Hopper.")
    else()
      message(STATUS "Not building W4A8 kernels as no compatible archs "
                     "found in CUDA target architectures")
    endif()
  endif()

713
  # Hadacore kernels
714
  cuda_archs_loose_intersection(HADACORE_ARCHS "8.0+PTX;9.0+PTX" "${CUDA_ARCHS}")
715
716
717
718
719
720
721
722
723
  if(HADACORE_ARCHS)
    set(SRCS "csrc/quantization/hadamard/hadacore/hadamard_transform_cuda.cu")
    set_gencode_flags_for_srcs(
      SRCS "${SRCS}"
      CUDA_ARCHS "${HADACORE_ARCHS}")
    list(APPEND VLLM_EXT_SRC "${SRCS}")
    message(STATUS "Building hadacore")
  endif()

724
# if CUDA endif
bnellnm's avatar
bnellnm committed
725
endif()
726
727
728
729
730
731
732
733

if (VLLM_GPU_LANG STREQUAL "HIP")
  # Add QuickReduce kernels
  list(APPEND VLLM_EXT_SRC
    "csrc/custom_quickreduce.cu"
  )
# if ROCM endif
endif()
bnellnm's avatar
bnellnm committed
734

735
message(STATUS "Enabling C extension.")
736
define_extension_target(
bnellnm's avatar
bnellnm committed
737
738
739
740
741
742
  _C
  DESTINATION vllm
  LANGUAGE ${VLLM_GPU_LANG}
  SOURCES ${VLLM_EXT_SRC}
  COMPILE_FLAGS ${VLLM_GPU_FLAGS}
  ARCHITECTURES ${VLLM_GPU_ARCHES}
743
  INCLUDE_DIRECTORIES ${CUTLASS_INCLUDE_DIR}
744
  INCLUDE_DIRECTORIES ${CUTLASS_TOOLS_UTIL_INCLUDE_DIR}
745
  USE_SABI 3
bnellnm's avatar
bnellnm committed
746
747
  WITH_SOABI)

748
749
# If CUTLASS is compiled on NVCC >= 12.5, it by default uses
# cudaGetDriverEntryPointByVersion as a wrapper to avoid directly calling the
750
751
752
753
# driver API. This causes problems when linking with earlier versions of CUDA.
# Setting this variable sidesteps the issue by calling the driver directly.
target_compile_definitions(_C PRIVATE CUTLASS_ENABLE_DIRECT_CUDA_DRIVER_CALL=1)

754
755
756
757
758
759
760
# add OR VLLM_GPU_LANG STREQUAL "HIP" here once
# https://github.com/vllm-project/vllm/issues/35163 is resolved
if(VLLM_GPU_LANG STREQUAL "CUDA")
  #
  # _C_stable_libtorch extension (ops registered via STABLE_TORCH_LIBRARY)
  #
  set(VLLM_STABLE_EXT_SRC
761
762
763
    "csrc/libtorch_stable/torch_bindings.cpp"
    "csrc/cutlass_extensions/common.cpp"
    "csrc/libtorch_stable/quantization/w8a8/cutlass/scaled_mm_entry.cu")
764
765

  if(VLLM_GPU_LANG STREQUAL "CUDA")
766
767
768
769
    list(APPEND VLLM_STABLE_EXT_SRC
      "csrc/libtorch_stable/permute_cols.cu"
      "csrc/libtorch_stable/quantization/w8a8/fp8/per_token_group_quant.cu"
      "csrc/libtorch_stable/quantization/w8a8/int8/per_token_group_quant.cu")
770
771
772
773
774
775
776
777
  endif()

  if(VLLM_GPU_LANG STREQUAL "CUDA")
    set_gencode_flags_for_srcs(
      SRCS "${VLLM_STABLE_EXT_SRC}"
      CUDA_ARCHS "${CUDA_ARCHS}")
  endif()

778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
  #
  # CUTLASS scaled_mm kernels (moved from _C to _C_stable_libtorch)
  #
  set(SCALED_MM_3X_ARCHS)
  # The cutlass_scaled_mm kernels for Hopper (c3x, i.e. CUTLASS 3.x) require
  # CUDA 12.0 or later
  cuda_archs_loose_intersection(SCALED_MM_ARCHS "9.0a;" "${CUDA_ARCHS}")
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.0 AND SCALED_MM_ARCHS)
    set(SRCS
       "csrc/libtorch_stable/quantization/w8a8/cutlass/scaled_mm_c3x_sm90.cu"
       "csrc/libtorch_stable/quantization/w8a8/cutlass/c3x/scaled_mm_sm90_fp8.cu"
       "csrc/libtorch_stable/quantization/w8a8/cutlass/c3x/scaled_mm_sm90_int8.cu"
       "csrc/libtorch_stable/quantization/w8a8/cutlass/c3x/scaled_mm_azp_sm90_int8.cu"
       "csrc/libtorch_stable/quantization/w8a8/cutlass/c3x/scaled_mm_blockwise_sm90_fp8.cu")
    set_gencode_flags_for_srcs(
      SRCS "${SRCS}"
      CUDA_ARCHS "${SCALED_MM_ARCHS}")
    list(APPEND VLLM_STABLE_EXT_SRC "${SRCS}")
    list(APPEND VLLM_GPU_FLAGS "-DENABLE_SCALED_MM_SM90=1")
    # Let scaled_mm_c2x know it doesn't need to build these arches
    list(APPEND SCALED_MM_3X_ARCHS "${SCALED_MM_ARCHS}")
    message(STATUS "Building scaled_mm_c3x_sm90 for archs: ${SCALED_MM_ARCHS}")
  else()
    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.0 AND SCALED_MM_ARCHS)
      message(STATUS "Not building scaled_mm_c3x_sm90 as CUDA Compiler version is "
                     "not >= 12.0, we recommend upgrading to CUDA 12.0 or "
                     "later if you intend on running FP8 quantized models on "
                     "Hopper.")
    else()
      message(STATUS "Not building scaled_mm_c3x_sm90 as no compatible archs found "
                     "in CUDA target architectures")
    endif()
  endif()


  # The cutlass_scaled_mm kernels for Blackwell SM12x (c3x, i.e. CUTLASS 3.x) require
  # CUDA 12.8 or later
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
    cuda_archs_loose_intersection(SCALED_MM_ARCHS "12.0f" "${CUDA_ARCHS}")
  else()
    cuda_archs_loose_intersection(SCALED_MM_ARCHS "12.0a;12.1a" "${CUDA_ARCHS}")
  endif()
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
    set(SRCS
      "csrc/libtorch_stable/quantization/w8a8/cutlass/scaled_mm_c3x_sm120.cu"
      "csrc/libtorch_stable/quantization/w8a8/cutlass/c3x/scaled_mm_sm120_fp8.cu"
      "csrc/libtorch_stable/quantization/w8a8/cutlass/c3x/scaled_mm_blockwise_sm120_fp8.cu"
    )
    set_gencode_flags_for_srcs(
      SRCS "${SRCS}"
      CUDA_ARCHS "${SCALED_MM_ARCHS}")
    list(APPEND VLLM_STABLE_EXT_SRC "${SRCS}")
    list(APPEND VLLM_GPU_FLAGS "-DENABLE_SCALED_MM_SM120=1")
    # Let scaled_mm_c2x know it doesn't need to build these arches
    list(APPEND SCALED_MM_3X_ARCHS "${SCALED_MM_ARCHS}")
    message(STATUS "Building scaled_mm_c3x_sm120 for archs: ${SCALED_MM_ARCHS}")
  else()
    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
      message(STATUS "Not building scaled_mm_c3x_sm120 as CUDA Compiler version is "
                     "not >= 12.8, we recommend upgrading to CUDA 12.8 or "
                     "later if you intend on running FP8 quantized models on "
                     "Blackwell.")
    else()
      message(STATUS "Not building scaled_mm_c3x_120 as no compatible archs found "
                     "in CUDA target architectures")
    endif()
  endif()


  # The cutlass_scaled_mm kernels for Blackwell SM100 (c3x, i.e. CUTLASS 3.x)
  # require CUDA 12.8 or later
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
    cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0f;11.0f" "${CUDA_ARCHS}")
  else()
    cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a;10.1a;10.3a" "${CUDA_ARCHS}")
  endif()
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
    set(SRCS
      "csrc/libtorch_stable/quantization/w8a8/cutlass/scaled_mm_c3x_sm100.cu"
      "csrc/libtorch_stable/quantization/w8a8/cutlass/c3x/scaled_mm_sm100_fp8.cu"
      "csrc/libtorch_stable/quantization/w8a8/cutlass/c3x/scaled_mm_blockwise_sm100_fp8.cu"
    )
    set_gencode_flags_for_srcs(
      SRCS "${SRCS}"
      CUDA_ARCHS "${SCALED_MM_ARCHS}")
    list(APPEND VLLM_STABLE_EXT_SRC "${SRCS}")
    list(APPEND VLLM_GPU_FLAGS "-DENABLE_SCALED_MM_SM100=1")
    # Let scaled_mm_c2x know it doesn't need to build these arches
    list(APPEND SCALED_MM_3X_ARCHS "${SCALED_MM_ARCHS}")
    message(STATUS "Building scaled_mm_c3x_sm100 for archs: ${SCALED_MM_ARCHS}")
  else()
    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
      message(STATUS "Not building scaled_mm_c3x_sm100 as CUDA Compiler version is "
                     "not >= 12.8, we recommend upgrading to CUDA 12.8 or "
                     "later if you intend on running FP8 quantized models on "
                     "Blackwell.")
    else()
      message(STATUS "Not building scaled_mm_c3x_100 as no compatible archs found "
                     "in CUDA target architectures")
    endif()
  endif()

  #
  # For the cutlass_scaled_mm kernels we want to build the c2x (CUTLASS 2.x)
  # kernels for the remaining archs that are not already built for 3x.
  # (Build 8.9 for FP8)
  cuda_archs_loose_intersection(SCALED_MM_2X_ARCHS
    "7.5;8.0;8.7;8.9+PTX" "${CUDA_ARCHS}")
  # subtract out the archs that are already built for 3x
  list(REMOVE_ITEM SCALED_MM_2X_ARCHS ${SCALED_MM_3X_ARCHS})
  if (SCALED_MM_2X_ARCHS)
    set(SRCS "csrc/libtorch_stable/quantization/w8a8/cutlass/scaled_mm_c2x.cu")
    set_gencode_flags_for_srcs(
      SRCS "${SRCS}"
      CUDA_ARCHS "${SCALED_MM_2X_ARCHS}")
    list(APPEND VLLM_STABLE_EXT_SRC "${SRCS}")
    list(APPEND VLLM_GPU_FLAGS "-DENABLE_SCALED_MM_C2X=1")
    message(STATUS "Building scaled_mm_c2x for archs: ${SCALED_MM_2X_ARCHS}")
  else()
    if (SCALED_MM_3X_ARCHS)
      message(STATUS "Not building scaled_mm_c2x as all archs are already built"
                     " for and covered by scaled_mm_c3x")
    else()
      message(STATUS "Not building scaled_mm_c2x as no compatible archs found "
                    "in CUDA target architectures")
    endif()
  endif()

  #
  # CUTLASS MoE kernels (moved from _C to _C_stable_libtorch)
  #

  # The MoE kernel cutlass_moe_mm requires CUDA 12.3 or later (and ONLY works
  # on Hopper). get_cutlass_(batched_)moe_mm_data should only be compiled
  # if it's possible to compile MoE kernels that use its output.
  cuda_archs_loose_intersection(SCALED_MM_ARCHS "9.0a" "${CUDA_ARCHS}")
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.3 AND SCALED_MM_ARCHS)
    set(SRCS "csrc/libtorch_stable/quantization/w8a8/cutlass/moe/grouped_mm_c3x_sm90.cu")
    set_gencode_flags_for_srcs(
      SRCS "${SRCS}"
      CUDA_ARCHS "${SCALED_MM_ARCHS}")
    list(APPEND VLLM_STABLE_EXT_SRC "${SRCS}")
    list(APPEND VLLM_GPU_FLAGS "-DENABLE_CUTLASS_MOE_SM90=1")
    message(STATUS "Building grouped_mm_c3x for archs: ${SCALED_MM_ARCHS}")
  else()
    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.3 AND SCALED_MM_ARCHS)
      message(STATUS "Not building grouped_mm_c3x kernels as CUDA Compiler version is "
                     "not >= 12.3, we recommend upgrading to CUDA 12.3 or later "
                     "if you intend on running FP8 quantized MoE models on Hopper.")
    else()
      message(STATUS "Not building grouped_mm_c3x as no compatible archs found "
                     "in CUDA target architectures.")
    endif()
  endif()

  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
    cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0f;11.0f" "${CUDA_ARCHS}")
  else()
    cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a;10.1a;10.3a" "${CUDA_ARCHS}")
  endif()
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
    set(SRCS "csrc/libtorch_stable/quantization/w8a8/cutlass/moe/grouped_mm_c3x_sm100.cu")
    set_gencode_flags_for_srcs(
      SRCS "${SRCS}"
      CUDA_ARCHS "${SCALED_MM_ARCHS}")
    list(APPEND VLLM_STABLE_EXT_SRC "${SRCS}")
    list(APPEND VLLM_GPU_FLAGS "-DENABLE_CUTLASS_MOE_SM100=1")
    message(STATUS "Building grouped_mm_c3x for archs: ${SCALED_MM_ARCHS}")
  else()
    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
      message(STATUS "Not building grouped_mm_c3x kernels as CUDA Compiler version is "
                     "not >= 12.8, we recommend upgrading to CUDA 12.8 or later "
                     "if you intend on running FP8 quantized MoE models on Blackwell.")
    else()
      message(STATUS "Not building grouped_mm_c3x as no compatible archs found "
                     "in CUDA target architectures.")
    endif()
  endif()

  # moe_data.cu is used by all CUTLASS MoE kernels.
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
    cuda_archs_loose_intersection(CUTLASS_MOE_DATA_ARCHS "9.0a;10.0f;11.0f;12.0f" "${CUDA_ARCHS}")
  else()
    cuda_archs_loose_intersection(CUTLASS_MOE_DATA_ARCHS "9.0a;10.0a;10.1a;10.3a;12.0a;12.1a" "${CUDA_ARCHS}")
  endif()
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.3 AND CUTLASS_MOE_DATA_ARCHS)
    set(SRCS "csrc/libtorch_stable/quantization/w8a8/cutlass/moe/moe_data.cu")
    set_gencode_flags_for_srcs(
      SRCS "${SRCS}"
      CUDA_ARCHS "${CUTLASS_MOE_DATA_ARCHS}")
    list(APPEND VLLM_STABLE_EXT_SRC "${SRCS}")
    message(STATUS "Building moe_data for archs: ${CUTLASS_MOE_DATA_ARCHS}")
  else()
    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.3 AND CUTLASS_MOE_DATA_ARCHS)
      message(STATUS "Not building moe_data as CUDA Compiler version is "
                     "not >= 12.3, we recommend upgrading to CUDA 12.3 or later "
                     "if you intend on running FP8 quantized MoE models on Hopper or Blackwell.")
    else()
      message(STATUS "Not building moe_data as no compatible archs found "
                     "in CUDA target architectures.")
    endif()
  endif()

981
982
983
984
985
986
987
988
  message(STATUS "Enabling C_stable extension.")
  define_extension_target(
    _C_stable_libtorch
    DESTINATION vllm
    LANGUAGE ${VLLM_GPU_LANG}
    SOURCES ${VLLM_STABLE_EXT_SRC}
    COMPILE_FLAGS ${VLLM_GPU_FLAGS}
    ARCHITECTURES ${VLLM_GPU_ARCHES}
989
    INCLUDE_DIRECTORIES ${CUTLASS_INCLUDE_DIR} ${CUTLASS_TOOLS_UTIL_INCLUDE_DIR}
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
    USE_SABI 3
    WITH_SOABI)

  # Set TORCH_TARGET_VERSION for stable ABI compatibility.
  # This ensures we only use C-shim APIs available in PyTorch 2.10.
  # _C_stable_libtorch is abi compatible with PyTorch >= TORCH_TARGET_VERSION
  # which is currently set to 2.10.
  target_compile_definitions(_C_stable_libtorch PRIVATE
    TORCH_TARGET_VERSION=0x020A000000000000ULL)

  # Needed to use cuda APIs from C-shim
  target_compile_definitions(_C_stable_libtorch PRIVATE
    USE_CUDA)
1003
1004
1005
1006

  # Needed by CUTLASS kernels
  target_compile_definitions(_C_stable_libtorch PRIVATE
    CUTLASS_ENABLE_DIRECT_CUDA_DRIVER_CALL=1)
1007
1008
endif()

bnellnm's avatar
bnellnm committed
1009
1010
1011
1012
1013
#
# _moe_C extension
#

set(VLLM_MOE_EXT_SRC
1014
  "csrc/moe/torch_bindings.cpp"
1015
  "csrc/moe/moe_align_sum_kernels.cu"
1016
  "csrc/moe/topk_softmax_kernels.cu")
bnellnm's avatar
bnellnm committed
1017

1018
if(VLLM_GPU_LANG STREQUAL "CUDA")
1019
1020
  list(APPEND VLLM_MOE_EXT_SRC
    "csrc/moe/moe_wna16.cu"
1021
    "csrc/moe/grouped_topk_kernels.cu"
1022
    "csrc/moe/gpt_oss_router_gemm.cu"
1023
    "csrc/moe/router_gemm.cu")
1024
1025
endif()

1026
1027
1028
1029
1030
1031
1032
1033
if(VLLM_GPU_LANG STREQUAL "CUDA")
  set(MOE_PERMUTE_SRC
      "csrc/moe/permute_unpermute_kernels/moe_permute_unpermute_kernel.cu"
      "csrc/moe/moe_permute_unpermute_op.cu")

  list(APPEND VLLM_MOE_EXT_SRC "${MOE_PERMUTE_SRC}")
endif()

1034
1035
1036
1037
set_gencode_flags_for_srcs(
  SRCS "${VLLM_MOE_EXT_SRC}"
  CUDA_ARCHS "${CUDA_ARCHS}")

1038
if(VLLM_GPU_LANG STREQUAL "CUDA")
1039
1040
1041
1042
1043
1044
1045
  set(VLLM_MOE_WNA16_SRC
    "csrc/moe/moe_wna16.cu")

  set_gencode_flags_for_srcs(
    SRCS "${VLLM_MOE_WNA16_SRC}"
    CUDA_ARCHS "${CUDA_ARCHS}")

1046
  list(APPEND VLLM_MOE_EXT_SRC "${VLLM_MOE_WNA16_SRC}")
1047
1048
1049
1050
  # moe marlin arches
  # note that we always set `use_atomic_add=False` for moe marlin now,
  # so we don't need 9.0 for bf16 atomicAdd PTX
  cuda_archs_loose_intersection(MARLIN_MOE_ARCHS "8.0+PTX" "${CUDA_ARCHS}")
1051
1052
  # moe marlin has limited support for turing
  cuda_archs_loose_intersection(MARLIN_MOE_SM75_ARCHS "7.5" "${CUDA_ARCHS}")
1053
1054
1055
1056
  # moe marlin arches for fp8 input
  # - sm80 doesn't support fp8 computation
  # - sm90 and sm100 don't support QMMA.16832.F32.E4M3.E4M3 SAAS instruction
  # so we only enable fp8 computation for SM89 (e.g. RTX 40x0)  and 12.0 (e.g. RTX 50x0)
Johnny's avatar
Johnny committed
1057
  cuda_archs_loose_intersection(MARLIN_MOE_FP8_ARCHS "8.9;12.0;12.1" "${CUDA_ARCHS}")
1058
1059
1060
  # moe marlin arches for other files
  cuda_archs_loose_intersection(MARLIN_MOE_OTHER_ARCHS "7.5;8.0+PTX" "${CUDA_ARCHS}")
  if (MARLIN_MOE_OTHER_ARCHS)
1061

1062
1063
1064
1065
1066
1067
1068
    #
    # For the Marlin MOE kernels we automatically generate sources for various
    # preselected input type pairs and schedules.
    # Generate sources:
    set(MOE_MARLIN_GEN_SCRIPT
      ${CMAKE_CURRENT_SOURCE_DIR}/csrc/moe/marlin_moe_wna16/generate_kernels.py)
    file(MD5 ${MOE_MARLIN_GEN_SCRIPT} MOE_MARLIN_GEN_SCRIPT_HASH)
1069
1070
    list(JOIN CUDA_ARCHS "," CUDA_ARCHS_STR)
    set(MOE_MARLIN_GEN_SCRIPT_HASH_AND_ARCH "${MOE_MARLIN_GEN_SCRIPT_HASH}(ARCH:${CUDA_ARCHS_STR})")
1071

1072
1073
    message(STATUS "Marlin MOE generation script hash with arch: ${MOE_MARLIN_GEN_SCRIPT_HASH_AND_ARCH}")
    message(STATUS "Last run Marlin MOE generate script hash with arch: $CACHE{MOE_MARLIN_GEN_SCRIPT_HASH_AND_ARCH}")
1074

1075
1076
    if (NOT DEFINED CACHE{MOE_MARLIN_GEN_SCRIPT_HASH_AND_ARCH}
        OR NOT $CACHE{MOE_MARLIN_GEN_SCRIPT_HASH_AND_ARCH} STREQUAL ${MOE_MARLIN_GEN_SCRIPT_HASH_AND_ARCH})
1077
1078
      execute_process(
        COMMAND ${CMAKE_COMMAND} -E env
1079
        PYTHONPATH=$ENV{PYTHONPATH}
1080
          ${Python_EXECUTABLE} ${MOE_MARLIN_GEN_SCRIPT} ${CUDA_ARCHS_STR}
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
        RESULT_VARIABLE moe_marlin_generation_result
        OUTPUT_VARIABLE moe_marlin_generation_output
        OUTPUT_FILE ${CMAKE_CURRENT_BINARY_DIR}/moe_marlin_generation.log
        ERROR_FILE ${CMAKE_CURRENT_BINARY_DIR}/moe_marlin_generation.log
      )

      if (NOT moe_marlin_generation_result EQUAL 0)
        message(FATAL_ERROR "Marlin MOE generation failed."
                            " Result: \"${moe_marlin_generation_result}\""
                            "\nCheck the log for details: "
                            "${CMAKE_CURRENT_BINARY_DIR}/moe_marlin_generation.log")
      else()
1093
        set(MOE_MARLIN_GEN_SCRIPT_HASH_AND_ARCH ${MOE_MARLIN_GEN_SCRIPT_HASH_AND_ARCH}
1094
1095
1096
1097
1098
1099
1100
            CACHE STRING "Last run Marlin MOE generate script hash" FORCE)
        message(STATUS "Marlin MOE generation completed successfully.")
      endif()
    else()
      message(STATUS "Marlin MOE generation script has not changed, skipping generation.")
    endif()

1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
    if (MARLIN_MOE_ARCHS)
      file(GLOB MARLIN_MOE_SRC "csrc/moe/marlin_moe_wna16/sm80_kernel_*.cu")
      set_gencode_flags_for_srcs(
        SRCS "${MARLIN_MOE_SRC}"
        CUDA_ARCHS "${MARLIN_MOE_ARCHS}")
      if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8)
        set_source_files_properties(${MARLIN_MOE_SRC}
          PROPERTIES COMPILE_FLAGS "-static-global-template-stub=false")
      endif()
      list(APPEND VLLM_MOE_EXT_SRC ${MARLIN_MOE_SRC})
    endif()

1113
    if (MARLIN_MOE_SM75_ARCHS)
1114
1115
1116
1117
1118
1119
1120
1121
1122
      file(GLOB MARLIN_MOE_SM75_SRC "csrc/moe/marlin_moe_wna16/sm75_kernel_*.cu")
      set_gencode_flags_for_srcs(
        SRCS "${MARLIN_MOE_SM75_SRC}"
        CUDA_ARCHS "${MARLIN_MOE_SM75_ARCHS}")
      if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8)
        set_source_files_properties(${MARLIN_MOE_SM75_SRC}
          PROPERTIES COMPILE_FLAGS "-static-global-template-stub=false")
      endif()
      list(APPEND VLLM_MOE_EXT_SRC ${MARLIN_MOE_SM75_SRC})
1123
    endif()
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135

    if (MARLIN_MOE_FP8_ARCHS)
      file(GLOB MARLIN_MOE_FP8_SRC "csrc/moe/marlin_moe_wna16/sm89_kernel_*.cu")
      set_gencode_flags_for_srcs(
        SRCS "${MARLIN_MOE_FP8_SRC}"
        CUDA_ARCHS "${MARLIN_MOE_FP8_ARCHS}")
      if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8)
        set_source_files_properties(${MARLIN_MOE_FP8_SRC}
          PROPERTIES COMPILE_FLAGS "-static-global-template-stub=false")
      endif()
      list(APPEND VLLM_MOE_EXT_SRC ${MARLIN_MOE_FP8_SRC})
    endif()
1136

1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
    set(MARLIN_MOE_OTHER_SRC "csrc/moe/marlin_moe_wna16/ops.cu")
    set_gencode_flags_for_srcs(
      SRCS "${MARLIN_MOE_OTHER_SRC}"
      CUDA_ARCHS "${MARLIN_MOE_OTHER_ARCHS}")
    if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8)
      set_source_files_properties(${MARLIN_MOE_OTHER_SRC}
        PROPERTIES COMPILE_FLAGS "-static-global-template-stub=false")
    endif()
    list(APPEND VLLM_MOE_EXT_SRC "${MARLIN_MOE_OTHER_SRC}")

    message(STATUS "Building Marlin MOE kernels for archs: ${MARLIN_MOE_OTHER_ARCHS}")
1148
1149
  else()
    message(STATUS "Not building Marlin MOE kernels as no compatible archs found"
1150
                   " in CUDA target architectures")
1151
  endif()
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172

  # DeepSeek V3 router GEMM kernel - requires SM90+
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
    cuda_archs_loose_intersection(DSV3_ROUTER_GEMM_ARCHS "9.0a;10.0f;11.0f" "${CUDA_ARCHS}")
  else()
    cuda_archs_loose_intersection(DSV3_ROUTER_GEMM_ARCHS "9.0a;10.0a;10.1a;10.3a" "${CUDA_ARCHS}")
  endif()
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.0 AND DSV3_ROUTER_GEMM_ARCHS)
    set(DSV3_ROUTER_GEMM_SRC
      "csrc/moe/dsv3_router_gemm_entry.cu"
      "csrc/moe/dsv3_router_gemm_float_out.cu"
      "csrc/moe/dsv3_router_gemm_bf16_out.cu")
    set_gencode_flags_for_srcs(
      SRCS "${DSV3_ROUTER_GEMM_SRC}"
      CUDA_ARCHS "${DSV3_ROUTER_GEMM_ARCHS}")
    list(APPEND VLLM_MOE_EXT_SRC "${DSV3_ROUTER_GEMM_SRC}")
    message(STATUS "Building DSV3 router GEMM kernel for archs: ${DSV3_ROUTER_GEMM_ARCHS}")
  else()
    message(STATUS "Not building DSV3 router GEMM kernel as no compatible archs found"
                   " (requires SM90+ and CUDA >= 12.0)")
  endif()
1173
1174
endif()

1175
message(STATUS "Enabling moe extension.")
1176
define_extension_target(
bnellnm's avatar
bnellnm committed
1177
1178
1179
1180
1181
1182
  _moe_C
  DESTINATION vllm
  LANGUAGE ${VLLM_GPU_LANG}
  SOURCES ${VLLM_MOE_EXT_SRC}
  COMPILE_FLAGS ${VLLM_GPU_FLAGS}
  ARCHITECTURES ${VLLM_GPU_ARCHES}
1183
1184
  INCLUDE_DIRECTORIES ${CUTLASS_INCLUDE_DIR}
  INCLUDE_DIRECTORIES ${CUTLASS_TOOLS_UTIL_INCLUDE_DIR}
1185
  USE_SABI 3
bnellnm's avatar
bnellnm committed
1186
1187
  WITH_SOABI)

1188
1189
1190
1191
1192
1193
if(VLLM_GPU_LANG STREQUAL "HIP")
  #
  # _rocm_C extension
  #
  set(VLLM_ROCM_EXT_SRC
    "csrc/rocm/torch_bindings.cpp"
1194
    "csrc/rocm/skinny_gemms.cu"
1195
1196
    "csrc/rocm/attention.cu")

1197
  define_extension_target(
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
    _rocm_C
    DESTINATION vllm
    LANGUAGE ${VLLM_GPU_LANG}
    SOURCES ${VLLM_ROCM_EXT_SRC}
    COMPILE_FLAGS ${VLLM_GPU_FLAGS}
    ARCHITECTURES ${VLLM_GPU_ARCHES}
    USE_SABI 3
    WITH_SOABI)
endif()

1208
1209
1210
1211
1212
# For CUDA and HIP builds also build the triton_kernels external package.
if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
    include(cmake/external_projects/triton_kernels.cmake)
endif()

1213
1214
1215
# For CUDA we also build and ship some external projects.
if (VLLM_GPU_LANG STREQUAL "CUDA")
    include(cmake/external_projects/flashmla.cmake)
1216
    include(cmake/external_projects/qutlass.cmake)
1217
1218

    # vllm-flash-attn should be last as it overwrites some CMake functions
1219
    include(cmake/external_projects/vllm_flash_attn.cmake)
1220
endif ()