CMakeLists.txt 34.1 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
# 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
18
message(STATUS "Build type: ${CMAKE_BUILD_TYPE}")
19
message(STATUS "Target device: ${VLLM_TARGET_DEVICE}")
bnellnm's avatar
bnellnm committed
20
21
22

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

23
24
25
# Suppress potential warnings about unused manually-specified variables
set(ignoreMe "${VLLM_PYTHON_PATH}")

26
27
28
# Prevent installation of dependencies (cutlass) by default.
install(CODE "set(CMAKE_INSTALL_LOCAL_ONLY TRUE)" ALL_COMPONENTS)

bnellnm's avatar
bnellnm committed
29
30
31
32
#
# 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.
#
33
set(PYTHON_SUPPORTED_VERSIONS "3.9" "3.10" "3.11" "3.12")
bnellnm's avatar
bnellnm committed
34
35

# Supported AMD GPU architectures.
36
set(HIP_SUPPORTED_ARCHS "gfx906;gfx908;gfx90a;gfx942;gfx950;gfx1030;gfx1100;gfx1101;gfx1200;gfx1201")
bnellnm's avatar
bnellnm committed
37
38
39
40
41
42
43
44
45

#
# 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
46
# versions are derived from docker/Dockerfile.rocm
bnellnm's avatar
bnellnm committed
47
#
Huy Do's avatar
Huy Do committed
48
49
set(TORCH_SUPPORTED_VERSION_CUDA "2.7.0")
set(TORCH_SUPPORTED_VERSION_ROCM "2.7.0")
bnellnm's avatar
bnellnm committed
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67

#
# 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")

68
69
# Ensure the 'nvcc' command is in the PATH
find_program(NVCC_EXECUTABLE nvcc)
70
if (CUDA_FOUND AND NOT NVCC_EXECUTABLE)
71
72
73
    message(FATAL_ERROR "nvcc not found")
endif()

bnellnm's avatar
bnellnm committed
74
75
76
77
78
79
80
81
#
# 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)

82
83
84
85
86
87
88
89
90
# 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
   CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL 12.8)
  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")
else()
  set(CUDA_SUPPORTED_ARCHS "7.0;7.2;7.5;8.0;8.6;8.7;8.9;9.0")
endif()

91
92
93
94
95
96
97
98
#
# 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()
99
        return()
100
101
102
103
    endif()
    return()
endif()

bnellnm's avatar
bnellnm committed
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
#
# 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)

123
124
125
  # ROCm 5.X and 6.X
  if (ROCM_VERSION_DEV_MAJOR GREATER_EQUAL 5 AND
      NOT Torch_VERSION VERSION_EQUAL ${TORCH_SUPPORTED_VERSION_ROCM})
126
    message(WARNING "Pytorch version >= ${TORCH_SUPPORTED_VERSION_ROCM} "
127
      "expected for ROCm build, saw ${Torch_VERSION} instead.")
bnellnm's avatar
bnellnm committed
128
129
130
131
132
  endif()
else()
  message(FATAL_ERROR "Can't find CUDA or HIP installation.")
endif()

133
134

if(VLLM_GPU_LANG STREQUAL "CUDA")
135
  #
136
  # For cuda we want to be able to control which architectures we compile for on
137
  # a per-file basis in order to cut down on compile time. So here we extract
138
  # the set of architectures we want to compile for and remove the from the
139
140
  # CMAKE_CUDA_FLAGS so that they are not applied globally.
  #
141
142
143
  clear_cuda_arches(CUDA_ARCH_FLAGS)
  extract_unique_cuda_archs_ascending(CUDA_ARCHS "${CUDA_ARCH_FLAGS}")
  message(STATUS "CUDA target architectures: ${CUDA_ARCHS}")
144
145
  # Filter the target architectures by the supported supported archs
  # since for some files we will build for all CUDA_ARCHS.
146
  cuda_archs_loose_intersection(CUDA_ARCHS
147
148
149
150
151
152
153
154
155
156
157
    "${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}")
158
159
endif()

bnellnm's avatar
bnellnm committed
160
161
162
163
164
165
166
167
168
169
170
171
172
173
#
# 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()

174
175
176
177
178
179
180
#
# Set nvcc fatbin compression.
#
if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND VLLM_GPU_LANG STREQUAL "CUDA")
  list(APPEND VLLM_GPU_FLAGS "-Xfatbin" "-compress-all" "-compress-mode=size")
endif()

181
182
183

#
# Use FetchContent for C++ dependencies that are compiled as part of vLLM's build process.
184
185
186
# 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>.
187
#
188
include(FetchContent)
189
file(MAKE_DIRECTORY ${FETCHCONTENT_BASE_DIR}) # Ensure the directory exists
190
message(STATUS "FetchContent base directory: ${FETCHCONTENT_BASE_DIR}")
191

192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
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
207
#
208
# Define other extension targets
bnellnm's avatar
bnellnm committed
209
210
#

211
212
213
214
215
216
217
218
219
220
221
222
223
224
#
# 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}")

if(VLLM_GPU_LANG STREQUAL "CUDA")
  message(STATUS "Enabling cumem allocator extension.")
  # link against cuda driver library
225
  list(APPEND CUMEM_LIBS CUDA::cuda_driver)
226
227
228
229
230
231
232
233
234
235
  define_gpu_extension_target(
    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
236
237
238
239
240
#
# _C extension
#

set(VLLM_EXT_SRC
241
  "csrc/mamba/mamba_ssm/selective_scan_fwd.cu"
bnellnm's avatar
bnellnm committed
242
  "csrc/cache_kernels.cu"
243
244
  "csrc/attention/paged_attention_v1.cu"
  "csrc/attention/paged_attention_v2.cu"
245
  "csrc/attention/merge_attn_states.cu"
246
  "csrc/attention/vertical_slash_index.cu"
bnellnm's avatar
bnellnm committed
247
248
249
  "csrc/pos_encoding_kernels.cu"
  "csrc/activation_kernels.cu"
  "csrc/layernorm_kernels.cu"
250
  "csrc/layernorm_quant_kernels.cu"
251
  "csrc/sampler.cu"
252
  "csrc/cuda_view.cu"
bnellnm's avatar
bnellnm committed
253
  "csrc/quantization/gptq/q_gemm.cu"
254
  "csrc/quantization/compressed_tensors/int8_quant_kernels.cu"
255
  "csrc/quantization/fp8/common.cu"
256
  "csrc/quantization/fused_kernels/fused_layernorm_dynamic_per_token_quant.cu"
257
  "csrc/quantization/gguf/gguf_kernel.cu"
258
  "csrc/quantization/activation_kernels.cu"
bnellnm's avatar
bnellnm committed
259
  "csrc/cuda_utils_kernels.cu"
260
  "csrc/prepare_inputs/advance_step.cu"
261
  "csrc/custom_all_reduce.cu"
262
  "csrc/torch_bindings.cpp")
bnellnm's avatar
bnellnm committed
263
264

if(VLLM_GPU_LANG STREQUAL "CUDA")
265
  SET(CUTLASS_ENABLE_HEADERS_ONLY ON CACHE BOOL "Enable only the header library")
266

267
  # Set CUTLASS_REVISION. Used for FetchContent. Also fixes some bogus messages when building.
268
  set(CUTLASS_REVISION "v4.0.0" CACHE STRING "CUTLASS revision to use")
269

270
271
272
273
274
275
276
277
278
279
280
281
282
  # 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(
283
        cutlass
284
        GIT_REPOSITORY https://github.com/nvidia/cutlass.git
285
        # Please keep this in sync with CUTLASS_REVISION line above.
286
        GIT_TAG ${CUTLASS_REVISION}
287
        GIT_PROGRESS TRUE
288
289
290
291

        # 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
292
        GIT_SHALLOW TRUE
293
294
    )
  endif()
295
296
  FetchContent_MakeAvailable(cutlass)

bnellnm's avatar
bnellnm committed
297
  list(APPEND VLLM_EXT_SRC
James Fleming's avatar
James Fleming committed
298
    "csrc/quantization/aqlm/gemm_kernels.cu"
bnellnm's avatar
bnellnm committed
299
    "csrc/quantization/awq/gemm_kernels.cu"
300
    "csrc/permute_cols.cu"
301
    "csrc/quantization/cutlass_w8a8/scaled_mm_entry.cu"
302
    "csrc/quantization/fp4/nvfp4_quant_entry.cu"
303
    "csrc/quantization/fp4/nvfp4_scaled_mm_entry.cu"
304
    "csrc/quantization/fp4/nvfp4_blockwise_moe_kernel.cu"
305
    "csrc/sparse/cutlass/sparse_scaled_mm_entry.cu"
306
307
    "csrc/cutlass_extensions/common.cpp"
    "csrc/attention/mla/cutlass_mla_entry.cu")
308
309
310
311
312
313
314
315

  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.
316
  # 9.0 for latest bf16 atomicAdd PTX
317
  cuda_archs_loose_intersection(MARLIN_ARCHS "8.0;8.7;9.0+PTX" "${CUDA_ARCHS}")
318
  if (MARLIN_ARCHS)
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363

    #
    # For the Marlin kernels we automatically generate sources for various
    # preselected input type pairs and schedules.
    # Generate sources:
    set(MARLIN_GEN_SCRIPT
      ${CMAKE_CURRENT_SOURCE_DIR}/csrc/quantization/gptq_marlin/generate_kernels.py)
    file(MD5 ${MARLIN_GEN_SCRIPT} MARLIN_GEN_SCRIPT_HASH)

    message(STATUS "Marlin generation script hash: ${MARLIN_GEN_SCRIPT_HASH}")
    message(STATUS "Last run Marlin generate script hash: $CACHE{MARLIN_GEN_SCRIPT_HASH}")

    if (NOT DEFINED CACHE{MARLIN_GEN_SCRIPT_HASH}
        OR NOT $CACHE{MARLIN_GEN_SCRIPT_HASH} STREQUAL ${MARLIN_GEN_SCRIPT_HASH})
      execute_process(
        COMMAND ${CMAKE_COMMAND} -E env
        PYTHONPATH=$PYTHONPATH
          ${Python_EXECUTABLE} ${MARLIN_GEN_SCRIPT}
        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()
        set(MARLIN_GEN_SCRIPT_HASH ${MARLIN_GEN_SCRIPT_HASH}
            CACHE STRING "Last run Marlin generate script hash" FORCE)
        message(STATUS "Marlin generation completed successfully.")
      endif()
    else()
      message(STATUS "Marlin generation script has not changed, skipping generation.")
    endif()

    file(GLOB MARLIN_TEMPLATE_KERNEL_SRC "csrc/quantization/gptq_marlin/kernel_*.cu")
    set_gencode_flags_for_srcs(
      SRCS "${MARLIN_TEMPLATE_KERNEL_SRC}"
      CUDA_ARCHS "${MARLIN_ARCHS}")

    list(APPEND VLLM_EXT_SRC ${MARLIN_TEMPLATE_KERNEL_SRC})

364
    set(MARLIN_SRCS
365
366
367
368
369
370
371
372
373
374
375
376
377
       "csrc/quantization/marlin/dense/marlin_cuda_kernel.cu"
       "csrc/quantization/marlin/sparse/marlin_24_cuda_kernel.cu"
       "csrc/quantization/marlin/qqq/marlin_qqq_gemm_kernel.cu"
       "csrc/quantization/gptq_marlin/gptq_marlin.cu"
       "csrc/quantization/gptq_marlin/gptq_marlin_repack.cu"
       "csrc/quantization/gptq_marlin/awq_marlin_repack.cu")
    set_gencode_flags_for_srcs(
      SRCS "${MARLIN_SRCS}"
      CUDA_ARCHS "${MARLIN_ARCHS}")
    list(APPEND VLLM_EXT_SRC "${MARLIN_SRCS}")
    message(STATUS "Building Marlin kernels for archs: ${MARLIN_ARCHS}")
  else()
    message(STATUS "Not building Marlin kernels as no compatible archs found"
378
                   " in CUDA target architectures")
379
380
  endif()

381
382
  # 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" "${CUDA_ARCHS}")
383
  if (ALLSPARK_ARCHS)
384
385
386
387
388
389
390
391
392
393
    set(ALLSPARK_SRCS
       "csrc/quantization/gptq_allspark/allspark_repack.cu"
       "csrc/quantization/gptq_allspark/allspark_qgemm_w8a16.cu")
    set_gencode_flags_for_srcs(
      SRCS "${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"
394
                   " in CUDA target architectures")
395
396
  endif()

397
398

  set(SCALED_MM_3X_ARCHS)
399
  # The cutlass_scaled_mm kernels for Hopper (c3x, i.e. CUTLASS 3.x) require
400
401
  # CUDA 12.0 or later
  cuda_archs_loose_intersection(SCALED_MM_ARCHS "9.0a;" "${CUDA_ARCHS}")
402
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.0 AND SCALED_MM_ARCHS)
403
    set(SRCS
404
       "csrc/quantization/cutlass_w8a8/scaled_mm_c3x_sm90.cu"
405
406
407
408
       "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm90_fp8.cu"
       "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm90_int8.cu"
       "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_azp_sm90_int8.cu"
       "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_blockwise_sm90_fp8.cu")
409
410
    set_gencode_flags_for_srcs(
      SRCS "${SRCS}"
411
      CUDA_ARCHS "${SCALED_MM_ARCHS}")
412
    list(APPEND VLLM_EXT_SRC "${SRCS}")
413
414
415
416
    list(APPEND VLLM_GPU_FLAGS "-DENABLE_SCALED_MM_SM90=1")
    # Let scaled_mm_c2x know it doesn't need to build these arches
    list(APPEND SCALED_MM_3X_ARCHS "${SCALED_MM_ARCHS}")
    message(STATUS "Building scaled_mm_c3x_sm90 for archs: ${SCALED_MM_ARCHS}")
417
  else()
418
    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.0 AND SCALED_MM_ARCHS)
419
      message(STATUS "Not building scaled_mm_c3x_sm90 as CUDA Compiler version is "
420
                     "not >= 12.0, we recommend upgrading to CUDA 12.0 or "
421
                     "later if you intend on running FP8 quantized models on "
422
423
                     "Hopper.")
    else()
424
      message(STATUS "Not building scaled_mm_c3x_sm90 as no compatible archs found "
425
426
                     "in CUDA target architectures")
    endif()
427
  endif()
428

429
430
431
432

  # The cutlass_scaled_mm kernels for Geforce Blackwell SM120 (c3x, i.e. CUTLASS 3.x) require
  # CUDA 12.8 or later
  cuda_archs_loose_intersection(SCALED_MM_ARCHS "12.0;12.0a" "${CUDA_ARCHS}")
433
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
434
435
436
437
438
439
440
441
442
443
444
445
446
    set(SRCS
      "csrc/quantization/cutlass_w8a8/scaled_mm_c3x_sm120.cu"
      "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm120_fp8.cu"
    )
    set_gencode_flags_for_srcs(
      SRCS "${SRCS}"
      CUDA_ARCHS "${SCALED_MM_ARCHS}")
    list(APPEND VLLM_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()
447
    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
448
449
450
451
452
453
454
455
456
457
458
      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()


459
460
461
  # The cutlass_scaled_mm kernels for Blackwell SM100 (c3x, i.e. CUTLASS 3.x)
  # require CUDA 12.8 or later
  cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a;10.1a" "${CUDA_ARCHS}")
462
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
463
464
465
    set(SRCS
      "csrc/quantization/cutlass_w8a8/scaled_mm_c3x_sm100.cu"
      "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm100_fp8.cu"
466
      "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_blockwise_sm100_fp8.cu"
467
468
469
470
471
472
473
474
475
476
    )
    set_gencode_flags_for_srcs(
      SRCS "${SRCS}"
      CUDA_ARCHS "${SCALED_MM_ARCHS}")
    list(APPEND VLLM_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()
477
    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
478
479
480
481
482
483
484
485
      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()
486
  endif()
487
488

  #
489
490
  # 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.
491
  # (Build 8.9 for FP8)
492
  cuda_archs_loose_intersection(SCALED_MM_2X_ARCHS
493
    "7.5;8.0;8.7;8.9+PTX" "${CUDA_ARCHS}")
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
  # subtract out the archs that are already built for 3x
  list(REMOVE_ITEM SCALED_MM_2X_ARCHS ${SCALED_MM_3X_ARCHS})
  if (SCALED_MM_2X_ARCHS)
    set(SRCS "csrc/quantization/cutlass_w8a8/scaled_mm_c2x.cu")
    set_gencode_flags_for_srcs(
      SRCS "${SRCS}"
      CUDA_ARCHS "${SCALED_MM_2X_ARCHS}")
    list(APPEND VLLM_EXT_SRC "${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")
511
512
513
514
515
516
517
    endif()
  endif()

  #
  # 2:4 Sparse Kernels

  # The 2:4 sparse kernels cutlass_scaled_sparse_mm and cutlass_compressor
518
519
  # require CUDA 12.2 or later (and only work on Hopper).
  cuda_archs_loose_intersection(SCALED_MM_ARCHS "9.0a;" "${CUDA_ARCHS}")
520
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.2 AND SCALED_MM_ARCHS)
521
    set(SRCS "csrc/sparse/cutlass/sparse_scaled_mm_c3x.cu")
522
523
    set_gencode_flags_for_srcs(
      SRCS "${SRCS}"
524
      CUDA_ARCHS "${SCALED_MM_ARCHS}")
525
526
    list(APPEND VLLM_EXT_SRC "${SRCS}")
    list(APPEND VLLM_GPU_FLAGS "-DENABLE_SPARSE_SCALED_MM_C3X=1")
527
    message(STATUS "Building sparse_scaled_mm_c3x for archs: ${SCALED_MM_ARCHS}")
528
  else()
529
    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.2 AND SCALED_MM_ARCHS)
530
531
532
533
534
535
      message(STATUS "Not building sparse_scaled_mm_c3x kernels as CUDA Compiler version is "
                     "not >= 12.2, we recommend upgrading to CUDA 12.2 or later "
                     "if you intend on running FP8 sparse quantized models on Hopper.")
    else()
      message(STATUS "Not building sparse_scaled_mm_c3x as no compatible archs found "
                     "in CUDA target architectures")
536
    endif()
537
  endif()
538

539
540
  # FP4 Archs and flags
  cuda_archs_loose_intersection(FP4_ARCHS "10.0a" "${CUDA_ARCHS}")
541
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND FP4_ARCHS)
542
    set(SRCS
543
      "csrc/quantization/fp4/nvfp4_quant_kernels.cu"
544
545
546
      "csrc/quantization/fp4/nvfp4_experts_quant.cu"
      "csrc/quantization/fp4/nvfp4_scaled_mm_kernels.cu"
      "csrc/quantization/fp4/nvfp4_blockwise_moe_kernel.cu")
547
548
549
550
551
    set_gencode_flags_for_srcs(
      SRCS "${SRCS}"
      CUDA_ARCHS "${FP4_ARCHS}")
    list(APPEND VLLM_EXT_SRC "${SRCS}")
    list(APPEND VLLM_GPU_FLAGS "-DENABLE_NVFP4=1")
552
    list(APPEND VLLM_GPU_FLAGS "-DENABLE_CUTLASS_MOE_SM100=1")
553
554
555
556
557
558
    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()
559

560
561
  # CUTLASS MLA Archs and flags
  cuda_archs_loose_intersection(MLA_ARCHS "10.0a" "${CUDA_ARCHS}")
562
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND MLA_ARCHS)
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
    set(SRCS
      "csrc/attention/mla/cutlass_mla_kernels.cu")
    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()

580
581
  # CUTLASS MoE kernels

582
  # The MoE kernel cutlass_moe_mm requires CUDA 12.3 or later (and ONLY works
583
584
  # on Hopper). get_cutlass_(pplx_)moe_mm_data should only be compiled
  # if it's possible to compile MoE kernels that use its output.
585
  cuda_archs_loose_intersection(SCALED_MM_ARCHS "9.0a" "${CUDA_ARCHS}")
586
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.3 AND SCALED_MM_ARCHS)
587
    set(SRCS "csrc/quantization/cutlass_w8a8/moe/grouped_mm_c3x.cu")
588
589
590
591
592
593
594
595
596
597
598
599
600
    set_gencode_flags_for_srcs(
      SRCS "${SRCS}"
      CUDA_ARCHS "${SCALED_MM_ARCHS}")
    list(APPEND VLLM_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 "
601
                     "in CUDA target architectures.")
602
603
604
    endif()
  endif()

605
606
607
608
609
610
611
612
  # moe_data.cu is used by all CUTLASS MoE kernels.
  cuda_archs_loose_intersection(CUTLASS_MOE_DATA_ARCHS "9.0a;10.0a" "${CUDA_ARCHS}")
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.3 AND CUTLASS_MOE_DATA_ARCHS)
    set(SRCS "csrc/quantization/cutlass_w8a8/moe/moe_data.cu")
    set_gencode_flags_for_srcs(
      SRCS "${SRCS}"
      CUDA_ARCHS "${CUTLASS_MOE_DATA_ARCHS}")
    list(APPEND VLLM_EXT_SRC "${SRCS}")
613
614
615
616
617
618
619
620
621
622
623
    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()
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
  
  cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a" "${CUDA_ARCHS}")
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
    set(SRCS "csrc/quantization/cutlass_w8a8/moe/blockwise_scaled_group_mm_sm100.cu")
    set_gencode_flags_for_srcs(
      SRCS "${SRCS}"
      CUDA_ARCHS "${SCALED_MM_ARCHS}")
    list(APPEND VLLM_EXT_SRC "${SRCS}")
    list(APPEND VLLM_GPU_FLAGS "-DENABLE_CUTLASS_MOE_SM100=1")
    message(STATUS "Building blockwise_scaled_group_mm_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 blockwise_scaled_group_mm_sm100 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 blockwise_scaled_group_mm_sm100 as no compatible archs found "
                     "in CUDA target architectures")
    endif()
  endif()
644

645
  #
646
  # Machete kernels
647

648
  # The machete kernels only work on hopper and require CUDA 12.0 or later.
649
650
  # Only build Machete kernels if we are building for something compatible with sm90a
  cuda_archs_loose_intersection(MACHETE_ARCHS "9.0a" "${CUDA_ARCHS}")
651
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.0 AND MACHETE_ARCHS)
652
    #
653
    # For the Machete kernels we automatically generate sources for various
654
655
    # preselected input type pairs and schedules.
    # Generate sources:
656
    set(MACHETE_GEN_SCRIPT
657
658
659
660
661
662
663
664
665
      ${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(
666
667
        COMMAND ${CMAKE_COMMAND} -E env
        PYTHONPATH=${CMAKE_CURRENT_SOURCE_DIR}/csrc/cutlass_extensions/:${CUTLASS_DIR}/python/:${VLLM_PYTHON_PATH}:$PYTHONPATH
668
669
670
671
672
673
674
675
676
          ${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."
677
                            " Result: \"${machete_generation_result}\""
678
679
680
                            "\nCheck the log for details: "
                            "${CMAKE_CURRENT_BINARY_DIR}/machete_generation.log")
      else()
681
        set(MACHETE_GEN_SCRIPT_HASH ${MACHETE_GEN_SCRIPT_HASH}
682
683
684
            CACHE STRING "Last run machete generate script hash" FORCE)
        message(STATUS "Machete generation completed successfully.")
      endif()
685
    else()
686
      message(STATUS "Machete generation script has not changed, skipping generation.")
687
    endif()
688

689
690
691
    # Add machete generated sources
    file(GLOB MACHETE_GEN_SOURCES "csrc/quantization/machete/generated/*.cu")
    list(APPEND VLLM_EXT_SRC ${MACHETE_GEN_SOURCES})
692

693
694
695
696
697
698
699
700
701
702
    # 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()
703
    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.0
704
705
706
707
708
709
710
711
712
        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()
713
  endif()
714
# if CUDA endif
bnellnm's avatar
bnellnm committed
715
endif()
716
717
718
719
720
721
722
723

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
724

725
message(STATUS "Enabling C extension.")
bnellnm's avatar
bnellnm committed
726
727
728
729
730
731
732
define_gpu_extension_target(
  _C
  DESTINATION vllm
  LANGUAGE ${VLLM_GPU_LANG}
  SOURCES ${VLLM_EXT_SRC}
  COMPILE_FLAGS ${VLLM_GPU_FLAGS}
  ARCHITECTURES ${VLLM_GPU_ARCHES}
733
  INCLUDE_DIRECTORIES ${CUTLASS_INCLUDE_DIR}
734
  INCLUDE_DIRECTORIES ${CUTLASS_TOOLS_UTIL_INCLUDE_DIR}
735
  USE_SABI 3
bnellnm's avatar
bnellnm committed
736
737
  WITH_SOABI)

738
739
# If CUTLASS is compiled on NVCC >= 12.5, it by default uses
# cudaGetDriverEntryPointByVersion as a wrapper to avoid directly calling the
740
741
742
743
# 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)

bnellnm's avatar
bnellnm committed
744
745
746
747
748
#
# _moe_C extension
#

set(VLLM_MOE_EXT_SRC
749
  "csrc/moe/torch_bindings.cpp"
750
  "csrc/moe/moe_align_sum_kernels.cu"
751
  "csrc/moe/topk_softmax_kernels.cu")
bnellnm's avatar
bnellnm committed
752

753
754
755
756
if(VLLM_GPU_LANG STREQUAL "CUDA")
  list(APPEND VLLM_MOE_EXT_SRC "csrc/moe/moe_wna16.cu")
endif()

757
758
759
760
set_gencode_flags_for_srcs(
  SRCS "${VLLM_MOE_EXT_SRC}"
  CUDA_ARCHS "${CUDA_ARCHS}")

761
if(VLLM_GPU_LANG STREQUAL "CUDA")
762
763
764
765
766
767
768
  set(VLLM_MOE_WNA16_SRC
    "csrc/moe/moe_wna16.cu")

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

769
  list(APPEND VLLM_MOE_EXT_SRC "${VLLM_MOE_WNA16_SRC}")
770
  # 9.0 for latest bf16 atomicAdd PTX
771
  cuda_archs_loose_intersection(MARLIN_MOE_ARCHS "8.0;8.7;9.0+PTX" "${CUDA_ARCHS}")
772
773
  if (MARLIN_MOE_ARCHS)

774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
    #
    # 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)

    message(STATUS "Marlin MOE generation script hash: ${MOE_MARLIN_GEN_SCRIPT_HASH}")
    message(STATUS "Last run Marlin MOE generate script hash: $CACHE{MOE_MARLIN_GEN_SCRIPT_HASH}")

    if (NOT DEFINED CACHE{MOE_MARLIN_GEN_SCRIPT_HASH}
        OR NOT $CACHE{MOE_MARLIN_GEN_SCRIPT_HASH} STREQUAL ${MOE_MARLIN_GEN_SCRIPT_HASH})
      execute_process(
        COMMAND ${CMAKE_COMMAND} -E env
789
        PYTHONPATH=$PYTHONPATH
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
          ${Python_EXECUTABLE} ${MOE_MARLIN_GEN_SCRIPT}
        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()
        set(MOE_MARLIN_GEN_SCRIPT_HASH ${MOE_MARLIN_GEN_SCRIPT_HASH}
            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()

    file(GLOB MOE_WNAA16_MARLIN_SRC "csrc/moe/marlin_moe_wna16/*.cu")
812
    set_gencode_flags_for_srcs(
813
      SRCS "${MOE_WNAA16_MARLIN_SRC}"
814
815
      CUDA_ARCHS "${MARLIN_MOE_ARCHS}")

816
817
    list(APPEND VLLM_MOE_EXT_SRC ${MOE_WNAA16_MARLIN_SRC})

818
819
820
    message(STATUS "Building Marlin MOE kernels for archs: ${MARLIN_MOE_ARCHS}")
  else()
    message(STATUS "Not building Marlin MOE kernels as no compatible archs found"
821
                   " in CUDA target architectures")
822
  endif()
823
824
endif()

825
826
827
828
829
830
831
832
833
834
835
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")

  set_gencode_flags_for_srcs(
    SRCS "${MARLIN_PERMUTE_SRC}"
    CUDA_ARCHS "${MOE_PERMUTE_ARCHS}")

  list(APPEND VLLM_MOE_EXT_SRC "${MOE_PERMUTE_SRC}")
endif()
836
message(STATUS "Enabling moe extension.")
bnellnm's avatar
bnellnm committed
837
838
839
840
841
842
843
define_gpu_extension_target(
  _moe_C
  DESTINATION vllm
  LANGUAGE ${VLLM_GPU_LANG}
  SOURCES ${VLLM_MOE_EXT_SRC}
  COMPILE_FLAGS ${VLLM_GPU_FLAGS}
  ARCHITECTURES ${VLLM_GPU_ARCHES}
844
845
  INCLUDE_DIRECTORIES ${CUTLASS_INCLUDE_DIR}
  INCLUDE_DIRECTORIES ${CUTLASS_TOOLS_UTIL_INCLUDE_DIR}
846
  USE_SABI 3
bnellnm's avatar
bnellnm committed
847
848
  WITH_SOABI)

849
850
851
852
853
854
if(VLLM_GPU_LANG STREQUAL "HIP")
  #
  # _rocm_C extension
  #
  set(VLLM_ROCM_EXT_SRC
    "csrc/rocm/torch_bindings.cpp"
855
    "csrc/rocm/skinny_gemms.cu"
856
857
858
859
860
861
862
863
864
865
866
867
868
    "csrc/rocm/attention.cu")

  define_gpu_extension_target(
    _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()

869
870
871
# For CUDA we also build and ship some external projects.
if (VLLM_GPU_LANG STREQUAL "CUDA")
    include(cmake/external_projects/flashmla.cmake)
872
873

    # vllm-flash-attn should be last as it overwrites some CMake functions
874
    include(cmake/external_projects/vllm_flash_attn.cmake)
875
endif ()