CMakeLists.txt 44.3 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;gfx1200;gfx1201;gfx1150;gfx1151;gfx928;gfx936")
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 "2.9.0")
set(TORCH_SUPPORTED_VERSION_ROCM "2.9.0")
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
97
98
   CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL 13.0)
  set(CUDA_SUPPORTED_ARCHS "7.5;8.0;8.6;8.7;8.9;9.0;10.0;11.0;12.0")
elseif(DEFINED CMAKE_CUDA_COMPILER_VERSION AND
99
100
101
102
103
104
   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()

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

if(VLLM_GPU_LANG STREQUAL "CUDA")
307
  SET(CUTLASS_ENABLE_HEADERS_ONLY ON CACHE BOOL "Enable only the header library")
308

309
  # Set CUTLASS_REVISION. Used for FetchContent. Also fixes some bogus messages when building.
310
  set(CUTLASS_REVISION "v4.2.1")
311

312
313
314
315
316
317
318
319
320
321
322
323
324
  # 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(
325
        cutlass
326
        GIT_REPOSITORY https://github.com/nvidia/cutlass.git
327
        # Please keep this in sync with CUTLASS_REVISION line above.
328
        GIT_TAG ${CUTLASS_REVISION}
329
        GIT_PROGRESS TRUE
330
331
332
333

        # 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
334
        GIT_SHALLOW TRUE
335
336
    )
  endif()
337
338
  FetchContent_MakeAvailable(cutlass)

bnellnm's avatar
bnellnm committed
339
340
  list(APPEND VLLM_EXT_SRC
    "csrc/quantization/awq/gemm_kernels.cu"
341
    "csrc/permute_cols.cu"
342
    "csrc/quantization/w8a8/cutlass/scaled_mm_entry.cu"
343
    "csrc/quantization/fp4/nvfp4_quant_entry.cu"
344
    "csrc/quantization/fp4/nvfp4_scaled_mm_entry.cu"
345
    "csrc/sparse/cutlass/sparse_scaled_mm_entry.cu"
346
    "csrc/cutlass_extensions/common.cpp"
347
348
    "csrc/quantization/w8a8/fp8/per_token_group_quant.cu"
    "csrc/quantization/w8a8/int8/per_token_group_quant.cu")
349
350
351
352
353
354
355
356

  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.
357
358
359
360
361
362
363
364
365
366
367

  # marlin arches for fp16 output
  cuda_archs_loose_intersection(MARLIN_ARCHS "8.0+PTX" "${CUDA_ARCHS}")
  # marlin arches for bf16 output (we need 9.0 for bf16 atomicAdd PTX)
  cuda_archs_loose_intersection(MARLIN_BF16_ARCHS "8.0+PTX;9.0+PTX" "${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)
  cuda_archs_loose_intersection(MARLIN_FP8_ARCHS "8.9;12.0" "${CUDA_ARCHS}")

368
  if (MARLIN_ARCHS)
369
370
371
372
373
374
375
376

    #
    # 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 ${MARLIN_GEN_SCRIPT} MARLIN_GEN_SCRIPT_HASH)
377
378
    list(JOIN CUDA_ARCHS "," CUDA_ARCHS_STR)
    set(MARLIN_GEN_SCRIPT_HASH_AND_ARCH "${MARLIN_GEN_SCRIPT_HASH}(ARCH:${CUDA_ARCHS_STR})")
379

380
381
    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}")
382

383
384
    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})
385
386
387
      execute_process(
        COMMAND ${CMAKE_COMMAND} -E env
        PYTHONPATH=$PYTHONPATH
388
          ${Python_EXECUTABLE} ${MARLIN_GEN_SCRIPT} ${CUDA_ARCHS_STR}
389
390
391
392
393
394
395
396
397
398
399
400
        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()
401
402
        set(MARLIN_GEN_SCRIPT_HASH_AND_ARCH ${MARLIN_GEN_SCRIPT_HASH_AND_ARCH}
            CACHE STRING "Last run Marlin generate script hash and arch" FORCE)
403
404
405
406
407
408
        message(STATUS "Marlin generation completed successfully.")
      endif()
    else()
      message(STATUS "Marlin generation script has not changed, skipping generation.")
    endif()

409
    file(GLOB MARLIN_TEMPLATE_KERNEL_SRC "csrc/quantization/gptq_marlin/sm80_kernel_*_float16.cu")
410
411
412
    set_gencode_flags_for_srcs(
      SRCS "${MARLIN_TEMPLATE_KERNEL_SRC}"
      CUDA_ARCHS "${MARLIN_ARCHS}")
413
414
415
416
    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()
417
418
    list(APPEND VLLM_EXT_SRC ${MARLIN_TEMPLATE_KERNEL_SRC})

419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
    file(GLOB MARLIN_TEMPLATE_BF16_KERNEL_SRC "csrc/quantization/gptq_marlin/sm80_kernel_*_bfloat16.cu")
    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})

    if (MARLIN_FP8_ARCHS) 
      file(GLOB MARLIN_TEMPLATE_FP8_KERNEL_SRC "csrc/quantization/gptq_marlin/sm89_kernel_*.cu")
      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()

441
    set(MARLIN_SRCS
442
443
       "csrc/quantization/marlin/sparse/marlin_24_cuda_kernel.cu"
       "csrc/quantization/gptq_marlin/gptq_marlin.cu"
444
       "csrc/quantization/gptq_marlin/marlin_int4_fp8_preprocess.cu"
445
446
447
448
449
       "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}")
450
451
452
453
    if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8)
      set_source_files_properties("csrc/quantization/gptq_marlin/gptq_marlin.cu"
        PROPERTIES COMPILE_FLAGS "-static-global-template-stub=false")
    endif()
454
    list(APPEND VLLM_EXT_SRC "${MARLIN_SRCS}")
455

456
457
458
    message(STATUS "Building Marlin kernels for archs: ${MARLIN_ARCHS}")
  else()
    message(STATUS "Not building Marlin kernels as no compatible archs found"
459
                   " in CUDA target architectures")
460
461
  endif()

462
463
  # 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" "${CUDA_ARCHS}")
464
  if (ALLSPARK_ARCHS)
465
466
467
468
469
470
471
472
473
474
    set(ALLSPARK_SRCS
       "csrc/quantization/gptq_allspark/allspark_repack.cu"
       "csrc/quantization/gptq_allspark/allspark_qgemm_w8a16.cu")
    set_gencode_flags_for_srcs(
      SRCS "${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"
475
                   " in CUDA target architectures")
476
477
  endif()

478
479

  set(SCALED_MM_3X_ARCHS)
480
  # The cutlass_scaled_mm kernels for Hopper (c3x, i.e. CUTLASS 3.x) require
481
482
  # CUDA 12.0 or later
  cuda_archs_loose_intersection(SCALED_MM_ARCHS "9.0a;" "${CUDA_ARCHS}")
483
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.0 AND SCALED_MM_ARCHS)
484
    set(SRCS
485
486
487
488
489
       "csrc/quantization/w8a8/cutlass/scaled_mm_c3x_sm90.cu"
       "csrc/quantization/w8a8/cutlass/c3x/scaled_mm_sm90_fp8.cu"
       "csrc/quantization/w8a8/cutlass/c3x/scaled_mm_sm90_int8.cu"
       "csrc/quantization/w8a8/cutlass/c3x/scaled_mm_azp_sm90_int8.cu"
       "csrc/quantization/w8a8/cutlass/c3x/scaled_mm_blockwise_sm90_fp8.cu")
490
491
    set_gencode_flags_for_srcs(
      SRCS "${SRCS}"
492
      CUDA_ARCHS "${SCALED_MM_ARCHS}")
493
    list(APPEND VLLM_EXT_SRC "${SRCS}")
494
495
496
497
    list(APPEND VLLM_GPU_FLAGS "-DENABLE_SCALED_MM_SM90=1")
    # Let scaled_mm_c2x know it doesn't need to build these arches
    list(APPEND SCALED_MM_3X_ARCHS "${SCALED_MM_ARCHS}")
    message(STATUS "Building scaled_mm_c3x_sm90 for archs: ${SCALED_MM_ARCHS}")
498
  else()
499
    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.0 AND SCALED_MM_ARCHS)
500
      message(STATUS "Not building scaled_mm_c3x_sm90 as CUDA Compiler version is "
501
                     "not >= 12.0, we recommend upgrading to CUDA 12.0 or "
502
                     "later if you intend on running FP8 quantized models on "
503
504
                     "Hopper.")
    else()
505
      message(STATUS "Not building scaled_mm_c3x_sm90 as no compatible archs found "
506
507
                     "in CUDA target architectures")
    endif()
508
  endif()
509

510
511

  # The cutlass_scaled_mm kernels for Geforce Blackwell SM120 (c3x, i.e. CUTLASS 3.x) require
512
  # CUDA 12.8 or later
Johnny's avatar
Johnny committed
513
514
515
516
517
  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" "${CUDA_ARCHS}")
  endif()
518
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
519
    set(SRCS
520
521
522
      "csrc/quantization/w8a8/cutlass/scaled_mm_c3x_sm120.cu"
      "csrc/quantization/w8a8/cutlass/c3x/scaled_mm_sm120_fp8.cu"
      "csrc/quantization/w8a8/cutlass/c3x/scaled_mm_blockwise_sm120_fp8.cu"
523
524
525
526
527
528
529
530
531
532
    )
    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()
533
    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
534
535
536
537
538
539
540
541
542
543
544
      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()


545
546
  # The cutlass_scaled_mm kernels for Blackwell SM100 (c3x, i.e. CUTLASS 3.x)
  # require CUDA 12.8 or later
Johnny's avatar
Johnny committed
547
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
548
    cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0f;11.0f" "${CUDA_ARCHS}")
Johnny's avatar
Johnny committed
549
  else()
550
    cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a;10.1a;10.3a" "${CUDA_ARCHS}")
Johnny's avatar
Johnny committed
551
  endif()
552
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
553
    set(SRCS
554
555
556
      "csrc/quantization/w8a8/cutlass/scaled_mm_c3x_sm100.cu"
      "csrc/quantization/w8a8/cutlass/c3x/scaled_mm_sm100_fp8.cu"
      "csrc/quantization/w8a8/cutlass/c3x/scaled_mm_blockwise_sm100_fp8.cu"
557
558
559
560
561
562
563
564
565
566
    )
    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()
567
    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
568
569
570
571
572
573
574
575
      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()
576
  endif()
577
578

  #
579
580
  # 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.
581
  # (Build 8.9 for FP8)
582
  cuda_archs_loose_intersection(SCALED_MM_2X_ARCHS
583
    "7.5;8.0;8.7;8.9+PTX" "${CUDA_ARCHS}")
584
585
586
  # subtract out the archs that are already built for 3x
  list(REMOVE_ITEM SCALED_MM_2X_ARCHS ${SCALED_MM_3X_ARCHS})
  if (SCALED_MM_2X_ARCHS)
587
    set(SRCS "csrc/quantization/w8a8/cutlass/scaled_mm_c2x.cu")
588
589
590
591
592
593
594
595
596
597
598
599
600
    set_gencode_flags_for_srcs(
      SRCS "${SRCS}"
      CUDA_ARCHS "${SCALED_MM_2X_ARCHS}")
    list(APPEND VLLM_EXT_SRC "${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")
601
602
603
604
605
606
607
    endif()
  endif()

  #
  # 2:4 Sparse Kernels

  # The 2:4 sparse kernels cutlass_scaled_sparse_mm and cutlass_compressor
608
609
  # require CUDA 12.2 or later (and only work on Hopper).
  cuda_archs_loose_intersection(SCALED_MM_ARCHS "9.0a;" "${CUDA_ARCHS}")
610
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.2 AND SCALED_MM_ARCHS)
611
    set(SRCS "csrc/sparse/cutlass/sparse_scaled_mm_c3x.cu")
612
613
    set_gencode_flags_for_srcs(
      SRCS "${SRCS}"
614
      CUDA_ARCHS "${SCALED_MM_ARCHS}")
615
616
    list(APPEND VLLM_EXT_SRC "${SRCS}")
    list(APPEND VLLM_GPU_FLAGS "-DENABLE_SPARSE_SCALED_MM_C3X=1")
617
    message(STATUS "Building sparse_scaled_mm_c3x for archs: ${SCALED_MM_ARCHS}")
618
  else()
619
    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.2 AND SCALED_MM_ARCHS)
620
621
622
623
624
625
      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")
626
    endif()
627
  endif()
628

629
630
  # The nvfp4_scaled_mm_sm120 kernels for Geforce Blackwell SM120 require
  # CUDA 12.8 or later
Johnny's avatar
Johnny committed
631
632
633
634
635
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
    cuda_archs_loose_intersection(FP4_ARCHS "12.0f" "${CUDA_ARCHS}")
  else()
    cuda_archs_loose_intersection(FP4_ARCHS "12.0a" "${CUDA_ARCHS}")
  endif()
636
637
638
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND FP4_ARCHS)
    set(SRCS
      "csrc/quantization/fp4/nvfp4_quant_kernels.cu"
639
      "csrc/quantization/fp4/activation_nvfp4_quant_fusion_kernels.cu"
640
641
642
      "csrc/quantization/fp4/nvfp4_experts_quant.cu"
      "csrc/quantization/fp4/nvfp4_scaled_mm_sm120_kernels.cu"
      "csrc/quantization/fp4/nvfp4_blockwise_moe_kernel.cu")
643
644
645
646
647
    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")
648
    list(APPEND VLLM_GPU_FLAGS "-DENABLE_CUTLASS_MOE_SM120=1")
649
650
651
652
653
654
655
    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()

656
  # FP4 Archs and flags
Johnny's avatar
Johnny committed
657
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
658
    cuda_archs_loose_intersection(FP4_ARCHS "10.0f;11.0f" "${CUDA_ARCHS}")
Johnny's avatar
Johnny committed
659
  else()
660
    cuda_archs_loose_intersection(FP4_ARCHS "10.0a;10.1a;10.3a" "${CUDA_ARCHS}")
Johnny's avatar
Johnny committed
661
  endif()
662
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND FP4_ARCHS)
663
    set(SRCS
664
      "csrc/quantization/fp4/nvfp4_quant_kernels.cu"
665
      "csrc/quantization/fp4/activation_nvfp4_quant_fusion_kernels.cu"
666
667
668
      "csrc/quantization/fp4/nvfp4_experts_quant.cu"
      "csrc/quantization/fp4/nvfp4_scaled_mm_kernels.cu"
      "csrc/quantization/fp4/nvfp4_blockwise_moe_kernel.cu")
669
670
671
672
    set_gencode_flags_for_srcs(
      SRCS "${SRCS}"
      CUDA_ARCHS "${FP4_ARCHS}")
    list(APPEND VLLM_EXT_SRC "${SRCS}")
673
    list(APPEND VLLM_GPU_FLAGS "-DENABLE_NVFP4_SM100=1")
674
    list(APPEND VLLM_GPU_FLAGS "-DENABLE_CUTLASS_MOE_SM100=1")
675
676
677
678
679
680
    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()
681

682
  # CUTLASS MLA Archs and flags
Johnny's avatar
Johnny committed
683
684
685
686
687
  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" "${CUDA_ARCHS}")
  endif()
688
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND MLA_ARCHS)
689
    set(SRCS
690
      "csrc/attention/mla/sm100_cutlass_mla_kernel.cu")
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
    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()

706
707
  # CUTLASS MoE kernels

708
  # The MoE kernel cutlass_moe_mm requires CUDA 12.3 or later (and ONLY works
709
710
  # on Hopper). get_cutlass_(pplx_)moe_mm_data should only be compiled
  # if it's possible to compile MoE kernels that use its output.
711
  cuda_archs_loose_intersection(SCALED_MM_ARCHS "9.0a" "${CUDA_ARCHS}")
712
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.3 AND SCALED_MM_ARCHS)
713
    set(SRCS "csrc/quantization/w8a8/cutlass/moe/grouped_mm_c3x_sm90.cu")
714
715
716
717
718
719
720
721
722
723
724
725
726
    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 "
727
                     "in CUDA target architectures.")
728
729
730
    endif()
  endif()

Johnny's avatar
Johnny committed
731
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
732
    cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0f;11.0f" "${CUDA_ARCHS}")
Johnny's avatar
Johnny committed
733
  else()
734
    cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a;10.1a;10.3a" "${CUDA_ARCHS}")
Johnny's avatar
Johnny committed
735
  endif()
736
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
737
    set(SRCS "csrc/quantization/w8a8/cutlass/moe/grouped_mm_c3x_sm100.cu")
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
    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 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()

755
  # moe_data.cu is used by all CUTLASS MoE kernels.
Johnny's avatar
Johnny committed
756
757
758
759
760
  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" "${CUDA_ARCHS}")
  endif()
761
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.3 AND CUTLASS_MOE_DATA_ARCHS)
762
    set(SRCS "csrc/quantization/w8a8/cutlass/moe/moe_data.cu")
763
764
765
766
    set_gencode_flags_for_srcs(
      SRCS "${SRCS}"
      CUDA_ARCHS "${CUTLASS_MOE_DATA_ARCHS}")
    list(APPEND VLLM_EXT_SRC "${SRCS}")
767
768
769
770
771
772
773
774
775
776
777
    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()
778

Johnny's avatar
Johnny committed
779
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
780
    cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0f;11.0f" "${CUDA_ARCHS}")
Johnny's avatar
Johnny committed
781
  else()
782
    cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a;10.1a;10.3a" "${CUDA_ARCHS}")
Johnny's avatar
Johnny committed
783
  endif()
784
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
785
    set(SRCS "csrc/quantization/w8a8/cutlass/moe/blockwise_scaled_group_mm_sm100.cu")
786
787
788
789
790
791
792
793
794
795
796
797
798
    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 "
799
800
801
802
                     "in CUDA target architectures")
    endif()
  endif()

803
  #
804
  # Machete kernels
805

806
  # The machete kernels only work on hopper and require CUDA 12.0 or later.
807
808
  # Only build Machete kernels if we are building for something compatible with sm90a
  cuda_archs_loose_intersection(MACHETE_ARCHS "9.0a" "${CUDA_ARCHS}")
809
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.0 AND MACHETE_ARCHS)
810
    #
811
    # For the Machete kernels we automatically generate sources for various
812
813
    # preselected input type pairs and schedules.
    # Generate sources:
814
    set(MACHETE_GEN_SCRIPT
815
816
817
818
819
820
821
822
823
      ${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(
824
825
        COMMAND ${CMAKE_COMMAND} -E env
        PYTHONPATH=${CMAKE_CURRENT_SOURCE_DIR}/csrc/cutlass_extensions/:${CUTLASS_DIR}/python/:${VLLM_PYTHON_PATH}:$PYTHONPATH
826
827
828
829
830
831
832
833
834
          ${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."
835
                            " Result: \"${machete_generation_result}\""
836
837
838
                            "\nCheck the log for details: "
                            "${CMAKE_CURRENT_BINARY_DIR}/machete_generation.log")
      else()
839
        set(MACHETE_GEN_SCRIPT_HASH ${MACHETE_GEN_SCRIPT_HASH}
840
841
842
            CACHE STRING "Last run machete generate script hash" FORCE)
        message(STATUS "Machete generation completed successfully.")
      endif()
843
    else()
844
      message(STATUS "Machete generation script has not changed, skipping generation.")
845
    endif()
846

847
848
849
    # Add machete generated sources
    file(GLOB MACHETE_GEN_SOURCES "csrc/quantization/machete/generated/*.cu")
    list(APPEND VLLM_EXT_SRC ${MACHETE_GEN_SOURCES})
850

851
852
853
854
855
856
857
858
859
860
    # 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()
861
    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.0
862
863
864
865
866
867
868
869
870
        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()
871
  endif()
872
873
874
875
876

  # 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
877
878
879
880
       "csrc/quantization/cutlass_w4a8/w4a8_mm_entry.cu"
       "csrc/quantization/cutlass_w4a8/w4a8_grouped_mm_entry.cu"
       "csrc/quantization/cutlass_w4a8/w4a8_utils.cu"
       )
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901

    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()

902
  # Hadacore kernels
903
  cuda_archs_loose_intersection(HADACORE_ARCHS "8.0+PTX;9.0+PTX" "${CUDA_ARCHS}")
904
905
906
907
908
909
910
911
912
  if(HADACORE_ARCHS)
    set(SRCS "csrc/quantization/hadamard/hadacore/hadamard_transform_cuda.cu")
    set_gencode_flags_for_srcs(
      SRCS "${SRCS}"
      CUDA_ARCHS "${HADACORE_ARCHS}")
    list(APPEND VLLM_EXT_SRC "${SRCS}")
    message(STATUS "Building hadacore")
  endif()

913
# if CUDA endif
bnellnm's avatar
bnellnm committed
914
915
endif()

916
917
918
919
920
921
922
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
923

924
message(STATUS "Enabling C extension.")
925
define_extension_target(
bnellnm's avatar
bnellnm committed
926
927
928
929
930
931
  _C
  DESTINATION vllm
  LANGUAGE ${VLLM_GPU_LANG}
  SOURCES ${VLLM_EXT_SRC}
  COMPILE_FLAGS ${VLLM_GPU_FLAGS}
  ARCHITECTURES ${VLLM_GPU_ARCHES}
932
  INCLUDE_DIRECTORIES ${CUTLASS_INCLUDE_DIR}
933
  INCLUDE_DIRECTORIES ${CUTLASS_TOOLS_UTIL_INCLUDE_DIR}
934
  USE_SABI 3
bnellnm's avatar
bnellnm committed
935
936
  WITH_SOABI)

937
938
# If CUTLASS is compiled on NVCC >= 12.5, it by default uses
# cudaGetDriverEntryPointByVersion as a wrapper to avoid directly calling the
939
940
941
942
# 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
943
944
945
946
947
#
# _moe_C extension
#

set(VLLM_MOE_EXT_SRC
948
  "csrc/moe/torch_bindings.cpp"
949
  "csrc/moe/moe_align_sum_kernels.cu"
950
  "csrc/moe/topk_softmax_kernels.cu")
bnellnm's avatar
bnellnm committed
951

952
if(VLLM_GPU_LANG STREQUAL "CUDA")
953
954
955
  list(APPEND VLLM_MOE_EXT_SRC
    "csrc/moe/moe_wna16.cu"
    "csrc/moe/grouped_topk_kernels.cu")
956
957
endif()

958
959
960
961
962
963
964
965
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()

966
967
968
969
set_gencode_flags_for_srcs(
  SRCS "${VLLM_MOE_EXT_SRC}"
  CUDA_ARCHS "${CUDA_ARCHS}")

970
if(VLLM_GPU_LANG STREQUAL "CUDA")
971
972
973
974
975
976
977
  set(VLLM_MOE_WNA16_SRC
    "csrc/moe/moe_wna16.cu")

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

978
  list(APPEND VLLM_MOE_EXT_SRC "${VLLM_MOE_WNA16_SRC}")
979
980
981
982
983
984
985
986
987
  # moe marlin arches
  # note that we always set `use_atomic_add=False` for moe marlin now,
  # so we don't need 9.0 for bf16 atomicAdd PTX
  cuda_archs_loose_intersection(MARLIN_MOE_ARCHS "8.0+PTX" "${CUDA_ARCHS}")
  # 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)
  cuda_archs_loose_intersection(MARLIN_MOE_FP8_ARCHS "8.9;12.0" "${CUDA_ARCHS}")
988
989
  if (MARLIN_MOE_ARCHS)

990
991
992
993
994
995
996
    #
    # 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 ${MOE_MARLIN_GEN_SCRIPT} MOE_MARLIN_GEN_SCRIPT_HASH)
997
998
    list(JOIN CUDA_ARCHS "," CUDA_ARCHS_STR)
    set(MOE_MARLIN_GEN_SCRIPT_HASH_AND_ARCH "${MOE_MARLIN_GEN_SCRIPT_HASH}(ARCH:${CUDA_ARCHS_STR})")
999

1000
1001
    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}")
1002

1003
1004
    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})
1005
1006
      execute_process(
        COMMAND ${CMAKE_COMMAND} -E env
1007
        PYTHONPATH=$PYTHONPATH
1008
          ${Python_EXECUTABLE} ${MOE_MARLIN_GEN_SCRIPT} ${CUDA_ARCHS_STR}
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
        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()
1021
        set(MOE_MARLIN_GEN_SCRIPT_HASH_AND_ARCH ${MOE_MARLIN_GEN_SCRIPT_HASH_AND_ARCH}
1022
1023
1024
1025
1026
1027
1028
            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()

1029
1030
    file(GLOB MARLIN_MOE_SRC "csrc/moe/marlin_moe_wna16/sm80_kernel_*.cu")
    list(APPEND MARLIN_MOE_SRC "csrc/moe/marlin_moe_wna16/ops.cu")
1031
    set_gencode_flags_for_srcs(
1032
      SRCS "${MARLIN_MOE_SRC}"
1033
      CUDA_ARCHS "${MARLIN_MOE_ARCHS}")
1034
    if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8)
1035
      set_source_files_properties(${MARLIN_MOE_SRC}
1036
1037
        PROPERTIES COMPILE_FLAGS "-static-global-template-stub=false")
    endif()
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
    list(APPEND VLLM_MOE_EXT_SRC ${MARLIN_MOE_SRC})

    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()
1051

1052
1053
1054
    message(STATUS "Building Marlin MOE kernels for archs: ${MARLIN_MOE_ARCHS}")
  else()
    message(STATUS "Not building Marlin MOE kernels as no compatible archs found"
1055
                   " in CUDA target architectures")
1056
  endif()
1057
1058
endif()

1059
message(STATUS "Enabling moe extension.")
1060
define_extension_target(
bnellnm's avatar
bnellnm committed
1061
1062
1063
1064
1065
1066
  _moe_C
  DESTINATION vllm
  LANGUAGE ${VLLM_GPU_LANG}
  SOURCES ${VLLM_MOE_EXT_SRC}
  COMPILE_FLAGS ${VLLM_GPU_FLAGS}
  ARCHITECTURES ${VLLM_GPU_ARCHES}
1067
1068
  INCLUDE_DIRECTORIES ${CUTLASS_INCLUDE_DIR}
  INCLUDE_DIRECTORIES ${CUTLASS_TOOLS_UTIL_INCLUDE_DIR}
1069
  USE_SABI 3
bnellnm's avatar
bnellnm committed
1070
1071
  WITH_SOABI)

1072
#[[
1073
1074
1075
1076
1077
1078
if(VLLM_GPU_LANG STREQUAL "HIP")
  #
  # _rocm_C extension
  #
  set(VLLM_ROCM_EXT_SRC
    "csrc/rocm/torch_bindings.cpp"
1079
    "csrc/rocm/skinny_gemms.cu"
1080
1081
    "csrc/rocm/attention.cu")

1082
  define_extension_target(
1083
1084
1085
1086
1087
1088
1089
1090
1091
    _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()
1092
]]
1093

1094
# For CUDA and HIP builds also build the triton_kernels external package.
zhuwenwen's avatar
zhuwenwen committed
1095
if(VLLM_GPU_LANG STREQUAL "CUDA")
1096
1097
1098
    include(cmake/external_projects/triton_kernels.cmake)
endif()

1099
1100
1101
# For CUDA we also build and ship some external projects.
if (VLLM_GPU_LANG STREQUAL "CUDA")
    include(cmake/external_projects/flashmla.cmake)
1102
    include(cmake/external_projects/qutlass.cmake)
1103
1104

    # vllm-flash-attn should be last as it overwrites some CMake functions
1105
    include(cmake/external_projects/vllm_flash_attn.cmake)
1106
endif ()