CMakeLists.txt 29.7 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}")

bnellnm's avatar
bnellnm committed
26
27
28
29
#
# 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.
#
30
set(PYTHON_SUPPORTED_VERSIONS "3.9" "3.10" "3.11" "3.12")
bnellnm's avatar
bnellnm committed
31
32

# Supported NVIDIA architectures.
33
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")
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
#
# 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()
90
        return()
91
92
93
94
    endif()
    return()
endif()

bnellnm's avatar
bnellnm committed
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
#
# 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)

114
115
116
  # ROCm 5.X and 6.X
  if (ROCM_VERSION_DEV_MAJOR GREATER_EQUAL 5 AND
      NOT Torch_VERSION VERSION_EQUAL ${TORCH_SUPPORTED_VERSION_ROCM})
117
    message(WARNING "Pytorch version >= ${TORCH_SUPPORTED_VERSION_ROCM} "
118
      "expected for ROCm build, saw ${Torch_VERSION} instead.")
bnellnm's avatar
bnellnm committed
119
120
121
122
123
  endif()
else()
  message(FATAL_ERROR "Can't find CUDA or HIP installation.")
endif()

124
125

if(VLLM_GPU_LANG STREQUAL "CUDA")
126
  #
127
  # For cuda we want to be able to control which architectures we compile for on
128
  # a per-file basis in order to cut down on compile time. So here we extract
129
  # the set of architectures we want to compile for and remove the from the
130
131
  # CMAKE_CUDA_FLAGS so that they are not applied globally.
  #
132
133
134
  clear_cuda_arches(CUDA_ARCH_FLAGS)
  extract_unique_cuda_archs_ascending(CUDA_ARCHS "${CUDA_ARCH_FLAGS}")
  message(STATUS "CUDA target architectures: ${CUDA_ARCHS}")
135
136
  # Filter the target architectures by the supported supported archs
  # since for some files we will build for all CUDA_ARCHS.
137
  cuda_archs_loose_intersection(CUDA_ARCHS
138
139
140
141
142
143
144
145
146
147
148
    "${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}")
149
150
endif()

bnellnm's avatar
bnellnm committed
151
152
153
154
155
156
157
158
159
160
161
162
163
164
#
# 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()

165
166
167

#
# Use FetchContent for C++ dependencies that are compiled as part of vLLM's build process.
168
169
170
# 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>.
171
#
172
include(FetchContent)
173
file(MAKE_DIRECTORY ${FETCHCONTENT_BASE_DIR}) # Ensure the directory exists
174
message(STATUS "FetchContent base directory: ${FETCHCONTENT_BASE_DIR}")
175

176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
#
# Set rocm version dev int.
#
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
195
#
196
# Define other extension targets
bnellnm's avatar
bnellnm committed
197
198
#

199
200
201
202
203
204
205
206
207
208
209
210
211
212
#
# 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
213
  list(APPEND CUMEM_LIBS CUDA::cuda_driver)
214
215
216
217
218
219
220
221
222
223
  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
224
225
226
227
228
229
#
# _C extension
#

set(VLLM_EXT_SRC
  "csrc/cache_kernels.cu"
230
231
  "csrc/attention/paged_attention_v1.cu"
  "csrc/attention/paged_attention_v2.cu"
232
  "csrc/attention/merge_attn_states.cu"
bnellnm's avatar
bnellnm committed
233
234
235
  "csrc/pos_encoding_kernels.cu"
  "csrc/activation_kernels.cu"
  "csrc/layernorm_kernels.cu"
236
  "csrc/layernorm_quant_kernels.cu"
237
  "csrc/cuda_view.cu"
bnellnm's avatar
bnellnm committed
238
  "csrc/quantization/gptq/q_gemm.cu"
239
  "csrc/quantization/compressed_tensors/int8_quant_kernels.cu"
240
  "csrc/quantization/fp8/common.cu"
241
  "csrc/quantization/fused_kernels/fused_layernorm_dynamic_per_token_quant.cu"
242
  "csrc/quantization/gguf/gguf_kernel.cu"
243
  "csrc/quantization/activation_kernels.cu"
bnellnm's avatar
bnellnm committed
244
  "csrc/cuda_utils_kernels.cu"
245
  "csrc/prepare_inputs/advance_step.cu"
246
  "csrc/custom_all_reduce.cu"
247
  "csrc/torch_bindings.cpp")
bnellnm's avatar
bnellnm committed
248
249

if(VLLM_GPU_LANG STREQUAL "CUDA")
250
  SET(CUTLASS_ENABLE_HEADERS_ONLY ON CACHE BOOL "Enable only the header library")
251

252
  # Set CUTLASS_REVISION. Used for FetchContent. Also fixes some bogus messages when building.
253
  set(CUTLASS_REVISION "v3.9.2" CACHE STRING "CUTLASS revision to use")
254

255
256
257
258
259
260
261
262
263
264
265
266
267
  # 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(
268
        cutlass
269
        GIT_REPOSITORY https://github.com/nvidia/cutlass.git
270
        # Please keep this in sync with CUTLASS_REVISION line above.
271
        GIT_TAG ${CUTLASS_REVISION}
272
        GIT_PROGRESS TRUE
273
274
275
276

        # 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
277
        GIT_SHALLOW TRUE
278
279
    )
  endif()
280
281
  FetchContent_MakeAvailable(cutlass)

bnellnm's avatar
bnellnm committed
282
  list(APPEND VLLM_EXT_SRC
283
284
    "csrc/mamba/mamba_ssm/selective_scan_fwd.cu"
    "csrc/mamba/causal_conv1d/causal_conv1d.cu"
James Fleming's avatar
James Fleming committed
285
    "csrc/quantization/aqlm/gemm_kernels.cu"
bnellnm's avatar
bnellnm committed
286
    "csrc/quantization/awq/gemm_kernels.cu"
287
    "csrc/permute_cols.cu"
288
    "csrc/quantization/cutlass_w8a8/scaled_mm_entry.cu"
289
    "csrc/quantization/fp4/nvfp4_quant_entry.cu"
290
    "csrc/quantization/fp4/nvfp4_scaled_mm_entry.cu"
291
    "csrc/quantization/fp4/nvfp4_blockwise_moe_kernel.cu"
292
    "csrc/sparse/cutlass/sparse_scaled_mm_entry.cu"
293
294
    "csrc/cutlass_extensions/common.cpp"
    "csrc/attention/mla/cutlass_mla_entry.cu")
295
296
297
298
299
300
301
302

  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.
303
  cuda_archs_loose_intersection(MARLIN_ARCHS "8.0;8.6;8.7;8.9;9.0;10.0;10.1;12.0" "${CUDA_ARCHS}")
304
  if (MARLIN_ARCHS)
305
306
307
308
309
310
311
312
313
314
315
316
317
318
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

    #
    # 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})

350
    set(MARLIN_SRCS
351
352
353
354
355
356
357
358
359
360
361
362
363
       "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"
364
                   " in CUDA target architectures")
365
366
  endif()

367
368
  # 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" "${CUDA_ARCHS}")
369
  if (ALLSPARK_ARCHS)
370
371
372
373
374
375
376
377
378
379
    set(ALLSPARK_SRCS
       "csrc/quantization/gptq_allspark/allspark_repack.cu"
       "csrc/quantization/gptq_allspark/allspark_qgemm_w8a16.cu")
    set_gencode_flags_for_srcs(
      SRCS "${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"
380
                   " in CUDA target architectures")
381
382
  endif()

383
384

  set(SCALED_MM_3X_ARCHS)
385
  # The cutlass_scaled_mm kernels for Hopper (c3x, i.e. CUTLASS 3.x) require
386
387
388
  # CUDA 12.0 or later
  cuda_archs_loose_intersection(SCALED_MM_ARCHS "9.0a;" "${CUDA_ARCHS}")
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.0 AND SCALED_MM_ARCHS)
389
    set(SRCS
390
       "csrc/quantization/cutlass_w8a8/scaled_mm_c3x_sm90.cu"
391
392
393
394
       "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")
395
396
    set_gencode_flags_for_srcs(
      SRCS "${SRCS}"
397
      CUDA_ARCHS "${SCALED_MM_ARCHS}")
398
    list(APPEND VLLM_EXT_SRC "${SRCS}")
399
400
401
402
    list(APPEND VLLM_GPU_FLAGS "-DENABLE_SCALED_MM_SM90=1")
    # Let scaled_mm_c2x know it doesn't need to build these arches
    list(APPEND SCALED_MM_3X_ARCHS "${SCALED_MM_ARCHS}")
    message(STATUS "Building scaled_mm_c3x_sm90 for archs: ${SCALED_MM_ARCHS}")
403
  else()
404
405
    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.0 AND SCALED_MM_ARCHS)
      message(STATUS "Not building scaled_mm_c3x_sm90 as CUDA Compiler version is "
406
                     "not >= 12.0, we recommend upgrading to CUDA 12.0 or "
407
                     "later if you intend on running FP8 quantized models on "
408
409
                     "Hopper.")
    else()
410
      message(STATUS "Not building scaled_mm_c3x_sm90 as no compatible archs found "
411
412
                     "in CUDA target architectures")
    endif()
413
  endif()
414

415
416
417
418
419
420
421
  # The cutlass_scaled_mm kernels for Blackwell (c3x, i.e. CUTLASS 3.x) require
  # CUDA 12.8 or later
  cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a;10.1a;12.0a" "${CUDA_ARCHS}")
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.8 AND SCALED_MM_ARCHS)
    set(SRCS
      "csrc/quantization/cutlass_w8a8/scaled_mm_c3x_sm100.cu"
      "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm100_fp8.cu"
422
      "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_blockwise_sm100_fp8.cu"
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
    )
    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()
    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 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()
442
  endif()
443
444

  #
445
446
  # 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.
447
  cuda_archs_loose_intersection(SCALED_MM_2X_ARCHS
448
    "7.5;8.0;8.6;8.7;8.9;9.0;10.0;10.1;12.0" "${CUDA_ARCHS}")
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
  # 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")
466
467
468
469
470
471
472
    endif()
  endif()

  #
  # 2:4 Sparse Kernels

  # The 2:4 sparse kernels cutlass_scaled_sparse_mm and cutlass_compressor
473
474
  # require CUDA 12.2 or later (and only work on Hopper).
  cuda_archs_loose_intersection(SCALED_MM_ARCHS "9.0a;" "${CUDA_ARCHS}")
475
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.2 AND SCALED_MM_ARCHS)
476
    set(SRCS "csrc/sparse/cutlass/sparse_scaled_mm_c3x.cu")
477
478
    set_gencode_flags_for_srcs(
      SRCS "${SRCS}"
479
      CUDA_ARCHS "${SCALED_MM_ARCHS}")
480
481
    list(APPEND VLLM_EXT_SRC "${SRCS}")
    list(APPEND VLLM_GPU_FLAGS "-DENABLE_SPARSE_SCALED_MM_C3X=1")
482
    message(STATUS "Building sparse_scaled_mm_c3x for archs: ${SCALED_MM_ARCHS}")
483
  else()
484
    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.2 AND SCALED_MM_ARCHS)
485
486
487
488
489
490
      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")
491
    endif()
492
  endif()
493

494
495
496
  # FP4 Archs and flags
  cuda_archs_loose_intersection(FP4_ARCHS "10.0a" "${CUDA_ARCHS}")
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.8 AND FP4_ARCHS)
497
    set(SRCS
498
      "csrc/quantization/fp4/nvfp4_quant_kernels.cu"
499
500
501
      "csrc/quantization/fp4/nvfp4_experts_quant.cu"
      "csrc/quantization/fp4/nvfp4_scaled_mm_kernels.cu"
      "csrc/quantization/fp4/nvfp4_blockwise_moe_kernel.cu")
502
503
504
505
506
507
508
509
510
511
512
    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")
    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()
513

514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
  # CUTLASS MLA Archs and flags
  cuda_archs_loose_intersection(MLA_ARCHS "10.0a" "${CUDA_ARCHS}")
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.8 AND MLA_ARCHS)
    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()

534
535
536
537
538
  # CUTLASS MoE kernels

  # The MoE kernel cutlass_moe_mm requires CUDA 12.3 or later (and only works
  # on Hopper). get_cutlass_moe_mm_data should only be compiled if it's possible
  # to compile MoE kernels that use its output.
539
  cuda_archs_loose_intersection(SCALED_MM_ARCHS "9.0a;10.0a" "${CUDA_ARCHS}")
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.3 AND SCALED_MM_ARCHS)
    set(SRCS "csrc/quantization/cutlass_w8a8/moe/grouped_mm_c3x.cu"
             "csrc/quantization/cutlass_w8a8/moe/moe_data.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_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()

560
  #
561
  # Machete kernels
562

563
  # The machete kernels only work on hopper and require CUDA 12.0 or later.
564
565
566
  # Only build Machete kernels if we are building for something compatible with sm90a
  cuda_archs_loose_intersection(MACHETE_ARCHS "9.0a" "${CUDA_ARCHS}")
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.0 AND MACHETE_ARCHS)
567
    #
568
    # For the Machete kernels we automatically generate sources for various
569
570
    # preselected input type pairs and schedules.
    # Generate sources:
571
    set(MACHETE_GEN_SCRIPT
572
573
574
575
576
577
578
579
580
      ${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(
581
582
        COMMAND ${CMAKE_COMMAND} -E env
        PYTHONPATH=${CMAKE_CURRENT_SOURCE_DIR}/csrc/cutlass_extensions/:${CUTLASS_DIR}/python/:${VLLM_PYTHON_PATH}:$PYTHONPATH
583
584
585
586
587
588
589
590
591
          ${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."
592
                            " Result: \"${machete_generation_result}\""
593
594
595
                            "\nCheck the log for details: "
                            "${CMAKE_CURRENT_BINARY_DIR}/machete_generation.log")
      else()
596
        set(MACHETE_GEN_SCRIPT_HASH ${MACHETE_GEN_SCRIPT_HASH}
597
598
599
            CACHE STRING "Last run machete generate script hash" FORCE)
        message(STATUS "Machete generation completed successfully.")
      endif()
600
    else()
601
      message(STATUS "Machete generation script has not changed, skipping generation.")
602
    endif()
603

604
605
606
    # Add machete generated sources
    file(GLOB MACHETE_GEN_SOURCES "csrc/quantization/machete/generated/*.cu")
    list(APPEND VLLM_EXT_SRC ${MACHETE_GEN_SOURCES})
607

608
609
610
611
612
613
614
615
616
617
    # 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()
618
    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.0
619
620
621
622
623
624
625
626
627
        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()
628
  endif()
629
# if CUDA endif
bnellnm's avatar
bnellnm committed
630
631
endif()

632
message(STATUS "Enabling C extension.")
bnellnm's avatar
bnellnm committed
633
634
635
636
637
638
639
define_gpu_extension_target(
  _C
  DESTINATION vllm
  LANGUAGE ${VLLM_GPU_LANG}
  SOURCES ${VLLM_EXT_SRC}
  COMPILE_FLAGS ${VLLM_GPU_FLAGS}
  ARCHITECTURES ${VLLM_GPU_ARCHES}
640
  INCLUDE_DIRECTORIES ${CUTLASS_INCLUDE_DIR}
641
  INCLUDE_DIRECTORIES ${CUTLASS_TOOLS_UTIL_INCLUDE_DIR}
642
  USE_SABI 3
bnellnm's avatar
bnellnm committed
643
644
  WITH_SOABI)

645
646
# If CUTLASS is compiled on NVCC >= 12.5, it by default uses
# cudaGetDriverEntryPointByVersion as a wrapper to avoid directly calling the
647
648
649
650
# 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
651
652
653
654
655
#
# _moe_C extension
#

set(VLLM_MOE_EXT_SRC
656
  "csrc/moe/torch_bindings.cpp"
657
  "csrc/moe/moe_align_sum_kernels.cu"
658
  "csrc/moe/topk_softmax_kernels.cu")
bnellnm's avatar
bnellnm committed
659

660
661
662
663
if(VLLM_GPU_LANG STREQUAL "CUDA")
  list(APPEND VLLM_MOE_EXT_SRC "csrc/moe/moe_wna16.cu")
endif()

664
665
666
667
set_gencode_flags_for_srcs(
  SRCS "${VLLM_MOE_EXT_SRC}"
  CUDA_ARCHS "${CUDA_ARCHS}")

668
if(VLLM_GPU_LANG STREQUAL "CUDA")
669
670
671
672
673
674
675
  set(VLLM_MOE_WNA16_SRC
    "csrc/moe/moe_wna16.cu")

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

676
  list(APPEND VLLM_MOE_EXT_SRC "${VLLM_MOE_WNA16_SRC}")
677
  cuda_archs_loose_intersection(MARLIN_MOE_ARCHS "8.0;8.6;8.7;8.9;9.0;10.0;10.1;12.0" "${CUDA_ARCHS}")
678
679
  if (MARLIN_MOE_ARCHS)

680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
    #
    # 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 ${CMAKE_COMMAND} -E env
695
        PYTHONPATH=$PYTHONPATH
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
          ${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")
718
    set_gencode_flags_for_srcs(
719
      SRCS "${MOE_WNAA16_MARLIN_SRC}"
720
721
      CUDA_ARCHS "${MARLIN_MOE_ARCHS}")

722
723
    list(APPEND VLLM_MOE_EXT_SRC ${MOE_WNAA16_MARLIN_SRC})

724
725
726
    message(STATUS "Building Marlin MOE kernels for archs: ${MARLIN_MOE_ARCHS}")
  else()
    message(STATUS "Not building Marlin MOE kernels as no compatible archs found"
727
                   " in CUDA target architectures")
728
  endif()
729
730
endif()

731
732
733
734
735
736
737
738
739
740
741
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()
742
message(STATUS "Enabling moe extension.")
bnellnm's avatar
bnellnm committed
743
744
745
746
747
748
749
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}
750
751
  INCLUDE_DIRECTORIES ${CUTLASS_INCLUDE_DIR}
  INCLUDE_DIRECTORIES ${CUTLASS_TOOLS_UTIL_INCLUDE_DIR}
752
  USE_SABI 3
bnellnm's avatar
bnellnm committed
753
754
  WITH_SOABI)

755
756
757
758
759
760
if(VLLM_GPU_LANG STREQUAL "HIP")
  #
  # _rocm_C extension
  #
  set(VLLM_ROCM_EXT_SRC
    "csrc/rocm/torch_bindings.cpp"
761
    "csrc/rocm/skinny_gemms.cu"
762
763
764
765
766
767
768
769
770
771
772
773
774
    "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()

775
776
777
778
# For CUDA we also build and ship some external projects.
if (VLLM_GPU_LANG STREQUAL "CUDA")
    include(cmake/external_projects/flashmla.cmake)
    include(cmake/external_projects/vllm_flash_attn.cmake)
779
endif ()