CMakeLists.txt 27.8 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")
18

zhangshao's avatar
zhangshao committed
19
20
set(CMAKE_BUILD_TYPE "Release")

bnellnm's avatar
bnellnm committed
21
message(STATUS "Build type: ${CMAKE_BUILD_TYPE}")
22
message(STATUS "Target device: ${VLLM_TARGET_DEVICE}")
bnellnm's avatar
bnellnm committed
23
24

include(${CMAKE_CURRENT_LIST_DIR}/cmake/utils.cmake)
zhangshao's avatar
zhangshao committed
25
add_compile_options(-w)
26

bnellnm's avatar
bnellnm committed
27

28
29
30
# Suppress potential warnings about unused manually-specified variables
set(ignoreMe "${VLLM_PYTHON_PATH}")

bnellnm's avatar
bnellnm committed
31
32
33
34
#
# 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.
#
35
set(PYTHON_SUPPORTED_VERSIONS "3.9" "3.10" "3.11" "3.12")
bnellnm's avatar
bnellnm committed
36
37

# Supported NVIDIA architectures.
38
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
39
40

# Supported AMD GPU architectures.
zhuwenwen's avatar
zhuwenwen committed
41
set(HIP_SUPPORTED_ARCHS "gfx908;gfx90a;gfx942;gfx950;gfx1030;gfx1100;gfx1101;gfx1200;gfx1201;gfx906;gfx926;gfx928;gfx936")
bnellnm's avatar
bnellnm committed
42
43
44
45
46
47
48
49
50

#
# 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
51
# versions are derived from docker/Dockerfile.rocm
bnellnm's avatar
bnellnm committed
52
#
Michael Goin's avatar
Michael Goin committed
53
54
set(TORCH_SUPPORTED_VERSION_CUDA "2.6.0")
set(TORCH_SUPPORTED_VERSION_ROCM "2.6.0")
bnellnm's avatar
bnellnm committed
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72

#
# 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")

73
74
# Ensure the 'nvcc' command is in the PATH
find_program(NVCC_EXECUTABLE nvcc)
75
if (CUDA_FOUND AND NOT NVCC_EXECUTABLE)
76
77
78
    message(FATAL_ERROR "nvcc not found")
endif()

bnellnm's avatar
bnellnm committed
79
80
81
82
83
84
85
86
#
# 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)

87
88
89
90
91
92
93
94
#
# 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()
95
        return()
96
97
98
99
    endif()
    return()
endif()

bnellnm's avatar
bnellnm committed
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
#
# 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)

119
120
121
  # ROCm 5.X and 6.X
  if (ROCM_VERSION_DEV_MAJOR GREATER_EQUAL 5 AND
      NOT Torch_VERSION VERSION_EQUAL ${TORCH_SUPPORTED_VERSION_ROCM})
122
    message(WARNING "Pytorch version >= ${TORCH_SUPPORTED_VERSION_ROCM} "
123
      "expected for ROCm build, saw ${Torch_VERSION} instead.")
bnellnm's avatar
bnellnm committed
124
125
126
127
128
  endif()
else()
  message(FATAL_ERROR "Can't find CUDA or HIP installation.")
endif()

129
130

if(VLLM_GPU_LANG STREQUAL "CUDA")
131
  #
132
  # For cuda we want to be able to control which architectures we compile for on
133
  # a per-file basis in order to cut down on compile time. So here we extract
134
  # the set of architectures we want to compile for and remove the from the
135
136
  # CMAKE_CUDA_FLAGS so that they are not applied globally.
  #
137
138
139
  clear_cuda_arches(CUDA_ARCH_FLAGS)
  extract_unique_cuda_archs_ascending(CUDA_ARCHS "${CUDA_ARCH_FLAGS}")
  message(STATUS "CUDA target architectures: ${CUDA_ARCHS}")
140
141
  # Filter the target architectures by the supported supported archs
  # since for some files we will build for all CUDA_ARCHS.
142
  cuda_archs_loose_intersection(CUDA_ARCHS
143
144
145
146
147
148
149
150
151
152
153
    "${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}")
154
155
endif()

bnellnm's avatar
bnellnm committed
156
157
158
159
160
161
162
163
164
165
166
167
168
169
#
# 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()

170
171
172

#
# Use FetchContent for C++ dependencies that are compiled as part of vLLM's build process.
173
174
175
# 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>.
176
#
177
include(FetchContent)
178
file(MAKE_DIRECTORY ${FETCHCONTENT_BASE_DIR}) # Ensure the directory exists
179
message(STATUS "FetchContent base directory: ${FETCHCONTENT_BASE_DIR}")
180

181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
#
# 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
200
#
201
# Define other extension targets
bnellnm's avatar
bnellnm committed
202
203
#

204
205
206
207
208
209
210
211
212
213
214
215
216
217
#
# 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
218
  list(APPEND CUMEM_LIBS CUDA::cuda_driver)
219
220
221
222
223
224
225
226
227
228
  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
229
230
231
232
233
234
#
# _C extension
#

set(VLLM_EXT_SRC
  "csrc/cache_kernels.cu"
235
236
  "csrc/attention/paged_attention_v1.cu"
  "csrc/attention/paged_attention_v2.cu"
237
  "csrc/attention/merge_attn_states.cu"
bnellnm's avatar
bnellnm committed
238
  "csrc/pos_encoding_kernels.cu"
huangwb's avatar
huangwb committed
239
  "csrc/pos_encoding_tgi_kernels.cu"
bnellnm's avatar
bnellnm committed
240
241
  "csrc/activation_kernels.cu"
  "csrc/layernorm_kernels.cu"
zhuwenwen's avatar
zhuwenwen committed
242
243
244
  "csrc/opt/transpose_kernels.cu"
  "csrc/opt/activation_kernels_opt.cu"
  "csrc/attention/attention_kernels_opt.cu"
245
  "csrc/attention/attention_kernels_opt_tc.cu"
zhuwenwen's avatar
zhuwenwen committed
246
247
248
  "csrc/attention/attention_with_mask_kernels.cu"
  "csrc/attention/attention_with_mask_kernels_opt.cu"
  "csrc/attention/attention_with_mask_kernels_opt_tc.cu"
zhuwenwen's avatar
zhuwenwen committed
249
  "csrc/opt/layernorm_kernels_opt.cu"
250
  # "csrc/layernorm_quant_kernels.cu"
251
  "csrc/cuda_view.cu"
252
  # "csrc/quantization/gptq/q_gemm.cu"
zhuwenwen's avatar
zhuwenwen committed
253
  "csrc/quantization/compressed_tensors/int8_quant_kernels.cu"
zhuwenwen's avatar
zhuwenwen committed
254
  # "csrc/quantization/fp8/common.cu"
255
  # "csrc/quantization/fused_kernels/fused_layernorm_dynamic_per_token_quant.cu"
256
  "csrc/quantization/gguf/gguf_kernel.cu"
bnellnm's avatar
bnellnm committed
257
  "csrc/cuda_utils_kernels.cu"
258
  "csrc/prepare_inputs/advance_step.cu"
259
  "csrc/custom_all_reduce.cu"
260
  "csrc/torch_bindings.cpp")
bnellnm's avatar
bnellnm committed
261
262

if(VLLM_GPU_LANG STREQUAL "CUDA")
263
  SET(CUTLASS_ENABLE_HEADERS_ONLY ON CACHE BOOL "Enable only the header library")
264
265

  # Set CUTLASS_REVISION manually -- its revision detection doesn't work in this case.
266
  # Please keep this in sync with FetchContent_Declare line below.
267
  set(CUTLASS_REVISION "v3.9.0" CACHE STRING "CUTLASS revision to use")
268

269
270
271
272
273
274
275
276
277
278
279
280
281
  # 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(
282
        cutlass
283
        GIT_REPOSITORY https://github.com/nvidia/cutlass.git
284
        # Please keep this in sync with CUTLASS_REVISION line above.
285
        GIT_TAG v3.9.0
286
        GIT_PROGRESS TRUE
287
288
289
290

        # 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
291
        GIT_SHALLOW TRUE
292
293
    )
  endif()
294
295
  FetchContent_MakeAvailable(cutlass)

bnellnm's avatar
bnellnm committed
296
  list(APPEND VLLM_EXT_SRC
297
298
    "csrc/mamba/mamba_ssm/selective_scan_fwd.cu"
    "csrc/mamba/causal_conv1d/causal_conv1d.cu"
James Fleming's avatar
James Fleming committed
299
    "csrc/quantization/aqlm/gemm_kernels.cu"
bnellnm's avatar
bnellnm committed
300
    "csrc/quantization/awq/gemm_kernels.cu"
301
    "csrc/permute_cols.cu"
302
    "csrc/quantization/cutlass_w8a8/scaled_mm_entry.cu"
303
    "csrc/quantization/fp4/nvfp4_quant_entry.cu"
304
    "csrc/quantization/fp4/nvfp4_scaled_mm_entry.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
  cuda_archs_loose_intersection(MARLIN_ARCHS "8.0;8.6;8.7;8.9;9.0;10.0;10.1;12.0" "${CUDA_ARCHS}")
317
  if (MARLIN_ARCHS)
318
    set(MARLIN_SRCS
319
320
321
322
323
324
325
326
327
328
329
330
331
332
       "csrc/quantization/fp8/fp8_marlin.cu"
       "csrc/quantization/marlin/dense/marlin_cuda_kernel.cu"
       "csrc/quantization/marlin/sparse/marlin_24_cuda_kernel.cu"
       "csrc/quantization/marlin/qqq/marlin_qqq_gemm_kernel.cu"
       "csrc/quantization/gptq_marlin/gptq_marlin.cu"
       "csrc/quantization/gptq_marlin/gptq_marlin_repack.cu"
       "csrc/quantization/gptq_marlin/awq_marlin_repack.cu")
    set_gencode_flags_for_srcs(
      SRCS "${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"
333
                   " in CUDA target architectures")
334
335
  endif()

336
337
  # 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" "${CUDA_ARCHS}")
338
  if (ALLSPARK_ARCHS)
339
340
341
342
343
344
345
346
347
348
    set(ALLSPARK_SRCS
       "csrc/quantization/gptq_allspark/allspark_repack.cu"
       "csrc/quantization/gptq_allspark/allspark_qgemm_w8a16.cu")
    set_gencode_flags_for_srcs(
      SRCS "${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"
349
                   " in CUDA target architectures")
350
351
  endif()

352
353

  set(SCALED_MM_3X_ARCHS)
354
  # The cutlass_scaled_mm kernels for Hopper (c3x, i.e. CUTLASS 3.x) require
355
356
357
  # 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)
358
    set(SRCS
359
       "csrc/quantization/cutlass_w8a8/scaled_mm_c3x_sm90.cu"
360
361
362
363
       "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")
364
365
    set_gencode_flags_for_srcs(
      SRCS "${SRCS}"
366
      CUDA_ARCHS "${SCALED_MM_ARCHS}")
367
    list(APPEND VLLM_EXT_SRC "${SRCS}")
368
369
370
371
    list(APPEND VLLM_GPU_FLAGS "-DENABLE_SCALED_MM_SM90=1")
    # Let scaled_mm_c2x know it doesn't need to build these arches
    list(APPEND SCALED_MM_3X_ARCHS "${SCALED_MM_ARCHS}")
    message(STATUS "Building scaled_mm_c3x_sm90 for archs: ${SCALED_MM_ARCHS}")
372
  else()
373
374
    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 "
375
376
377
378
                     "not >= 12.0, we recommend upgrading to CUDA 12.0 or "
                     "later if you intend on running FP8 quantized models on "
                     "Hopper.")
    else()
379
      message(STATUS "Not building scaled_mm_c3x_sm90 as no compatible archs found "
380
381
                     "in CUDA target architectures")
    endif()
382
  endif()
383

384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
  # 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"
    )
    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()
410
  endif()
411
412

  #
413
414
  # 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.
415
  cuda_archs_loose_intersection(SCALED_MM_2X_ARCHS
416
    "7.5;8.0;8.6;8.7;8.9;9.0;10.0;10.1;12.0" "${CUDA_ARCHS}")
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
  # 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")
    endif()
435
  endif()
436

437
438
439
440
  #
  # 2:4 Sparse Kernels

  # The 2:4 sparse kernels cutlass_scaled_sparse_mm and cutlass_compressor
441
442
  # require CUDA 12.2 or later (and only work on Hopper).
  cuda_archs_loose_intersection(SCALED_MM_ARCHS "9.0a;" "${CUDA_ARCHS}")
443
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.2 AND SCALED_MM_ARCHS)
444
    set(SRCS "csrc/sparse/cutlass/sparse_scaled_mm_c3x.cu")
445
446
    set_gencode_flags_for_srcs(
      SRCS "${SRCS}"
447
      CUDA_ARCHS "${SCALED_MM_ARCHS}")
448
449
    list(APPEND VLLM_EXT_SRC "${SRCS}")
    list(APPEND VLLM_GPU_FLAGS "-DENABLE_SPARSE_SCALED_MM_C3X=1")
450
    message(STATUS "Building sparse_scaled_mm_c3x for archs: ${SCALED_MM_ARCHS}")
451
  else()
452
    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.2 AND SCALED_MM_ARCHS)
453
454
455
456
457
458
      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")
459
    endif()
460
  endif()
461

462
463
464
  # 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)
465
    set(SRCS
466
      "csrc/quantization/fp4/nvfp4_quant_kernels.cu"
467
      "csrc/quantization/fp4/nvfp4_scaled_mm_kernels.cu")
468
469
470
471
472
473
474
475
476
477
478
    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()
479

480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
  # 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()

500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
  # 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.
  cuda_archs_loose_intersection(SCALED_MM_ARCHS "9.0a;" "${CUDA_ARCHS}")
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.3 AND SCALED_MM_ARCHS)
    set(SRCS "csrc/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()

526
  #
527
  # Machete kernels
528

529
  # The machete kernels only work on hopper and require CUDA 12.0 or later.
530
531
532
  # 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)
533
    #
534
    # For the Machete kernels we automatically generate sources for various
535
536
    # preselected input type pairs and schedules.
    # Generate sources:
537
    set(MACHETE_GEN_SCRIPT
538
539
540
541
542
543
544
545
546
      ${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(
547
548
        COMMAND ${CMAKE_COMMAND} -E env
        PYTHONPATH=${CMAKE_CURRENT_SOURCE_DIR}/csrc/cutlass_extensions/:${CUTLASS_DIR}/python/:${VLLM_PYTHON_PATH}:$PYTHONPATH
549
550
551
552
553
554
555
556
557
          ${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."
558
                            " Result: \"${machete_generation_result}\""
559
560
561
                            "\nCheck the log for details: "
                            "${CMAKE_CURRENT_BINARY_DIR}/machete_generation.log")
      else()
562
        set(MACHETE_GEN_SCRIPT_HASH ${MACHETE_GEN_SCRIPT_HASH}
563
564
565
            CACHE STRING "Last run machete generate script hash" FORCE)
        message(STATUS "Machete generation completed successfully.")
      endif()
566
    else()
567
      message(STATUS "Machete generation script has not changed, skipping generation.")
568
    endif()
569

570
571
572
    # Add machete generated sources
    file(GLOB MACHETE_GEN_SOURCES "csrc/quantization/machete/generated/*.cu")
    list(APPEND VLLM_EXT_SRC ${MACHETE_GEN_SOURCES})
573

574
575
576
577
578
579
580
581
582
583
    # 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()
584
    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.0
585
586
587
588
589
590
591
592
593
        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()
594
  endif()
595
# if CUDA endif
bnellnm's avatar
bnellnm committed
596
597
endif()

598
message(STATUS "Enabling C extension.")
bnellnm's avatar
bnellnm committed
599
600
601
602
603
604
605
define_gpu_extension_target(
  _C
  DESTINATION vllm
  LANGUAGE ${VLLM_GPU_LANG}
  SOURCES ${VLLM_EXT_SRC}
  COMPILE_FLAGS ${VLLM_GPU_FLAGS}
  ARCHITECTURES ${VLLM_GPU_ARCHES}
606
  INCLUDE_DIRECTORIES ${CUTLASS_INCLUDE_DIR}
607
  INCLUDE_DIRECTORIES ${CUTLASS_TOOLS_UTIL_INCLUDE_DIR}
608
  USE_SABI 3
bnellnm's avatar
bnellnm committed
609
610
  WITH_SOABI)

611
612
# If CUTLASS is compiled on NVCC >= 12.5, it by default uses
# cudaGetDriverEntryPointByVersion as a wrapper to avoid directly calling the
613
614
615
616
# 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
617
618
619
620
621
#
# _moe_C extension
#

set(VLLM_MOE_EXT_SRC
622
  "csrc/moe/torch_bindings.cpp"
623
  "csrc/moe/moe_align_sum_kernels.cu"
王敏's avatar
王敏 committed
624
625
  "csrc/moe/topk_softmax_kernels.cu"
  "csrc/moe/moe_fused_gate.cu")
bnellnm's avatar
bnellnm committed
626

627
628
629
630
if(VLLM_GPU_LANG STREQUAL "CUDA")
  list(APPEND VLLM_MOE_EXT_SRC "csrc/moe/moe_wna16.cu")
endif()

631
632
633
634
set_gencode_flags_for_srcs(
  SRCS "${VLLM_MOE_EXT_SRC}"
  CUDA_ARCHS "${CUDA_ARCHS}")

635
if(VLLM_GPU_LANG STREQUAL "CUDA")
636
637
638
639
640
641
642
  set(VLLM_MOE_WNA16_SRC
    "csrc/moe/moe_wna16.cu")

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

643
  list(APPEND VLLM_MOE_EXT_SRC "${VLLM_MOE_WNA16_SRC}")
644
  cuda_archs_loose_intersection(MARLIN_MOE_ARCHS "8.0;8.6;8.7;8.9;9.0;10.0;10.1;12.0" "${CUDA_ARCHS}")
645
646
  if (MARLIN_MOE_ARCHS)

647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
    #
    # 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 ${CMAKE_COMMAND} -E env
        PYTHONPATH=${CMAKE_CURRENT_SOURCE_DIR}/csrc/cutlass_extensions/:${CUTLASS_DIR}/python/:${VLLM_PYTHON_PATH}:$PYTHONPATH
          ${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")
685
    set_gencode_flags_for_srcs(
686
      SRCS "${MOE_WNAA16_MARLIN_SRC}"
687
688
      CUDA_ARCHS "${MARLIN_MOE_ARCHS}")

689
690
    list(APPEND VLLM_MOE_EXT_SRC ${MOE_WNAA16_MARLIN_SRC})

691
692
693
    message(STATUS "Building Marlin MOE kernels for archs: ${MARLIN_MOE_ARCHS}")
  else()
    message(STATUS "Not building Marlin MOE kernels as no compatible archs found"
694
                   " in CUDA target architectures")
695
  endif()
696
697
endif()

698
message(STATUS "Enabling moe extension.")
bnellnm's avatar
bnellnm committed
699
700
701
702
703
704
705
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}
706
  USE_SABI 3
bnellnm's avatar
bnellnm committed
707
708
  WITH_SOABI)

709
#[[  
710
711
712
713
714
715
if(VLLM_GPU_LANG STREQUAL "HIP")
  #
  # _rocm_C extension
  #
  set(VLLM_ROCM_EXT_SRC
    "csrc/rocm/torch_bindings.cpp"
716
    "csrc/rocm/skinny_gemms.cu"
717
718
719
720
721
722
723
724
725
726
727
728
    "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()
729
]]
bnellnm's avatar
bnellnm committed
730

731
732
733
734
# 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)
735
endif ()