CMakeLists.txt 20.6 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}")

31
32
33
# Prevent installation of dependencies (cutlass) by default.
install(CODE "set(CMAKE_INSTALL_LOCAL_ONLY TRUE)" ALL_COMPONENTS)

bnellnm's avatar
bnellnm committed
34
35
36
37
#
# 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.
#
38
set(PYTHON_SUPPORTED_VERSIONS "3.8" "3.9" "3.10" "3.11" "3.12")
bnellnm's avatar
bnellnm committed
39
40
41
42
43

# Supported NVIDIA architectures.
set(CUDA_SUPPORTED_ARCHS "7.0;7.5;8.0;8.6;8.9;9.0")

# Supported AMD GPU architectures.
zhuwenwen's avatar
zhuwenwen committed
44
set(HIP_SUPPORTED_ARCHS "gfx906;gfx908;gfx90a;gfx940;gfx941;gfx942;gfx1030;gfx1100;gfx926;gfx928")
bnellnm's avatar
bnellnm committed
45
46
47
48
49
50
51
52
53
54
55

#
# 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
# versions are derived from Dockerfile.rocm
#
56
set(TORCH_SUPPORTED_VERSION_CUDA "2.4.0")
57
set(TORCH_SUPPORTED_VERSION_ROCM "2.5.0")
bnellnm's avatar
bnellnm committed
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75

#
# 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")

76
77
# Ensure the 'nvcc' command is in the PATH
find_program(NVCC_EXECUTABLE nvcc)
78
if (CUDA_FOUND AND NOT NVCC_EXECUTABLE)
79
80
81
    message(FATAL_ERROR "nvcc not found")
endif()

bnellnm's avatar
bnellnm committed
82
83
84
85
86
87
88
89
#
# 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)

90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
#
message(STATUS "Enabling core extension.")

# Define _core_C extension
#  built for (almost) every target platform, (excludes TPU and Neuron)

set(VLLM_EXT_SRC
  "csrc/core/torch_bindings.cpp")

define_gpu_extension_target(
  _core_C
  DESTINATION vllm
  LANGUAGE CXX
  SOURCES ${VLLM_EXT_SRC}
  COMPILE_FLAGS ${CXX_COMPILE_FLAGS}
  USE_SABI 3
  WITH_SOABI)

108
109
110
111
112
113
114
115
#
# 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()
116
        return()
117
118
119
120
    endif()
    return()
endif()

bnellnm's avatar
bnellnm committed
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
#
# 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)

140
141
142
  # ROCm 5.X and 6.X
  if (ROCM_VERSION_DEV_MAJOR GREATER_EQUAL 5 AND
      NOT Torch_VERSION VERSION_EQUAL ${TORCH_SUPPORTED_VERSION_ROCM})
143
    message(WARNING "Pytorch version >= ${TORCH_SUPPORTED_VERSION_ROCM} "
144
      "expected for ROCm build, saw ${Torch_VERSION} instead.")
bnellnm's avatar
bnellnm committed
145
146
147
148
149
  endif()
else()
  message(FATAL_ERROR "Can't find CUDA or HIP installation.")
endif()

150
151

if(VLLM_GPU_LANG STREQUAL "CUDA")
152
153
154
155
156
157
  #
  # For cuda we want to be able to control which architectures we compile for on 
  # a per-file basis in order to cut down on compile time. So here we extract
  # the set of architectures we want to compile for and remove the from the 
  # CMAKE_CUDA_FLAGS so that they are not applied globally.
  #
158
159
160
  clear_cuda_arches(CUDA_ARCH_FLAGS)
  extract_unique_cuda_archs_ascending(CUDA_ARCHS "${CUDA_ARCH_FLAGS}")
  message(STATUS "CUDA target architectures: ${CUDA_ARCHS}")
161
162
163
164
165
166
167
168
169
170
171
172
173
174
  # Filter the target architectures by the supported supported archs
  # since for some files we will build for all CUDA_ARCHS.
  cuda_archs_loose_intersection(CUDA_ARCHS 
    "${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}")
175
176
endif()

bnellnm's avatar
bnellnm committed
177
178
179
180
181
182
183
184
185
186
187
188
189
190
#
# 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()

191
192
193
194
195

#
# Use FetchContent for C++ dependencies that are compiled as part of vLLM's build process.
# Configure it to place files in vllm/.deps, in order to play nicely with sccache.
#
196
include(FetchContent)
197
198
199
200
get_filename_component(PROJECT_ROOT_DIR "${CMAKE_CURRENT_SOURCE_DIR}" ABSOLUTE)
file(MAKE_DIRECTORY "${FETCHCONTENT_BASE_DIR}")
set(FETCHCONTENT_BASE_DIR "${PROJECT_ROOT_DIR}/.deps")
message(STATUS "FetchContent base directory: ${FETCHCONTENT_BASE_DIR}")
201

bnellnm's avatar
bnellnm committed
202
#
203
# Define other extension targets
bnellnm's avatar
bnellnm committed
204
205
206
207
208
209
210
211
212
213
#

#
# _C extension
#

set(VLLM_EXT_SRC
  "csrc/cache_kernels.cu"
  "csrc/attention/attention_kernels.cu"
  "csrc/pos_encoding_kernels.cu"
huangwb's avatar
huangwb committed
214
  "csrc/pos_encoding_tgi_kernels.cu"
bnellnm's avatar
bnellnm committed
215
216
  "csrc/activation_kernels.cu"
  "csrc/layernorm_kernels.cu"
zhuwenwen's avatar
zhuwenwen committed
217
218
219
  "csrc/opt/transpose_kernels.cu"
  "csrc/opt/activation_kernels_opt.cu"
  "csrc/attention/attention_kernels_opt.cu"
220
  "csrc/attention/attention_kernels_opt_tc.cu"
zhuwenwen's avatar
zhuwenwen committed
221
  "csrc/opt/layernorm_kernels_opt.cu"
222
  # "csrc/quantization/gptq/q_gemm.cu"
zhuwenwen's avatar
zhuwenwen committed
223
  "csrc/quantization/compressed_tensors/int8_quant_kernels.cu"
zhuwenwen's avatar
zhuwenwen committed
224
  # "csrc/quantization/fp8/common.cu"
bnellnm's avatar
bnellnm committed
225
226
  "csrc/cuda_utils_kernels.cu"
  "csrc/moe_align_block_size_kernels.cu"
227
  "csrc/prepare_inputs/advance_step.cu"
228
  "csrc/torch_bindings.cpp")
bnellnm's avatar
bnellnm committed
229
230

if(VLLM_GPU_LANG STREQUAL "CUDA")
231
  SET(CUTLASS_ENABLE_HEADERS_ONLY ON CACHE BOOL "Enable only the header library")
232
233
234
235

  # Set CUTLASS_REVISION manually -- its revision detection doesn't work in this case.
  set(CUTLASS_REVISION "v3.5.1" CACHE STRING "CUTLASS revision to use")

236
  FetchContent_Declare(
237
        cutlass
238
        GIT_REPOSITORY https://github.com/nvidia/cutlass.git
239
        GIT_TAG v3.5.1
240
        GIT_PROGRESS TRUE
241
242
243
244
245

        # 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
        GIT_SHALLOW TRUE
246
247
248
  )
  FetchContent_MakeAvailable(cutlass)

bnellnm's avatar
bnellnm committed
249
  list(APPEND VLLM_EXT_SRC
250
251
    "csrc/mamba/mamba_ssm/selective_scan_fwd.cu"
    "csrc/mamba/causal_conv1d/causal_conv1d.cu"
James Fleming's avatar
James Fleming committed
252
    "csrc/quantization/aqlm/gemm_kernels.cu"
bnellnm's avatar
bnellnm committed
253
    "csrc/quantization/awq/gemm_kernels.cu"
254
    "csrc/quantization/gguf/gguf_kernel.cu"
255
    "csrc/custom_all_reduce.cu"
256
    "csrc/permute_cols.cu"
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
    "csrc/quantization/cutlass_w8a8/scaled_mm_entry.cu")

  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.
  cuda_archs_loose_intersection(MARLIN_ARCHS "8.0;8.6;8.9;9.0" ${CUDA_ARCHS})
  if (MARLIN_ARCHS)
    set(MARLIN_SRCS 
       "csrc/quantization/fp8/fp8_marlin.cu"
       "csrc/quantization/marlin/dense/marlin_cuda_kernel.cu"
       "csrc/quantization/marlin/sparse/marlin_24_cuda_kernel.cu"
       "csrc/quantization/marlin/qqq/marlin_qqq_gemm_kernel.cu"
       "csrc/quantization/gptq_marlin/gptq_marlin.cu"
       "csrc/quantization/gptq_marlin/gptq_marlin_repack.cu"
       "csrc/quantization/gptq_marlin/awq_marlin_repack.cu")
    set_gencode_flags_for_srcs(
      SRCS "${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"
                   "in CUDA target architectures")
  endif()

  #
  # The cutlass_scaled_mm kernels for Hopper (c3x, i.e. CUTLASS 3.x) require
  # CUDA 12.0 or later (and only work on Hopper, 9.0/9.0a for now).
  cuda_archs_loose_intersection(SCALED_MM_3X_ARCHS "9.0;9.0a" "${CUDA_ARCHS}")
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.0 AND SCALED_MM_3X_ARCHS)
    set(SRCS "csrc/quantization/cutlass_w8a8/scaled_mm_c3x.cu")
    set_gencode_flags_for_srcs(
      SRCS "${SRCS}"
      CUDA_ARCHS "${SCALED_MM_3X_ARCHS}")
    list(APPEND VLLM_EXT_SRC "${SRCS}")
    list(APPEND VLLM_GPU_FLAGS "-DENABLE_SCALED_MM_C3X=1")
    message(STATUS "Building scaled_mm_c3x for archs: ${SCALED_MM_3X_ARCHS}")
  else()
    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.0 AND SCALED_MM_3X_ARCHS)
      message(STATUS "Not building scaled_mm_c3x as CUDA Compiler version is "
                     "not >= 12.0, we recommend upgrading to CUDA 12.0 or "
                     "later if you intend on running FP8 quantized models on "
                     "Hopper.")
    else()
      message(STATUS "Not building scaled_mm_c3x as no compatible archs found "
                     "in CUDA target architectures")
    endif()
308
309
310
311

    # clear SCALED_MM_3X_ARCHS so the scaled_mm_c2x kernels know we didn't 
    # build any 3x kernels
    set(SCALED_MM_3X_ARCHS)
312
  endif()
313
314

  #
315
316
317
  # 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.
  cuda_archs_loose_intersection(SCALED_MM_2X_ARCHS 
318
    "7.5;8.0;8.6;8.9;9.0" "${CUDA_ARCHS}")
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
  # 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()
337
  endif()
338

339

340
  #
341
  # Machete kernels
342

343
  # The machete kernels only work on hopper and require CUDA 12.0 or later.
344
345
346
  # 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)
347
348
349
350
    #
    # For the Machete kernels we automatically generate sources for various 
    # preselected input type pairs and schedules.
    # Generate sources:
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
    set(MACHETE_GEN_SCRIPT 
      ${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(
        COMMAND ${CMAKE_COMMAND} -E env 
        PYTHONPATH=${CMAKE_CURRENT_SOURCE_DIR}/csrc/cutlass_extensions/:${CUTLASS_DIR}/python/:${VLLM_PYTHON_PATH}:$PYTHONPATH 
          ${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."
                            " Result: \"${machete_generation_result}\"" 
                            "\nCheck the log for details: "
                            "${CMAKE_CURRENT_BINARY_DIR}/machete_generation.log")
      else()
        set(MACHETE_GEN_SCRIPT_HASH ${MACHETE_GEN_SCRIPT_HASH} 
            CACHE STRING "Last run machete generate script hash" FORCE)
        message(STATUS "Machete generation completed successfully.")
      endif()
380
    else()
381
      message(STATUS "Machete generation script has not changed, skipping generation.")
382
    endif()
383

384
385
386
    # Add machete generated sources
    file(GLOB MACHETE_GEN_SOURCES "csrc/quantization/machete/generated/*.cu")
    list(APPEND VLLM_EXT_SRC ${MACHETE_GEN_SOURCES})
387

388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
    # 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()
    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.0 
        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()
408
  endif()
409
# if CUDA endif
bnellnm's avatar
bnellnm committed
410
411
endif()

412
message(STATUS "Enabling C extension.")
bnellnm's avatar
bnellnm committed
413
414
415
416
417
418
419
define_gpu_extension_target(
  _C
  DESTINATION vllm
  LANGUAGE ${VLLM_GPU_LANG}
  SOURCES ${VLLM_EXT_SRC}
  COMPILE_FLAGS ${VLLM_GPU_FLAGS}
  ARCHITECTURES ${VLLM_GPU_ARCHES}
420
  INCLUDE_DIRECTORIES ${CUTLASS_INCLUDE_DIR}
421
  USE_SABI 3
bnellnm's avatar
bnellnm committed
422
423
  WITH_SOABI)

424
425
426
427
428
429
# If CUTLASS is compiled on NVCC >= 12.5, it by default uses 
# cudaGetDriverEntryPointByVersion as a wrapper to avoid directly calling the 
# 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
430
431
432
433
434
#
# _moe_C extension
#

set(VLLM_MOE_EXT_SRC
435
  "csrc/moe/torch_bindings.cpp"
bnellnm's avatar
bnellnm committed
436
437
  "csrc/moe/topk_softmax_kernels.cu")

438
439
440
441
set_gencode_flags_for_srcs(
  SRCS "${VLLM_MOE_EXT_SRC}"
  CUDA_ARCHS "${CUDA_ARCHS}")

442
if(VLLM_GPU_LANG STREQUAL "CUDA")
443
444
445
446
447
448
449
450
  cuda_archs_loose_intersection(MARLIN_MOE_ARCHS "8.0;8.6;8.9;9.0" "${CUDA_ARCHS}")
  if (MARLIN_MOE_ARCHS)
    set(MARLIN_MOE_SRC
        "csrc/moe/marlin_kernels/marlin_moe_kernel.h"
        "csrc/moe/marlin_kernels/marlin_moe_kernel_ku4b8.h"
        "csrc/moe/marlin_kernels/marlin_moe_kernel_ku4b8.cu"
        "csrc/moe/marlin_kernels/marlin_moe_kernel_ku8b128.h"
        "csrc/moe/marlin_kernels/marlin_moe_kernel_ku8b128.cu"
451
452
        "csrc/moe/marlin_kernels/marlin_moe_kernel_ku4.h"
        "csrc/moe/marlin_kernels/marlin_moe_kernel_ku4.cu"
453
454
455
456
457
458
459
460
461
462
463
464
        "csrc/moe/marlin_moe_ops.cu")

    set_gencode_flags_for_srcs(
      SRCS "${MARLIN_MOE_SRC}"
      CUDA_ARCHS "${MARLIN_MOE_ARCHS}")

    list(APPEND VLLM_MOE_EXT_SRC "${MARLIN_MOE_SRC}")
    message(STATUS "Building Marlin MOE kernels for archs: ${MARLIN_MOE_ARCHS}")
  else()
    message(STATUS "Not building Marlin MOE kernels as no compatible archs found"
                   "in CUDA target architectures")
  endif()
465
466
endif()

467
message(STATUS "Enabling moe extension.")
bnellnm's avatar
bnellnm committed
468
469
470
471
472
473
474
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}
475
  USE_SABI 3
bnellnm's avatar
bnellnm committed
476
477
  WITH_SOABI)

478
#[[  
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
if(VLLM_GPU_LANG STREQUAL "HIP")
  #
  # _rocm_C extension
  #
  set(VLLM_ROCM_EXT_SRC
    "csrc/rocm/torch_bindings.cpp"
    "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()
497
]]
bnellnm's avatar
bnellnm committed
498

499
500
501
502
# vllm-flash-attn currently only supported on CUDA
if (NOT VLLM_TARGET_DEVICE STREQUAL "cuda")
  return()
endif ()
bnellnm's avatar
bnellnm committed
503

504
505
506
507
508
509
510
511
512
513
514
# vLLM flash attention requires VLLM_GPU_ARCHES to contain the set of target  
# arches in the CMake syntax (75-real, 89-virtual, etc), since we clear the 
# arches in the CUDA case (and instead set the gencodes on a per file basis) 
# we need to manually set VLLM_GPU_ARCHES here.
if(VLLM_GPU_LANG STREQUAL "CUDA")
  foreach(_ARCH ${CUDA_ARCHS})
    string(REPLACE "." "" _ARCH "${_ARCH}")
    list(APPEND VLLM_GPU_ARCHES "${_ARCH}-real")
  endforeach()
endif()

515
516
517
518
519
520
521
522
523
524
#
# Build vLLM flash attention from source
#
# IMPORTANT: This has to be the last thing we do, because vllm-flash-attn uses the same macros/functions as vLLM.
# Because functions all belong to the global scope, vllm-flash-attn's functions overwrite vLLMs.
# They should be identical but if they aren't, this is a massive footgun.
#
# The vllm-flash-attn install rules are nested under vllm to make sure the library gets installed in the correct place.
# To only install vllm-flash-attn, use --component vllm_flash_attn_c.
# If no component is specified, vllm-flash-attn is still installed.
bnellnm's avatar
bnellnm committed
525

526
527
528
529
530
531
# If VLLM_FLASH_ATTN_SRC_DIR is set, vllm-flash-attn is installed from that directory instead of downloading.
# This is to enable local development of vllm-flash-attn within vLLM.
# It can be set as an environment variable or passed as a cmake argument.
# The environment variable takes precedence.
if (DEFINED ENV{VLLM_FLASH_ATTN_SRC_DIR})
  set(VLLM_FLASH_ATTN_SRC_DIR $ENV{VLLM_FLASH_ATTN_SRC_DIR})
532
endif()
533

534
535
if(VLLM_FLASH_ATTN_SRC_DIR)
  FetchContent_Declare(vllm-flash-attn SOURCE_DIR ${VLLM_FLASH_ATTN_SRC_DIR})
536
#[[ 
537
538
539
540
541
542
543
else()
  FetchContent_Declare(
          vllm-flash-attn
          GIT_REPOSITORY https://github.com/vllm-project/flash-attention.git
          GIT_TAG 013f0c4fc47e6574060879d9734c1df8c5c273bd
          GIT_PROGRESS TRUE
  )
544
]]
bnellnm's avatar
bnellnm committed
545
endif()
546
547
548
549

# Set the parent build flag so that the vllm-flash-attn library does not redo compile flag and arch initialization.
set(VLLM_PARENT_BUILD ON)

550
#[[ 
551
552
553
# Ensure the vllm/vllm_flash_attn directory exists before installation
install(CODE "file(MAKE_DIRECTORY \"\${CMAKE_INSTALL_PREFIX}/vllm/vllm_flash_attn\")" COMPONENT vllm_flash_attn_c)

554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
# Make sure vllm-flash-attn install rules are nested under vllm/
install(CODE "set(CMAKE_INSTALL_LOCAL_ONLY FALSE)" COMPONENT vllm_flash_attn_c)
install(CODE "set(OLD_CMAKE_INSTALL_PREFIX \"\${CMAKE_INSTALL_PREFIX}\")" COMPONENT vllm_flash_attn_c)
install(CODE "set(CMAKE_INSTALL_PREFIX \"\${CMAKE_INSTALL_PREFIX}/vllm/\")" COMPONENT vllm_flash_attn_c)

# Fetch the vllm-flash-attn library
FetchContent_MakeAvailable(vllm-flash-attn)
message(STATUS "vllm-flash-attn is available at ${vllm-flash-attn_SOURCE_DIR}")

# Restore the install prefix
install(CODE "set(CMAKE_INSTALL_PREFIX \"\${OLD_CMAKE_INSTALL_PREFIX}\")" COMPONENT vllm_flash_attn_c)
install(CODE "set(CMAKE_INSTALL_LOCAL_ONLY TRUE)" COMPONENT vllm_flash_attn_c)

# Copy over the vllm-flash-attn python files
install(
        DIRECTORY ${vllm-flash-attn_SOURCE_DIR}/vllm_flash_attn/
        DESTINATION vllm/vllm_flash_attn
        COMPONENT vllm_flash_attn_c
        FILES_MATCHING PATTERN "*.py"
)

# Nothing after vllm-flash-attn, see comment about macros above
576
]]