@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / 94144e726c
Fork 0

A high-throughput and memory-efficient inference and serving engine for LLMs

Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git

Synced Nov 16, 17:00 UTC

[CI/Build][Kernel] Update CUTLASS to 3.5.1 tag (#8043)

Tyler Michael Smith committed Sep 10, 2024 at 19:51 UTC 94144e726cfeeba0c1758751b7fd46a20b6bd3b4
1 file changed +13 -2
CMakeLists.txt
+13 -2
@@ -195,9 +195,13 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
195 FetchContent_Declare(
196 cutlass
197 GIT_REPOSITORY https://github.com/nvidia/cutlass.git
198 - # CUTLASS 3.5.1
199 - GIT_TAG 06b21349bcf6ddf6a1686a47a137ad1446579db9
198 + GIT_TAG v3.5.1
199 GIT_PROGRESS TRUE
200 +
201 + # Speed up CUTLASS download by retrieving only the specified GIT_TAG instead of the history.
202 + # Important: If GIT_SHALLOW is enabled then GIT_TAG works only with branch names and tags.
203 + # So if the GIT_TAG above is updated to a commit hash, GIT_SHALLOW must be set to FALSE
204 + GIT_SHALLOW TRUE
205 )
206 FetchContent_MakeAvailable(cutlass)
207
@@ -231,6 +235,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
235 "-gencode arch=compute_90a,code=sm_90a")
236 endif()
237
238 +
239 #
240 # Machete kernels
241
@@ -289,6 +294,12 @@ define_gpu_extension_target(
294 USE_SABI 3
295 WITH_SOABI)
296
297 +# If CUTLASS is compiled on NVCC >= 12.5, it by default uses
298 +# cudaGetDriverEntryPointByVersion as a wrapper to avoid directly calling the
299 +# driver API. This causes problems when linking with earlier versions of CUDA.
300 +# Setting this variable sidesteps the issue by calling the driver directly.
301 +target_compile_definitions(_C PRIVATE CUTLASS_ENABLE_DIRECT_CUDA_DRIVER_CALL=1)
302 +
303 #
304 # _moe_C extension
305 #