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 13, 08:15 UTC
[Bugfix] Adjust Marlin CUDA arch selection to 8.0+PTX;9.0+PTX (#28294)
Michael Goin committed
Nov 12, 2025 at 18:14 UTC
10f01d5a3a5c1570e36e15cd57eb3cc1534483d6
1 file changed
+2
-2
CMakeLists.txt
+2
-2
@@ -331,7 +331,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
331
# Keep building Marlin for 9.0 as there are some group sizes and shapes that
332
# are not supported by Machete yet.
333
# 9.0 for latest bf16 atomicAdd PTX
334
- cuda_archs_loose_intersection(MARLIN_ARCHS "8.0;8.7;9.0+PTX" "${CUDA_ARCHS}")
334
+ cuda_archs_loose_intersection(MARLIN_ARCHS "8.0+PTX;9.0+PTX" "${CUDA_ARCHS}")
335
if (MARLIN_ARCHS)
336
337
#
@@ -915,7 +915,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
915
916
list(APPEND VLLM_MOE_EXT_SRC "${VLLM_MOE_WNA16_SRC}")
917
# 9.0 for latest bf16 atomicAdd PTX
918
- cuda_archs_loose_intersection(MARLIN_MOE_ARCHS "8.0;8.7;9.0+PTX" "${CUDA_ARCHS}")
918
+ cuda_archs_loose_intersection(MARLIN_MOE_ARCHS "8.0+PTX;9.0+PTX" "${CUDA_ARCHS}")
919
if (MARLIN_MOE_ARCHS)
920
921
#