biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Kernel] CUTLASS grouped gemm fp8 MoE kernel (#13972)

Browse Source 
Signed-off-by: ElizaWszola <eliza@neuralmagic.com>
Signed-off-by: ElizaWszola <ewszola@redhat.com>
Co-authored-by: Lucas Wilkinson <wilkinson.lucas@gmail.com>
This commit is contained in:
ElizaWszola
2025-03-27 01:54:44 +01:00
committed by GitHub
parent 7a6d45bc8a
commit 9239bf718e
22 changed files with 2317 additions and 15 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
vllm/model_executor/layers/quantization/utils/w8a8_utils.py
View File

@@ -50,6 +50,16 @@ def cutlass_block_fp8_supported() -> bool:
return ops.cutlass_scaled_mm_supports_block_fp8(capability)
def cutlass_group_gemm_supported() -> bool:
if not current_platform.is_cuda():
return False
capability_tuple = current_platform.get_device_capability()
capability = -1 if capability_tuple is None else capability_tuple.to_int()
return ops.cutlass_group_gemm_supported(capability)
CUTLASS_FP8_SUPPORTED = cutlass_fp8_supported()
CUTLASS_BLOCK_FP8_SUPPORTED = cutlass_block_fp8_supported()