[Feature] Integrate SM100 DeepGEMM support (#20087)

This commit is contained in:
Wentao Ye
2025-07-10 20:18:05 -07:00
committed by GitHub
parent 5b032352cc
commit e2de455c34
16 changed files with 397 additions and 114 deletions
@@ -6,10 +6,8 @@ import torch
from vllm.platforms import current_platform
from vllm.triton_utils import triton
from vllm.utils import direct_register_custom_op, has_deep_gemm
if has_deep_gemm():
import deep_gemm
from vllm.utils import direct_register_custom_op
from vllm.utils.deep_gemm import fp8_gemm_nt
logger = logging.getLogger(__name__)
@@ -57,7 +55,7 @@ def w8a8_block_fp8_matmul_deepgemm(
output_dtype)
# Deepgemm only supports output tensor type as bfloat16
assert C.dtype == torch.bfloat16
deep_gemm.gemm_fp8_fp8_bf16_nt((A, As), (B, Bs), C)
fp8_gemm_nt((A, As), (B, Bs), C)
return C