[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
@@ -15,6 +15,8 @@ from vllm.model_executor.layers.quantization.utils.mxfp4_utils import (
from vllm.platforms import current_platform
from vllm.triton_utils import tl, triton
from vllm.utils import cdiv
from vllm.utils.deep_gemm import (is_blackwell_deep_gemm_used,
per_token_group_cast_to_fp8)
@triton.jit
@@ -115,7 +117,10 @@ def _fp8_quantize(
assert not per_act_token
assert len(block_shape) == 2
_, block_k = block_shape[0], block_shape[1]
A, A_scale = per_token_group_quant_fp8(A, block_k)
if is_blackwell_deep_gemm_used():
A, A_scale = per_token_group_cast_to_fp8(A, block_k)
else:
A, A_scale = per_token_group_quant_fp8(A, block_k)
assert cdiv(A.size(-1), block_k) == A_scale.size(-1)
return A, A_scale