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

[Feature] Add Hopper DeepGEMM E8M0 for DeepSeekV3.1 scale_fmt (#23666)

Browse Source 
Signed-off-by: yewentao256 <zhyanwentao@126.com>
Signed-off-by: youkaichao <youkaichao@gmail.com>
Co-authored-by: youkaichao <youkaichao@gmail.com>
This commit is contained in:
Wentao Ye
2025-08-27 10:09:08 -04:00
committed by GitHub
parent 513c1fe255
commit 3af47c3cc6
10 changed files with 68 additions and 53 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
vllm/model_executor/layers/fused_moe/fused_moe.py
View File

@@ -40,7 +40,7 @@ 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 direct_register_custom_op, is_torch_equal_or_newer
from vllm.utils.deep_gemm import is_blackwell_deep_gemm_e8m0_used
from vllm.utils.deep_gemm import is_deep_gemm_e8m0_used
from .rocm_aiter_fused_moe import is_rocm_aiter_moe_enabled
@@ -1431,9 +1431,8 @@ def fused_experts(hidden_states: torch.Tensor,
# E8M0 scale, which means we requantize the weight and input to the specific
# scale. Fallen back to cutlass or triton for some cases would cause
# accuracy issue.
if (allow_deep_gemm and use_fp8_w8a8
and (is_blackwell_deep_gemm_e8m0_used()
or _valid_deep_gemm(hidden_states, w1, w2))):
if (allow_deep_gemm and use_fp8_w8a8 and
(is_deep_gemm_e8m0_used() or _valid_deep_gemm(hidden_states, w1, w2))):
assert apply_router_weight_on_input is False
assert is_act_and_mul, (
"DeepGemm only supports is_act_and_mul=True for now.")