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

4
vllm/model_executor/layers/quantization/utils/fp8_utils.py
View File

@@ -20,7 +20,7 @@ from vllm.model_executor.layers.quantization.utils.w8a8_utils import (
from vllm.platforms import current_platform
from vllm.triton_utils import tl, triton
from vllm.utils import cdiv, direct_register_custom_op
from vllm.utils.deep_gemm import (is_blackwell_deep_gemm_e8m0_used,
from vllm.utils.deep_gemm import (is_deep_gemm_e8m0_used,
should_use_deepgemm_for_fp8_linear)
logger = init_logger(__name__)
@@ -385,7 +385,7 @@ def per_token_group_quant_fp8(
scaling factor.
"""
if use_ue8m0 is None:
use_ue8m0 = is_blackwell_deep_gemm_e8m0_used()
use_ue8m0 = is_deep_gemm_e8m0_used()
dtype = current_platform.fp8_dtype() if dtype is None else dtype
assert (x.shape[-1] % group_size == 0), (
f"the last dimension of `x` {x.shape[-1]} must be divisible "