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

[Bugfix] Fix gpt-oss w4a8 DP/EP on B200 (#26729)

Browse Source 
Signed-off-by: Varun Sundar Rabindranath <vsundarr@redhat.com>
Co-authored-by: Varun Sundar Rabindranath <vsundarr@redhat.com>
Co-authored-by: Michael Goin <mgoin64@gmail.com>
This commit is contained in:
Varun Sundar Rabindranath
2025-10-21 01:51:14 -04:00
committed by GitHub
parent f95da13c3d
commit 5ff5d94e77
5 changed files with 82 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

18
vllm/model_executor/layers/quantization/mxfp4.py
View File

@@ -18,6 +18,7 @@ from vllm.model_executor.layers.fused_moe import (
from vllm.model_executor.layers.fused_moe import modular_kernel as mk
from vllm.model_executor.layers.fused_moe.config import (
FusedMoEQuantConfig,
mxfp4_mxfp8_moe_quant_config,
mxfp4_w4a16_moe_quant_config,
ocp_mx_moe_quant_config,
)
@@ -747,6 +748,23 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
w1_scale=w1_scale,
w2_scale=w2_scale,
)
elif self.mxfp4_backend in [
Mxfp4Backend.SM100_FI_MXFP4_MXFP8_TRTLLM,
Mxfp4Backend.SM100_FI_MXFP4_MXFP8_CUTLASS,
]:
return mxfp4_mxfp8_moe_quant_config(
w1_bias=layer.w13_bias,
w2_bias=layer.w2_bias,
w1_scale=layer.w13_weight_scale,
w2_scale=layer.w2_weight_scale,
)
elif self.mxfp4_backend in [Mxfp4Backend.SM100_FI_MXFP4_BF16]:
return mxfp4_w4a16_moe_quant_config(
w1_bias=layer.w13_bias,
w2_bias=layer.w2_bias,
w1_scale=layer.w13_weight_scale,
w2_scale=layer.w2_weight_scale,
)
else:
w1_scale = layer.w13_weight_scale
w2_scale = layer.w2_weight_scale