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

[Quantization] fix attention quantization of gpt_oss model (#27334)

Browse Source 
Signed-off-by: xuebwang-amd <xuebwang@amd.com>
This commit is contained in:
xuebwang-amd
2025-11-12 01:06:00 +08:00
committed by GitHub
parent 05576df85c
commit 5a1271d83a
3 changed files with 101 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

15
vllm/model_executor/layers/quantization/mxfp4.py
View File

@@ -190,14 +190,25 @@ class Mxfp4Config(QuantizationConfig):
fused_mapping=self.packed_modules_mapping,
):
return UnquantizedLinearMethod()
raise NotImplementedError("Mxfp4 linear layer is not implemented")
# TODO: Add support for MXFP4 Linear Method.
# MXFP4 LinearMethod is available in AMD-Quark, refer to that implementation
# if you are interested in enabling MXFP4 here.
logger.warning_once(
"MXFP4 linear layer is not implemented - falling back to "
"UnquantizedLinearMethod."
)
return UnquantizedLinearMethod()
elif isinstance(layer, FusedMoE):
if current_platform.is_xpu():
return IpexMxfp4MoEMethod(layer.moe_config)
else:
return Mxfp4MoEMethod(layer.moe_config)
elif isinstance(layer, Attention):
raise NotImplementedError("Mxfp4 attention layer is not implemented")
# TODO: Add support for MXFP4 Attention.
logger.warning_once(
"MXFP4 attention layer is not implemented. "
"Skipping quantization for this layer."
)
return None