EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

fix the fp8 topk_config.correction_bias is none bug (#10040)

Browse Source
This commit is contained in:
Rain Jiang
2025-09-07 20:28:14 -07:00
committed by GitHub
parent bc5fc332f7
commit 7802586cab
1 changed files with 6 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
python/sglang/srt/layers/quantization/fp8.py
View File

@@ -1132,10 +1132,12 @@ class Fp8MoEMethod(FusedMoEMethodBase):
and topk_config.topk_group is not None and topk_config.topk_group is not None
), "Current trtllm_fp8_block_scale_moe kernel does not support these two arguments as None" ), "Current trtllm_fp8_block_scale_moe kernel does not support these two arguments as None"
if topk_config.correction_bias is None: correction_bias = (
correction_bias = topk_config.correction_bias.to(x.dtype) None
else: if topk_config.correction_bias is None
correction_bias = None else topk_config.correction_bias.to(x.dtype)
)
return trtllm_fp8_block_scale_moe( return trtllm_fp8_block_scale_moe(
routing_logits=router_logits.to(torch.float32), routing_logits=router_logits.to(torch.float32),
routing_bias=correction_bias, routing_bias=correction_bias,