minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[hotfix] Fix deepep w4a8 bug (#12642)

Browse Source
This commit is contained in:
Baizhou Zhang
2025-11-04 13:55:59 -08:00
committed by GitHub
parent 211f4070e5
commit 42889acbd0
1 changed files with 3 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/layers/moe/token_dispatcher/deepep.py
+3 -3
View File
@@ -357,9 +357,9 @@ class _DeepEPDispatcherImplNormal(_DeepEPDispatcherImplBase):
): ):
topk_weights, topk_ids = topk_output.topk_weights, topk_output.topk_ids topk_weights, topk_ids = topk_output.topk_weights, topk_output.topk_ids
topk_ids = topk_ids.to(torch.int64) topk_ids = topk_ids.to(torch.int64)
if deep_gemm_wrapper.ENABLE_JIT_DEEPGEMM and not ( if (
get_moe_runner_backend().is_cutlass() deep_gemm_wrapper.ENABLE_JIT_DEEPGEMM
and self.quant_config.get_name() == "w4afp8" and not get_moe_runner_backend().is_cutlass()
): ):
# TODO hard code 128 block quant,use fp8 communication # TODO hard code 128 block quant,use fp8 communication
hidden_states = sglang_per_token_group_quant_fp8( hidden_states = sglang_per_token_group_quant_fp8(