From 42889acbd0d0303158518cb947ae3911eff290e0 Mon Sep 17 00:00:00 2001 From: Baizhou Zhang Date: Tue, 4 Nov 2025 13:55:59 -0800 Subject: [PATCH] [hotfix] Fix deepep w4a8 bug (#12642) --- python/sglang/srt/layers/moe/token_dispatcher/deepep.py | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/python/sglang/srt/layers/moe/token_dispatcher/deepep.py b/python/sglang/srt/layers/moe/token_dispatcher/deepep.py index 26921a92e..12fccaacb 100644 --- a/python/sglang/srt/layers/moe/token_dispatcher/deepep.py +++ b/python/sglang/srt/layers/moe/token_dispatcher/deepep.py @@ -357,9 +357,9 @@ class _DeepEPDispatcherImplNormal(_DeepEPDispatcherImplBase): ): topk_weights, topk_ids = topk_output.topk_weights, topk_output.topk_ids topk_ids = topk_ids.to(torch.int64) - if deep_gemm_wrapper.ENABLE_JIT_DEEPGEMM and not ( - get_moe_runner_backend().is_cutlass() - and self.quant_config.get_name() == "w4afp8" + if ( + deep_gemm_wrapper.ENABLE_JIT_DEEPGEMM + and not get_moe_runner_backend().is_cutlass() ): # TODO hard code 128 block quant,use fp8 communication hidden_states = sglang_per_token_group_quant_fp8(