From 8e2142c15aacde4dd77c795acd3c3c222770e2c9 Mon Sep 17 00:00:00 2001 From: Ming Yang Date: Sun, 10 May 2026 21:57:43 -0700 Subject: [PATCH] [MoE] Fix NaN in flashinfer TRT-LLM A2A dispatch by sanitizing padding slots (#24850) --- python/sglang/srt/layers/moe/token_dispatcher/flashinfer.py | 1 + 1 file changed, 1 insertion(+) diff --git a/python/sglang/srt/layers/moe/token_dispatcher/flashinfer.py b/python/sglang/srt/layers/moe/token_dispatcher/flashinfer.py index 7b5080bb8..1c2710c24 100644 --- a/python/sglang/srt/layers/moe/token_dispatcher/flashinfer.py +++ b/python/sglang/srt/layers/moe/token_dispatcher/flashinfer.py @@ -219,6 +219,7 @@ class FlashinferDispatcher(BaseDispatcher): self.dummy_topk_ids_current_rank if self.has_dummy_token else topk_ids, payloads, self.runtime_max_tokens_per_rank, + invalid_token_expert_id=-1, expert_id_payload_index=expert_id_payload_index, ) if x_sf is not None: