From 0a64139c94c4d6bcd69d46f25b5818eb4ef92611 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?=E7=8E=8B=E9=B9=A4=E7=94=B7?= Date: Thu, 16 Jul 2026 14:40:01 +0800 Subject: [PATCH] Fix --moe-a2a-backend silently ignored for LongCat-2.0 (moe_topk missing from gate) (#30975) Co-authored-by: whn09 Co-authored-by: Claude Opus 4.8 --- python/sglang/srt/managers/scheduler.py | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/python/sglang/srt/managers/scheduler.py b/python/sglang/srt/managers/scheduler.py index ea11a15a3..02581abe0 100644 --- a/python/sglang/srt/managers/scheduler.py +++ b/python/sglang/srt/managers/scheduler.py @@ -729,12 +729,14 @@ class Scheduler( ) # Different MoE architectures expose the per-token expert count under - # different attribute names (e.g. Gemma4 uses ``top_k_experts``). + # different attribute names (e.g. Gemma4 uses ``top_k_experts``, + # LongCat-2.0 uses ``moe_topk``). moe_topk_attrs = ( "num_experts_per_tok", "num_experts_per_token", "top_k_experts", "moe_top_k", + "moe_topk", ) if any(hasattr(config_to_check, attr) for attr in moe_topk_attrs): initialize_moe_config(self.server_args)