From 3c7886ec4cbef98f1908ca3e12f8323d0280f248 Mon Sep 17 00:00:00 2001 From: Chen1022 Date: Sun, 7 Dec 2025 14:03:34 +0800 Subject: [PATCH] Fix attention backend logic for Qwen3-Next on SM100 (#14560) --- python/sglang/srt/server_args.py | 15 +++++++++++++++ 1 file changed, 15 insertions(+) diff --git a/python/sglang/srt/server_args.py b/python/sglang/srt/server_args.py index 5a56c9484..6e3af4623 100644 --- a/python/sglang/srt/server_args.py +++ b/python/sglang/srt/server_args.py @@ -1282,6 +1282,21 @@ class ServerArgs: logger.info( "Use flashinfer_trtllm as MoE runner backend on sm100 for Qwen3NextForCausalLM" ) + if self.attention_backend is None: + self.attention_backend = "triton" + logger.info( + "Use triton as attention backend on sm100 for Qwen3NextForCausalLM" + ) + if ( + not self.disable_radix_cache + and self.attention_backend == "trtllm_mha" + ): + logger.warning( + "Disabling radix cache since trtllm_mha does not support page_size = 1, which is required by MambaRadixCache. " + "Try to use --attention-backend triton if radix cache is necessary." + ) + self.disable_radix_cache = True + self.disable_overlap_schedule = False elif model_arch in [ "NemotronHForCausalLM", "FalconH1ForCausalLM",