From 47e12e082e07e9804dad29b19b3ce44c9a6f4eff Mon Sep 17 00:00:00 2001 From: b8zhong Date: Thu, 23 Oct 2025 18:22:58 -0700 Subject: [PATCH] Enable Llama 4 + TRTLLM MHA (#12003) --- python/sglang/srt/server_args.py | 8 +++++++- 1 file changed, 7 insertions(+), 1 deletion(-) diff --git a/python/sglang/srt/server_args.py b/python/sglang/srt/server_args.py index 7935de6f6..bd14f4b33 100644 --- a/python/sglang/srt/server_args.py +++ b/python/sglang/srt/server_args.py @@ -966,7 +966,13 @@ class ServerArgs: "fa3", "aiter", "triton", - }, "fa3, aiter, or triton is required for Llama4 model" + "trtllm_mha", + }, "fa3, aiter, triton, or trtllm_mha is required for Llama4 model" + if is_sm100_supported() and self.attention_backend is None: + self.attention_backend = "trtllm_mha" + logger.warning( + "Use trtllm_mha as attention backend on sm100 for Llama4 model" + ) elif model_arch in [ "Gemma2ForCausalLM", "Gemma3ForCausalLM",