From fa9021b21f927ba0f4f3378747841891268b207b Mon Sep 17 00:00:00 2001 From: Xiaoyu Zhang <35585791+BBuf@users.noreply.github.com> Date: Mon, 1 Dec 2025 17:54:23 +0800 Subject: [PATCH] fix: Increase FlashInfer workspace size for Qwen3VL models (#14173) --- python/sglang/srt/layers/attention/flashinfer_backend.py | 4 ++++ 1 file changed, 4 insertions(+) diff --git a/python/sglang/srt/layers/attention/flashinfer_backend.py b/python/sglang/srt/layers/attention/flashinfer_backend.py index 302d88f7c..ecebcf76d 100644 --- a/python/sglang/srt/layers/attention/flashinfer_backend.py +++ b/python/sglang/srt/layers/attention/flashinfer_backend.py @@ -161,6 +161,10 @@ class FlashInferAttnBackend(AttentionBackend): "Qwen2ForCausalLM" in model_runner.model_config.hf_config.architectures or "Qwen3ForCausalLM" in model_runner.model_config.hf_config.architectures or "MiMoForCausalLM" in model_runner.model_config.hf_config.architectures + or "Qwen3VLForConditionalGeneration" + in model_runner.model_config.hf_config.architectures + or "Qwen3VLMoeForConditionalGeneration" + in model_runner.model_config.hf_config.architectures ): envs.SGLANG_FLASHINFER_WORKSPACE_SIZE.set(512 * 1024 * 1024)