minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

fix: Increase FlashInfer workspace size for Qwen3VL models (#14173)

Browse Source
This commit is contained in:
Xiaoyu Zhang
2025-12-01 17:54:23 +08:00
committed by GitHub
parent 9c80072845
commit fa9021b21f
1 changed files with 4 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/layers/attention/flashinfer_backend.py
+4
View File
@@ -161,6 +161,10 @@ class FlashInferAttnBackend(AttentionBackend):
"Qwen2ForCausalLM" in model_runner.model_config.hf_config.architectures "Qwen2ForCausalLM" in model_runner.model_config.hf_config.architectures
or "Qwen3ForCausalLM" 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 "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) envs.SGLANG_FLASHINFER_WORKSPACE_SIZE.set(512 * 1024 * 1024)