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

vlm: Use fa3 as the default backend for qwen3 vl (#14634)

Browse Source
This commit is contained in:
Mick
2025-12-08 12:56:20 -08:00
committed by GitHub
parent 7bf16c6339
commit 8810152e88
1 changed files with 3 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/utils/common.py
+3 -2
View File
@@ -315,7 +315,6 @@ try:
except: except:
is_intel_amx_backend_available = False is_intel_amx_backend_available = False
try: try:
# move torch._C._cpu._is_amx_tile_supported() from cpu_has_amx_support # move torch._C._cpu._is_amx_tile_supported() from cpu_has_amx_support
# to support torch compile # to support torch compile
@@ -2707,14 +2706,16 @@ def is_fa3_default_architecture(hf_config):
if not isinstance(architectures, list) or not architectures: if not isinstance(architectures, list) or not architectures:
return False return False
default_archs = { default_archs = {
"Qwen2ForCausalLM",
"Llama4ForConditionalGeneration", "Llama4ForConditionalGeneration",
"LlamaForCausalLM", "LlamaForCausalLM",
"Olmo2ForCausalLM", "Olmo2ForCausalLM",
"Gemma2ForCausalLM", "Gemma2ForCausalLM",
"Gemma3ForConditionalGeneration", "Gemma3ForConditionalGeneration",
"Qwen2ForCausalLM",
"Qwen3ForCausalLM", "Qwen3ForCausalLM",
"Qwen3MoeForCausalLM", "Qwen3MoeForCausalLM",
"Qwen3VLForConditionalGeneration",
"Qwen3VLMoeForConditionalGeneration",
"Glm4MoeForCausalLM", "Glm4MoeForCausalLM",
"Glm4vForConditionalGeneration", "Glm4vForConditionalGeneration",
"Glm4vMoeForConditionalGeneration", "Glm4vMoeForConditionalGeneration",