Skip to content

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

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merrymercy merged 9 commits into from
Dec 8, 2025
Merged

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

Changes from all commits
Commits
Show all changes
9 commits
Select commit Hold shift + click to select a range
6830aff
vlm: fix attention backend selection for causal part of VLMs
mickqian Dec 8, 2025
4fbf5b3
fix
mickqian Dec 8, 2025
f97d6c7
fix
mickqian Dec 8, 2025
2e07cf4
fix
mickqian Dec 8, 2025
afd45ee
fix
mickqian Dec 8, 2025
f8bfa52
fix
mickqian Dec 8, 2025
6a7a780
fix
mickqian Dec 8, 2025
fca49f7
fix
mickqian Dec 8, 2025
7087574
Refactor is_fa3_default_architecture function
merrymercy Dec 8, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
5 changes: 3 additions & 2 deletions python/sglang/srt/utils/common.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -315,7 +315,6 @@ def support_triton(backend: str) -> bool:
except:
is_intel_amx_backend_available = False


try:
# move torch._C._cpu._is_amx_tile_supported() from cpu_has_amx_support
# to support torch compile
Expand Down Expand Up @@ -2707,14 +2706,16 @@ def is_fa3_default_architecture(hf_config):
if not isinstance(architectures, list) or not architectures:
return False
default_archs = {
"Qwen2ForCausalLM",
"Llama4ForConditionalGeneration",
"LlamaForCausalLM",
"Olmo2ForCausalLM",
"Gemma2ForCausalLM",
"Gemma3ForConditionalGeneration",
"Qwen2ForCausalLM",
"Qwen3ForCausalLM",
"Qwen3MoeForCausalLM",
"Qwen3VLForConditionalGeneration",
"Qwen3VLMoeForConditionalGeneration",
"Glm4MoeForCausalLM",
"Glm4vForConditionalGeneration",
"Glm4vMoeForConditionalGeneration",
Expand Down
Loading