Set the default attention backend for GLM-4.5v to fa3 (#9245)

84b30d9e · zifeitong · GitHub · ff0cf51c · 84b30d9e
Unverified Commit 84b30d9e authored Aug 17, 2025 by zifeitong Committed by GitHub Aug 17, 2025
Show whitespace changes
Inline Side-by-side

Showing with 1 addition and 0 deletions

python/sglang/srt/utils.py python/sglang/srt/utils.py +1 -0

No files found.
--- a/python/sglang/srt/utils.py
+++ b/python/sglang/srt/utils.py
@@ -2345,6 +2345,7 @@ def is_fa3_default_architecture(hf_config):
        "Qwen3ForCausalLM",
        "Qwen3MoeForCausalLM",
        "Glm4MoeForCausalLM",
+        "Glm4vMoeForConditionalGeneration",
        "Step3VLForConditionalGeneration",
    }
    return architectures[0] in default_archs