EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

Support FA3 backend for gpt-oss (#9028)

Browse Source
This commit is contained in:
Ke Bao
2025-08-14 01:41:50 +08:00
committed by GitHub
parent 4a16a71c36
commit 0ff6d1fce1
4 changed files with 24 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
python/sglang/srt/server_args.py
View File

@@ -2106,10 +2106,10 @@ class ServerArgs:
if model_arch in ["GptOssForCausalLM"]:
if self.attention_backend is None:
self.attention_backend = "triton"
assert self.attention_backend in [
"triton",
"trtllm_mha",
], f"GptOssForCausalLM requires 'triton' or 'trtllm_mha' attention backend, but got {self.attention_backend}"
supported_backends = ["triton", "trtllm_mha", "fa3"]
assert (
self.attention_backend in supported_backends
), f"GptOssForCausalLM requires one of {supported_backends} attention backend, but got '{self.attention_backend}'"
quantization_config = getattr(hf_config, "quantization_config", None)
is_mxfp4_quant_format = (
quantization_config is not None