A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 05:30 UTC
[Bugfix] Fix failing GPT-OSS initialization test (#22557)
Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
Isotr0py committed
Aug 9, 2025 at 15:03 UTC
7920e9b1c5e168fe6218d2d147bdb9acf6bc993d
2 files changed
+6
-1
tests/models/registry.py
+1
-1
@@ -200,7 +200,7 @@ _TEXT_GENERATION_EXAMPLE_MODELS = {
200
{"6b": "EleutherAI/gpt-j-6b"}),
201
"GPTNeoXForCausalLM": _HfExamplesInfo("EleutherAI/pythia-70m",
202
{"1b": "EleutherAI/pythia-1.4b"}),
203
- "GptOssForCausalLM": _HfExamplesInfo("openai/gpt-oss-20b"),
203
+ "GptOssForCausalLM": _HfExamplesInfo("lmsys/gpt-oss-20b-bf16"),
204
"GraniteForCausalLM": _HfExamplesInfo("ibm/PowerLM-3b"),
205
"GraniteMoeForCausalLM": _HfExamplesInfo("ibm/PowerMoE-3b"),
206
"GraniteMoeHybridForCausalLM": _HfExamplesInfo("ibm-granite/granite-4.0-tiny-preview"), # noqa: E501
tests/models/test_initialization.py
+5
@@ -68,6 +68,11 @@ def can_initialize(model_arch: str, monkeypatch: pytest.MonkeyPatch,
68
if model_arch == "Phi4FlashForCausalLM":
69
# Phi4FlashForCausalLM only supports DIFFERENTIAL_FLASH_ATTN backend
70
m.setenv("VLLM_ATTENTION_BACKEND", "DIFFERENTIAL_FLASH_ATTN")
71
+ if model_arch == "GptOssForCausalLM":
72
+ # FIXME: A hack to bypass FA3 assertion because our CI's L4 GPU
73
+ # has cc==8.9 which hasn't supported FA3 yet. Remove this hack when
74
+ # L4 supports FA3.
75
+ m.setenv("VLLM_ATTENTION_BACKEND", "TRITON_ATTN_VLLM_V1")
76
LLM(
77
model_info.default,
78
tokenizer=model_info.tokenizer,