Add CI for gpt-oss model on hopper (#8851)

python/sglang/test/run_eval.py

@@ -65,9 +65,10 @@ def run_eval(args):
 
     sampler = ChatCompletionSampler(
         model=args.model,
-        max_tokens=2048,
+        max_tokens=getattr(args, "max_tokens", 2048),
         base_url=base_url,
         temperature=getattr(args, "temperature", 0.0),
+        reasoning_effort=getattr(args, "reasoning_effort", None),
     )
 
     # Run eval
@@ -120,7 +121,9 @@ if __name__ == "__main__":
     parser.add_argument("--eval-name", type=str, default="mmlu")
     parser.add_argument("--num-examples", type=int)
     parser.add_argument("--num-threads", type=int, default=512)
+    parser.add_argument("--max-tokens", type=int, default=2048)
     parser.add_argument("--temperature", type=float, default=0.0)
+    parser.add_argument("--reasoning-effort", type=str)
     args = parser.parse_args()
 
     run_eval(args)

python/sglang/test/simple_eval_common.py

@@ -91,6 +91,7 @@ class ChatCompletionSampler(SamplerBase):
         model: Optional[str] = None,
         system_message: Optional[str] = None,
         temperature: float = 0.0,
+        reasoning_effort: Optional[str] = None,
         max_tokens: int = 2048,
     ):
         self.client = OpenAI(base_url=base_url, http_client=LargerHttpxClient())
@@ -102,7 +103,11 @@ class ChatCompletionSampler(SamplerBase):
         self.system_message = system_message
         self.temperature = temperature
         self.max_tokens = max_tokens
+        self.reasoning_effort = reasoning_effort
         self.image_format = "url"
+        print(
+            f"ChatCompletionSampler initialized with {self.system_message=} {self.temperature=} {self.max_tokens=} {self.reasoning_effort=}"
+        )
 
     def _handle_image(
         self,
@@ -138,6 +143,7 @@ class ChatCompletionSampler(SamplerBase):
                     messages=message_list,
                     temperature=self.temperature,
                     max_tokens=self.max_tokens,
+                    reasoning_effort=self.reasoning_effort,
                 )
                 return response.choices[0].message.content
             # NOTE: BadRequestError is triggered once for MMMU, please uncomment if you are rerunning MMMU

python/sglang/test/simple_eval_gpqa.py

@@ -71,6 +71,8 @@ class GPQAEval(Eval):
                 )
             ]
             response_text = sampler(prompt_messages)
+            if response_text is None:
+                response_text = ""
             match = re.search(ANSWER_PATTERN_MULTICHOICE, response_text)
             extracted_answer = match.group(1) if match else None
             score = 1.0 if extracted_answer == correct_answer else 0.0