ministral3 (#14251)

Signed-off-by: Xinyuan Tong <xinyuantong.cs@gmail.com>
Co-authored-by: Yueming Yuan <yy28@illinois.edu>

benchmark/mmmu/eval_utils.py

@@ -36,7 +36,8 @@ class EvalArgs:
     profile: bool = False
     profile_number: int = 5
     concurrency: int = 1
-    max_new_tokens: int = 30
+    max_new_tokens: Optional[int] = None
+    temperature: Optional[float] = None
     response_answer_regex: str = "(.*)"
     lora_path: Optional[str] = None
 
@@ -101,6 +102,12 @@ class EvalArgs:
             default=EvalArgs.max_new_tokens,
             help="Maximum number of new tokens to generate per sample.",
         )
+        parser.add_argument(
+            "--temperature",
+            type=float,
+            default=EvalArgs.temperature,
+            help="Sampling temperature for generation.",
+        )
         parser.add_argument(
             "--response-answer-regex",
             type=str,
@@ -241,19 +248,21 @@ def prepare_samples(eval_args: EvalArgs):
 
 
 def get_sampling_params(eval_args):
-    max_new_tokens = eval_args.max_new_tokens
-    temperature = 0.001
-
     extra_request_body = {}
     if eval_args.extra_request_body:
         extra_request_body = json.loads(eval_args.extra_request_body)
-
-    return {
-        "temperature": temperature,
-        "max_new_tokens": max_new_tokens,
+    sampling_params = {
         **extra_request_body,
     }
 
+    if eval_args.max_new_tokens is not None and eval_args.max_new_tokens > 0:
+        sampling_params.update({"max_completion_tokens": eval_args.max_new_tokens})
+
+    if eval_args.temperature is not None:
+        sampling_params.update({"temperature": eval_args.temperature})
+
+    return sampling_params
+
 
 # ----------- Process Multi-choice -------------
 def parse_multi_choice_response(response, all_choices, index2ans):