Update MLX model patterns and reduce max_tokens in eval script

Added '-mlx-' to the list of MLX model patterns in should_use_mlx for broader matching. Reduced max_tokens from 32768 to 8192 in get_llm_response within eval_math500_benchmark.py to limit token usage.

Author: codelion

optillm/inference.py

@@ -189,7 +189,8 @@ def should_use_mlx(model_id: str) -> bool:
     # Models that should use MLX
     mlx_patterns = [
         "mlx-community/",
-        "mlx-"
+        "mlx-",
+        "-mlx-"
     ]
 
     # Known problematic models that should prefer MLX on Apple Silicon

scripts/eval_math500_benchmark.py

@@ -692,7 +692,7 @@ def get_llm_response(problem: str, model: str) -> str:
             messages=[
                 {"role": "user", "content": SYSTEM_PROMPT + "\n" + problem}
             ],
-            max_tokens=32768, # for thinking models, we need to use a lot more tokens
+            max_tokens=8192, # for thinking models, we need to use a lot more tokens
             # extra_body = {
             #     "decoding" : "thinkdeeper",
             # }