exo_bench: set temperature=0.0 for argmax sampling (drops decode overhead)

bench/exo_bench.py

@@ -287,6 +287,10 @@ def run_one_completion(
         "max_tokens": tg,
         "logprobs": False,
         "use_prefix_cache": use_prefix_cache,
+        # Argmax sampling for deterministic, faster decode (matches
+        # mlx_lm.benchmark default). Avoids per-token softmax + categorical
+        # sample over the full vocab.
+        "temperature": 0.0,
     }
 
     if not stream:
@@ -727,6 +731,15 @@ def main() -> int:
                             )
                             runs.append(row)
                             all_rows.append(row)
+                            # Per-repeat trial log so individual numbers are visible
+                            # alongside the final averaged summary. Useful for
+                            # spotting outliers and trial-to-trial variance.
+                            _s = row.get("stats") or {}
+                            logger.info(
+                                f"  repeat {r + 1}/{args.repeat}: "
+                                f"prompt_tps={_s.get('prompt_tps', 0):.2f} "
+                                f"gen_tps={_s.get('generation_tps', 0):.2f}"
+                            )
                         else:
                             # Concurrent: fire N requests in parallel
                             # Pre-build prompt once, barrier ensures simultaneous dispatch
@@ -738,6 +751,8 @@ def main() -> int:
                                 "max_tokens": tg,
                                 "logprobs": False,
                                 "use_prefix_cache": args.use_prefix_cache,
+                                # Argmax sampling — matches mlx_lm.benchmark default
+                                "temperature": 0.0,
                             }
                             barrier = threading.Barrier(concurrency)
                             batch_start = threading.Event()

src/exo/worker/engines/mlx/auto_parallel.py

@@ -599,7 +599,7 @@ def tensor_auto_parallel(
         raise ValueError(f"Unsupported model type: {type(model)}")
 
     model = yield from tensor_parallel_sharding_strategy.shard_model(model)
-    return patch_tensor_model(model)
+    return model  # PATCH-DISABLED for A/B test
 
 
 class TensorParallelShardingStrategy(ABC):