updates

llms/mlx_lm/tuner/grpo_trainer.py

@@ -175,6 +175,8 @@ def generate_grpo(
     try:
         import time
 
+        model.freeze()
+
         start_time = time.time()
 
         if len(prompts.shape) == 1:
@@ -229,8 +231,9 @@ def generate_grpo(
                     # Check for EOS token
                     if token == tokenizer.eos_token_id:
                         break
-                    
+
                     current_tokens.append(token)
+                    mx.eval(current_tokens[-1])
 
                     # Check for end token
                     if len(current_tokens) >= len(end_sequence) and mx.array_equal(
@@ -268,7 +271,9 @@ def generate_grpo(
         )
         print(f"Average generation speed: {avg_tokens_per_second:.2f} tokens/sec")
 
+        results = [mx.stop_gradient(r) for r in results]
         mx.eval(results)
+        model.unfreeze()
         return results
 
     except Exception as e:
@@ -307,6 +312,7 @@ def grpo_loss(
     temperature: float = 0.8,
     reward_weights: Optional[List[float]] = None,
     batch_size: int = 1,
+    is_validation: bool = False
 ):
     prompt_tokens, _, prompt_text, answer_text = batch
     total_samples = len(prompt_tokens)
@@ -321,6 +327,7 @@ def grpo_loss(
 
     # Set model to eval mode for generation
     model.eval()
+    print(f"Is model now in training mode: {model.training}")
 
     # Process in smaller batches
     for i in range(0, total_samples, batch_size):
@@ -340,6 +347,7 @@ def grpo_loss(
         prompt_tensor = mx.array(padded_prompts)
 
         try:
+            mx.metal.clear_cache()
             completions = generate_grpo(
                 model,
                 prompt_tensor,
@@ -363,6 +371,7 @@ def grpo_loss(
                         mx.eval(completion_ids)
         except Exception as e:
             print(f"Generation error: {e}")
+            print(f"Is model in training mode after generation: {model.training}")
             continue
 
     # Restore original training state if we're not in validation mode
@@ -712,7 +721,8 @@ def evaluate_grpo(
             epsilon=epsilon,
             ref_model=ref_model,
             temperature=temperature,
-            max_tokens=max_tokens
+            max_tokens=max_tokens,
+            is_validation=True
         )
 
         all_losses += losses * toks