Update grpo_train.py

grpo_train.py

@@ -302,7 +302,7 @@ model, tokenizer = FastLanguageModel.from_pretrained(
     model_name="unsloth/Llama-3.1-8B-Instruct",
     load_in_4bit=True, # Strictly True for L4 24GB
     max_seq_length=2048,
-    dtype=torch.bfloat16, # L4 Native Support
+    dtype=torch.float16, # PERFECT ALIGNMENT: 4-bit uses fp16 math natively
 )
 
 model = FastLanguageModel.get_peft_model(
@@ -339,8 +339,8 @@ trainer = GRPOTrainer(
         max_completion_length=128,
         logging_steps=5,
         warmup_ratio=0.1,
-        bf16=True,
-        fp16=False,
+        bf16=False, # DISABLED TO PREVENT CLASH
+        fp16=True,  # ENABLED TO MATCH MODEL DTYPE
         report_to="none",
     ),
     train_dataset=dataset,
@@ -371,7 +371,7 @@ if __name__ == "__main__":
     tokenizer.save_pretrained(LORA_DIR)
     print(f"LoRA adapter saved to {LORA_DIR}")
 
-    print("Merging adapter into base model (bf16)...")
+    print("Merging adapter into base model (fp16)...")
     merged_model, merged_tokenizer = FastLanguageModel.from_pretrained(
         model_name=LORA_DIR,
         load_in_4bit=False,