换用github jwjohns/unsloth-GRPO-qwen2.5 验证GRPO训练模型

src/qwen_notebook_clone.py

@@ -17,11 +17,11 @@ PatchFastRL("GRPO", FastLanguageModel)
 model_name = f"../models/pretrained/DeepSeek-R1-Distill-Qwen-1.5B"
 model, tokenizer = FastLanguageModel.from_pretrained(
     model_name=model_name,
-    max_seq_length=2048,
+    max_seq_length=1024, # 2028
     load_in_4bit=True,
     fast_inference=True,
     max_lora_rank=128,
-    gpu_memory_utilization=0.80,
+    gpu_memory_utilization=0.60,
 )
 
 model = FastLanguageModel.get_peft_model(
@@ -145,7 +145,7 @@ training_args = GRPOConfig(
     bf16 = is_bfloat16_supported(),
     fp16 = not is_bfloat16_supported(),
     per_device_train_batch_size = 1,
-    gradient_accumulation_steps = 2,  # Increased for better stability
+    gradient_accumulation_steps = 1,  # 2 Increased for better stability
     num_generations = 8,
     max_prompt_length = 256,
     max_completion_length = 200,