GRPO Training Test: Qwen3-4B¶

# Environment Setup
import os

# FIX: Set ACCELERATE_MIXED_PRECISION BEFORE importing unsloth
# This ensures autocast dtype matches model dtype (bfloat16)
os.environ['ACCELERATE_MIXED_PRECISION'] = 'bf16'

from dotenv import load_dotenv
load_dotenv()

# Force text-based progress instead of HTML widgets
os.environ["TQDM_NOTEBOOK"] = "false"

# CRITICAL: Import unsloth FIRST for proper TRL patching
import unsloth
from unsloth import FastLanguageModel, is_bf16_supported

import torch
from trl import GRPOConfig, GRPOTrainer
from datasets import Dataset

# Environment summary
gpu = torch.cuda.get_device_name(0) if torch.cuda.is_available() else "CPU"
print(f"Environment: unsloth {unsloth.__version__}, PyTorch {torch.__version__}, {gpu}")
print(f"ACCELERATE_MIXED_PRECISION: {os.environ.get('ACCELERATE_MIXED_PRECISION', 'not set')}")
print(f"HF_TOKEN loaded: {'Yes' if os.environ.get('HF_TOKEN') else 'No'}")

# Load Qwen3-4B with 4-bit quantization
MODEL_NAME = "unsloth/Qwen3-4B-unsloth-bnb-4bit"
print(f"\nLoading {MODEL_NAME.split('/')[-1]}...")

model, tokenizer = FastLanguageModel.from_pretrained(
    MODEL_NAME,
    max_seq_length=512,
    load_in_4bit=True,
    dtype=None,  # Auto-detect - let Unsloth manage dtype
)

# Ensure pad token is set
if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.eos_token
    tokenizer.pad_token_id = tokenizer.eos_token_id

print(f"Model loaded: {type(model).__name__}")

# Apply LoRA adapters for GRPO training
model = FastLanguageModel.get_peft_model(
    model,
    r=16,
    lora_alpha=16,
    lora_dropout=0,
    target_modules=["q_proj", "k_proj", "v_proj", "o_proj",
                    "gate_proj", "up_proj", "down_proj"],
    bias="none",
    use_gradient_checkpointing="unsloth",  # Use Unsloth-optimized checkpointing
    random_state=42,
)

trainable = sum(p.numel() for p in model.parameters() if p.requires_grad)
total = sum(p.numel() for p in model.parameters())
print(f"LoRA applied: {trainable:,} trainable / {total:,} total ({100*trainable/total:.2f}%)")

# Create minimal synthetic prompt dataset for GRPO (5 prompts)
# GRPO requires prompts only - completions are generated during training

prompts = [
    "Explain the concept of recursion in programming.",
    "What are the benefits of using version control?",
    "Describe how a hash table works.",
    "What is the difference between a stack and a queue?",
    "Explain what an API is to a beginner.",
]

# Format prompts for GRPO (requires "prompt" field)
dataset = Dataset.from_dict({
    "prompt": [
        tokenizer.apply_chat_template(
            [{"role": "user", "content": p}],
            tokenize=False,
            add_generation_prompt=True
        ) for p in prompts
    ]
})

print(f"Dataset created: {len(dataset)} prompts")
print(f"Sample prompt:\n{dataset[0]['prompt'][:150]}...")

# Define a simple length-based reward function for testing
# In production, this would be a learned reward model or rule-based system

def simple_reward_fn(completions, prompts=None, **kwargs):
    """
    Simple reward function for testing GRPO.
    Rewards longer, more informative responses.
    """
    rewards = []
    for completion in completions:
        # Basic heuristics for testing:
        # - Reward longer responses (up to a point)
        # - Penalize very short responses
        length = len(completion.split())
        
        if length < 5:
            reward = -1.0  # Too short
        elif length < 20:
            reward = 0.5   # Acceptable
        elif length < 50:
            reward = 1.0   # Good length
        else:
            reward = 0.8   # Slightly penalize very long
        
        rewards.append(reward)
    
    return rewards

print("Reward function defined: simple_reward_fn")
print("Rewards based on response length (testing only)")

# Post-training inference test
FastLanguageModel.for_inference(model)

test_prompt = "Explain what machine learning is in simple terms."
messages = [{"role": "user", "content": test_prompt}]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")

with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=64,
        temperature=0.7,
        top_p=0.9,
        do_sample=True,
        pad_token_id=tokenizer.pad_token_id,
    )

response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print("=" * 60)
print("GRPO Training Pipeline Test PASSED")
print("=" * 60)
print(f"Sample generation:\n{response[-200:]}")