hf_grpo_train.py

# /// script
# dependencies = [
#   "torch>=2.2.0",
#   "transformers>=4.40.0",
#   "trl>=1.2.0",
#   "peft>=0.10.0",
#   "accelerate>=0.27.0",
#   "bitsandbytes>=0.43.0",
#   "datasets>=2.18.0",
#   "huggingface-hub>=0.22.0",
#   "trackio",
#   "pydantic>=2.0",
#   "numpy>=1.24",
#   "pandas>=2.0",
#   "matplotlib>=3.8",
#   "tqdm>=4.60",
#   "networkx>=3.0",
#   "scipy>=1.10",
#   "fastapi>=0.100",
#   "uvicorn>=0.22",
#   "httpx>=0.24",
#   "pyyaml>=6.0",
# ]
# ///
"""
PolyGuard GRPO Training Job — runs on Hugging Face Jobs cloud GPU.

This script:
  1. Clones the project from GitHub
  2. Installs the polyguard-rl package
  3. Loads the SFT adapter from Hub (or base model if no adapter)
  4. Generates GRPO rollout prompts from the PolyGuard environment
  5. Runs GRPO training with 4 grouped reward functions
  6. Pushes the final adapter to the Hugging Face Hub

Submit via CLI:
  hf jobs uv run \
    --flavor a10g-large \
    --timeout 4h \
    --secrets HF_TOKEN \
    "https://huggingface.co/TheJackBright/polyguard-training-scripts/resolve/main/hf_grpo_train.py"

Environment variables:
  HF_TOKEN          : HF write token (required, passed via --secrets)
  SFT_MODEL_ID      : SFT adapter on Hub (default: TheJackBright/polyguard-qwen-sft)
  HUB_MODEL_ID      : output GRPO model repo (default: TheJackBright/polyguard-qwen-grpo)
  N_EPISODES        : GRPO rollout episodes (default: 256)
  EPOCHS            : training epochs (default: 2)
  BATCH_SIZE        : per-device train batch size (default: 2)
  GRAD_ACCUM        : gradient accumulation steps (default: 8)
  MAX_PROMPT_LEN    : max prompt tokens (default: 512)
  MAX_COMPLETION_LEN: max completion tokens (default: 512)
  GROUP_SIZE        : GRPO group size (default: 4)
"""
from __future__ import annotations

import inspect
import json
import os
import subprocess
import sys
from pathlib import Path
from typing import Any, Callable, Dict, List, Optional

# ─── Config from environment ────────────────────────────────────────────────
GITHUB_REPO        = os.environ.get("GITHUB_REPO", "https://github.com/Vishwa-docs/Meta_PyTorch_Scalar_OpenEnv-Hackathon.git")
SFT_MODEL_ID       = os.environ.get("SFT_MODEL_ID", "TheJackBright/polyguard-qwen-sft")
BASE_MODEL_NAME    = os.environ.get("BASE_MODEL_NAME", "Qwen/Qwen2.5-1.5B-Instruct")
HUB_MODEL_ID       = os.environ.get("HUB_MODEL_ID", "TheJackBright/polyguard-qwen-grpo")
N_EPISODES         = int(os.environ.get("N_EPISODES", "256"))
EPOCHS             = int(os.environ.get("EPOCHS", "2"))
BATCH_SIZE         = int(os.environ.get("BATCH_SIZE", "2"))
GRAD_ACCUM         = int(os.environ.get("GRAD_ACCUM", "8"))
MAX_PROMPT_LEN     = int(os.environ.get("MAX_PROMPT_LEN", "512"))
MAX_COMPLETION_LEN = int(os.environ.get("MAX_COMPLETION_LEN", "512"))
GROUP_SIZE         = int(os.environ.get("GROUP_SIZE", "4"))
SEED               = 42
OUTPUT_DIR         = "/tmp/polyguard_grpo_output"
PROMPTS_PATH       = "/tmp/polyguard_grpo_prompts.jsonl"

print("=" * 60)
print("PolyGuard GRPO Training on HF Jobs")
print(f"  SFT checkpoint: {SFT_MODEL_ID}")
print(f"  Hub output:     {HUB_MODEL_ID}")
print(f"  Episodes:       {N_EPISODES}")
print(f"  Epochs:         {EPOCHS}")
print(f"  Group size:     {GROUP_SIZE}")
print("=" * 60)


# ─── Step 1: Clone repo and install polyguard-rl ────────────────────────────
print("\n[1/5] Cloning project from GitHub...")
clone_dir = Path("/tmp/polyguard_project")
if not clone_dir.exists():
    subprocess.run(
        ["git", "clone", "--depth=1", GITHUB_REPO, str(clone_dir)],
        check=True,
    )
else:
    print("  Already cloned.")

polyguard_rl_dir = clone_dir / "polyguard-rl"
print(f"\n[2/5] Installing polyguard-rl...")
subprocess.run(
    [sys.executable, "-m", "pip", "install", "-e", str(polyguard_rl_dir), "--quiet"],
    check=True,
)
if str(polyguard_rl_dir) not in sys.path:
    sys.path.insert(0, str(polyguard_rl_dir))
os.chdir(polyguard_rl_dir)


# ─── Step 2: Build GRPO prompt dataset ──────────────────────────────────────
print(f"\n[3/5] Building GRPO rollout dataset ({N_EPISODES} episodes)...")

from app.env.env_core import PolyGuardEnv  # noqa: E402
from app.models.policy.prompt_templates import build_planner_prompt  # noqa: E402
from app.training.openenv_wrapper import PolyGuardEnvWrapper  # noqa: E402

env = PolyGuardEnv()
prompts: List[Dict[str, Any]] = []

for ep_idx in range(N_EPISODES):
    seed_i = SEED + ep_idx
    obs, _ = env.reset(seed=seed_i)
    prompt_text = build_planner_prompt(obs)
    prompts.append({
        "prompt": prompt_text,
        "seed": seed_i,
        "episode": ep_idx,
        "difficulty": getattr(obs, "difficulty", "medium"),
    })

print(f"  Built {len(prompts)} GRPO prompts.")
with open(PROMPTS_PATH, "w") as f:
    for row in prompts:
        f.write(json.dumps(row) + "\n")


# ─── Step 3: Load model from SFT checkpoint ─────────────────────────────────
print(f"\n[4/5] Loading SFT model from {SFT_MODEL_ID}...")

import torch
from datasets import Dataset
from huggingface_hub import HfApi
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
from trl import GRPOConfig, GRPOTrainer

dtype = torch.bfloat16 if (torch.cuda.is_available() and torch.cuda.is_bf16_supported()) else torch.float16
print(f"  CUDA: {torch.cuda.is_available()}")
if torch.cuda.is_available():
    print(f"  GPU: {torch.cuda.get_device_name(0)}")

tokenizer = AutoTokenizer.from_pretrained(SFT_MODEL_ID, trust_remote_code=True)
if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.eos_token

# Try loading SFT adapter, fall back to base model
try:
    base_model = AutoModelForCausalLM.from_pretrained(
        BASE_MODEL_NAME,
        torch_dtype=dtype,
        device_map="auto",
        trust_remote_code=True,
    )
    model = PeftModel.from_pretrained(base_model, SFT_MODEL_ID)
    print("  Loaded SFT adapter on top of base model.")
except Exception as e:
    print(f"  SFT load failed ({e}), falling back to base model...")
    model = AutoModelForCausalLM.from_pretrained(
        BASE_MODEL_NAME,
        torch_dtype=dtype,
        device_map="auto",
        trust_remote_code=True,
    )


# ─── Step 4: Build grouped reward function ──────────────────────────────────
env_wrapper = PolyGuardEnvWrapper(score_timeout_s=30.0)

def build_polyguard_reward_fn(group_name: str) -> Callable:
    from app.training.grpo_rewards import make_group_reward_fn
    return make_group_reward_fn(env_wrapper=env_wrapper, group_name=group_name)

# TRL GRPO expects a single reward function that returns List[float]
# We combine the 4 reward groups into a weighted aggregate
from app.training.grpo_rewards import REWARD_GROUPS  # noqa: E402

_group_fns = {g: build_polyguard_reward_fn(g) for g in REWARD_GROUPS}

def combined_reward_fn(
    prompts: List[str],
    completions: List[str],
    **kwargs,
) -> List[float]:
    weights = {
        "format_schema": 0.20,
        "clinical_safety": 0.45,
        "process_grounding": 0.20,
        "anti_hack": 0.15,
    }
    n = len(prompts)
    totals = [0.0] * n
    for group, fn in _group_fns.items():
        group_rewards = fn(prompts=prompts, completions=completions, **kwargs)
        w = weights.get(group, 0.25)
        for i, r in enumerate(group_rewards):
            totals[i] += w * float(r)
    # Clamp to [0.001, 0.999]
    return [max(0.001, min(0.999, round(v, 3))) for v in totals]


# ─── Step 5: GRPO Training ──────────────────────────────────────────────────
ds = Dataset.from_list([{"prompt": p["prompt"]} for p in prompts]).shuffle(seed=SEED)

grpo_config_kwargs: Dict[str, Any] = {
    "output_dir": OUTPUT_DIR,
    "num_train_epochs": EPOCHS,
    "per_device_train_batch_size": BATCH_SIZE,
    "gradient_accumulation_steps": GRAD_ACCUM,
    "learning_rate": 1e-5,
    "bf16": dtype == torch.bfloat16,
    "fp16": dtype == torch.float16,
    "logging_steps": 5,
    "save_steps": 50,
    "save_total_limit": 2,
    "seed": SEED,
    "report_to": ["trackio"],
    "run_name": "polyguard-grpo-qwen",
    "project": "polyguard-training",
    "push_to_hub": True,
    "hub_model_id": HUB_MODEL_ID,
    "hub_strategy": "every_save",
    "num_generations": GROUP_SIZE,
    "max_prompt_length": MAX_PROMPT_LEN,
    "max_completion_length": MAX_COMPLETION_LEN,
    "temperature": 0.9,
    "beta": 0.1,
}

grpo_params = set(inspect.signature(GRPOConfig).parameters)
grpo_config = GRPOConfig(**{k: v for k, v in grpo_config_kwargs.items() if k in grpo_params})

trainer_kwargs: Dict[str, Any] = {
    "model": model,
    "args": grpo_config,
    "train_dataset": ds,
    "reward_funcs": combined_reward_fn,
}
trainer_params = set(inspect.signature(GRPOTrainer).parameters)
if "processing_class" in trainer_params:
    trainer_kwargs["processing_class"] = tokenizer
elif "tokenizer" in trainer_params:
    trainer_kwargs["tokenizer"] = tokenizer

trainer = GRPOTrainer(**{k: v for k, v in trainer_kwargs.items() if k in trainer_params})

print(f"\n  GRPO dataset size:  {len(ds)}")
print(f"  Group size:         {GROUP_SIZE}")
print(f"  Starting GRPO training...\n")

train_result = trainer.train()

print("\n  GRPO training complete.")
print(f"  Steps: {train_result.global_step}")

print(f"\n[5/5] Pushing GRPO model to Hub: {HUB_MODEL_ID}...")
trainer.push_to_hub()
tokenizer.push_to_hub(HUB_MODEL_ID)

print("\n" + "=" * 60)
print("GRPO training complete!")
print(f"Model saved to: https://huggingface.co/{HUB_MODEL_ID}")
print("=" * 60)