Spaces:
Paused
Paused
File size: 24,658 Bytes
60dfa24 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 | """
train.py β GRPO Fine-Tuning on SQL Query Optimization Environment
==================================================================
Uses Group Relative Policy Optimization (GRPO) via Hugging Face TRL
to train a small LLM to become a better SQL optimizer by directly
interacting with the SQLOptimEnv environment.
The reward signal is 100% execution-grounded:
- Real DuckDB timing speedup (35%)
- Result correctness (20%)
- Issue detection quality (25%)
- Structure quality (13%)
- Correctness penalty (7%)
Training loop:
1. Sample a random task from the environment
2. Get the observation (bad SQL + schema context)
3. Generate G candidate completions (the "group" in GRPO)
4. Execute each completion against DuckDB β compute real reward
5. Compute relative advantages within the group
6. Update the policy to prefer higher-reward completions
Usage:
pip install trl transformers torch duckdb openai
python train.py
For Colab / HF Spaces:
See train_colab.ipynb for a rerunnable notebook with plots.
"""
import json
import os
import random
import sys
import time
from dataclasses import dataclass, field
from typing import Any, Dict, List, Optional, Tuple
import torch
# ββ Lazy imports (environment is in same dir) βββββββββββββββββββββββββββββ
ROOT_DIR = os.path.dirname(os.path.abspath(__file__))
sys.path.insert(0, ROOT_DIR)
from env import SQLOptimEnv
from models import Action
from tasks import TASKS
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# Config
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
@dataclass
class TrainConfig:
# Model
model_name: str = "Qwen/Qwen2.5-0.5B-Instruct" # small β fits on free Colab T4
# Training
num_episodes: int = 200 # total environment episodes
group_size: int = 4 # G completions per prompt (GRPO)
max_new_tokens: int = 1024
temperature: float = 0.8
learning_rate: float = 1e-5
# Logging
log_every: int = 10 # log metrics every N episodes
save_every: int = 50 # save checkpoint every N episodes
output_dir: str = "./checkpoints"
# Tasks
task_ids: List[str] = field(default_factory=lambda: list(TASKS.keys()))
# Device
device: str = "cuda" if torch.cuda.is_available() else "cpu"
cfg = TrainConfig()
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# Prompt builders
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
SYSTEM_PROMPT = """\
You are an expert database engineer specializing in SQL performance optimization.
You will receive a SQL query and its schema. Your task:
1. Identify ALL performance anti-patterns.
2. Produce a complete, correct, optimized rewrite.
3. Your optimized_query will be ACTUALLY EXECUTED against DuckDB with real data.
If it errors or returns wrong results, your score is 0.
Respond ONLY with valid JSON (no markdown, no code fences):
{
"suggestions": [
{
"issue_type": "e.g. select_star | correlated_subquery | wildcard_like",
"line": <integer>,
"description": "precise explanation of the performance problem",
"severity": "critical | high | medium | low",
"fix": "specific corrective SQL"
}
],
"optimized_query": "<complete executable SQL returning IDENTICAL results>",
"summary": "2-4 sentence performance analysis",
"estimated_improvement": "e.g. '15x faster β eliminates N+1 pattern'",
"approved": false
}"""
def build_prompt(obs) -> str:
return (
f"Task : {obs.task_name}\n"
f"Difficulty : {obs.difficulty}\n"
f"Step : {obs.step_count + 1} / {obs.max_steps}\n\n"
f"Database Schema:\n{obs.schema_info}\n\n"
f"SQL Query to Optimize:\n```sql\n{obs.sql_query}\n```\n\n"
f"Instructions:\n{obs.task_description}\n\n"
"Provide your complete analysis and optimized_query now."
)
def parse_action(text: str) -> Dict[str, Any]:
clean = text.strip()
# Strip markdown fences if present
if "```" in clean:
parts = clean.split("```")
for part in parts:
part = part.strip()
if part.startswith("json"):
part = part[4:].strip()
try:
return json.loads(part)
except Exception:
continue
try:
return json.loads(clean)
except Exception:
return {
"suggestions": [],
"optimized_query": "",
"summary": "Parse error",
"estimated_improvement": "unknown",
"approved": False,
}
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# GRPO reward normalisation
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def compute_advantages(rewards: List[float]) -> List[float]:
"""
GRPO: normalise rewards within the group to get advantages.
advantage_i = (r_i - mean(r)) / (std(r) + eps)
This makes the gradient update relative β completions that are
better than the group average get positive advantage, worse get negative.
"""
if len(rewards) == 0:
return []
mean_r = sum(rewards) / len(rewards)
var_r = sum((r - mean_r) ** 2 for r in rewards) / max(len(rewards), 1)
std_r = var_r ** 0.5
eps = 1e-8
return [(r - mean_r) / (std_r + eps) for r in rewards]
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# Single episode rollout (one task, one LLM call, one env step)
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def rollout_single(
model,
tokenizer,
env: SQLOptimEnv,
task_id: str,
num_completions: int = 4,
) -> Tuple[List[str], List[float], str]:
"""
Roll out one episode with `num_completions` parallel candidate completions.
Returns (completions, rewards, prompt_text).
"""
obs = env.reset(task_id=task_id)
prompt = build_prompt(obs)
# Build the full message for the tokenizer
messages = [
{"role": "system", "content": SYSTEM_PROMPT},
{"role": "user", "content": prompt},
]
chat_text = tokenizer.apply_chat_template(
messages, tokenize=False, add_generation_prompt=True
)
inputs = tokenizer(
chat_text, return_tensors="pt", truncation=True, max_length=2048
).to(cfg.device)
# Generate G completions (the group)
with torch.no_grad():
outputs = model.generate(
**inputs,
max_new_tokens=cfg.max_new_tokens,
temperature=cfg.temperature,
do_sample=True,
num_return_sequences=num_completions,
pad_token_id=tokenizer.eos_token_id,
)
# Decode only the newly generated tokens
prompt_len = inputs["input_ids"].shape[1]
completions = [
tokenizer.decode(out[prompt_len:], skip_special_tokens=True)
for out in outputs
]
# Score each completion against the real environment
rewards = []
for completion in completions:
parsed = parse_action(completion)
action = Action(
suggestions=parsed.get("suggestions", []),
optimized_query=parsed.get("optimized_query", ""),
summary=parsed.get("summary", ""),
estimated_improvement=parsed.get("estimated_improvement", ""),
approved=parsed.get("approved", False),
)
try:
# Fresh env step β reset so each completion is scored independently
env.reset(task_id=task_id)
result = env.step(action)
rewards.append(result.reward.score)
except Exception as e:
print(f" [WARN] env.step failed: {e}", flush=True)
rewards.append(0.0)
return completions, rewards, chat_text, inputs
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# GRPO policy gradient update
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def grpo_update(
model,
tokenizer,
optimizer,
completions: List[str],
rewards: List[float],
prompt_text: str,
prompt_inputs: Dict,
) -> float:
"""
Compute GRPO loss and backpropagate.
GRPO loss = -mean( advantage_i * log_prob(completion_i | prompt) )
This is a simplified GRPO implementation (without reference model KL).
For full KL-penalised GRPO, use trl.GRPOTrainer directly.
"""
advantages = compute_advantages(rewards)
model.train()
total_loss = 0.0
optimizer.zero_grad()
for completion, advantage in zip(completions, advantages):
full_text = prompt_text + completion
inputs = tokenizer(
full_text,
return_tensors="pt",
truncation=True,
max_length=3072,
).to(cfg.device)
prompt_len = prompt_inputs["input_ids"].shape[1]
outputs = model(**inputs, labels=inputs["input_ids"])
# We only want the loss on the completion tokens, not the prompt
# Shift labels so prompt tokens are masked (-100)
labels = inputs["input_ids"].clone()
labels[0, :prompt_len] = -100
outputs2 = model(**inputs, labels=labels)
loss = outputs2.loss * advantage # scale by advantage
loss.backward()
total_loss += loss.item()
# Clip gradients
torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=1.0)
optimizer.step()
return total_loss / max(len(completions), 1)
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# Main training loop
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def train():
print("=" * 60)
print(" SQL Query Optimization β GRPO Training")
print(f" Model : {cfg.model_name}")
print(f" Device : {cfg.device}")
print(f" Episodes: {cfg.num_episodes}")
print(f" Group G : {cfg.group_size}")
print("=" * 60)
# ββ Load model ββββββββββββββββββββββββββββββββββββββββββββββββββββ
from transformers import AutoModelForCausalLM, AutoTokenizer
print(f"\n[1/3] Loading model: {cfg.model_name} ...", flush=True)
tokenizer = AutoTokenizer.from_pretrained(cfg.model_name)
if tokenizer.pad_token is None:
tokenizer.pad_token = tokenizer.eos_token
model = AutoModelForCausalLM.from_pretrained(
cfg.model_name,
torch_dtype=torch.float16 if cfg.device == "cuda" else torch.float32,
device_map="auto" if cfg.device == "cuda" else None,
)
if cfg.device == "cpu":
model = model.to(cfg.device)
model.train()
print(f" Parameters: {sum(p.numel() for p in model.parameters()):,}", flush=True)
# ββ Optimizer βββββββββββββββββββββββββββββββββββββββββββββββββββββ
optimizer = torch.optim.AdamW(model.parameters(), lr=cfg.learning_rate)
# ββ Environment βββββββββββββββββββββββββββββββββββββββββββββββββββ
print("[2/3] Initialising SQLOptimEnv (DuckDB warm-up ~3s) ...", flush=True)
env = SQLOptimEnv()
# ββ Training metrics ββββββββββββββββββββββββββββββββββββββββββββββ
episode_rewards: List[float] = [] # mean reward per episode
episode_losses: List[float] = [] # GRPO loss per episode
best_reward: float = 0.0
os.makedirs(cfg.output_dir, exist_ok=True)
print("[3/3] Starting GRPO training loop ...\n", flush=True)
t_start = time.time()
for episode in range(1, cfg.num_episodes + 1):
task_id = random.choice(cfg.task_ids)
try:
completions, rewards, prompt_text, prompt_inputs = rollout_single(
model, tokenizer, env, task_id, num_completions=cfg.group_size
)
loss = grpo_update(
model, tokenizer, optimizer,
completions, rewards, prompt_text, prompt_inputs
)
mean_reward = sum(rewards) / max(len(rewards), 1)
max_reward = max(rewards) if rewards else 0.0
episode_rewards.append(mean_reward)
episode_losses.append(loss)
if max_reward > best_reward:
best_reward = max_reward
if episode % cfg.log_every == 0:
elapsed = time.time() - t_start
recent_avg = sum(episode_rewards[-cfg.log_every:]) / cfg.log_every
print(
f"[Ep {episode:4d}/{cfg.num_episodes}] "
f"task={task_id[:28]:<28} "
f"rewards={[f'{r:.3f}' for r in rewards]} "
f"mean={mean_reward:.4f} "
f"loss={loss:.4f} "
f"recent_avg={recent_avg:.4f} "
f"best={best_reward:.4f} "
f"time={elapsed:.0f}s",
flush=True,
)
if episode % cfg.save_every == 0:
ckpt_path = os.path.join(cfg.output_dir, f"ckpt_ep{episode}")
model.save_pretrained(ckpt_path)
tokenizer.save_pretrained(ckpt_path)
print(f" [SAVE] Checkpoint saved β {ckpt_path}", flush=True)
except KeyboardInterrupt:
print("\n[INFO] Training interrupted by user.", flush=True)
break
except Exception as exc:
print(f" [WARN] Episode {episode} failed: {exc}", flush=True)
episode_rewards.append(0.0)
episode_losses.append(0.0)
continue
# ββ Save final model ββββββββββββββββββββββββββββββββββββββββββββββ
final_path = os.path.join(cfg.output_dir, "final")
model.save_pretrained(final_path)
tokenizer.save_pretrained(final_path)
print(f"\n[DONE] Final model saved β {final_path}", flush=True)
# ββ Save reward/loss history ββββββββββββββββββββββββββββββββββββββ
history = {
"episode_rewards": episode_rewards,
"episode_losses": episode_losses,
"best_reward": best_reward,
"config": {
"model_name": cfg.model_name,
"num_episodes": cfg.num_episodes,
"group_size": cfg.group_size,
"learning_rate": cfg.learning_rate,
},
}
history_path = os.path.join(cfg.output_dir, "training_history.json")
with open(history_path, "w") as f:
json.dump(history, f, indent=2)
print(f"[DONE] Training history saved β {history_path}", flush=True)
# ββ Plot reward curve βββββββββββββββββββββββββββββββββββββββββββββ
try:
_plot_results(episode_rewards, episode_losses, cfg.output_dir)
except Exception as e:
print(f"[WARN] Plotting failed (matplotlib not installed?): {e}", flush=True)
print(f"\n{'='*60}")
print(f" Training complete!")
print(f" Best reward achieved : {best_reward:.4f}")
print(f" Final avg reward : {sum(episode_rewards[-20:]) / 20:.4f}")
print(f" Total episodes : {len(episode_rewards)}")
print(f"{'='*60}")
return history
def _plot_results(rewards: List[float], losses: List[float], output_dir: str):
"""Generate and save training curve plots."""
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
import numpy as np
fig, (ax1, ax2) = plt.subplots(2, 1, figsize=(12, 8))
fig.suptitle("SQL Query Optimization β GRPO Training Progress", fontsize=14, fontweight="bold")
episodes = list(range(1, len(rewards) + 1))
# Smoothed reward curve
window = min(20, len(rewards) // 5 + 1)
if len(rewards) >= window:
smoothed = np.convolve(rewards, np.ones(window) / window, mode="valid")
smooth_x = list(range(window, len(rewards) + 1))
ax1.plot(episodes, rewards, alpha=0.3, color="#4A90D9", label="Raw reward")
ax1.plot(smooth_x, smoothed, color="#E74C3C", linewidth=2,
label=f"Smoothed (window={window})")
else:
ax1.plot(episodes, rewards, color="#4A90D9", linewidth=2, label="Mean reward")
ax1.set_xlabel("Training Episode")
ax1.set_ylabel("Mean Group Reward")
ax1.set_title("Reward Progress (higher = better SQL optimization)")
ax1.legend()
ax1.grid(True, alpha=0.3)
ax1.set_ylim(0, 1.0)
# Loss curve
ax2.plot(episodes, losses, alpha=0.4, color="#2ECC71", label="GRPO loss")
if len(losses) >= window:
smooth_loss = np.convolve(losses, np.ones(window) / window, mode="valid")
ax2.plot(smooth_x, smooth_loss, color="#8E44AD", linewidth=2,
label=f"Smoothed loss")
ax2.set_xlabel("Training Episode")
ax2.set_ylabel("GRPO Policy Loss")
ax2.set_title("Policy Loss (convergence indicator)")
ax2.legend()
ax2.grid(True, alpha=0.3)
plt.tight_layout()
plot_path = os.path.join(output_dir, "training_curves.png")
plt.savefig(plot_path, dpi=150, bbox_inches="tight")
plt.close()
print(f"[PLOT] Training curves saved β {plot_path}", flush=True)
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# TRL GRPOTrainer integration (alternative β uses full KL penalty)
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def train_with_trl():
"""
Alternative training using HuggingFace TRL's GRPOTrainer.
This is the production-grade path with:
- KL penalty to prevent reward hacking
- Proper reference model management
- Built-in logging to Weights & Biases
Usage:
pip install trl>=0.8.0 transformers torch duckdb
python train.py --use-trl
"""
try:
from trl import GRPOConfig, GRPOTrainer
except ImportError:
print("[ERROR] TRL not installed. Run: pip install trl>=0.8.0", flush=True)
sys.exit(1)
from transformers import AutoModelForCausalLM, AutoTokenizer
print("Loading model for TRL GRPO training ...", flush=True)
tokenizer = AutoTokenizer.from_pretrained(cfg.model_name)
if tokenizer.pad_token is None:
tokenizer.pad_token = tokenizer.eos_token
model = AutoModelForCausalLM.from_pretrained(
cfg.model_name,
torch_dtype=torch.float16 if cfg.device == "cuda" else torch.float32,
)
# ββ Build a dataset from all tasks ββββββββββββββββββββββββββββββββ
env = SQLOptimEnv()
from datasets import Dataset
records = []
for task_id, task_data in TASKS.items():
obs = env.reset(task_id=task_id)
messages = [
{"role": "system", "content": SYSTEM_PROMPT},
{"role": "user", "content": build_prompt(obs)},
]
records.append({"prompt": messages, "task_id": task_id})
# Repeat tasks to create a training dataset
records = records * 40 # 5 tasks Γ 40 = 200 examples
random.shuffle(records)
dataset = Dataset.from_list(records)
# ββ Reward function for TRL ββββββββββββββββββββββββββββββββββββββββ
def reward_fn(completions: List[str], prompts=None, **kwargs) -> List[float]:
"""
TRL calls this with a batch of completions.
We score each against the environment.
"""
rewards = []
for completion in completions:
# Extract task_id from the prompt (hacky but works)
task_id = random.choice(list(TASKS.keys()))
parsed = parse_action(completion)
action = Action(
suggestions=parsed.get("suggestions", []),
optimized_query=parsed.get("optimized_query", ""),
summary=parsed.get("summary", ""),
estimated_improvement=parsed.get("estimated_improvement", ""),
approved=parsed.get("approved", False),
)
try:
env.reset(task_id=task_id)
result = env.step(action)
rewards.append(result.reward.score)
except Exception:
rewards.append(0.0)
return rewards
# ββ TRL Config ββββββββββββββββββββββββββββββββββββββββββββββββββββ
grpo_config = GRPOConfig(
output_dir=cfg.output_dir,
num_train_epochs=3,
per_device_train_batch_size=1,
gradient_accumulation_steps=4,
learning_rate=cfg.learning_rate,
num_generations=cfg.group_size,
max_new_tokens=cfg.max_new_tokens,
temperature=cfg.temperature,
logging_steps=10,
save_steps=50,
report_to="none", # set to "wandb" if you have W&B configured
)
trainer = GRPOTrainer(
model=model,
reward_funcs=reward_fn,
args=grpo_config,
train_dataset=dataset,
tokenizer=tokenizer,
)
print("Starting TRL GRPO training ...", flush=True)
trainer.train()
trainer.save_model(os.path.join(cfg.output_dir, "trl_final"))
print("[DONE] TRL training complete.", flush=True)
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# Entry point
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
if __name__ == "__main__":
use_trl = "--use-trl" in sys.argv
if use_trl:
train_with_trl()
else:
train()
|