File size: 13,969 Bytes
367dc36 |
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 |
# ---
# jupyter:
# jupytext:
# text_representation:
# extension: .py
# format_name: percent
# kernelspec:
# display_name: Python 3
# language: python
# name: python3
# ---
# %% [markdown]
# # Training LLMs to Write Fast GPU Kernels with GRPO
#
# This notebook demonstrates how to train a language model to write optimized CUDA/Triton
# kernels using TRL's GRPOTrainer and the kernrl OpenEnv environment.
#
# **What is kernrl?**
# - An RL environment for GPU kernel optimization
# - Agents receive PyTorch reference implementations
# - Must write faster CUDA/Triton kernels that produce correct outputs
# - Rewards based on compilation success, correctness, and speedup
#
# **What is GRPO?**
# - Group Relative Policy Optimization
# - Efficient RL algorithm for training LLMs
# - Uses multiple generations per prompt to estimate advantages
# - Works well with environment-based reward signals
# %% [markdown]
# ## Installation
#
# First, install the required packages:
# %%
# !pip install torch triton trl transformers accelerate
# !pip install git+https://github.com/meta-pytorch/OpenEnv.git
# %% [markdown]
# ## Setup
#
# Import necessary libraries and configure the environment.
# %%
import torch
from datasets import Dataset
from transformers import AutoTokenizer
from trl import GRPOConfig, GRPOTrainer
from trl.experimental.openenv import generate_rollout_completions
# Import kernrl environment
from kernrl import kernrl_env, KernelAction, KernelObservation
# %%
# Configuration
MODEL_ID = "Qwen/Qwen2.5-Coder-1.5B-Instruct" # Good for code generation
ENV_URL = "http://localhost:8000" # kernrl server URL
# Initialize tokenizer
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
if tokenizer.pad_token is None:
tokenizer.pad_token = tokenizer.eos_token
# %% [markdown]
# ## Connect to kernrl Environment
#
# The kernrl environment evaluates submitted kernels for:
# 1. **Compilation**: Does the code compile?
# 2. **Correctness**: Does output match reference (within tolerance)?
# 3. **Performance**: Is it faster than PyTorch baseline?
# %%
# Connect to the kernrl server
# Option 1: Connect to running server
env = kernrl_env(base_url=ENV_URL)
# Option 2: Load from HuggingFace Hub (requires GPU)
# env = kernrl_env.from_hub("Infatoshi/kernrl")
# Option 3: Local Docker
# env = kernrl_env.from_docker_image("kernrl:latest")
# Test the connection
obs = env.reset(problem_id="L1_23_Softmax")
print(f"Problem: {obs.problem_id}")
print(f"GPU: {obs.gpu_info}")
print(f"Max turns: {obs.max_turns}")
# %% [markdown]
# ## Reward Functions
#
# We define multiple reward signals to guide the model:
# - **Compilation reward**: +0.1 for successful compilation
# - **Correctness reward**: +0.3 for matching reference output
# - **Speedup reward**: Scaled reward for beating baseline performance
# %%
import math
def reward_compilation(completions: list[str], **kwargs) -> list[float]:
"""Reward for successful compilation."""
compilation_success = kwargs.get("compilation_success", [])
return [0.1 if success else 0.0 for success in compilation_success]
def reward_correctness(completions: list[str], **kwargs) -> list[float]:
"""Reward for correct output."""
correctness_pass = kwargs.get("correctness_pass", [])
return [0.3 if correct else 0.0 for correct in correctness_pass]
def reward_speedup(completions: list[str], **kwargs) -> list[float]:
"""Reward scaled by speedup achieved."""
speedups = kwargs.get("speedup", [])
rewards = []
for speedup in speedups:
if speedup is None or speedup <= 0:
rewards.append(0.0)
elif speedup <= 1.0:
# Below baseline: small penalty
rewards.append(-0.1)
else:
# Above baseline: reward scales with log2(speedup)
# 2x speedup = 0.3, 4x = 0.6, 8x = 0.9
bonus = min(0.3 * math.log2(speedup), 0.6)
rewards.append(0.3 + bonus)
return rewards
def reward_combined(completions: list[str], **kwargs) -> list[float]:
"""Combined reward from all signals."""
comp_rewards = reward_compilation(completions, **kwargs)
corr_rewards = reward_correctness(completions, **kwargs)
speed_rewards = reward_speedup(completions, **kwargs)
return [c + r + s for c, r, s in zip(comp_rewards, corr_rewards, speed_rewards)]
# %% [markdown]
# ## System Prompt
#
# The system prompt provides context about the task and expected output format.
# %%
SYSTEM_PROMPT = """You are an expert GPU kernel engineer specializing in CUDA and Triton.
Your task is to optimize PyTorch operations by writing custom GPU kernels.
Guidelines:
1. Analyze the reference PyTorch implementation carefully
2. Identify optimization opportunities (memory access patterns, parallelism, fusion)
3. Write a Triton or CUDA kernel that computes the same result
4. Ensure numerical correctness (outputs must match within tolerance)
Output format:
- Provide a complete Python file
- Include a Model class with the same interface as the reference
- The Model.forward() method should use your optimized kernel
- Include all necessary imports (torch, triton, triton.language)
Focus on:
- Coalesced memory access
- Efficient use of shared memory
- Minimizing thread divergence
- Optimal block/grid dimensions"""
# %% [markdown]
# ## Rollout Function
#
# The rollout function generates kernel code and evaluates it in the environment.
# %%
def make_prompt(problem_description: str, feedback: str = "") -> str:
"""Create the user prompt for the model."""
prompt = f"{problem_description}\n"
if feedback:
prompt += f"\n## Previous Attempt Feedback\n{feedback}\n"
prompt += "\nProvide your optimized kernel implementation:"
return prompt
def extract_code(completion: str) -> str:
"""Extract code from model completion."""
# Handle markdown code blocks
if "```python" in completion:
start = completion.find("```python") + 9
end = completion.find("```", start)
if end > start:
return completion[start:end].strip()
if "```" in completion:
start = completion.find("```") + 3
end = completion.find("```", start)
if end > start:
return completion[start:end].strip()
# Return as-is if no code blocks
return completion.strip()
def rollout_func(prompts: list[str], trainer: GRPOTrainer) -> dict[str, list]:
"""
Custom rollout function for kernrl environment.
Generates kernel code and evaluates it to get rewards.
"""
# Generate completions
outputs = generate_rollout_completions(trainer, prompts)
completions_text = [
tokenizer.decode(out["completion_ids"], skip_special_tokens=True)
for out in outputs
]
# Evaluate each completion in the environment
compilation_success = []
correctness_pass = []
speedups = []
for completion in completions_text:
# Reset environment for each evaluation
obs = env.reset()
# Extract code and submit
code = extract_code(completion)
action = KernelAction(code=code)
try:
result = env.step(action)
obs = result.observation
compilation_success.append(obs.compilation_success)
correctness_pass.append(obs.correctness_pass or False)
speedups.append(obs.speedup)
except Exception as e:
print(f"Evaluation error: {e}")
compilation_success.append(False)
correctness_pass.append(False)
speedups.append(None)
return {
"prompt_ids": [out["prompt_ids"] for out in outputs],
"completion_ids": [out["completion_ids"] for out in outputs],
"logprobs": [out["logprobs"] for out in outputs],
# Pass reward signals to reward functions
"compilation_success": compilation_success,
"correctness_pass": correctness_pass,
"speedup": speedups,
}
# %% [markdown]
# ## Create Training Dataset
#
# We create a dataset from kernrl problems. Each problem becomes a training prompt.
# %%
def create_dataset(env: kernrl_env, levels: list[int] = [1, 2]) -> Dataset:
"""Create training dataset from kernrl problems."""
prompts = []
problem_ids = []
# Get all problem IDs
all_problems = env.list_problems()
for problem_id in all_problems:
# Filter by level
level = int(problem_id.split("_")[0][1:]) # Extract level from "L1_..."
if level not in levels:
continue
# Reset to get problem description
obs = env.reset(problem_id=problem_id)
# Create prompt
messages = [
{"role": "system", "content": SYSTEM_PROMPT},
{"role": "user", "content": make_prompt(obs.problem_description)},
]
prompt = tokenizer.apply_chat_template(
messages,
add_generation_prompt=True,
tokenize=False,
)
prompts.append(prompt)
problem_ids.append(problem_id)
return Dataset.from_dict({
"prompt": prompts,
"problem_id": problem_ids,
})
# Create dataset from Level 1 and 2 problems
dataset = create_dataset(env, levels=[1, 2])
print(f"Created dataset with {len(dataset)} problems")
# %% [markdown]
# ## Configure Training
#
# Set up GRPOTrainer with our custom rollout function and reward signals.
# %%
# Training configuration
config = GRPOConfig(
output_dir="./kernrl_grpo_output",
# vLLM settings
use_vllm=True,
vllm_mode="colocate", # Use "server" mode for multi-GPU
# Generation settings
num_generations=4, # Generations per prompt
max_completion_length=2048, # Kernel code can be long
temperature=0.7,
# Training settings
num_train_epochs=3,
per_device_train_batch_size=2,
gradient_accumulation_steps=4,
learning_rate=1e-5,
# Logging
logging_steps=10,
save_steps=100,
report_to="wandb", # Optional: log to Weights & Biases
)
# %% [markdown]
# ## Initialize Trainer
# %%
trainer = GRPOTrainer(
model=MODEL_ID,
processing_class=tokenizer,
reward_funcs=[
reward_compilation,
reward_correctness,
reward_speedup,
],
train_dataset=dataset,
rollout_func=rollout_func,
args=config,
)
# %% [markdown]
# ## Train!
#
# Start the training loop. The model will learn to write faster kernels through
# environment feedback.
# %%
# Start training
trainer.train()
# Save the final model
trainer.save_model("./kernrl_trained_model")
# %% [markdown]
# ## Evaluate the Trained Model
#
# Test the trained model on some problems to see how well it learned.
# %%
def evaluate_model(model_path: str, problem_ids: list[str]) -> dict:
"""Evaluate a trained model on kernel optimization problems."""
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained(model_path)
model.eval()
results = []
for problem_id in problem_ids:
obs = env.reset(problem_id=problem_id)
# Generate kernel code
messages = [
{"role": "system", "content": SYSTEM_PROMPT},
{"role": "user", "content": make_prompt(obs.problem_description)},
]
prompt = tokenizer.apply_chat_template(
messages,
add_generation_prompt=True,
tokenize=False,
)
inputs = tokenizer(prompt, return_tensors="pt")
with torch.no_grad():
outputs = model.generate(
**inputs,
max_new_tokens=2048,
temperature=0.3, # Lower temp for evaluation
do_sample=True,
)
completion = tokenizer.decode(outputs[0], skip_special_tokens=True)
code = extract_code(completion)
# Evaluate
result = env.step(KernelAction(code=code))
obs = result.observation
results.append({
"problem_id": problem_id,
"compilation": obs.compilation_success,
"correctness": obs.correctness_pass,
"speedup": obs.speedup,
})
print(f"{problem_id}: compile={obs.compilation_success}, "
f"correct={obs.correctness_pass}, speedup={obs.speedup:.2f}x"
if obs.speedup else f"{problem_id}: compile={obs.compilation_success}")
return results
# Evaluate on a few problems
# eval_results = evaluate_model("./kernrl_trained_model", ["L1_23_Softmax", "L1_26_GELU_"])
# %% [markdown]
# ## Running with Server Mode (Multi-GPU)
#
# For larger models or faster training, use vLLM in server mode:
#
# ```bash
# # Terminal 1: Start vLLM server
# CUDA_VISIBLE_DEVICES=0 trl vllm-serve --model Qwen/Qwen2.5-Coder-7B-Instruct
#
# # Terminal 2: Start kernrl environment
# CUDA_VISIBLE_DEVICES=1 uvicorn kernrl.server.app:app --host 0.0.0.0 --port 8000
#
# # Terminal 3: Run training
# CUDA_VISIBLE_DEVICES=2 python train_kernrl.py --vllm-mode server --vllm-server-url http://localhost:8000
# ```
#
# Update the config:
# ```python
# config = GRPOConfig(
# use_vllm=True,
# vllm_mode="server",
# vllm_server_base_url="http://localhost:8000",
# ...
# )
# ```
# %% [markdown]
# ## Tips for Better Results
#
# 1. **Start with simpler problems**: Level 1 problems (matmul, softmax) are easier
# 2. **Use code-focused models**: Qwen2.5-Coder, DeepSeek-Coder work well
# 3. **Increase generations**: More generations per prompt = better advantage estimates
# 4. **Multi-turn training**: Let the model iterate based on feedback
# 5. **Curriculum learning**: Start with L1, add harder problems gradually
# %% [markdown]
# ## Resources
#
# - [kernrl HuggingFace Space](https://huggingface.co/spaces/Infatoshi/kernrl)
# - [OpenEnv Repository](https://github.com/meta-pytorch/OpenEnv)
# - [TRL Documentation](https://huggingface.co/docs/trl)
# - [Triton Tutorial](https://triton-lang.org/main/getting-started/tutorials/)
|