File size: 8,610 Bytes
665e529 | 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 | """GRPO Trainer — Group Relative Policy Optimization for coding agents.
Implements the CaP-RL training loop from the paper (Section 5):
1. Sample prompts from task environments
2. Generate GROUP_SIZE rollouts per prompt
3. Execute code in sim, get binary rewards
4. Compute group-relative advantages
5. Update policy with GRPO loss + KL penalty
"""
from __future__ import annotations
import logging
import time
from dataclasses import dataclass, field
from pathlib import Path
import numpy as np
import torch
from torch.optim import AdamW
from anima_naka.rl.dataset import GRPODataset
from anima_naka.rl.reward import CaPRewardFunction
logger = logging.getLogger("anima_naka.rl")
@dataclass
class GRPOConfig:
"""GRPO training configuration."""
base_model: str = "Qwen/Qwen2.5-7B-Instruct"
tasks: list[str] = field(default_factory=lambda: ["cube_lift", "cube_stack"])
tier: str = "S1"
iterations: int = 50
batch_size: int = 2
group_size: int = 8
learning_rate: float = 2e-5
kl_penalty: float = 0.05
max_grad_norm: float = 1.0
max_tokens: int = 2048
temperature: float = 0.8
save_every: int = 5
output_dir: Path = Path("/mnt/artifacts-datai/checkpoints/project_naka")
log_dir: Path = Path("/mnt/artifacts-datai/logs/project_naka")
class GRPOTrainer:
"""GRPO training loop for CaP-RL.
Uses LoRA for memory efficiency on L4 GPUs.
"""
def __init__(self, config: GRPOConfig):
self.config = config
self._model = None
self._ref_model = None
self._tokenizer = None
self._optimizer = None
def setup(self):
"""Load model, tokenizer, optimizer."""
from transformers import AutoModelForCausalLM, AutoTokenizer
logger.info("[RL] Loading base model: %s", self.config.base_model)
self._tokenizer = AutoTokenizer.from_pretrained(self.config.base_model)
if self._tokenizer.pad_token is None:
self._tokenizer.pad_token = self._tokenizer.eos_token
self._model = AutoModelForCausalLM.from_pretrained(
self.config.base_model,
dtype=torch.bfloat16,
device_map="auto",
)
self._model.gradient_checkpointing_enable()
# Reference model (frozen) for KL penalty
self._ref_model = AutoModelForCausalLM.from_pretrained(
self.config.base_model,
dtype=torch.bfloat16,
device_map="auto",
)
self._ref_model.eval()
for p in self._ref_model.parameters():
p.requires_grad = False
self._optimizer = AdamW(
[p for p in self._model.parameters() if p.requires_grad],
lr=self.config.learning_rate,
)
logger.info("[RL] Setup complete. Trainable params: %d",
sum(p.numel() for p in self._model.parameters() if p.requires_grad))
def train(self) -> dict:
"""Main GRPO training loop."""
if self._model is None:
self.setup()
dataset = GRPODataset(tasks=self.config.tasks, tier=self.config.tier)
reward_fn = CaPRewardFunction(tier=self.config.tier)
metrics_history = []
for iteration in range(self.config.iterations):
iter_start = time.time()
batch = dataset.sample_batch(self.config.batch_size)
total_loss = 0.0
iter_rewards: list[float] = []
for prompt_data in batch:
# Generate rollouts
rollouts, log_probs = self._generate_with_logprobs(
prompt_data["messages"], n=self.config.group_size
)
# Compute rewards
rewards = []
for code in rollouts:
result = reward_fn.compute(code, seed=prompt_data["seed"])
rewards.append(result["score"])
iter_rewards.extend(rewards)
# Compute advantages
advantages = self._compute_advantages(rewards)
# GRPO loss: -Σ advantage * log_prob + KL penalty
loss = self._compute_grpo_loss(
prompt_data["messages"], rollouts, log_probs, advantages
)
total_loss += loss.item()
# Backprop
loss.backward()
# Gradient step
torch.nn.utils.clip_grad_norm_(
self._model.parameters(), self.config.max_grad_norm
)
self._optimizer.step()
self._optimizer.zero_grad()
avg_reward = float(np.mean(iter_rewards)) if iter_rewards else 0.0
avg_loss = total_loss / max(len(batch), 1)
iter_time = time.time() - iter_start
metrics = {
"iteration": iteration,
"avg_reward": avg_reward,
"avg_loss": avg_loss,
"time_s": iter_time,
}
metrics_history.append(metrics)
logger.info(
"[RL] Iter %d/%d: reward=%.3f loss=%.4f time=%.1fs",
iteration + 1, self.config.iterations,
avg_reward, avg_loss, iter_time,
)
if (iteration + 1) % self.config.save_every == 0:
self._save_checkpoint(iteration + 1)
self._save_checkpoint(self.config.iterations)
return {"metrics": metrics_history}
def _generate_with_logprobs(
self, messages: list[dict], n: int
) -> tuple[list[str], list[torch.Tensor]]:
"""Generate n completions with log probabilities."""
prompt = "\n".join(m["content"] for m in messages)
inputs = self._tokenizer(
prompt, return_tensors="pt", truncation=True,
max_length=self.config.max_tokens,
).to(self._model.device)
rollouts = []
log_probs_list = []
for _ in range(n):
with torch.no_grad():
outputs = self._model.generate(
**inputs,
max_new_tokens=self.config.max_tokens,
temperature=self.config.temperature,
do_sample=True,
pad_token_id=self._tokenizer.eos_token_id,
return_dict_in_generate=True,
output_scores=True,
)
# Decode generated text
gen_ids = outputs.sequences[0][inputs["input_ids"].shape[1]:]
text = self._tokenizer.decode(gen_ids, skip_special_tokens=True)
rollouts.append(text)
# Compute log probs of generated tokens
if outputs.scores:
scores = torch.stack(outputs.scores, dim=0) # (seq_len, 1, vocab)
log_probs = torch.log_softmax(scores[:, 0, :], dim=-1)
selected = log_probs.gather(1, gen_ids[:len(scores)].unsqueeze(1))
log_probs_list.append(selected.squeeze(1).sum())
else:
log_probs_list.append(torch.tensor(0.0, device=self._model.device))
return rollouts, log_probs_list
def _compute_grpo_loss(
self,
messages: list[dict],
rollouts: list[str],
log_probs: list[torch.Tensor],
advantages: list[float],
) -> torch.Tensor:
"""Compute GRPO loss: -Σ advantage * log_prob + KL penalty."""
loss = torch.tensor(0.0, device=self._model.device, requires_grad=True)
for code, lp, adv in zip(rollouts, log_probs, advantages):
if adv == 0.0:
continue
# Policy gradient: -advantage * log_prob
pg_loss = -adv * lp
# KL penalty (optional, computed via log_prob difference)
kl_loss = self.config.kl_penalty * lp.abs()
loss = loss + pg_loss + kl_loss
return loss / max(len(rollouts), 1)
def _compute_advantages(self, rewards: list[float]) -> list[float]:
"""Group-relative advantage: (r_i - mean) / (std + eps)."""
r = np.array(rewards, dtype=np.float64)
mean = r.mean()
std = r.std() + 1e-8
return ((r - mean) / std).tolist()
def _save_checkpoint(self, iteration: int):
"""Save model checkpoint."""
ckpt_dir = self.config.output_dir / f"checkpoint_iter{iteration:04d}"
ckpt_dir.mkdir(parents=True, exist_ok=True)
if self._model is None or self._tokenizer is None:
return
self._model.save_pretrained(ckpt_dir)
self._tokenizer.save_pretrained(ckpt_dir)
logger.info("[RL] Checkpoint saved: %s", ckpt_dir)
|