File size: 11,352 Bytes
2abcc30
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from __future__ import annotations

import json
from pathlib import Path

import torch

from .constants import (
    DEFAULT_EXPORT_DIR,
    DEFAULT_PACK_DIR,
    DEFAULT_RL_ADAPTER_DIR,
    LORA_ALPHA,
    LORA_RANK,
)
from .reward import score_texts
from .sft import _load_pack, default_pack, lora_target_modules


def train_grpo(
    *,
    pack_path: Path | None = None,
    model_dir: Path = DEFAULT_EXPORT_DIR,
    output_dir: Path = DEFAULT_RL_ADAPTER_DIR,
    max_steps: int = 40,
    max_completion_len: int = 512,
    num_generations: int = 2,
    per_device_batch_size: int = 1,
    lr: float = 5e-6,
    lora_rank: int = LORA_RANK,
    smoke: bool = False,
) -> Path:
    """Light on-policy GRPO. Reward is gate/edit/submit — not proxy_score.

    Custom loop (not TRL GRPOTrainer): this checkpoint is Qwen3.5-MoE VL and
    TRL's generate path feeds float `input_ids` into `embed_tokens`.
    """
    from local_eval.cuda_env import apply as apply_cuda

    apply_cuda()
    pack_path = Path(pack_path or default_pack(DEFAULT_PACK_DIR))
    output_dir = Path(output_dir)
    output_dir.mkdir(parents=True, exist_ok=True)
    rows = _load_pack(pack_path)
    if smoke:
        rows = rows[:16]
        max_steps = min(max_steps, 8)
        max_completion_len = min(max_completion_len, 768)
        num_generations = min(num_generations, 2)
    if not rows:
        raise ValueError(f"empty pack: {pack_path}")

    from peft import LoraConfig, get_peft_model
    from transformers import AutoModelForCausalLM, AutoTokenizer

    local_rank = int(__import__("os").environ.get("LOCAL_RANK", 0))
    world = int(__import__("os").environ.get("WORLD_SIZE", 1))
    if world > 1 and not torch.distributed.is_initialized():
        torch.distributed.init_process_group(backend="nccl")
    if torch.cuda.is_available():
        torch.cuda.set_device(local_rank)
        device = torch.device(f"cuda:{local_rank}")
    else:
        device = torch.device("cpu")

    tokenizer = AutoTokenizer.from_pretrained(str(model_dir), trust_remote_code=False)
    if tokenizer.pad_token is None:
        tokenizer.pad_token = tokenizer.eos_token
    tokenizer.padding_side = "left"

    model = AutoModelForCausalLM.from_pretrained(
        str(model_dir),
        torch_dtype=torch.bfloat16,
        trust_remote_code=False,
        attn_implementation="sdpa",
    )
    model.config.use_cache = False
    if hasattr(model, "enable_input_require_grads"):
        model.enable_input_require_grads()
    if hasattr(model, "gradient_checkpointing_enable"):
        model.gradient_checkpointing_enable()
    if not _has_lora(model):
        model = get_peft_model(
            model,
            LoraConfig(
                r=lora_rank,
                lora_alpha=LORA_ALPHA,
                lora_dropout=0.05,
                bias="none",
                task_type="CAUSAL_LM",
                target_modules=lora_target_modules(model),
            ),
        )
    model.to(device)
    model.train()
    if world > 1:
        model = torch.nn.parallel.DistributedDataParallel(
            model,
            device_ids=[local_rank],
            output_device=local_rank,
            find_unused_parameters=True,
        )
    optimizer = torch.optim.AdamW((p for p in model.parameters() if p.requires_grad), lr=lr)

    steps_done = 0
    updated = 0
    last_stats: dict = {}
    while steps_done < max_steps:
        batch = [rows[(steps_done * world + local_rank + i) % len(rows)] for i in range(per_device_batch_size)]
        loss, stats = _grpo_step(
            model=model,
            tokenizer=tokenizer,
            batch=batch,
            num_generations=num_generations,
            max_completion_len=max_completion_len,
            device=device,
        )
        last_stats = stats
        # Always backward so DDP ranks stay in lockstep even when advantages are 0.
        optimizer.zero_grad(set_to_none=True)
        loss.backward()
        torch.nn.utils.clip_grad_norm_((p for p in model.parameters() if p.requires_grad), 1.0)
        optimizer.step()
        if stats.get("signaled"):
            updated += 1
        if local_rank == 0:
            print(
                f"rl step={steps_done + 1}/{max_steps} loss={float(loss.detach()):.4f} "
                f"mean_r={stats['mean_r']:.3f} std_r={stats['std_r']:.3f} "
                f"fatal={stats['n_fatal']}/{stats['n']} bash={stats['n_bash']}/{stats['n']}",
                flush=True,
            )
            snippet = (stats.get("sample") or "").replace("\n", " ")
            if snippet:
                print(f"  on-policy: {snippet[:160]!r}", flush=True)
        steps_done += 1

    raw = model.module if hasattr(model, "module") else model
    if local_rank == 0:
        report = {
            "pack": str(pack_path),
            "model": str(model_dir),
            "n": len(rows),
            "max_steps": max_steps,
            "num_generations": num_generations,
            "updated_steps": updated,
            "smoke": smoke,
            "reward": "gate/edit/exact-submit (not proxy_score)",
            "last_stats": last_stats,
        }
        (output_dir / "rl-report.json").write_text(json.dumps(report, indent=2) + "\n")
        if updated:
            raw.save_pretrained(str(output_dir))
            print(f"rl adapter: {output_dir} updated_steps={updated}", flush=True)
        else:
            print(f"rl skipped save (no advantage signal): {output_dir}", flush=True)
    if world > 1:
        torch.distributed.barrier()
    return output_dir


def _grpo_step(*, model, tokenizer, batch, num_generations, max_completion_len, device):
    prompts = [row["prompt"] for row in batch]
    encoded = tokenizer(
        prompts,
        return_tensors="pt",
        padding=True,
        truncation=True,
        max_length=2048,
        add_special_tokens=False,
    )
    prompt_ids = encoded["input_ids"].to(device=device, dtype=torch.long)
    prompt_mask = encoded["attention_mask"].to(device=device)
    prompt_ids = prompt_ids.repeat_interleave(num_generations, dim=0)
    prompt_mask = prompt_mask.repeat_interleave(num_generations, dim=0)
    unwrapped = model.module if hasattr(model, "module") else model
    with torch.no_grad():
        was_training = unwrapped.training
        unwrapped.eval()
        unwrapped.config.use_cache = True
        generated = unwrapped.generate(
            input_ids=prompt_ids,
            attention_mask=prompt_mask,
            max_new_tokens=max_completion_len,
            do_sample=True,
            temperature=1.1,
            top_p=0.95,
            pad_token_id=tokenizer.pad_token_id,
            eos_token_id=tokenizer.eos_token_id,
        )
        unwrapped.config.use_cache = False
        if was_training:
            unwrapped.train()
    prompt_len = prompt_ids.size(1)
    generated = _inject_gold_group(
        generated,
        batch=batch,
        prompts=prompts,
        tokenizer=tokenizer,
        prompt_len=prompt_len,
        num_generations=num_generations,
    )
    completion_ids = generated[:, prompt_len:]
    texts = tokenizer.batch_decode(completion_ids, skip_special_tokens=True)
    rewards = []
    n_fatal = 0
    n_bash = 0
    for index, text in enumerate(texts):
        row = batch[index // num_generations]
        br = score_texts(
            [text],
            submit_command=row.get("submit_command") or "",
            gold_paths=row.get("gold_paths") or [],
        )
        rewards.append(br.reward)
        n_fatal += int(br.fatal)
        n_bash += int("```bash" in text)
    reward_t = torch.tensor(rewards, device=device, dtype=torch.float32)
    advantages = group_advantages(reward_t, num_generations)
    signaled = bool(not torch.allclose(advantages, torch.zeros_like(advantages)))

    full_ids = generated.to(device=device, dtype=torch.long)
    attn = (full_ids != (tokenizer.pad_token_id or -1)).long() if tokenizer.pad_token_id is not None else torch.ones_like(full_ids)
    outputs = model(input_ids=full_ids, attention_mask=attn)
    logp = torch.nn.functional.log_softmax(outputs.logits[:, :-1, :], dim=-1)
    target = full_ids[:, 1:]
    token_logp = logp.gather(-1, target.unsqueeze(-1)).squeeze(-1)
    comp_mask = torch.zeros_like(token_logp)
    if prompt_len > 0:
        comp_mask[:, prompt_len - 1 :] = 1.0
    pad_id = tokenizer.pad_token_id
    if pad_id is not None:
        comp_mask = comp_mask * (target != pad_id).float()
    seq_logp = (token_logp * comp_mask).sum(dim=1) / comp_mask.sum(dim=1).clamp(min=1.0)
    # Zero advantages still produce a graph-connected 0 loss so DDP allreduces.
    loss = -(advantages * seq_logp).mean()
    if not signaled:
        loss = loss * 0.0 + seq_logp.mean() * 0.0
    stats = {
        "mean_r": float(reward_t.mean()),
        "std_r": float(reward_t.std(unbiased=False)),
        "n_fatal": n_fatal,
        "n_bash": n_bash,
        "n": len(texts),
        "signaled": signaled,
        "rewards": [round(r, 4) for r in rewards],
        "sample": texts[1] if len(texts) > 1 else (texts[0] if texts else ""),
    }
    return loss, stats


def gold_continuation(prompt: str, completion: str) -> str:
    """Drop a duplicated <think> open — the chat template already started it."""
    if not completion:
        return ""
    if prompt.endswith("<think>\n") and completion.startswith("<think>\n"):
        return completion[len("<think>\n") :]
    return completion


def _inject_gold_group(generated, *, batch, prompts, tokenizer, prompt_len, num_generations):
    """Replace generation 0 in each group with the gold continuation (protocol teacher)."""
    pad_id = tokenizer.pad_token_id
    if pad_id is None:
        pad_id = tokenizer.eos_token_id or 0
    for index, row in enumerate(batch):
        gold = gold_continuation(prompts[index], row.get("completion") or "")
        if not gold.strip():
            continue
        gold_ids = tokenizer(gold, add_special_tokens=False, return_tensors="pt")["input_ids"][0]
        gold_ids = gold_ids.to(device=generated.device, dtype=generated.dtype)
        need = prompt_len + int(gold_ids.numel())
        if need > generated.size(1):
            extra = torch.full(
                (generated.size(0), need - generated.size(1)),
                pad_id,
                device=generated.device,
                dtype=generated.dtype,
            )
            generated = torch.cat([generated, extra], dim=1)
        slot = index * num_generations
        generated[slot, prompt_len:] = pad_id
        n = min(int(gold_ids.numel()), generated.size(1) - prompt_len)
        generated[slot, prompt_len : prompt_len + n] = gold_ids[:n]
    return generated


def group_advantages(rewards: torch.Tensor, num_generations: int) -> torch.Tensor:
    """Within-group z-score; fall back to batch baseline when a group is tied."""
    if rewards.numel() < 2:
        return torch.zeros_like(rewards)
    grouped = rewards.view(-1, num_generations)
    adv = (grouped - grouped.mean(dim=1, keepdim=True)) / (grouped.std(dim=1, keepdim=True) + 1e-6)
    flat = adv.reshape(-1)
    if torch.allclose(flat, torch.zeros_like(flat)):
        flat = (rewards - rewards.mean()) / (rewards.std() + 1e-6)
    return flat.detach()


def _has_lora(model) -> bool:
    return any("lora_" in name for name, _ in model.named_parameters())