File size: 20,697 Bytes
f2df60e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""DriftCall training orchestrator β€” local GPU run. **DEPRECATED.**

This script wires the legacy TRL ``GRPOTrainer`` + Unsloth ``UnslothGRPOTrainer``
chain to ``cells/step_15/16/17_train_stage*.train()``. That stack hits Unsloth's
broken ``Linear4bit.forward`` patch on Gemma 3n's ``per_layer_model_projection``
inside TRL's training loop (our lazy fix-up patch fires too late) and breaks on
every TRL/Unsloth version bump.

**Use the new self-contained loop instead:**

    scripts/train_driftcall_grpo.py        # single stage
    scripts/train_full_gemma3n.sh          # all three stages

That loop bypasses TRL entirely and drives rollouts/rewards/updates directly,
with controlled patch ordering. See ``docs/modules/training.md`` Β§3.2.

This file is kept on disk because some environments still reference it.
Do not extend it. New training work goes in ``scripts/train_driftcall_grpo.py``.
"""

from __future__ import annotations

import argparse
import hashlib
import json
import os
import sys
from pathlib import Path
from typing import Any

# Add project root to PYTHONPATH so cells/ is importable without install.
_REPO_ROOT = Path(__file__).resolve().parent.parent
if str(_REPO_ROOT) not in sys.path:
    sys.path.insert(0, str(_REPO_ROOT))


# transformers 5.x removed the legacy ``TRANSFORMERS_CACHE`` symbol that
# ``llm_blender`` (a TRL 0.24 transitive dep) imports at module load time.
# Restore it BEFORE any ``from trl import ...`` so GRPOTrainer can boot.
def _patch_transformers_cache() -> None:
    try:
        import transformers.utils.hub as _hub
    except Exception:
        return
    if not hasattr(_hub, "TRANSFORMERS_CACHE"):
        _hub.TRANSFORMERS_CACHE = os.environ.get(
            "HF_HOME",
            os.path.expanduser("~/.cache/huggingface"),
        )


_patch_transformers_cache()


# Unsloth 2026.4.x ships a buggy ``Gemma3nRMSNorm_forward`` patch that
# unconditionally reads ``self.weight``; for Gemma3n's
# ``embedding_post_projection_norm`` (constructed with ``with_scale=False``)
# this raises AttributeError during model.generate(). Override the patched
# ``Gemma3nMultimodalEmbedder.forward`` with a with_scale-aware version.
def _patch_unsloth_gemma3n_rmsnorm() -> None:
    try:
        import torch
        from transformers.models.gemma3n.modeling_gemma3n import (
            Gemma3nMultimodalEmbedder,
        )
    except Exception:
        return

    def _safe_rmsnorm(norm_module: Any, x: Any) -> Any:
        # Mirror the canonical Gemma3n RMSNorm forward but respect with_scale.
        normed = norm_module._norm(x.float())
        if getattr(norm_module, "with_scale", True):
            normed = normed * norm_module.weight.float()
        return normed.type_as(x)

    def _patched_forward(
        self: Any,
        input_ids: Any = None,
        inputs_embeds: Any = None,
    ) -> Any:
        if (input_ids is None) ^ (inputs_embeds is not None):
            raise ValueError(
                "You must specify exactly one of input_ids or inputs_embeds"
            )
        if inputs_embeds is not None:
            emb_norm = _safe_rmsnorm(self.soft_embedding_norm, inputs_embeds)
        else:
            hard_emb = self.embedding(input_ids - self.vocab_offset)
            emb_norm = _safe_rmsnorm(self.hard_embedding_norm, hard_emb)
        old_dtype = emb_norm.dtype
        emb_norm = emb_norm.to(torch.float32)
        with torch.autocast(device_type="cuda", dtype=torch.float32, enabled=True):
            emb_norm_proj = self.embedding_projection(emb_norm)
        emb_norm_proj = emb_norm_proj.to(old_dtype)
        return _safe_rmsnorm(self.embedding_post_projection_norm, emb_norm_proj)

    Gemma3nMultimodalEmbedder.forward = _patched_forward


def _patch_unsloth_bnb_linear4bit_quant_state() -> None:
    """Trigger ``fix_4bit_weight_quant_state_from_module`` on packed 4-bit
    weights that bnb stored with ``shape[0] == 1`` (transposed packing).

    Unsloth's stock patch only checks ``weight.shape[-1] == 1``; some Gemma3n
    layers (notably ``per_layer_model_projection``) ship with the packed
    dim on axis 0 instead, so the auto-fix never fires and we crash with
    ``mat1 and mat2 shapes cannot be multiplied (..., 1xPACKED)``.
    """
    try:
        import torch
        import bitsandbytes
        from unsloth_zoo.temporary_patches.bitsandbytes import (
            fix_4bit_weight_quant_state_from_module,
        )
    except Exception:
        return

    Linear4bit = bitsandbytes.nn.modules.Linear4bit
    _orig_forward = Linear4bit.forward

    def _safe_forward(self: Any, x: torch.Tensor) -> torch.Tensor:
        weight = self.weight
        # Detect packed 4-bit tensors with no quant_state and a flat shape
        # in either orientation. Trigger the fix routine which restores
        # quant_state from the ``module`` attribute and reshapes the weight.
        try:
            if (
                getattr(weight, "quant_state", None) is None
                and weight.dim() == 2
                and (weight.shape[0] == 1 or weight.shape[-1] == 1)
            ):
                fix_4bit_weight_quant_state_from_module(self)
        except Exception:
            pass
        return _orig_forward(self, x)

    Linear4bit.forward = _safe_forward


# ---------------------------------------------------------------------------
# Action parser β€” extract DriftCallAction from model text output
# ---------------------------------------------------------------------------


def _parse_action(text: str) -> Any:
    """Parse a DriftCallAction from the model's assistant turn text.

    Tries JSON extraction first; falls back to an ABORT action on parse
    failure so the episode terminates cleanly rather than hanging.
    """
    from cells.step_04_models import ActionType, DriftCallAction

    text = text.strip()
    # Try to extract the first JSON object in the text
    try:
        start = text.index("{")
        depth, end = 0, -1
        for i, ch in enumerate(text[start:], start):
            if ch == "{":
                depth += 1
            elif ch == "}":
                depth -= 1
                if depth == 0:
                    end = i + 1
                    break
        if end > start:
            obj = json.loads(text[start:end])
            action_type_str = obj.get("action_type", "abort")
            try:
                atype = ActionType(action_type_str)
            except ValueError:
                atype = ActionType.ABORT
            return DriftCallAction(
                action_type=atype,
                tool_name=obj.get("tool_name"),
                tool_args=obj.get("tool_args"),
                message=obj.get("message"),
                confidence=obj.get("confidence"),
                rationale=obj.get("rationale"),
            )
    except (ValueError, json.JSONDecodeError, KeyError):
        pass
    # Fallback: submit if the model said something confidence-like, else abort
    lower = text.lower()
    if "submit" in lower:
        return DriftCallAction(
            action_type=ActionType.SUBMIT,
            message=text[:200],
            confidence=0.5,
        )
    return DriftCallAction(action_type=ActionType.ABORT, message=text[:200])


# ---------------------------------------------------------------------------
# Observation serializer β€” obs β†’ messages list (training.md Β§3.2.1)
# ---------------------------------------------------------------------------


def _obs_to_messages(
    goal: Any,
    obs: Any,
    history: list[dict[str, str]],
    is_turn_zero: bool,
) -> list[dict[str, str]]:
    """Append the latest observation fields to the message history.

    training.md Β§3.2.1 β€” returns the updated history list (mutated in-place
    for efficiency; callers can deepcopy if needed).
    """
    from cells.step_14_custom_trainer import PINNED_SYSTEM_PROMPT

    if is_turn_zero:
        # Build the system prompt with available tools appended.
        tool_schemas = getattr(obs, "available_tools", [])
        system_content = PINNED_SYSTEM_PROMPT
        if tool_schemas:
            system_content += "\nAvailable tools: " + json.dumps(
                tool_schemas, ensure_ascii=False, sort_keys=True
            )
        history.clear()
        history.append({"role": "system", "content": system_content})
        history.append(
            {"role": "user", "content": getattr(goal, "seed_utterance", "")}
        )
    else:
        # Append any new tool results from the last step.
        tool_results = getattr(obs, "tool_results", [])
        if tool_results:
            for tr in tool_results[-1:]:  # only the latest tool result
                history.append(
                    {
                        "role": "tool",
                        "content": json.dumps(
                            {
                                "tool": getattr(tr, "tool_name", ""),
                                "status": getattr(tr, "status", ""),
                                "response": getattr(tr, "response", {}),
                            },
                            ensure_ascii=False,
                            sort_keys=True,
                        ),
                    }
                )
        # Append drift events if any.
        drift_log = getattr(obs, "drift_log", [])
        if drift_log:
            drift_json = json.dumps(
                [
                    {
                        "turn": getattr(d, "turn", 0),
                        "type": getattr(d, "drift_type", ""),
                        "domain": getattr(d, "domain", ""),
                        "description": getattr(d, "description", ""),
                    }
                    for d in drift_log[-3:]  # last 3 drifts to cap token budget
                ],
                ensure_ascii=False,
                sort_keys=True,
            )
            # Append as user message so the model sees the drift signal.
            history.append(
                {"role": "user", "content": f"[drift] {drift_json}"}
            )

    return history


def _derive_rollout_seed(goal: Any, g_index: int, episode_seed: int) -> int:
    """Deterministic seed per rollout within a group (training.md Β§3.2)."""
    payload = f"{episode_seed}:{getattr(goal, 'seed_utterance', '')}:{g_index}".encode()
    digest = hashlib.blake2b(payload, digest_size=8).digest()
    return int.from_bytes(digest, "little") & 0x7FFF_FFFF


# ---------------------------------------------------------------------------
# rollout_group_fn β€” the core multi-turn inference loop
# ---------------------------------------------------------------------------


def build_rollout_group_fn(
    *,
    max_turns: int = 8,
    max_new_tokens: int = 512,
    temperature: float = 0.9,
    top_p: float = 0.95,
    hardware: str = "v100",
) -> Any:
    """Build and return a RolloutGroupFn (training.md Β§3.2).

    Runs G independent multi-turn episodes with the live model. Each rollout:
      1. env.reset(seed=derived_seed)
      2. Serialise obs β†’ messages, generate one action token-by-token
      3. Parse β†’ DriftCallAction β†’ env.step(action)
      4. Repeat until obs.done or max_turns reached
    Returns (tuple[Episode, ...], tuple[str, ...]) of length G.
    """

    def rollout_group_fn(
        *,
        model: Any,
        tokenizer: Any,
        goal: Any,
        episode_seed: int,
        num_generations: int,
        env_factory: Any,
    ) -> tuple[tuple[Any, ...], tuple[str, ...]]:
        import torch
        from cells.step_04_models import ActionType, DriftCallAction

        # Apply Gemma3n + bnb patches on first rollout call, AFTER Unsloth has
        # already monkey-patched its (broken) versions.
        _patch_unsloth_gemma3n_rmsnorm()
        _patch_unsloth_bnb_linear4bit_quant_state()

        device = next(model.parameters()).device
        episodes_out: list[Any] = []
        completions_out: list[str] = []

        for g in range(num_generations):
            seed = _derive_rollout_seed(goal, g, episode_seed)
            env = env_factory()
            obs = env.reset(seed=seed)

            history: list[dict[str, str]] = []
            all_responses: list[str] = []
            is_turn_zero = True

            for _turn in range(max_turns):
                # Build messages for this turn.
                _obs_to_messages(goal, obs, history, is_turn_zero)
                is_turn_zero = False

                # Tokenize the conversation so far.
                try:
                    prompt_str = tokenizer.apply_chat_template(
                        history,
                        tokenize=False,
                        add_generation_prompt=True,
                    )
                except Exception:
                    prompt_str = " ".join(
                        m.get("content", "") for m in history
                    )

                # Gemma 3n's processor is multimodal β€” pass `text=` explicitly
                # so the call dispatches to the text-only branch.
                inputs = tokenizer(
                    text=prompt_str,
                    return_tensors="pt",
                    truncation=True,
                    max_length=1024,
                ).to(device)

                # Generate the assistant response.
                with torch.no_grad():
                    gen_kwargs: dict[str, Any] = {
                        "max_new_tokens": max_new_tokens,
                        "do_sample": True,
                        "temperature": temperature,
                        "top_p": top_p,
                        "pad_token_id": tokenizer.eos_token_id,
                    }
                    output_ids = model.generate(
                        **inputs, **gen_kwargs
                    )
                new_token_ids = output_ids[0][inputs["input_ids"].shape[1]:]
                response_text = tokenizer.decode(
                    new_token_ids, skip_special_tokens=True
                ).strip()

                all_responses.append(response_text)
                history.append({"role": "assistant", "content": response_text})

                # Parse and step the environment.
                action = _parse_action(response_text)
                obs = env.step(action)

                if obs.done:
                    break

            # Collect the completed episode from the env.
            episode = env.episode()
            episodes_out.append(episode)
            completions_out.append("\n".join(all_responses))

        return tuple(episodes_out), tuple(completions_out)

    return rollout_group_fn


# ---------------------------------------------------------------------------
# CLI
# ---------------------------------------------------------------------------


def _parse_args(argv: list[str] | None = None) -> argparse.Namespace:
    p = argparse.ArgumentParser(
        prog="run_driftcall_train",
        description="DriftCall 3-stage GRPO training on a local GPU.",
    )
    p.add_argument("--stage1-steps", type=int, default=150)
    p.add_argument("--stage2-steps", type=int, default=200)
    p.add_argument("--stage3-steps", type=int, default=150)
    p.add_argument("--hardware", choices=["v100", "h100"], default="v100")
    p.add_argument("--output-dir", type=Path, default=Path("/workspace/checkpoints"))
    p.add_argument("--eval-episodes", type=int, default=50)
    p.add_argument("--probe-episodes", type=int, default=200)
    p.add_argument(
        "--skip-eval", action="store_true", help="Skip baseline/final eval + probe."
    )
    p.add_argument(
        "--push-to-hub",
        action="store_true",
        help="Push trained LoRA to HF Hub after stage 3.",
    )
    p.add_argument("--hf-repo", type=str, default=os.environ.get("DRIFTCALL_HF_REPO", ""))
    return p.parse_args(argv)


def main(argv: list[str] | None = None) -> int:
    args = _parse_args(argv)
    out_dir = args.output_dir

    print(f"[train] hardware={args.hardware}  CUDA_VISIBLE_DEVICES={os.environ.get('CUDA_VISIBLE_DEVICES', 'unset')}")
    print(f"[train] steps: stage1={args.stage1_steps}  stage2={args.stage2_steps}  stage3={args.stage3_steps}")
    print(f"[train] output_dir={out_dir}")

    from cells.step_07_task_generator import generate as task_gen_fn
    from cells.step_10_env import DriftCallEnv
    from cells.step_12_gemma_boot import BootConfig

    boot_config = BootConfig(hardware=args.hardware)
    rollout_group_fn = build_rollout_group_fn(hardware=args.hardware)

    # ---------------------------------------------------------------------------
    # Stage 1
    # ---------------------------------------------------------------------------
    print("\n[train] === Stage 1 ===")
    from cells.step_15_train_stage1 import train as train_stage1

    def env_factory_s1() -> DriftCallEnv:
        return DriftCallEnv(config={"curriculum_stage": 1, "audio_boundary_enabled": False})

    ckpt_s1 = train_stage1(
        num_steps=args.stage1_steps,
        output_dir=out_dir / "stage1" / "final",
        boot_config=boot_config,
        task_gen=task_gen_fn,
        env_factory=env_factory_s1,
        rollout_group_fn=rollout_group_fn,
    )
    print(f"[train] Stage 1 complete β†’ {ckpt_s1}")

    # ---------------------------------------------------------------------------
    # Stage 2
    # ---------------------------------------------------------------------------
    print("\n[train] === Stage 2 ===")
    from cells.step_16_train_stage2 import train as train_stage2

    def env_factory_s2() -> DriftCallEnv:
        return DriftCallEnv(config={"curriculum_stage": 2, "audio_boundary_enabled": False})

    ckpt_s2 = train_stage2(
        num_steps=args.stage2_steps,
        resume_from=ckpt_s1,
        output_dir=out_dir / "stage2" / "final",
        boot_config=boot_config,
        task_gen=task_gen_fn,
        env_factory=env_factory_s2,
        rollout_group_fn=rollout_group_fn,
    )
    print(f"[train] Stage 2 complete β†’ {ckpt_s2}")

    # ---------------------------------------------------------------------------
    # Stage 3
    # ---------------------------------------------------------------------------
    print("\n[train] === Stage 3 ===")
    from cells.step_17_train_stage3 import train as train_stage3

    def env_factory_s3() -> DriftCallEnv:
        return DriftCallEnv(config={"curriculum_stage": 3, "audio_boundary_enabled": False})

    ckpt_s3 = train_stage3(
        num_steps=args.stage3_steps,
        resume_from=ckpt_s2,
        output_dir=out_dir / "stage3" / "final",
        boot_config=boot_config,
        task_gen=task_gen_fn,
        env_factory=env_factory_s3,
        rollout_group_fn=rollout_group_fn,
    )
    print(f"[train] Stage 3 complete β†’ {ckpt_s3}")

    if not args.skip_eval:
        # ---------------------------------------------------------------------------
        # Baseline + Final eval
        # ---------------------------------------------------------------------------
        print("\n[train] === Baseline eval ===")
        from cells.step_18_eval_baseline import eval_baseline
        eval_dir = Path(os.environ.get("DRIFTCALL_EVAL_DIR", "/workspace/eval_reports"))
        eval_dir.mkdir(parents=True, exist_ok=True)

        baseline_report = eval_baseline(
            n_episodes=args.eval_episodes,
            output_path=eval_dir / "baseline.json",
            env_factory=env_factory_s3,
            task_gen=task_gen_fn,
        )
        print(f"[train] Baseline eval: R1={getattr(baseline_report, 'r1_mean', '?'):.3f}")

        print("\n[train] === Final eval ===")
        from cells.step_19_eval_final import eval_final
        final_report = eval_final(
            checkpoint_path=ckpt_s3,
            n_episodes=args.eval_episodes,
            output_path=eval_dir / "final.json",
            env_factory=env_factory_s3,
            task_gen=task_gen_fn,
        )
        print(f"[train] Final eval: R1={getattr(final_report, 'r1_mean', '?'):.3f}")

    if args.push_to_hub and args.hf_repo:
        print(f"\n[train] === Pushing LoRA to HF Hub: {args.hf_repo} ===")
        from cells.step_24_deploy_hf import push_lora_to_hub
        result = push_lora_to_hub(
            ckpt_s3, repo_id=args.hf_repo, token=os.environ.get("HF_TOKEN")
        )
        if result.success:
            print(f"[train] Pushed to hub: {args.hf_repo}")
        else:
            print(f"[train] Push failed (rc={result.return_code}): {result.stderr[:200]}")

    print("\n[train] === COMPLETE ===")
    print(f"[train] Checkpoints at: {out_dir}/stage{{1,2,3}}/final")
    return 0


if __name__ == "__main__":
    raise SystemExit(main())