File size: 15,657 Bytes
7f7a72e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d3df1cb
7f7a72e
 
 
 
d3df1cb
 
7f7a72e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Kaggle/Unsloth training entrypoint for Manthan-T1 (TinyLLaVA-style).

This script is intended to be copied into a Kaggle notebook and run on 2×T4.
It supports two stages:
- stage1: projector alignment pretraining (e.g., LLaVA-CC3M-Pretrain-595K)
- stage2: instruction tuning (e.g., LLaVA-Instruct-150K)

Notes:
- We follow MicroLLaVA/TinyLLaVA convention: IMAGE_TOKEN_INDEX = -200 is inserted
  into input_ids for <image> placeholders.
- Labels are IGNORE_INDEX for everything except assistant tokens.
- This script trains:
  - the multimodal projector (always)
  - LoRA adapters on the text model (optional, recommended)
  - vision tower is frozen by default

You still need a *real* base model + vision tower weights. Stub exports will run
but won't learn useful vision-language alignment.
"""

from __future__ import annotations

import argparse
import os
from dataclasses import dataclass
from typing import Any, Dict, List, Optional, Tuple

import torch
from torch import nn
from torch.utils.data import Dataset

try:
    # Fallback for non-Unsloth environments
    from peft import LoraConfig, get_peft_model
except Exception:  # pragma: no cover
    LoraConfig = None
    get_peft_model = None

try:
    # Kaggle + Unsloth
    import unsloth  # noqa: F401
    from unsloth import FastLanguageModel
except Exception:  # pragma: no cover
    FastLanguageModel = None

from transformers import AutoModelForCausalLM, AutoTokenizer, get_cosine_schedule_with_warmup

try:
    from datasets import load_dataset
except Exception as e:  # pragma: no cover
    raise RuntimeError(
        "Missing dependency `datasets`. Install with `pip install datasets` (Kaggle: add to notebook)."
    ) from e


IMAGE_TOKEN_INDEX = -200
IGNORE_INDEX = -100


def tokenizer_image_token(prompt: str, tokenizer, image_token_index: int = IMAGE_TOKEN_INDEX) -> List[int]:
    """MicroLLaVA/TinyLLaVA tokenizer: split on '<image>' and insert a negative id."""

    def _insert_separator(X, sep):
        return [ele for sublist in zip(X, [sep] * len(X)) for ele in sublist][:-1]

    prompt_chunks = [tokenizer(chunk).input_ids for chunk in prompt.split("<image>")]

    input_ids: List[int] = []
    offset = 0
    if (
        len(prompt_chunks) > 0
        and len(prompt_chunks[0]) > 0
        and tokenizer.bos_token_id is not None
        and prompt_chunks[0][0] == tokenizer.bos_token_id
    ):
        offset = 1
        input_ids.append(prompt_chunks[0][0])

    for x in _insert_separator(prompt_chunks, [image_token_index] * (offset + 1)):
        input_ids.extend(x[offset:])

    return input_ids


def build_prompt_from_conversations(conversations: List[Dict[str, str]]) -> Tuple[str, str]:
    """Return (full_prompt, assistant_answer_text).

    LLaVA datasets are 2-turn: human then gpt.
    We map to the string template used in `ManthanForCausalLM.format_chat_prompt`.
    """

    # Expect 2 turns
    human = conversations[0]["value"]
    assistant = conversations[1]["value"]

    system = (
        "A chat between a curious user and an artificial intelligence assistant. "
        "The assistant gives helpful, detailed, and polite answers to the user's questions. "
    )
    # IMPORTANT: no trailing space after ASSISTANT:
    full = system + f"USER: {human.strip()} ASSISTANT:" + assistant
    return full, assistant


@dataclass
class TrainExample:
    input_ids: torch.LongTensor
    labels: torch.LongTensor
    image_path: str


class LlavaLikeDataset(Dataset):
    def __init__(
        self,
        ds_name: str,
        split: str,
        tokenizer,
        max_length: int,
        limit: Optional[int] = None,
    ) -> None:
        self.tokenizer = tokenizer
        self.max_length = max_length

        # Streaming keeps Kaggle disk usage low.
        self.ds = load_dataset(ds_name, split=split, streaming=True)
        self.limit = limit

        # Materialize a small index for non-streaming dataloader behavior.
        self._cache: List[Dict[str, Any]] = []
        for i, ex in enumerate(self.ds):
            self._cache.append(ex)
            if limit is not None and i + 1 >= limit:
                break

    def __len__(self) -> int:
        return len(self._cache)

    def __getitem__(self, idx: int) -> TrainExample:
        ex = self._cache[idx]
        image_path = ex["image"]
        conversations = ex["conversations"]

        full_prompt, _assistant = build_prompt_from_conversations(conversations)
        ids = tokenizer_image_token(full_prompt, self.tokenizer, IMAGE_TOKEN_INDEX)

        # Truncate
        ids = ids[: self.max_length]

        # Labels: only learn on assistant answer tokens.
        # Simple heuristic: find the last occurrence of " ASSISTANT:" marker.
        marker = " ASSISTANT:"
        marker_ids = self.tokenizer(marker).input_ids

        # Find marker in tokenized ids (best-effort).
        start = 0
        for j in range(0, len(ids) - len(marker_ids) + 1):
            if ids[j : j + len(marker_ids)] == marker_ids:
                start = j + len(marker_ids)

        labels = [IGNORE_INDEX] * len(ids)
        for j in range(start, len(ids)):
            if ids[j] == IMAGE_TOKEN_INDEX:
                labels[j] = IGNORE_INDEX
            else:
                labels[j] = ids[j]

        return TrainExample(
            input_ids=torch.tensor(ids, dtype=torch.long),
            labels=torch.tensor(labels, dtype=torch.long),
            image_path=image_path,
        )


def load_image_tensor(image_path: str, image_size: int) -> torch.FloatTensor:
    """Load image from local path in dataset.

    In Kaggle, LLaVA datasets provide image paths relative to the dataset repo.
    Hugging Face datasets streaming yields paths that resolve via HF cache.
    """

    from PIL import Image
    import torchvision.transforms as T

    img = Image.open(image_path).convert("RGB")
    tfm = T.Compose([T.Resize((image_size, image_size)), T.ToTensor()])
    return tfm(img)


def collate_fn(batch: List[TrainExample], image_size: int) -> Dict[str, torch.Tensor]:
    # Pad to max length
    max_len = max(x.input_ids.numel() for x in batch)
    input_ids = torch.full((len(batch), max_len), 0, dtype=torch.long)
    labels = torch.full((len(batch), max_len), IGNORE_INDEX, dtype=torch.long)
    attention_mask = torch.zeros((len(batch), max_len), dtype=torch.long)

    for i, ex in enumerate(batch):
        L = ex.input_ids.numel()
        input_ids[i, :L] = ex.input_ids
        labels[i, :L] = ex.labels
        attention_mask[i, :L] = 1

    # Images
    pixel_values = torch.stack([load_image_tensor(ex.image_path, image_size) for ex in batch], dim=0)

    return {
        "input_ids": input_ids,
        "labels": labels,
        "attention_mask": attention_mask,
        "pixel_values": pixel_values,
    }


def set_requires_grad(module: nn.Module, requires_grad: bool) -> None:
    for p in module.parameters():
        p.requires_grad = requires_grad


def save_projector(model, output_dir: str) -> None:
    os.makedirs(output_dir, exist_ok=True)
    if not hasattr(model, "projector"):
        return
    torch.save(model.projector.state_dict(), os.path.join(output_dir, "projector.pt"))


def maybe_add_lora_to_model(model, args) -> None:
    """Attach LoRA adapters (Unsloth preferred; PEFT fallback)."""

    if not args.use_lora:
        return

    # If the model already has adapters (e.g., loaded via Unsloth), skip.
    if hasattr(model, "peft_config"):
        return

    if get_peft_model is None or LoraConfig is None:
        raise RuntimeError("PEFT not installed, and Unsloth not available. Install `peft` or enable Unsloth.")

    target_modules = [
        # Qwen-like
        "q_proj",
        "k_proj",
        "v_proj",
        "o_proj",
        "gate_proj",
        "up_proj",
        "down_proj",
        # GPT-like fallback
        "c_attn",
        "c_proj",
    ]
    cfg = LoraConfig(
        r=args.lora_r,
        lora_alpha=args.lora_alpha,
        lora_dropout=args.lora_dropout,
        bias="none",
        task_type="CAUSAL_LM",
        target_modules=target_modules,
    )

    # Wrap the language model inside Manthan
    model.language_model = get_peft_model(model.language_model, cfg)



def main() -> int:
    ap = argparse.ArgumentParser()
    ap.add_argument("--stage", choices=["stage1", "stage2"], required=True)
    ap.add_argument("--text_model", type=str, default="Qwen/Qwen3-0.6B-Base")
    ap.add_argument("--vision_model", type=str, default="google/siglip-so400m-patch14-384")
    ap.add_argument("--dataset", type=str, required=True)
    ap.add_argument("--output_dir", type=str, default="./outputs")
    ap.add_argument("--max_length", type=int, default=2048)
    ap.add_argument("--image_size", type=int, default=384)
    ap.add_argument("--limit", type=int, default=2048, help="For debugging: number of samples to materialize")

    # Training
    ap.add_argument("--epochs", type=int, default=1)
    ap.add_argument("--batch_size", type=int, default=1)
    ap.add_argument("--grad_accum", type=int, default=16)
    ap.add_argument("--lr", type=float, default=1e-4)
    ap.add_argument("--warmup_ratio", type=float, default=0.03)
    ap.add_argument("--use_lora", action="store_true")
    ap.add_argument("--lora_r", type=int, default=16)
    ap.add_argument("--lora_alpha", type=int, default=32)
    ap.add_argument("--lora_dropout", type=float, default=0.05)

    ap.add_argument(
        "--manthan_model",
        type=str,
        required=True,
        help="HF repo id or local path that contains Manthan remote-code (the thing you push to HF).",
    )
    ap.add_argument("--save_every", type=int, default=500)
    ap.add_argument("--dry_run", action="store_true", help="Run a single synthetic step (no datasets).")

    args = ap.parse_args()

    os.makedirs(args.output_dir, exist_ok=True)

    device = "cuda" if torch.cuda.is_available() else "cpu"
    if device != "cuda":
        print("WARNING: This script is designed for CUDA (Kaggle). Running on CPU will be extremely slow.")

    # Tokenizer (use the LLM tokenizer)
    tok = AutoTokenizer.from_pretrained(args.text_model, trust_remote_code=True, use_fast=False)
    if tok.pad_token_id is None:
        tok.pad_token = tok.eos_token

    # Load Manthan remote-code model
    # (This should contain config that points to your desired text_model_id & vision_model_id.)
    model = AutoModelForCausalLM.from_pretrained(
        args.manthan_model,
        trust_remote_code=True,
        torch_dtype=torch.float16 if device == "cuda" else None,
    )

    model.train()
    model.to(device)

    # Make sure we don't train the vision tower (T4-friendly)
    if hasattr(model, "vision_model") and model.vision_model is not None:
        set_requires_grad(model.vision_model, False)
    if hasattr(model, "vision_tower") and model.vision_tower is not None:
        set_requires_grad(model.vision_tower, False)

    # Train projector always
    if hasattr(model, "projector"):
        set_requires_grad(model.projector, True)

    # Add LoRA to the language model (recommended)
    maybe_add_lora_to_model(model, args)

    # Optimizer params = trainable only
    trainable_params = [p for p in model.parameters() if p.requires_grad]
    if len(trainable_params) == 0:
        raise RuntimeError("No trainable parameters. Did you freeze everything?")

    optim = torch.optim.AdamW(trainable_params, lr=args.lr, betas=(0.9, 0.95), weight_decay=0.01)

    # Data
    if args.dry_run:
        # Minimal synthetic batch (no images on disk). This just validates loss pathway.
        B, T = 1, min(64, args.max_length)

        # IMPORTANT: some tokenizers report an imprecise `vocab_size`; `len(tok)` is the safe upper bound.
        tok_vocab = int(len(tok))
        input_ids = torch.randint(low=0, high=max(tok_vocab - 1, 1), size=(B, T), dtype=torch.long)
        labels = input_ids.clone()
        attn = torch.ones_like(input_ids)
        pixel_values = torch.randn(B, 3, args.image_size, args.image_size)

        # Insert one image placeholder
        input_ids[0, 5] = IMAGE_TOKEN_INDEX
        labels[0, :10] = IGNORE_INDEX

        # If tokenizer vocab > model vocab (common in dry_run), clamp to avoid CE index errors.
        lm_vocab = None
        try:
            if hasattr(model, "language_model") and hasattr(model.language_model, "config"):
                lm_vocab = int(getattr(model.language_model.config, "vocab_size", 0) or 0)
        except Exception:
            lm_vocab = None

        if lm_vocab and lm_vocab > 0:
            safe_ids = input_ids.clone()
            mask = safe_ids >= 0
            safe_ids[mask] = safe_ids[mask].clamp(min=0, max=lm_vocab - 1)
            input_ids = safe_ids

            safe_labels = labels.clone()
            mask = safe_labels >= 0
            safe_labels[mask] = safe_labels[mask].clamp(min=0, max=lm_vocab - 1)
            labels = safe_labels

        batch = {
            "input_ids": input_ids.to(device),
            "labels": labels.to(device),
            "attention_mask": attn.to(device),
            "pixel_values": pixel_values.to(device),
        }
        out = model(**batch)
        print("dry_run loss:", float(out.loss))
        out.loss.backward()
        optim.step()
        optim.zero_grad(set_to_none=True)
        save_projector(model, args.output_dir)
        if hasattr(model, "language_model") and hasattr(model.language_model, "save_pretrained"):
            # Save adapters if present
            try:
                model.language_model.save_pretrained(args.output_dir)
            except Exception:
                pass
        return 0

    ds = LlavaLikeDataset(args.dataset, split="train", tokenizer=tok, max_length=args.max_length, limit=args.limit)
    from torch.utils.data import DataLoader

    dl = DataLoader(
        ds,
        batch_size=args.batch_size,
        shuffle=True,
        num_workers=2,
        collate_fn=lambda b: collate_fn(b, args.image_size),
    )

    total_steps = (len(dl) * args.epochs) // max(1, args.grad_accum)
    warmup_steps = max(1, int(total_steps * args.warmup_ratio))
    sched = get_cosine_schedule_with_warmup(optim, warmup_steps, total_steps)

    step = 0
    optim.zero_grad(set_to_none=True)
    for epoch in range(args.epochs):
        for micro_idx, batch in enumerate(dl):
            batch = {k: v.to(device) for k, v in batch.items()}

            # Mixed precision on Kaggle
            with torch.autocast(device_type="cuda", dtype=torch.float16, enabled=(device == "cuda")):
                out = model(**batch)
                loss = out.loss / max(1, args.grad_accum)

            loss.backward()

            if (micro_idx + 1) % args.grad_accum == 0:
                torch.nn.utils.clip_grad_norm_(trainable_params, 1.0)
                optim.step()
                sched.step()
                optim.zero_grad(set_to_none=True)
                step += 1

                if step % 10 == 0:
                    print(f"epoch={epoch} step={step}/{total_steps} loss={float(out.loss):.4f}")

                if step % args.save_every == 0:
                    save_projector(model, args.output_dir)
                    # Save adapters if any
                    try:
                        model.save_pretrained(args.output_dir)
                    except Exception:
                        pass

            if step >= total_steps:
                break

    save_projector(model, args.output_dir)
    try:
        model.save_pretrained(args.output_dir)
    except Exception:
        pass

    print("DONE")
    return 0


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