File size: 26,568 Bytes
ffcfc75
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
import os
import re

import pathlib
from argparse import ArgumentParser
from typing import List, Dict, Optional
from dataclasses import dataclass, field

import torch
from torch import nn
import torch.nn.functional as F
from torch.optim import AdamW
from torch.utils.data import DataLoader, Dataset
from transformers import get_cosine_schedule_with_warmup, AutoTokenizer

from transformers import (
    AutoTokenizer,
    AutoModelForCausalLM,
    AutoModelForMaskedLM,
    AutoProcessor,
)

from datasets import load_dataset, DatasetDict

from peft import get_peft_model, LoraConfig, prepare_model_for_kbit_training
from transformers import BitsAndBytesConfig

import pytorch_lightning as pl
from pytorch_lightning.callbacks import ModelCheckpoint, LearningRateMonitor
from pytorch_lightning.loggers import WandbLogger

from trl import GRPOConfig, GRPOTrainer, ModelConfig, ScriptArguments, TrlParser, get_peft_config
#from unsloth import FastLanguageModel, is_bfloat16_supported

from bioreason.models.dna_llm import DNALLMModel
from bioreason.dna_modules import NucleotideDNAModule
from bioreason.models.dl.processing_dl import DLProcessor
from bioreason.trainer import DNALLMGRPOTrainer, DNALLMGRPOConfig
from bioreason.models.evo2_tokenizer import Evo2Tokenizer, register_evo2_tokenizer
register_evo2_tokenizer()

# Custom TrainerCallback to override the saving mechanism
from transformers import TrainerCallback, TrainerState, TrainerControl
from transformers.trainer_utils import PREFIX_CHECKPOINT_DIR

class SaveWithPyTorchCallback(TrainerCallback):
    """Custom callback to save models with PyTorch's native save mechanism instead of safetensors"""
    def on_save(self, args, state, control, **kwargs):
        # Get the checkpoint folder
        checkpoint_folder = os.path.join(
            args.output_dir, f"{PREFIX_CHECKPOINT_DIR}-{state.global_step}"
        )
        os.makedirs(checkpoint_folder, exist_ok=True)
        
        # Save with PyTorch instead of safetensors
        checkpoint_path = os.path.join(checkpoint_folder, "pytorch_model.bin")
        model = kwargs.get("model")
        
        # Get model unwrapped from accelerator etc.
        unwrapped_model = model.module if hasattr(model, "module") else model
        
        # Save using PyTorch directly
        torch.save(unwrapped_model.state_dict(), checkpoint_path)
        
        # DNALLMModel doesn't have a direct config attribute, so we need to save
        # the configs of its sub-models
        if hasattr(unwrapped_model, "text_model"):
            if hasattr(unwrapped_model.text_model, "config"):
                unwrapped_model.text_model.config.save_pretrained(checkpoint_folder)
            # Handle PEFT models which might have base_model
            elif hasattr(unwrapped_model.text_model, "base_model") and hasattr(unwrapped_model.text_model.base_model, "config"):
                unwrapped_model.text_model.base_model.config.save_pretrained(checkpoint_folder)
        
        # Print info about what's being saved
        print(f"Saved model checkpoint to {checkpoint_folder}")
        lora_params = [k for k in unwrapped_model.state_dict().keys() if "lora" in k]
        print(f"Checkpoint contains {len(lora_params)} LoRA parameters")
        
        # Signal that we've saved
        control.should_save = False
        return control

def _get_target_modules(model: DNALLMModel):
    # Apply LoRA to all linear layers in the text model
    target_modules = []

    # Get all unique linear layer names
    seen_names = set()
    for name, module in model.text.named_modules():
        if isinstance(module, torch.nn.Linear):
            names = name.split(".")
            target_name = names[-1]  # Use the last part of the name

            # Skip output head but include all other linear layers
            if target_name != "lm_head" and target_name not in seen_names:
                target_modules.append(target_name)
                seen_names.add(target_name)

    # Add attention-specific layers
    attention_patterns = [
        "q_proj",
        "k_proj",
        "v_proj",
        "out_proj",
        "query",
        "key",
        "value",
    ]
    for pattern in attention_patterns:
        if pattern not in seen_names:
            target_modules.append(pattern)

    # Return all unique layer names to apply LoRA to all layers
    return list(target_modules)


def extract_xml_answer(text: str) -> str:
    # answer = text.split("<answer>")[-1]
    # answer = answer.split("</answer>")[0]
    answer = text.split("</think>")[-1]
    return answer.strip()

def extract_hash_answer(text: str) -> str | None:
    if "####" not in text:
        return None
    return text.split("####")[1].strip()

def get_kegg_questions() -> Dataset:
    data = load_dataset('wanglab/kegg', 'default') # type: ignore
    example_dna_sequences = ["ATCTACATGCAT", "CAGCAGCTACAG", "CATCACATCGACATCGAC"]
    num_dna_sequences = 2 # TODO: Change to 2!

    data = data.map(lambda x: { # type: ignore
        'prompt': [
     
            {
                'role': 'user',
                'content': [
                    *({'type': 'dna', 'text': None} for _ in range(num_dna_sequences)),
                    {'type': 'text', 'text': x['question']},
                ],
            },
        ],
        'dna_sequences': [x['reference_sequence'], x['variant_sequence']],
        'answer': x['answer'],
    })  # type: ignore

    return data

# uncomment middle messages for 1-shot prompting
def get_gsm8k_questions(question_prompt: str) -> Dataset:
    data = load_dataset('openai/gsm8k', 'main') # type: ignore

    example_dna_sequences = ["ATCTACATGCAT", "CAGCAGCTACAG", "CATCACATCGACATCGAC"]
    data = data.map(lambda x: { # type: ignore
        'prompt': [ 

            {
                'role': 'user',
                'content': [
                    *({'type': 'dna', 'text': None} for _ in range(len(example_dna_sequences))),
                    {'type': 'text', 'text': 'Give me a short introduction to large language model.'}
                ]
            },
        ],
        'dna_sequences': [dna for dna in example_dna_sequences],
        'answer': extract_hash_answer(x['answer']),
    }) # type: ignore
    
    return data # type: ignore

def get_gsm8k_questions_old(question_prompt: str) -> Dataset:
    data = load_dataset('openai/gsm8k', 'main') # type: ignore

    example_dna_sequences = ["ATCTACATGCAT", "CAGCAGCTACAG", "CATCACATCGACATCGAC"]
    data = data.map(lambda x: { # type: ignore
        'prompt': [
            {
                'role': 'user',
                'content': [
                    *({'type': 'dna', 'text': None} for _ in range(len(example_dna_sequences))),
                    {'type': 'text', 'text': question_prompt.format(Question=x['question'])}
                ]
            },
        ],
        'dna_sequences': [dna for dna in example_dna_sequences],
        'answer': extract_hash_answer(x['answer']),
    }) # type: ignore
    
    return data # type: ignore

# Reward functions
def correctness_reward_func(prompts, completions, answer, **kwargs) -> list[float]:
    responses = [completion[0]['content'] for completion in completions]
    q = prompts[0][-1]['content']
    extracted_responses = [extract_xml_answer(r) for r in responses]
    # extracted_responses = [r.lower().replace("answer:", "").strip() for r in extracted_responses]
    print('-'*20, f"Question:\n{q}", f"\nAnswer:\n{answer[0]}", f"\nResponse:\n{responses[0]}", f"\nExtracted:\n{extracted_responses[0]}")
    return [2.0 if a.lower() in r.lower() else 0.0 for r, a in zip(extracted_responses, answer[0])]

def less_than_4_reward_func(completions, **kwargs) -> list[float]:
    responses = [completion[0]['content'] for completion in completions]
    extracted_responses = [extract_xml_answer(r) for r in responses]
    return [0.5 if len(r.split(' ')) <= 4 else 0.0 for r in extracted_responses]

def strict_format_reward_func(completions, **kwargs) -> list[float]:
    """Reward function that checks if the completion has a specific format."""
    pattern = r"^<think>\n.*?\n</think>\n.*?\n$"
    responses = [completion[0]["content"] for completion in completions]
    matches = [re.match(pattern, r) for r in responses]
    return [0.5 if match else 0.0 for match in matches]

def soft_format_reward_func(completions, **kwargs) -> list[float]:
    """Reward function that checks if the completion has a specific format."""
    pattern = r"<think>.*?</think>\s*.*?"
    responses = [completion[0]["content"] for completion in completions]
    matches = [re.match(pattern, r) for r in responses]
    return [0.5 if match else 0.0 for match in matches]

def count_xml(text) -> float:
    count = 0.0
    if text.count("<think>\n") == 1:
        count += 0.125
    if text.count("\n</think>\n") == 1:
        count += 0.125
    return count

def xmlcount_reward_func(completions, **kwargs) -> list[float]:
    contents = [completion[0]["content"] for completion in completions]
    return [count_xml(c) for c in contents]

# Format into conversation
def make_conversation(example):
    return {
        "prompt": [
            {"role": "system", "content": SYSTEM_PROMPT},
            {"role": "user", "content": example["problem"]},
        ],
    }

def make_conversation_image(example):
    return {
        "prompt": [
            {
                "role": "user",
                "content": [
                    {"type": "image"},
                ],
            },
        ],
    }

@dataclass
# class GRPOModelConfig(ModelConfig):
    
#     # "HuggingFaceTB/SmolLM-135M-Instruct"
#     # "Qwen/Qwen2.5-0.5B-Instruct"
#     model_name_or_path: str = field(default="Qwen/Qwen3-0.6B", metadata={"help": "Model checkpoint for weights initialization."})
#     dna_model_name_or_path: str = field(default="InstaDeepAI/nucleotide-transformer-v2-100m-multi-species", metadata={"help": "Model checkpoint for weights initialization."})
#     cache_dir: str = field(default=None, metadata={"help": "Path to model cache directory."})
#     max_length_text: int = field(default=800, metadata={"help": "Maximum length of text sequences."})
#     max_length_dna: int = field(default=800, metadata={"help": "Maximum length of DNA sequences, in groups of 6 nucleotides."})
#     sft_checkpoint: str = field(default=None, metadata={"help": "Path to the checkpoint for SFT."})
#     lora_r: int = field(default=32, metadata={"help": "LoRA R value."})
#     lora_alpha: int = field(default=64, metadata={"help": "LoRA alpha."})
#     lora_dropout: float = field(default=0.05, metadata={"help": "LoRA dropout."})
#     lora_modules_to_save: Optional[list[str]] = field(
#         default="embed_tokens",
#         metadata={"help": "Model layers to unfreeze & train."},
#     )
#     freeze_dna_modules: bool = False

class GRPOModelConfig(ModelConfig):

    model_name_or_path: str = field(default="Qwen/Qwen3-0.6B", metadata={"help": "Model checkpoint for LLM weights initialization."})
    protein_model_name_or_path: str = field(default="esm2_t33_650M_UR50D", metadata={"help": "Model checkpoint for ESM-2 protein weights initialization."})
    cache_dir: str = field(default=None, metadata={"help": "Path to model cache directory."})
    max_length_text: int = field(default=800, metadata={"help": "Maximum length of text sequences."})
    max_length_protein: int = field(default=800, metadata={"help": "Maximum length of protein sequences (number of amino acids)."})
    sft_checkpoint: str = field(default=None, metadata={"help": "Path to the checkpoint for SFT."})
    lora_r: int = field(default=32, metadata={"help": "LoRA R value."})
    lora_alpha: int = field(default=64, metadata={"help": "LoRA alpha."})
    lora_dropout: float = field(default=0.05, metadata={"help": "LoRA dropout."})
    lora_modules_to_save: Optional[list[str]] = field(
        default_factory=lambda: ["embed_tokens", "lm_head"],
        metadata={"help": "Model layers to unfreeze & train with LoRA."},
    )
    
    # Updated: Renamed `freeze_dna_modules` to `freeze_protein_model`
    freeze_protein_model: bool = field(default=True, metadata={"help": "Whether to freeze the ESM-2 protein model during training."})
    num_query_tokens: int = field(default=32, metadata={"help": "The number of query tokens used by the Q-Former to summarize protein features. These tokens will be injected into the LLM input."})
    # New: Parameters for the projector layer
    projector_hidden_size: int = field(default=1280, metadata={"help": "Hidden size of the projector layer. It should match the ESM-2's output hidden size."})
    projector_output_size: int = field(default=1024, metadata={"help": "Output size of the projector layer. It should match the LLM's hidden size."})
    
    # New: Parameter to control projector training
    freeze_projector: bool = field(default=False, metadata={"help": "Whether to freeze the projector layer during training."})

@dataclass
class GRPOScriptArguments(ScriptArguments):
    """
    Script arguments for the GRPO training script.
    """
    dataset_name: str = field(default="wanglab/kegg", metadata={"help": "Dataset name with default."})
    data_file_paths: str = field(
        default=None,
        metadata={"help": "Paths to data files, separated by ':'"},
    )
    arrow_cache_dir: str = field(
        default=None,
        metadata={"help": "Path to arrow cache directory"},
    )
    val_split_ratio: float = field(
        default=0.0,
        metadata={"help": "Ratio of validation split, default 0.0"},
    )
    reward_funcs: list[str] = field(
        #default_factory=lambda: ["accuracy", "format"],
        default_factory=lambda: ["xmlcount", "soft_format", "strict_format", "less_than_4", "correctness"],
        #metadata={"help": "List of reward functions. Possible values: 'accuracy', 'format'"},
        metadata={"help": "List of reward functions. Possible values: 'accuracy', 'xmlcount', 'soft_format', 'strict_format', 'less_than_4', 'correctness'"},
    )
    # max_pixels: Optional[int] = field(
    #     default=12845056,
    #     metadata={"help": "Maximum number of pixels for the image (for QwenVL)"},
    # )
    # min_pixels: Optional[int] = field(
    #     default=3136,
    #     metadata={"help": "Minimum number of pixels for the image (for QwenVL)"},
    # )
    # task_type: Optional[str] = field(
    #     default=None,
    #     metadata={"help": "Choose task type: 'default', 'gui', ..."},
    # )



reward_funcs_registry = {
    # "accuracy": accuracy_reward,
    # "format": format_reward,
    "xmlcount": xmlcount_reward_func,
    "soft_format": soft_format_reward_func,
    "strict_format": strict_format_reward_func,
    "less_than_4": less_than_4_reward_func,
    "correctness": correctness_reward_func,
}

def get_vlm_module(model_name_or_path):
    if any(mini_name in model_name_or_path.lower() for mini_name in ["qwen", "smol"]):
        return NucleotideDNAModule
    else:
        raise ValueError(f"Unsupported model: {model_name_or_path}")
    
def _get_target_modules(model):
    # Apply LoRA to all linear layers in the text model
    target_modules = []

    # Get all unique linear layer names
    seen_names = set()
    for name, module in model.text_model.named_modules():
        if isinstance(module, torch.nn.Linear):
            names = name.split(".")
            target_name = names[-1]  # Use the last part of the name

            # Skip output head but include all other linear layers
            if target_name != "lm_head" and target_name not in seen_names:
                target_modules.append(target_name)
                seen_names.add(target_name)

    # Add attention-specific layers
    attention_patterns = [
        "q_proj",
        "k_proj",
        "v_proj",
        "out_proj",
        "query",
        "key",
        "value",
    ]
    for pattern in attention_patterns:
        if pattern not in seen_names:
            target_modules.append(pattern)

    # Return all unique layer names to apply LoRA to all layers
    return list(target_modules)


def _prep_for_training(model, training_args, dna_model_finetune: bool = False) -> LoraConfig:
    """
    Load and configure the DNALLMModel.
    """

    # Freeze DNA encoder parameters
    if dna_model_finetune:
        pass
    else:
        for param in model.dna_model.parameters():
            param.requires_grad = False

    target_modules = _get_target_modules(model)

    lora_config = LoraConfig(
        r=training_args.lora_r,
        lora_alpha=training_args.lora_alpha,
        lora_dropout=training_args.lora_dropout,
        target_modules=target_modules,
        init_lora_weights="gaussian",
        bias="none",
        task_type="CAUSAL_LM",
    )

    # Prepare text model for training
    model.text_model = prepare_model_for_kbit_training(model.text_model)
    model.text_model = get_peft_model(model.text_model, lora_config)

    # Make projection layer trainable
    for param in model.dna_projection.parameters():
        param.requires_grad = True

    return lora_config

def main(script_args, training_args, model_args):

    print(training_args.output_dir)
    #pl.seed_everything(args.seed)
    # os.environ["CUDA_LAUNCH_BLOCKING"] = "1"
    torch.cuda.empty_cache()
    torch.set_float32_matmul_precision("medium")

    # Initialize model
    # Load tokenizer for target text
    # tokenizer = AutoTokenizer.from_pretrained(model_args.model_name_or_path)
    # tokenizer.pad_token = tokenizer.eos_token

    # Load model
    model = DNALLMModel(
        text_model_name=model_args.model_name_or_path,
        dna_model_name=model_args.dna_model_name_or_path,
        cache_dir=model_args.cache_dir,
        max_length_text=model_args.max_length_text,
        max_length_dna=model_args.max_length_dna,
        text_model_finetune=True,
        dna_model_finetune=not model_args.freeze_dna_modules,
        debug=False,
    )

    # load checkpoint
    if model_args.sft_checkpoint is not None:
        print(f"Loading SFT checkpoint from {model_args.sft_checkpoint}")
        
        # Determine if it's a directory (PEFT format) or file (PyTorch state dict)
        is_directory = os.path.isdir(model_args.sft_checkpoint)
        
        if is_directory:
            # It's a PEFT checkpoint directory - load properly with PEFT
            from peft import PeftModel
            
            # First initialize the text model with PEFT
            print("Loading as PEFT checkpoint directory")
            model.text_model = PeftModel.from_pretrained(
                model.text_model,
                model_args.sft_checkpoint,
                is_trainable=True
            )
            
            # Verify loaded adapters
            print("Loaded LoRA adapters:", model.text_model.active_adapter)
            
            # Optional: Merge weights into base model
            print("Merging SFT LoRA weights into base model...")
            model.text_model = model.text_model.merge_and_unload()
            print("Successfully merged SFT knowledge into base model")
            
        else:
            # It's a PyTorch state dict file
            print("Loading as PyTorch state dict file")
            checkpoint = torch.load(model_args.sft_checkpoint)
            
            # replace model.text_model with text_model for all in state dict
            def new_key(k):
                if k.startswith("=model."): return k[6:]
                elif k.startswith("_forward_module."): return k[len("_forward_module."):]
                else: return k
            
            if "state_dict" in checkpoint:
                magic = {new_key(k): v for k, v in checkpoint["state_dict"].items()}
            elif "module" in checkpoint:
                magic = {new_key(k): v for k, v in checkpoint["module"].items()}
            elif isinstance(checkpoint, dict) and all(isinstance(k, str) for k in checkpoint.keys()):
                # Direct state dict - the checkpoint itself is the state dict
                print("Detected direct state dict format")
                magic = {new_key(k): v for k, v in checkpoint.items()}
            else:
                raise ValueError(f"Unsupported checkpoint format: {model_args.sft_checkpoint}")
            
            # Handle prefix mapping for different model architectures
            lora_prefix = False
            for key in magic.keys():
                if "lora" in key:
                    lora_prefix = True
                    break
            
            if lora_prefix:
                print("Detected LoRA weights in state dict")
                # First prepare model for LoRA training
                _prep_for_training(model, model_args, dna_model_finetune=model_args.freeze_dna_modules)
                
                # Print some diagnostic info about the keys
                model_keys = set(model.state_dict().keys())
                checkpoint_keys = set(magic.keys())
                print(f"Model has {len(model_keys)} keys")
                print(f"Checkpoint has {len(checkpoint_keys)} keys")
                
                # Try to map LoRA keys more intelligently
                new_magic = {}
                for k, v in magic.items():
                    # Try different prefix mappings based on common patterns
                    if "base_model.model" in k and k not in model_keys:
                        new_k = k.replace("text_model.base_model.model", "text_model")
                        if new_k in model_keys:
                            new_magic[new_k] = v
                            continue
                    
                    # Try removing common prefixes
                    if k.startswith("text_model.") and k not in model_keys:
                        new_k = "text_model.base_model.model." + k[len("text_model."):]
                        if new_k in model_keys:
                            new_magic[new_k] = v
                            continue
                    
                    # Keep original key if no mapping found
                    new_magic[k] = v
                
                # Include missing target modules in diagnostic info
                magic = new_magic
                print(f"After key mapping: {len(magic)} keys")
                
                # Then load weights, allowing missing/extra keys
                result = model.load_state_dict(magic, strict=False)
                
                if len(result.unexpected_keys) > 0:
                    print(f"Sample unexpected keys: {result.unexpected_keys[:5]}")
                if len(result.missing_keys) > 0:
                    print(f"Sample missing keys: {result.missing_keys[:5]}")
                    
                print(f"Loaded checkpoint with {len(result.missing_keys)} missing keys and {len(result.unexpected_keys)} unexpected keys")
            else:
                print("Standard weights detected - remapping keys")
                # Map keys to model structure
                magic = {k.replace("text_model", "text_model.base_model.model"): v for k, v in magic.items()}
                magic = {k.replace("dna_model", "dna_model"): v for k, v in magic.items()}
                
                # Fix the shared memory tensors issue by making a copy of weights
                for key in list(magic.keys()):
                    if 'lm_head.weight' in key:
                        magic[key] = magic[key].clone()
                
                # Load weights before setting up LoRA
                result = model.load_state_dict(magic, strict=False)
                print(f"Loaded checkpoint with {len(result.missing_keys)} missing keys and {len(result.unexpected_keys)} unexpected keys")
                
                # Now prepare for LoRA training
                _prep_for_training(model, model_args, dna_model_finetune=model_args.freeze_dna_modules)
    else:
        # No checkpoint, just prepare for training
        _prep_for_training(model, model_args, dna_model_finetune=model_args.freeze_dna_modules)

    # Get reward functions
    reward_funcs = [reward_funcs_registry[func] for func in script_args.reward_funcs]
    # reward_funcs = [
    #         xmlcount_reward_func,
    #         soft_format_reward_func,
    #         strict_format_reward_func,
    #         int_reward_func,
    #         correctness_reward_func,
    #     ]
    print("reward_funcs:", reward_funcs)

    vlm_module_cls = get_vlm_module(model_args.model_name_or_path)
    print("using vlm module:", vlm_module_cls.__name__)
    question_prompt = vlm_module_cls.get_question_template()


    dataset = get_kegg_questions()
    
    #dataset = get_gsm8k_questions(question_prompt)

    print(dataset)

    #print('ITEM ONE OF THE DATASET', dataset['train'][0])

    # Custom callback to handle saving with PyTorch's native mechanism
    custom_save_callback = SaveWithPyTorchCallback()

    # Initialize the GRPO trainer with custom callback
    trainer = DNALLMGRPOTrainer(
        model=model,
        reward_funcs=reward_funcs,
        args=training_args,
        dna_module=vlm_module_cls(),
        train_dataset=dataset['train'],
        eval_dataset=dataset['val'] if training_args.eval_strategy != "no" else None,
        peft_config=get_peft_config(model_args),
        attn_implementation=model_args.attn_implementation,
        torch_dtype=model_args.torch_dtype,
        callbacks=[custom_save_callback],  # Add our custom callback
    )

    # Set the trainer to save in PyTorch format instead of safetensors
    training_args.save_safetensors = False

    # Train and push the model to the Hub
    # if list(pathlib.Path(training_args.output_dir).glob("checkpoint-*")):
    #     trainer.train(resume_from_checkpoint=True)
    # else:
    #     trainer.train()

    # Train and push the model to the Hub
    trainer.train()


if __name__ == "__main__":
    # os.environ["CUDA_VISIBLE_DEVICES"] = "0,1"
    print(f"CUDA_VISIBLE_DEVICES: {os.environ.get('CUDA_VISIBLE_DEVICES')}")
    parser = TrlParser((GRPOScriptArguments, DNALLMGRPOConfig, GRPOModelConfig))
    script_args, training_args, model_args = parser.parse_args_and_config()
    
    # Ensure we use PyTorch's save mechanism instead of safetensors
    training_args.save_safetensors = False

    main(script_args, training_args, model_args)
    
    # parser.add_argument("--wandb_project", type=str, default="dna-text-finetune")
    # parser.add_argument("--wandb_entity", type=str, default="adibvafa")

    # args = parser.parse_args()