File size: 29,701 Bytes
6ecf14b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Multiple GPUS"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "\n",
    "import torch\n",
    "import torch.nn.functional as F\n",
    "from torch.utils.data import Dataset, DataLoader\n",
    "from datautils import MyTrainDataset\n",
    "import torch.multiprocessing as mp\n",
    "from torch.utils.data.distributed import DistributedSampler\n",
    "from torch.nn.parallel import DistributedDataParallel as DDP\n",
    "from torch.distributed import init_process_group, destroy_process_group\n",
    "import os\n",
    "import argparse\n",
    "from datasets import load_dataset\n",
    "from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig\n",
    "from peft import LoraConfig\n",
    "from trl import SFTTrainer\n",
    "from transformers import TrainingArguments\n",
    "\n",
    "def ddp_setup(rank, world_size):\n",
    "    os.environ[\"MASTER_ADDR\"] = \"localhost\"\n",
    "    os.environ[\"MASTER_PORT\"] = \"12355\"\n",
    "    init_process_group(backend=\"nccl\", rank=rank, world_size=world_size)\n",
    "    torch.cuda.set_device(rank)\n",
    "\n",
    "class Trainer:\n",
    "    def __init__(self, model, train_data, optimizer, gpu_id, save_every):\n",
    "        self.gpu_id = gpu_id\n",
    "        self.model = model.to(gpu_id)\n",
    "        self.train_data = train_data\n",
    "        self.optimizer = optimizer\n",
    "        self.save_every = save_every\n",
    "      f.model = DDP(model, device_ids=[gpu_id])\n",
    "\n",
    "    def _run_batch(self, source, targets):\n",
    "        self.optimizer.zero_grad()\n",
    "        output = self.model(source)\n",
    "        loss = F.cross_entropy(output, targets)\n",
    "        loss.backward()\n",
    "        self.optimizer.step()\n",
    "\n",
    "    def _run_epoch(self, epoch):\n",
    "        b_sz = len(next(iter(self.train_data))[0])\n",
    "        print(f\"[GPU{self.gpu_id}] Epoch {epoch} | Batchsize: {b_sz} | Steps: {len(self.train_data)}\")\n",
    "        self.train_data.sampler.set_epoch(epoch)\n",
    "        for source, targets in self.train_data:\n",
    "            source = source.to(self.gpu_id)\n",
    "            targets = targets.to(self.gpu_id)\n",
    "            self._run_batch(source, targets)\n",
    "\n",
    "    def _save_checkpoint(self, epoch):\n",
    "        ckp = self.model.module.state_dict()\n",
    "        PATH = \"checkpoint.pt\"\n",
    "        torch.save(ckp, PATH)\n",
    "        print(f\"Epoch {epoch} | Training checkpoint saved at {PATH}\")\n",
    "\n",
    "    def train(self, max_epochs):\n",
    "        for epoch in range(max_epochs):\n",
    "            self._run_epoch(epoch)\n",
    "            if self.gpu_id == 0 and epoch % self.save_every == 0:\n",
    "                self._save_checkpoint(epoch)\n",
    "\n",
    "def load_train_objs():\n",
    "    dataset_name = \"ruslanmv/ai-medical-dataset\"\n",
    "    dataset = load_dataset(dataset_name, split=\"train\")\n",
    "    dataset = dataset.select(range(100))\n",
    "\n",
    "    model_name = \"meta-llama/Meta-Llama-3-8B-Instruct\"\n",
    "    tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)\n",
    "    tokenizer.pad_token = tokenizer.eos_token\n",
    "\n",
    "    bnb_config = BitsAndBytesConfig(\n",
    "        load_in_4bit=True,\n",
    "        bnb_4bit_quant_type=\"nf4\",\n",
    "        bnb_4bit_compute_dtype=torch.float16,\n",
    "    )\n",
    "    model = AutoModelForCausalLM.from_pretrained(\n",
    "        model_name,\n",
    "        quantization_config=bnb_config,\n",
    "        trust_remote_code=True,\n",
    "        use_cache=False,\n",
    "        device_map=\"auto\",\n",
    "    )\n",
    "\n",
    "    lora_alpha = 16\n",
    "    lora_dropout = 0.1\n",
    "    lora_r = 32\n",
    "    peft_config = LoraConfig(\n",
    "        lora_alpha=lora_alpha,\n",
    "        lora_dropout=lora_dropout,\n",
    "        r=lora_r,\n",
    "        bias=\"none\",\n",
    "        task_type=\"CAUSAL_LM\",\n",
    "        target_modules=[\"k_proj\", \"q_proj\", \"v_proj\", \"up_proj\", \"down_proj\", \"gate_proj\"],\n",
    "        modules_to_save=[\"embed_tokens\", \"input_layernorm\", \"post_attention_layernorm\", \"norm\"],\n",
    "    )\n",
    "\n",
    "    max_seq_length = 512\n",
    "    output_dir = \"./results\"\n",
    "    per_device_train_batch_size = 2\n",
    "    gradient_accumulation_steps = 2\n",
    "    optim = \"adamw_torch\"\n",
    "    save_steps = 10\n",
    "    logging_steps = 1\n",
    "    learning_rate = 2e-4\n",
    "    max_grad_norm = 0.3\n",
    "    max_steps = 1\n",
    "    warmup_ratio = 0.1\n",
    "    lr_scheduler_type = \"cosine\"\n",
    "\n",
    "    training_arguments = TrainingArguments(\n",
    "        output_dir=output_dir,\n",
    "        per_device_train_batch_size=per_device_train_batch_size,\n",
    "        gradient_accumulation_steps=gradient_accumulation_steps,\n",
    "        optim=optim,\n",
    "        save_steps=save_steps,\n",
    "        logging_steps=logging_steps,\n",
    "        learning_rate=learning_rate,\n",
    "        fp16=True,\n",
    "        max_grad_norm=max_grad_norm,\n",
    "        max_steps=max_steps,\n",
    "        warmup_ratio=warmup_ratio,\n",
    "        group_by_length=True,\n",
    "        lr_scheduler_type=lr_scheduler_type,\n",
    "        gradient_checkpointing=True,\n",
    "    )\n",
    "\n",
    "    return dataset, model, peft_config, tokenizer, training_arguments\n",
    "\n",
    "def prepare_dataloader(dataset, batch_size):\n",
    "    return DataLoader(\n",
    "        dataset,\n",
    "        batch_size=batch_size,\n",
    "        pin_memory=True,\n",
    "        shuffle=False,\n",
    "        sampler=DistributedSampler(dataset),\n",
    "    )\n",
    "\n",
    "def main(rank, world_size, save_every, total_epochs, batch_size):\n",
    "    ddp_setup(rank, world_size)\n",
    "    dataset, model, peft_config, tokenizer, training_arguments = load_train_objs()\n",
    "    train_data = prepare_dataloader(dataset, batch_size)\n",
    "    trainer = SFTTrainer(\n",
    "        model=model,\n",
    "        train_dataset=dataset,\n",
    "        peft_config=peft_config,\n",
    "        dataset_text_field=\"context\",\n",
    "        max_seq_length=max_seq_length,\n",
    "        tokenizer=tokenizer,\n",
    "        args=training_arguments,\n",
    "    )\n",
    "    trainer = Trainer(model, train_data, optimizer=trainer.optimizer, gpu_id=rank, save_every=save_every)\n",
    "    trainer.train(total_epochs)\n",
    "    destroy_process_group()\n",
    "\n",
    "TOTAL_EPOCHS = 10\n",
    "SAVE_EVERY = 2\n",
    "BATCH_SIZE = 32\n",
    "\n",
    "if __name__ == \"__main__\":\n",
    "    world_size = torch.cuda.device_count()\n",
    "    mp.set_start_method(\"spawn\", force=True)  # Add this line\n",
    "    mp.spawn(main, args=(world_size, SAVE_EVERY, TOTAL_EPOCHS, BATCH_SIZE), nprocs=world_size)\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import torch\n",
    "import torch.nn.functional as F\n",
    "from torch.utils.data import Dataset, DataLoader\n",
    "import torch.multiprocessing as mp\n",
    "from torch.utils.data.distributed import DistributedSampler\n",
    "from torch.nn.parallel import DistributedDataParallel as DDP\n",
    "from torch.distributed import init_process_group, destroy_process_group\n",
    "import os\n",
    "import argparse\n",
    "from datasets import load_dataset\n",
    "from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig, AutoTokenizer\n",
    "from peft import LoraConfig\n",
    "from trl import SFTTrainer\n",
    "from transformers import TrainingArguments\n",
    "def ddp_setup(rank, world_size):\n",
    "    \"\"\"\n",
    "    Args:\n",
    "        rank: Unique identifier of each process\n",
    "        world_size: Total number of processes\n",
    "    \"\"\"\n",
    "    os.environ[\"MASTER_ADDR\"] = \"localhost\"\n",
    "    os.environ[\"MASTER_PORT\"] = \"12355\"\n",
    "    init_process_group(backend=\"nccl\", rank=rank, world_size=world_size)\n",
    "    torch.cuda.set_device(rank)\n",
    "\n",
    "class Trainer:\n",
    "    def __init__(self, model, train_data, optimizer, gpu_id, save_every):\n",
    "        self.gpu_id = gpu_id\n",
    "        self.model = model.to(gpu_id)\n",
    "        self.train_data = train_data\n",
    "        self.optimizer = optimizer\n",
    "        self.save_every = save_every\n",
    "        self.model = DDP(model, device_ids=[gpu_id])\n",
    "\n",
    "    def _run_batch(self, source, targets):\n",
    "        self.optimizer.zero_grad()\n",
    "        output = self.model(source)\n",
    "        loss = F.cross_entropy(output, targets)\n",
    "        loss.backward()\n",
    "        self.optimizer.step()\n",
    "\n",
    "    def _run_epoch(self, epoch):\n",
    "        b_sz = len(next(iter(self.train_data))[0])\n",
    "        print(f\"[GPU{self.gpu_id}] Epoch {epoch} | Batchsize: {b_sz} | Steps: {len(self.train_data)}\")\n",
    "        self.train_data.sampler.set_epoch(epoch)\n",
    "        for source, targets in self.train_data:\n",
    "            source = source.to(self.gpu_id)\n",
    "            targets = targets.to(self.gpu_id)\n",
    "            self._run_batch(source, targets)\n",
    "\n",
    "    def _save_checkpoint(self, epoch):\n",
    "        ckp = self.model.module.statt()\n",
    "        PATH = \"checkpoint.pt\"\n",
    "        torch.save(ckp, PATH)\n",
    "        print(f\"Epoch {epoch} | Training checkpoint saved at {PATH}\")\n",
    "\n",
    "    def train(self, max_epochs):\n",
    "        for epoch in range(max_epochs):\n",
    "            self._run_epoch(epoch)\n",
    "            if self.gpu_id == 0 and epoch % self.save_every == 0:\n",
    "                self._save_checkpoint(epoch)\n",
    "\n",
    "def load_train_objs():\n",
    "    dataset_name = \"ruslanmv/ai-medical-dataset\"\n",
    "    dataset = load_dataset(dataset_name, split=\"train\")\n",
    "    dataset = dataset.select(range(100))\n",
    "\n",
    "    model_name = \"meta-llama/Meta-Llama-3-8B-Instruct\"\n",
    "    tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)\n",
    "    tokenizer.pad_token = tokenizer.eos_token\n",
    "\n",
    "    bnb_config = BitsAndBytesConfig(\n",
    "        load_in_4bit=True,\n",
    "        bnb_4bit_quant_type=\"nf4\",\n",
    "        bnb_4bit_compute_dtype=torch.float16,\n",
    "    )\n",
    "    model = AutoModelForCausalLM.from_pretrained(\n",
    "        model_name,\n",
    "        quantization_config=bnb_config,\n",
    "        trust_remote_code=True,\n",
    "        use_cache=False,\n",
    "        device_map=\"auto\",\n",
    "    )\n",
    "\n",
    "    lora_alpha = 16\n",
    "    lora_dropout = 0.1\n",
    "    lora_r = 32\n",
    "    peft_config = LoraConfig(\n",
    "        lora_alpha=lora_alpha,\n",
    "        lora_dropout=lora_dropout,\n",
    "        r=lora_r,\n",
    "        bias=\"none\",\n",
    "        task_type=\"CAUSAL_LM\",\n",
    "        target_modules=[\"k_proj\", \"q_proj\", \"v_proj\", \"up_proj\", \"down_proj\", \"gate_proj\"],\n",
    "        modules_to_save=[\"embed_tokens\", \"input_layernorm\", \"post_attention_layernorm\", \"norm\"],\n",
    "    )\n",
    "\n",
    "    max_seq_length = 512\n",
    "    output_dir = \"./results\"\n",
    "    per_device_train_batch_size = 2\n",
    "    gradient_accumulation_steps = 2\n",
    "    optim = \"adamw_torch\"\n",
    "    save_steps = 10\n",
    "    logging_steps = 1\n",
    "    learning_rate = 2e-4\n",
    "    max_grad_norm = 0.3\n",
    "    max_steps = 1\n",
    "    warmup_ratio = 0.1\n",
    "    lr_scheduler_type = \"cosine\"\n",
    "\n",
    "    training_arguments = TrainingArguments(\n",
    "        output_dir=output_dir,\n",
    "        per_device_train_batch_size=per_device_train_batch_size,\n",
    "        gradient_accumulation_steps=gradient_accumulation_steps,\n",
    "        optim=optim,\n",
    "        save_steps=save_steps,\n",
    "        logging_steps=logging_steps,\n",
    "        learning_rate=learning_rate,\n",
    "        fp16=True,\n",
    "        max_grad_norm=max_grad_norm,\n",
    "        max_steps=max_steps,\n",
    "        warmup_ratio=warmup_ratio,\n",
    "        group_by_length=True,\n",
    "        lr_scheduler_type=lr_scheduler_type,\n",
    "        gradient_checkpointing=True,\n",
    "    )\n",
    "\n",
    "    return dataset, model, peft_config, tokenizer, training_arguments\n",
    "\n",
    "def prepare_dataloader(dataset, batch_size):\n",
    "    return DataLoader(\n",
    "        dataset,\n",
    "        batch_size=batch_size,\n",
    "        pin_memory=True,\n",
    "        shuffle=False,\n",
    "        sampler=DistributedSampler(dataset),\n",
    "    )"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import torch\n",
    "import torch.multiprocessing as mp\n",
    "\n",
    "def main(rank, world_size):\n",
    "    # Define the parameters as constants\n",
    "    TOTAL_EPOCHS = 10\n",
    "    SAVE_EVERY = 2\n",
    "    BATCH_SIZE = 32\n",
    "    torch.cuda.init()\n",
    "    ddp_setup(rank, world_size)  \n",
    "    dataset, model, peft_config, tokenizer, training_arguments = load_train_objs()\n",
    "    train_data = prepare_dataloader(dataset, BATCH_SIZE)  # Corrected batch_size variable\n",
    "    trainer = SFTTrainer(\n",
    "        model=model,\n",
    "        train_dataset=dataset,\n",
    "        peft_config=peft_config,\n",
    "        dataset_text_field=\"context\",\n",
    "        max_seq_length=max_seq_length,\n",
    "        tokenizer=tokenizer,\n",
    "        args=training_arguments,\n",
    "    )\n",
    "    trainer = Trainer(model, train_data, optimizer=trainer.optimizer, gpu_id=rank, save_every=SAVE_EVERY)\n",
    "    trainer.train(TOTAL_EPOCHS)\n",
    "    destroy_process_group()\n",
    "\n",
    "if __name__ == \"__main__\":\n",
    "    mp.set_start_method('spawn')  # Set start method to 'spawn'\n",
    "    world_size = torch.cuda.device_count()\n",
    "\n",
    "    # Workaround for Jupyter Notebook and interactive environments\n",
    "    processes = []\n",
    "    for rank in range(world_size):\n",
    "        p = mp.Process(target=main, args=(rank, world_size))\n",
    "        p.start()\n",
    "        processes.append(p)\n",
    "\n",
    "    for p in processes:\n",
    "        p.join()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import torch\n",
    "from transformers import AutoModelForCausalLM, AutoTokenizer\n",
    "from datasets import load_dataset\n",
    "from trl import SFTTrainer\n",
    "from transformers import TrainingArguments\n",
    "import os\n",
    "import socket\n",
    "\n",
    "# Distributed training setup (assuming all GPUs are available on a single machine)\n",
    "def init_distributed(rank, world_size):\n",
    "    \"\"\"Initializes distributed training using `nccl` backend.\"\"\"\n",
    "    if rank == 0:\n",
    "        os.environ[\"MASTER_ADDR\"] = socket.gethostname()  # Set MASTER_ADDR using rank 0's hostname\n",
    "    else:\n",
    "        # Wait a bit to ensure MASTER_ADDR is set before other ranks try to use it\n",
    "        import time\n",
    "        time.sleep(5)\n",
    "    os.environ[\"MASTER_PORT\"] = \"12345\"  # Set MASTER_PORT environment variable\n",
    "    os.environ[\"RANK\"] = str(rank)  # Set RANK environment variable\n",
    "    os.environ[\"WORLD_SIZE\"] = str(world_size)  # Set WORLD_SIZE environment variable\n",
    "    torch.distributed.init_process_group(backend='nccl', init_method='env://')\n",
    "\n",
    "# Cleanup after training\n",
    "def cleanup_distributed():\n",
    "    if torch.distributed.is_initialized():\n",
    "        torch.distributed.destroy_process_group()\n",
    "\n",
    "# Model and tokenizer selection\n",
    "model_name = \"facebook/bart-base\"  # Replace with your desired model\n",
    "tokenizer = AutoTokenizer.from_pretrained(model_name)\n",
    "model = AutoModelForCausalLM.from_pretrained(model_name)\n",
    "\n",
    "# Dataset loading (replace with your dataset and field names)\n",
    "dataset = load_dataset(\"glue\", \"mnli\", split=\"train\")\n",
    "text_field = \"premise\"  # Assuming premise is the field containing text for prediction\n",
    "\n",
    "# Training arguments (adjust hyperparameters as needed)\n",
    "training_args = TrainingArguments(\n",
    "    output_dir=\"./results\",\n",
    "    per_device_train_batch_size=2,  # Adjust based on GPU memory (might need to adjust)\n",
    "    save_steps=500,\n",
    "    save_total_limit=2,\n",
    "    num_train_epochs=3,  # Adjust training time as needed\n",
    ")\n",
    "\n",
    "world_size = torch.cuda.device_count()\n",
    "if world_size > 1:\n",
    "    # Initialize distributed training\n",
    "    init_distributed(rank=0, world_size=world_size)  # Rank is assumed to be 0 here\n",
    "\n",
    "    # Wrap model in DDP for distributed training\n",
    "    model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[torch.cuda.current_device()])\n",
    "\n",
    "    # Create SFTTrainer with distributed settings\n",
    "    trainer = SFTTrainer(\n",
    "        model=model,\n",
    "        args=training_args,  # Pass training_args as 'args' instead of 'training_args'\n",
    "        train_dataset=dataset,\n",
    "        dataset_text_field=text_field,\n",
    "        compute_metrics=None,  # You can define your custom metrics here\n",
    "    )\n",
    "    print(\"Trainer For distributed training loaded\")\n",
    "else:\n",
    "    # For single-GPU training\n",
    "    trainer = SFTTrainer(\n",
    "        model=model,\n",
    "        args=training_args,  # Pass training_args as 'args' instead of 'training_args'\n",
    "        train_dataset=dataset,\n",
    "        dataset_text_field=text_field,\n",
    "        compute_metrics=None,  # You can define your custom metrics here\n",
    "    )\n",
    "    print(\"Trainer For single-GPU loaded\")\n",
    "\n",
    "# Start training\n",
    "trainer.train()\n",
    "\n",
    "# Cleanup after training\n",
    "cleanup_distributed()\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import os\n",
    "import torch\n",
    "import torch.multiprocessing as mp\n",
    "from datasets import load_dataset\n",
    "import torch\n",
    "from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig, AutoTokenizer\n",
    "from peft import LoraConfig\n",
    "from trl import SFTTrainer\n",
    "from transformers import TrainingArguments\n",
    "from torch.nn.parallel import DistributedDataParallel as DDP\n",
    "\n",
    "\n",
    "# Distributed training setup\n",
    "def init_distributed():\n",
    "    os.environ[\"MASTER_ADDR\"] = \"localhost\"\n",
    "    os.environ[\"MASTER_PORT\"] = \"12345\"\n",
    "    torch.distributed.init_process_group(backend='nccl', world_size=torch.cuda.device_count(), rank=rank)\n",
    "\n",
    "def cleanup_distributed():\n",
    "    torch.distributed.destroy_process_group()\n",
    "\n",
    "def main_worker(rank, world_size):\n",
    "    init_distributed()\n",
    "\n",
    "    # Your model training and fine-tuning code goes here\n",
    "    # Load the dataset\n",
    "    dataset_name = \"ruslanmv/ai-medical-dataset\"\n",
    "    dataset = load_dataset(dataset_name, split=\"train\")\n",
    "    # Select the first 1M rows of the dataset\n",
    "    dataset = dataset.select(range(100))\n",
    "\n",
    "    # Load the model + tokenizer\n",
    "    model_name = \"meta-llama/Meta-Llama-3-8B-Instruct\"\n",
    "    tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)\n",
    "    tokenizer.pad_token = tokenizer.eos_token\n",
    "    bnb_config = BitsAndBytesConfig(\n",
    "        load_in_4bit=True,\n",
    "        bnb_4bit_quant_type=\"nf4\",\n",
    "        bnb_4bit_compute_dtype=torch.float16,\n",
    "    )\n",
    "    model = AutoModelForCausalLM.from_pretrained(\n",
    "        model_name,\n",
    "        quantization_config=bnb_config,\n",
    "        trust_remote_code=True,\n",
    "        use_cache=False,\n",
    "    )\n",
    "\n",
    "    # Check for available GPUs\n",
    "    device = torch.device(f\"cuda:{rank}\" if torch.cuda.is_available() else \"cpu\")\n",
    "\n",
    "    # PEFT config\n",
    "    lora_alpha = 1\n",
    "    lora_dropout = 0.1\n",
    "    lora_r = 32  # 64\n",
    "    peft_config = LoraConfig(\n",
    "        lora_alpha=lora_alpha,\n",
    "        lora_dropout=lora_dropout,\n",
    "        task_type=\"CAUSAL_LM\",\n",
    "        target_modules=[\"k_proj\", \"q_proj\", \"v_proj\", \"up_proj\", \"down_proj\", \"gate_proj\"],\n",
    "        modules_to_save=[\"embed_tokens\", \"input_layernorm\", \"post_attention_layernorm\", \"norm\"],\n",
    "    )\n",
    "\n",
    "    # Args\n",
    "    max_seq_length = 512\n",
    "    output_dir = \"./results\"\n",
    "    per_device_train_batch_size = 2  # reduced batch size to avoid OOM\n",
    "    gradient_accumulation_steps = 2\n",
    "    optim = \"adamw_torch\"\n",
    "    save_steps = 10\n",
    "    logging_steps = 1\n",
    "    learning_rate = 2e-4\n",
    "    max_grad_norm = 0.3\n",
    "    max_steps = 1  # 300 Approx the size of guanaco at bs 8, ga 2, 2 GPUs.\n",
    "    warmup_ratio = 0.1\n",
    "    lr_scheduler_type = \"cosine\"\n",
    "    training_arguments = TrainingArguments(\n",
    "        output_dir=output_dir,\n",
    "        per_device_train_batch_size=per_device_train_batch_size,\n",
    "        gradient_accumulation_steps=gradient_accumulation_steps,\n",
    "        optim=optim,\n",
    "        save_steps=save_steps,\n",
    "        logging_steps=logging_steps,\n",
    "        learning_rate=learning_rate,\n",
    "        fp16=True,\n",
    "        max_grad_norm=max_grad_norm,\n",
    "        max_steps=max_steps,\n",
    "        warmup_ratio=warmup_ratio,\n",
    "        group_by_length=True,\n",
    "        lr_scheduler_type=lr_scheduler_type,\n",
    "        gradient_checkpointing=True,  # gradient checkpointing\n",
    "        #report_to=\"wandb\",\n",
    "    )\n",
    "\n",
    "    # Trainer\n",
    "    trainer = SFTTrainer(\n",
    "        model=model,\n",
    "        train_dataset=dataset,\n",
    "        peft_config=peft_config,\n",
    "        dataset_text_field=\"context\",\n",
    "        max_seq_length=max_seq_length,\n",
    "        tokenizer=tokenizer,\n",
    "        args=training_arguments,\n",
    "    )\n",
    "\n",
    "    # Train :)\n",
    "    trainer.train()\n",
    "    cleanup_distributed()\n",
    "\n",
    "\n",
    "if __name__ == \"__main__\":\n",
    "    world_size = torch.cuda.device_count()\n",
    "    mp.set_start_method('spawn')  # Add this line to fix the error\n",
    "    processes = []\n",
    "    for rank in range(world_size):\n",
    "        p = mp.Process(target=main_worker, args=(rank, world_size))\n",
    "        p.start()\n",
    "        processes.append(p)\n",
    "    for p in processes:\n",
    "        p.join()\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "def finetune():\n",
    "    from datasets import load_dataset\n",
    "    import torch\n",
    "    from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig, AutoTokenizer\n",
    "    from peft import LoraConfig\n",
    "    from trl import SFTTrainer\n",
    "    from transformers import TrainingArguments\n",
    "    from torch.nn.parallel import DistributedDataParallel as DDP\n",
    "    # Load the dataset\n",
    "    dataset_name = \"ruslanmv/ai-medical-dataset\"\n",
    "    dataset = load_dataset(dataset_name, split=\"train\")\n",
    "    # Select the first 1M rows of the dataset\n",
    "    dataset = dataset.select(range(100))\n",
    "    # Load the model + tokenizer\n",
    "    model_name = \"meta-llama/Meta-Llama-3-8B-Instruct\"\n",
    "    tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)\n",
    "    tokenizer.pad_token = tokenizer.eos_token\n",
    "    bnb_config = BitsAndBytesConfig(\n",
    "      load_in_4bit=True,\n",
    "      bnb_4bit_quant_type=\"nf4\",\n",
    "      bnb_4bit_compute_dtype=torch.float16,\n",
    "    )\n",
    "    model = AutoModelForCausalLM.from_pretrained(\n",
    "      model_name,\n",
    "      quantization_config=bnb_config,\n",
    "      trust_remote_code=True,\n",
    "      use_cache=False,\n",
    "    )\n",
    "    # Check for available GPUs\n",
    "    if torch.cuda.device_count() > 1:\n",
    "      print(\"Multiple GPUs detected, enabling DataParallel...\")\n",
    "      model = DDP(model)  # Wrap the model with DDP\n",
    "    else:\n",
    "      print(\"Using single GPU...\")\n",
    "    # PEFT config\n",
    "    lora_alpha = 16\n",
    "    lora_dropout = 0.1\n",
    "    lora_r = 32  # 64\n",
    "    peft_config = LoraConfig(\n",
    "      lora_alpha=lora_alpha,\n",
    "      lora_dropout=lora_dropout,\n",
    "      r=lora_r,\n",
    "      bias=\"none\",\n",
    "      task_type=\"CAUSAL_LM\",\n",
    "      target_modules=[\"k_proj\", \"q_proj\", \"v_proj\", \"up_proj\", \"down_proj\", \"gate_proj\"],\n",
    "      modules_to_save=[\"embed_tokens\", \"input_layernorm\", \"post_attention_layernorm\", \"norm\"],\n",
    "    )\n",
    "    # Args\n",
    "    max_seq_length = 512\n",
    "    output_dir = \"./results\"\n",
    "    per_device_train_batch_size = 2  # reduced batch size to avoid OOM\n",
    "    gradient_accumulation_steps = 2\n",
    "    optim = \"adamw_torch\"\n",
    "    save_steps = 10\n",
    "    logging_steps = 1\n",
    "    learning_rate = 2e-4\n",
    "    max_grad_norm = 0.3\n",
    "    max_steps = 1  # 300 Approx the size of guanaco at bs 8, ga 2, 2 GPUs.\n",
    "    warmup_ratio = 0.1\n",
    "    lr_scheduler_type = \"cosine\"\n",
    "\n",
    "    training_arguments = TrainingArguments(\n",
    "      output_dir=output_dir,\n",
    "      per_device_train_batch_size=per_device_train_batch_size,\n",
    "      gradient_accumulation_steps=gradient_accumulation_steps,\n",
    "      optim=optim,\n",
    "      save_steps=save_steps,\n",
    "      logging_steps=logging_steps,\n",
    "      learning_rate=learning_rate,\n",
    "      fp16=True,\n",
    "      max_grad_norm=max_grad_norm,\n",
    "      max_steps=max_steps,\n",
    "      warmup_ratio=warmup_ratio,\n",
    "      group_by_length=True,\n",
    "      lr_scheduler_type=lr_scheduler_type,\n",
    "      gradient_checkpointing=True,  # gradient checkpointing\n",
    "      #report_to=\"wandb\",\n",
    "    )\n",
    "    # Trainer\n",
    "    trainer = SFTTrainer(\n",
    "      model=model,\n",
    "      train_dataset=dataset,\n",
    "      peft_config=peft_config,\n",
    "      dataset_text_field=\"context\",\n",
    "      max_seq_length=max_seq_length,\n",
    "      tokenizer=tokenizer,\n",
    "      args=training_arguments,\n",
    "    )\n",
    "    # Train :)\n",
    "    trainer.train()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import os\n",
    "import torch\n",
    "import torch.multiprocessing as mp\n",
    "\n",
    "def init_distributed(rank, world_size, local_rank=0):  # Add local_rank argument\n",
    "    os.environ[\"MASTER_ADDR\"] = \"localhost\"\n",
    "    os.environ[\"MASTER_PORT\"] = \"12345\"  # Adjust port if needed\n",
    "    if rank == 0:\n",
    "        print(\"Initializing distributed process group...\")\n",
    "    torch.distributed.init_process_group(backend='nccl', world_size=world_size, rank=rank)\n",
    "    torch.cuda.set_device(local_rank)  # Set unique GPU device for each process\n",
    "\n",
    "def cleanup_distributed():\n",
    "    torch.distributed.destroy_process_group()\n",
    "\n",
    "def main_worker(rank, world_size):\n",
    "    local_rank = rank % torch.cuda.device_count()  # Assign unique local rank\n",
    "    init_distributed(rank, world_size, local_rank)\n",
    "    # Your model training and fine-tuning code goes here with model on local_rank GPU\n",
    "    finetune()  # Move model to assigned GPU\n",
    "    cleanup_distributed()\n",
    "if __name__ == \"__main__\":\n",
    "    world_size = torch.cuda.device_count()\n",
    "\n",
    "    # Workaround for Jupyter Notebook and interactive environments\n",
    "    processes = []\n",
    "    for rank in range(world_size):\n",
    "        p = mp.Process(target=main_worker, args=(rank, world_size))\n",
    "        p.start()\n",
    "        processes.append(p)\n",
    "\n",
    "    for p in processes:\n",
    "        p.join()\n"
   ]
  }
 ],
 "metadata": {
  "language_info": {
   "name": "python"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}