{{ message }}
Repository navigation
Expand file tree
/
Copy pathunsloth-cli.py
More file actions
604 lines (535 loc) · 19.5 KB
/
Copy pathunsloth-cli.py
File metadata and controls
604 lines (535 loc) · 19.5 KB
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
#!/usr/bin/env python3
"""
🦥 Starter Script for Fine-Tuning FastLanguageModel with Unsloth
Configurable options for model loading, PEFT, training, and saving/pushing.
Customize the dataset loading/preprocessing and the save/push config for your case.
Usage (most options have sensible defaults; this is an extended example):
python unsloth-cli.py --model_name "unsloth/llama-3-8b" --max_seq_length 8192 --dtype None --load_in_4bit \
--r 64 --lora_alpha 32 --lora_dropout 0.1 --bias "none" --use_gradient_checkpointing "unsloth" \
--random_state 3407 --use_rslora --per_device_train_batch_size 4 --gradient_accumulation_steps 8 \
--warmup_steps 5 --max_steps 400 --learning_rate 2e-6 --logging_steps 1 --optim "adamw_8bit" \
--weight_decay 0.005 --lr_scheduler_type "linear" --seed 3407 --output_dir "outputs" \
--report_to "tensorboard" --save_model --save_path "model" --quantization "f16" \
--push_model --hub_path "hf/model" --hub_token "your_hf_token"
Run `python unsloth-cli.py --help` for the full list of options.
"""
import argparse
import os
def _is_mlx_backend(unsloth_module):
return bool(getattr(unsloth_module, "_IS_MLX", False))
def _normalize_dtype(dtype, is_mlx):
if is_mlx and isinstance(dtype, str) and dtype.strip().lower() in {"", "none", "auto"}:
return None
return dtype
def _prepare_device_map(is_mlx):
if is_mlx:
return None, False
from unsloth.models.loader_utils import prepare_device_map
return prepare_device_map()
class _CallableTokenizerProxy:
def __init__(self, tokenizer):
self._tokenizer = tokenizer
def __getattr__(self, name):
return getattr(self._tokenizer, name)
def __call__(self, text, *args, **kwargs):
# MLX/torch-free: never request torch tensors; keep plain python ids.
kwargs.pop("return_tensors", None)
wrapped = getattr(self._tokenizer, "_tokenizer", None)
if callable(wrapped):
return wrapped(text, *args, **kwargs)
add_special_tokens = kwargs.get("add_special_tokens", False)
input_ids = self._tokenizer.encode(text, add_special_tokens = add_special_tokens)
return {"input_ids": input_ids}
def _tokenizer_for_raw_text_loader(tokenizer, is_mlx):
if not is_mlx or callable(tokenizer):
return tokenizer
return _CallableTokenizerProxy(tokenizer)
def _raw_text_loader_for_backend(
RawTextDataLoader,
tokenizer,
is_mlx,
chunk_size = 2048,
stride = 512,
):
return RawTextDataLoader(
_tokenizer_for_raw_text_loader(tokenizer, is_mlx),
chunk_size,
stride,
return_tokenized = not is_mlx,
)
def _train_with_legacy_save_control(trainer, is_mlx):
if not is_mlx:
return trainer.train()
original_save_model = getattr(trainer, "save_model", None)
if original_save_model is None:
return trainer.train()
def skip_internal_final_save(*args, **kwargs):
raise ValueError("legacy unsloth-cli.py owns final save")
trainer.save_model = skip_internal_final_save
try:
return trainer.train()
finally:
trainer.save_model = original_save_model
def _iter_quantization_methods(quantization):
if isinstance(quantization, list):
return quantization
return [quantization]
def _save_or_push_model(model, tokenizer, args, is_mlx):
if not args.save_model:
print("Warning: The model is not saved!")
return
# Enter the GGUF branch when saving or pushing GGUF, so --push_gguf works without --save_gguf.
if args.save_gguf or args.push_gguf:
if not args.save_gguf:
print("Warning: --save_gguf not set, pushing GGUF to hub without saving locally.")
for quantization_method in _iter_quantization_methods(args.quantization):
if args.save_gguf:
print(f"Saving model with quantization method: {quantization_method}")
model.save_pretrained_gguf(
args.save_path,
tokenizer,
quantization_method = quantization_method,
)
if args.push_model or args.push_gguf:
model.push_to_hub_gguf(
args.hub_path,
tokenizer,
quantization_method = quantization_method,
token = args.hub_token,
)
return
if is_mlx:
model.save_pretrained_merged(
args.save_path,
tokenizer,
save_method = args.save_method,
push_to_hub = args.push_model,
repo_id = args.hub_path if args.push_model else None,
token = args.hub_token,
)
return
model.save_pretrained_merged(args.save_path, tokenizer, save_method = args.save_method)
if args.push_model:
model.push_to_hub_merged(args.hub_path, tokenizer, args.save_method, token = args.hub_token)
def _save_or_push_model_with_mlx_ddp(model, tokenizer, args, is_mlx, trainer):
if not is_mlx or int(trainer.distributed_world_size) <= 1:
return _save_or_push_model(model, tokenizer, args, is_mlx)
raise_distributed_failure = getattr(trainer, "_raise_distributed_failure", None)
if not callable(raise_distributed_failure):
raise RuntimeError(
"Unsloth MLX DDP CLI requires MLXTrainer failure coordination. "
"Upgrade unsloth-zoo to a compatible version."
)
error = None
if bool(trainer.is_main_process):
try:
_save_or_push_model(model, tokenizer, args, is_mlx)
except BaseException as exc:
error = exc
raise_distributed_failure(error is not None, "legacy CLI final save", error)
def _build_sft_config(SFTConfig, args, is_mlx, bf16_supported):
config_kwargs = dict(
per_device_train_batch_size = args.per_device_train_batch_size,
gradient_accumulation_steps = args.gradient_accumulation_steps,
warmup_steps = args.warmup_steps,
max_steps = args.max_steps,
learning_rate = args.learning_rate,
fp16 = not bf16_supported,
bf16 = bf16_supported,
logging_steps = args.logging_steps,
optim = args.optim,
weight_decay = args.weight_decay,
lr_scheduler_type = args.lr_scheduler_type,
seed = args.seed,
output_dir = args.output_dir,
report_to = args.report_to,
max_length = args.max_seq_length,
dataset_num_proc = 2,
packing = args.packing,
)
if getattr(args, "context_parallel_size", 1) > 1:
config_kwargs["context_parallel_size"] = args.context_parallel_size
config_kwargs["context_parallel_backend"] = getattr(
args, "context_parallel_backend", "ring"
)
config_kwargs["context_parallel_rotate_method"] = getattr(
args, "context_parallel_rotate_method", "allgather"
)
# Required under CP; older transformers releases default it to False.
config_kwargs["average_tokens_across_devices"] = True
if is_mlx:
if args.per_device_eval_batch_size != 4:
print("Warning: --per_device_eval_batch_size is ignored on MLX without eval data.")
else:
config_kwargs["per_device_eval_batch_size"] = args.per_device_eval_batch_size
return SFTConfig(**config_kwargs)
def run(args):
import unsloth
from unsloth import FastLanguageModel
from datasets import load_dataset
from transformers.utils import strtobool
from trl import SFTTrainer, SFTConfig
from unsloth import is_bfloat16_supported
import logging
from unsloth import RawTextDataLoader
logging.getLogger("hf-to-gguf").setLevel(logging.WARNING)
is_mlx = _is_mlx_backend(unsloth)
# MLX routes get_peft_model to FastMLXModel, which ignores use_dora (plain LoRA).
if args.use_dora and is_mlx:
raise NotImplementedError("DoRA is not supported for MLX training yet.")
device_map, distributed = _prepare_device_map(is_mlx)
model, tokenizer = FastLanguageModel.from_pretrained(
model_name = args.model_name,
max_seq_length = args.max_seq_length,
dtype = _normalize_dtype(args.dtype, is_mlx),
load_in_4bit = args.load_in_4bit,
device_map = device_map,
)
model = FastLanguageModel.get_peft_model(
model,
r = args.r,
target_modules = [
"q_proj",
"k_proj",
"v_proj",
"o_proj",
"gate_proj",
"up_proj",
"down_proj",
],
lora_alpha = args.lora_alpha,
lora_dropout = args.lora_dropout,
bias = args.bias,
use_gradient_checkpointing = args.use_gradient_checkpointing,
random_state = args.random_state,
use_rslora = args.use_rslora,
loftq_config = args.loftq_config,
use_dora = args.use_dora,
)
alpaca_prompt = """Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.
### Instruction:
{}
### Input:
{}
### Response:
{}"""
EOS_TOKEN = tokenizer.eos_token
def formatting_prompts_func(examples):
instructions = examples["instruction"]
inputs = examples["input"]
outputs = examples["output"]
texts = []
for instruction, input, output in zip(instructions, inputs, outputs):
text = alpaca_prompt.format(instruction, input, output) + EOS_TOKEN
texts.append(text)
return {"text": texts}
def load_dataset_smart(args):
from transformers.utils import strtobool
if args.raw_text_file:
loader = _raw_text_loader_for_backend(
RawTextDataLoader, tokenizer, is_mlx, args.chunk_size, args.stride
)
dataset = loader.load_from_file(args.raw_text_file)
elif args.dataset.endswith((".txt", ".md", ".json", ".jsonl")):
loader = _raw_text_loader_for_backend(RawTextDataLoader, tokenizer, is_mlx)
dataset = loader.load_from_file(args.dataset)
else:
use_modelscope = strtobool(os.environ.get("UNSLOTH_USE_MODELSCOPE", "False"))
if use_modelscope:
from modelscope import MsDataset
dataset = MsDataset.load(args.dataset, split = "train")
else:
dataset = load_dataset(args.dataset, split = "train")
dataset = dataset.map(formatting_prompts_func, batched = True)
return dataset
dataset = load_dataset_smart(args)
print("Data is formatted and ready!")
training_args = _build_sft_config(SFTConfig, args, is_mlx, is_bfloat16_supported())
if distributed:
training_args.ddp_find_unused_parameters = False
trainer = SFTTrainer(
model = model,
processing_class = tokenizer,
train_dataset = dataset,
args = training_args,
)
_train_with_legacy_save_control(trainer, is_mlx)
_save_or_push_model_with_mlx_ddp(model, tokenizer, args, is_mlx, trainer)
if __name__ == "__main__":
parser = argparse.ArgumentParser(description = "🦥 Fine-tune your llm faster using unsloth!")
model_group = parser.add_argument_group("🤖 Model Options")
model_group.add_argument(
"--model_name",
type = str,
default = "unsloth/llama-3-8b",
help = "Model name to load",
)
model_group.add_argument(
"--max_seq_length",
type = int,
default = 2048,
help = "Maximum sequence length, default is 2048. We auto support RoPE Scaling internally!",
)
model_group.add_argument(
"--dtype",
type = str,
default = None,
help = "Data type for model (None for auto detection)",
)
model_group.add_argument(
"--load_in_4bit",
action = "store_true",
help = "Use 4bit quantization to reduce memory usage",
)
model_group.add_argument(
"--dataset",
type = str,
default = "yahma/alpaca-cleaned",
help = "Huggingface dataset to use for training",
)
lora_group = parser.add_argument_group(
"🧠 LoRA Options",
"These options are used to configure the LoRA model.",
)
lora_group.add_argument(
"--r",
type = int,
default = 16,
help = "Rank for Lora model, default is 16. (common values: 8, 16, 32, 64, 128)",
)
lora_group.add_argument(
"--lora_alpha",
type = int,
default = 16,
help = "LoRA alpha parameter, default is 16. (common values: 8, 16, 32, 64, 128)",
)
lora_group.add_argument(
"--lora_dropout",
type = float,
default = 0.0,
help = "LoRA dropout rate, default is 0.0 which is optimized.",
)
lora_group.add_argument(
"--bias",
type = str,
default = "none",
help = "Bias setting for LoRA",
)
lora_group.add_argument(
"--use_gradient_checkpointing",
type = str,
default = "unsloth",
help = "Use gradient checkpointing",
)
lora_group.add_argument(
"--random_state",
type = int,
default = 3407,
help = "Random state for reproducibility, default is 3407.",
)
lora_group.add_argument(
"--use_rslora",
action = "store_true",
help = "Use rank stabilized LoRA",
)
lora_group.add_argument(
"--loftq_config",
type = str,
default = None,
help = "Configuration for LoftQ",
)
lora_group.add_argument(
"--use_dora",
action = "store_true",
help = "Use DoRA (Weight-Decomposed LoRA)",
)
training_group = parser.add_argument_group("🎓 Training Options")
training_group.add_argument(
"--per_device_train_batch_size",
type = int,
default = 2,
help = "Batch size per device during training, default is 2.",
)
training_group.add_argument(
"--per_device_eval_batch_size",
type = int,
default = 4,
help = "Batch size per device during evaluation, default is 4.",
)
training_group.add_argument(
"--gradient_accumulation_steps",
type = int,
default = 4,
help = "Number of gradient accumulation steps, default is 4.",
)
training_group.add_argument(
"--warmup_steps",
type = int,
default = 5,
help = "Number of warmup steps, default is 5.",
)
training_group.add_argument(
"--max_steps",
type = int,
default = 400,
help = "Maximum number of training steps.",
)
training_group.add_argument(
"--learning_rate",
type = float,
default = 2e-4,
help = "Learning rate, default is 2e-4.",
)
training_group.add_argument(
"--optim",
type = str,
default = "adamw_8bit",
help = "Optimizer type.",
)
training_group.add_argument(
"--weight_decay",
type = float,
default = 0.01,
help = "Weight decay, default is 0.01.",
)
training_group.add_argument(
"--lr_scheduler_type",
type = str,
default = "linear",
help = "Learning rate scheduler type, default is 'linear'.",
)
training_group.add_argument(
"--seed",
type = int,
default = 3407,
help = "Seed for reproducibility, default is 3407.",
)
training_group.add_argument(
"--packing",
action = "store_true",
help = "Enable padding-free sample packing via TRL's bin packer.",
)
training_group.add_argument(
"--context_parallel_size",
type = int,
default = 1,
help = "Ranks per context parallel group (torchrun, PyTorch >= 2.7).",
)
training_group.add_argument(
"--context_parallel_backend",
type = str,
default = "ring",
choices = ["ring", "ulysses"],
help = "Context parallel attention backend: ring (SDPA ring attention) or ulysses "
"(all-to-all over heads; keeps packing and the usual attention kernels, needs heads "
"divisible by the size).",
)
training_group.add_argument(
"--context_parallel_rotate_method",
type = str,
default = "allgather",
choices = ["allgather", "alltoall"],
help = "How ring attention rotates KV shards between ranks.",
)
report_group = parser.add_argument_group("📊 Report Options")
report_group.add_argument(
"--report_to",
type = str,
default = "tensorboard",
choices = [
"azure_ml",
"clearml",
"codecarbon",
"comet_ml",
"dagshub",
"dvclive",
"flyte",
"mlflow",
"neptune",
"tensorboard",
"wandb",
"all",
"none",
],
help = (
"The list of integrations to report the results and logs to. Supported platforms are:\n\t\t "
"'azure_ml', 'clearml', 'codecarbon', 'comet_ml', 'dagshub', 'dvclive', 'flyte', "
"'mlflow', 'neptune', 'tensorboard', and 'wandb'. Use 'all' to report to all integrations "
"installed, 'none' for no integrations."
),
)
report_group.add_argument(
"--logging_steps",
type = int,
default = 1,
help = "Logging steps, default is 1",
)
save_group = parser.add_argument_group("💾 Save Model Options")
save_group.add_argument(
"--output_dir",
type = str,
default = "outputs",
help = "Output directory",
)
save_group.add_argument(
"--save_model",
action = "store_true",
help = "Save the model after training",
)
save_group.add_argument(
"--save_method",
type = str,
default = "merged_16bit",
choices = ["merged_16bit", "merged_4bit", "lora"],
help = "Save method for the model, default is 'merged_16bit'",
)
save_group.add_argument(
"--save_gguf",
action = "store_true",
help = "Convert the model to GGUF after training",
)
save_group.add_argument(
"--save_path",
type = str,
default = "model",
help = "Path to save the model",
)
save_group.add_argument(
"--quantization",
type = str,
default = "q8_0",
nargs = "+",
help = (
"Quantization method for saving the model. common values ('f16', 'q4_k_m', 'q8_0'), "
"Check our wiki for all quantization methods https://github.com/unslothai/unsloth/wiki#saving-to-gguf"
),
)
push_group = parser.add_argument_group("🚀 Push Model Options")
push_group.add_argument(
"--push_model",
action = "store_true",
help = "Push the model to Hugging Face hub after training",
)
push_group.add_argument(
"--push_gguf",
action = "store_true",
help = "Push the model as GGUF to Hugging Face hub after training",
)
push_group.add_argument(
"--hub_path",
type = str,
default = "hf/model",
help = "Path on Hugging Face hub to push the model",
)
push_group.add_argument(
"--hub_token",
type = str,
help = "Token for pushing the model to Hugging Face hub",
)
parser.add_argument("--raw_text_file", type = str, help = "Path to raw text file for training")
parser.add_argument(
"--chunk_size", type = int, default = 2048, help = "Size of text chunks for training"
)
parser.add_argument("--stride", type = int, default = 512, help = "Overlap between chunks")
args = parser.parse_args()
run(args)
You can’t perform that action at this time.