riaz.somc/transformers
1
0
Fork 0
mirror of https://github.com/huggingface/transformers.git synced 2025-07-03 04:40:06 +06:00
Code Issues Actions 28 Packages Projects Releases Wiki Activity

Simplify and update trl examples (#38772)

Browse Source 
* Simplify and update trl examples

* Remove optim_args from SFTConfig in Trainer documentation

* Update docs/source/en/trainer.md

* Apply suggestions from code review

* Update docs/source/en/trainer.md

Co-authored-by: Steven Liu <59462357+stevhliu@users.noreply.github.com>

---------

Co-authored-by: Quentin Gallouédec <qgallouedec@Quentins-MacBook-Pro.local>
Co-authored-by: Steven Liu <59462357+stevhliu@users.noreply.github.com>
This commit is contained in:
Quentin Gallouédec 2025-06-13 14:03:49 +02:00 committed by GitHub
parent de24fb63ed
commit c989ddd294
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
7 changed files with 114 additions and 347 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

165
docs/source/ar/trainer.md
View File

@ -306,75 +306,45 @@ pip install galore-torch
ثم أضف ببساطة أحد `["galore_adamw"، "galore_adafactor"، "galore_adamw_8bit"]` في `optim` جنبًا إلى جنب مع `optim_target_modules`، والتي يمكن أن تكون قائمة من السلاسل أو التعبيرات النمطية regex أو المسار الكامل المطابق لأسماء الوحدات المستهدفة التي تريد تكييفها. فيما يلي مثال على النص البرمجي كامل(تأكد من `pip install trl datasets`): ثم أضف ببساطة أحد `["galore_adamw"، "galore_adafactor"، "galore_adamw_8bit"]` في `optim` جنبًا إلى جنب مع `optim_target_modules`، والتي يمكن أن تكون قائمة من السلاسل أو التعبيرات النمطية regex أو المسار الكامل المطابق لأسماء الوحدات المستهدفة التي تريد تكييفها. فيما يلي مثال على النص البرمجي كامل(تأكد من `pip install trl datasets`):
```python ```python
import torch
import datasets import datasets
import trl from trl import SFTConfig, SFTTrainer
from transformers import TrainingArguments, AutoConfig, AutoTokenizer, AutoModelForCausalLM
train_dataset = datasets.load_dataset('imdb', split='train') train_dataset = datasets.load_dataset('imdb', split='train')
args = SFTConfig(
args = TrainingArguments( output_dir="./test-galore",
output_dir="./test-galore"،
max_steps=100, max_steps=100,
per_device_train_batch_size=2, optim="galore_adamw",
optim="galore_adamw"، optim_target_modules=[r".*.attn.*", r".*.mlp.*"],
optim_target_modules=[r".*.attn.*"، r".*.mlp.*"] gradient_checkpointing=True,
) )
trainer = SFTTrainer(
model_id = "google/gemma-2b" model="google/gemma-2b",
config = AutoConfig.from_pretrained(model_id)
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_config(config).to(0)
trainer = trl.SFTTrainer(
model=model,
args=args, args=args,
train_dataset=train_dataset, train_dataset=train_dataset,
dataset_text_field='text',
max_seq_length=512,
) )
trainer.train() trainer.train()
``` ```
لتمرير معامﻻت إضافية يدعمها GaLore، يجب عليك تمرير `optim_args` بشكل صحيح، على سبيل المثال: لتمرير معامﻻت إضافية يدعمها GaLore، يجب عليك تمرير `optim_args` بشكل صحيح، على سبيل المثال:
```python ```python
import torch
import datasets import datasets
import trl from trl import SFTConfig, SFTTrainer
from transformers import TrainingArguments, AutoConfig, AutoTokenizer, AutoModelForCausalLM
train_dataset = datasets.load_dataset('imdb', split='train') train_dataset = datasets.load_dataset('imdb', split='train')
args = SFTConfig(
args = TrainingArguments(
output_dir="./test-galore", output_dir="./test-galore",
max_steps=100, max_steps=100,
per_device_train_batch_size=2,
optim="galore_adamw", optim="galore_adamw",
optim_target_modules=[r".*.attn.*", r".*.mlp.*"], optim_target_modules=[r".*.attn.*", r".*.mlp.*"],
optim_args="rank=64, update_proj_gap=100, scale=0.10", optim_args="rank=64, update_proj_gap=100, scale=0.10",
gradient_checkpointing=True,
) )
trainer = SFTTrainer(
model_id = "google/gemma-2b" model="google/gemma-2b",
config = AutoConfig.from_pretrained(model_id)
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_config(config).to(0)
trainer = trl.SFTTrainer(
model=model,
args=args, args=args,
train_dataset=train_dataset, train_dataset=train_dataset,
dataset_text_field='text',
max_seq_length=512,
) )
trainer.train() trainer.train()
``` ```
يمكنك قراءة المزيد حول الطريقة في [المستودع الأصلي](https://github.com/jiaweizzhao/GaLore) أو [الورقة البحثية](https://huggingface.co/papers/2403.03507). يمكنك قراءة المزيد حول الطريقة في [المستودع الأصلي](https://github.com/jiaweizzhao/GaLore) أو [الورقة البحثية](https://huggingface.co/papers/2403.03507).
@ -386,37 +356,22 @@ trainer.train()
يمكنك أيضًا إجراء تحسين طبقة تلو الأخرى عن طريق إضافة `layerwise` إلى اسم المُحسِّن كما هو موضح أدناه: يمكنك أيضًا إجراء تحسين طبقة تلو الأخرى عن طريق إضافة `layerwise` إلى اسم المُحسِّن كما هو موضح أدناه:
```python ```python
import torch
import datasets import datasets
import trl from trl import SFTConfig, SFTTrainer
from transformers import TrainingArguments، AutoConfig، AutoTokenizer، AutoModelForCausalLM train_dataset = datasets.load_dataset('imdb', split='train')
args = SFTConfig(
train_dataset = datasets.load_dataset('imdb'، split='train') output_dir="./test-galore",
max_steps=100,
args = TrainingArguments( optim="galore_adamw_layerwise",
output_dir="./test-galore"، optim_target_modules=[r".*.attn.*", r".*.mlp.*"],
max_steps=100، gradient_checkpointing=True,
per_device_train_batch_size=2،
optim="galore_adamw_layerwise"،
optim_target_modules=[r".*.attn.*"، r".*.mlp.*"]
) )
trainer = SFTTrainer(
model_id = "google/gemma-2b" model="google/gemma-2b",
args=args,
config = AutoConfig.from_pretrained(model_id) train_dataset=train_dataset,
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_config(config).to(0)
trainer = trl.SFTTrainer(
model=model،
args=args،
train_dataset=train_dataset،
dataset_text_field='text'،
max_seq_length=512،
) )
trainer.train() trainer.train()
``` ```
@ -436,39 +391,21 @@ trainer.train()
فيما يلي نص برمجي بسيط يوضح كيفية ضبط نموذج [google/gemma-2b](https://huggingface.co/google/gemma-2b) على مجموعة بيانات IMDB في الدقة الكاملة: فيما يلي نص برمجي بسيط يوضح كيفية ضبط نموذج [google/gemma-2b](https://huggingface.co/google/gemma-2b) على مجموعة بيانات IMDB في الدقة الكاملة:
```python ```python
import torch
import datasets import datasets
from transformers import TrainingArguments، AutoTokenizer، AutoModelForCausalLM from trl import SFTConfig, SFTTrainer
import trl
train_dataset = datasets.load_dataset('imdb'، split='train') train_dataset = datasets.load_dataset('imdb', split='train')
args = SFTConfig(
args = TrainingArguments( output_dir="./test-lomo",
output_dir="./test-lomo"، max_steps=100,
max_steps=100، optim="adalomo",
per_device_train_batch_size=4، gradient_checkpointing=True,
optim="adalomo"،
gradient_checkpointing=True،
logging_strategy="steps"،
logging_steps=1،
learning_rate=2e-6،
save_strategy="no"،
run_name="lomo-imdb"،
) )
trainer = SFTTrainer(
model_id = "google/gemma-2b" model="google/gemma-2b",
args=args,
tokenizer = AutoTokenizer.from_pretrained(model_id) train_dataset=train_dataset,
model = AutoModelForCausalLM.from_pretrained(model_id).to(0)
trainer = trl.SFTTrainer(
model=model،
args=args،
train_dataset=train_dataset،
dataset_text_field='text'،
max_seq_length=1024،
) )
trainer.train() trainer.train()
``` ```
@ -524,39 +461,21 @@ trainer.train()
فيما يلي نص برمجى بسيط لشرح كيفية ضبط [google/gemma-2b](https://huggingface.co/google/gemma-2b) بدقة على مجموعة بيانات IMDB بدقة كاملة: فيما يلي نص برمجى بسيط لشرح كيفية ضبط [google/gemma-2b](https://huggingface.co/google/gemma-2b) بدقة على مجموعة بيانات IMDB بدقة كاملة:
```python ```python
import torch
import datasets import datasets
from transformers import TrainingArguments, AutoTokenizer, AutoModelForCausalLM from trl import SFTConfig, SFTTrainer
import trl
train_dataset = datasets.load_dataset('imdb', split='train') train_dataset = datasets.load_dataset('imdb', split='train')
args = SFTConfig(
args = TrainingArguments( output_dir="./test-galore",
output_dir="./test-schedulefree", max_steps=100,
max_steps=1000,
per_device_train_batch_size=4,
optim="schedule_free_adamw", optim="schedule_free_adamw",
gradient_checkpointing=True, gradient_checkpointing=True,
logging_strategy="steps",
logging_steps=1,
learning_rate=2e-6,
save_strategy="no",
run_name="sfo-imdb",
) )
trainer = SFTTrainer(
model_id = "google/gemma-2b" model="google/gemma-2b",
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id).to(0)
trainer = trl.SFTTrainer(
model=model,
args=args, args=args,
train_dataset=train_dataset, train_dataset=train_dataset,
dataset_text_field='text',
max_seq_length=1024,
) )
trainer.train() trainer.train()
``` ```
## تسريع ومدرب ## تسريع ومدرب

45
docs/source/en/model_doc/mamba.md
View File

@ -97,39 +97,22 @@ print(tokenizer.decode(output[0], skip_special_tokens=True))
- Mamba stacks `mixer` layers which are equivalent to `Attention` layers. You can find the main logic of Mamba in the `MambaMixer` class. - Mamba stacks `mixer` layers which are equivalent to `Attention` layers. You can find the main logic of Mamba in the `MambaMixer` class.
- The example below demonstrates how to fine-tune Mamba with [PEFT](https://huggingface.co/docs/peft). - The example below demonstrates how to fine-tune Mamba with [PEFT](https://huggingface.co/docs/peft).
```py ```py
from datasets import load_dataset from datasets import load_dataset
from trl import SFTTrainer from trl import SFTConfig, SFTTrainer
from peft import LoraConfig from peft import LoraConfig
from transformers import AutoTokenizer, AutoModelForCausalLM, TrainingArguments
model_id = "state-spaces/mamba-130m-hf" model_id = "state-spaces/mamba-130m-hf"
tokenizer = AutoTokenizer.from_pretrained(model_id) dataset = load_dataset("Abirate/english_quotes", split="train")
model = AutoModelForCausalLM.from_pretrained(model_id) training_args = SFTConfig(dataset_text_field="quote")
dataset = load_dataset("Abirate/english_quotes", split="train") lora_config = LoraConfig(target_modules=["x_proj", "embeddings", "in_proj", "out_proj"])
training_args = TrainingArguments( trainer = SFTTrainer(
output_dir="./results", model=model_id,
num_train_epochs=3,
per_device_train_batch_size=4,
logging_dir='./logs',
logging_steps=10,
learning_rate=2e-3
)
lora_config = LoraConfig(
r=8,
target_modules=["x_proj", "embeddings", "in_proj", "out_proj"],
task_type="CAUSAL_LM",
bias="none"
)
trainer = SFTTrainer(
model=model,
processing_class=tokenizer,
args=training_args, args=training_args,
peft_config=lora_config, train_dataset=dataset,
train_dataset=dataset, peft_config=lora_config,
dataset_text_field="quote", )
) trainer.train()
trainer.train()
``` ```
## MambaConfig ## MambaConfig

35
docs/source/en/model_doc/mamba2.md
View File

@ -103,40 +103,19 @@ print(tokenizer.decode(output[0], skip_special_tokens=True))
- The example below demonstrates how to fine-tune Mamba 2 with [PEFT](https://huggingface.co/docs/peft). - The example below demonstrates how to fine-tune Mamba 2 with [PEFT](https://huggingface.co/docs/peft).
```python ```python
from trl import SFTTrainer from datasets import load_dataset
from peft import LoraConfig from peft import LoraConfig
from transformers import AutoTokenizer, Mamba2ForCausalLM, TrainingArguments from trl import SFTConfig, SFTTrainer
model_id = 'mistralai/Mamba-Codestral-7B-v0.1'
tokenizer = AutoTokenizer.from_pretrained(model_id, revision='refs/pr/9', from_slow=True, legacy=False)
tokenizer.pad_token = tokenizer.eos_token
tokenizer.padding_side = "left" #enforce padding side left
model = Mamba2ForCausalLM.from_pretrained(model_id, revision='refs/pr/9') model_id = "mistralai/Mamba-Codestral-7B-v0.1"
dataset = load_dataset("Abirate/english_quotes", split="train") dataset = load_dataset("Abirate/english_quotes", split="train")
# Without CUDA kernels, batch size of 2 occupies one 80GB device training_args = SFTConfig(dataset_text_field="quote", gradient_checkpointing=True, per_device_train_batch_size=4)
# but precision can be reduced. lora_config = LoraConfig(target_modules=["x_proj", "embeddings", "in_proj", "out_proj"])
# Experiments and trials welcome!
training_args = TrainingArguments(
output_dir="./results",
num_train_epochs=3,
per_device_train_batch_size=2,
logging_dir='./logs',
logging_steps=10,
learning_rate=2e-3
)
lora_config = LoraConfig(
r=8,
target_modules=["embeddings", "in_proj", "out_proj"],
task_type="CAUSAL_LM",
bias="none"
)
trainer = SFTTrainer( trainer = SFTTrainer(
model=model, model=model_id,
tokenizer=tokenizer,
args=training_args, args=training_args,
peft_config=lora_config,
train_dataset=dataset, train_dataset=dataset,
dataset_text_field="quote", peft_config=lora_config,
) )
trainer.train() trainer.train()
``` ```

42
docs/source/en/trainer.md
View File

@ -392,15 +392,15 @@ training_args = TrainingArguments(
[Gradient Low-Rank Projection (GaLore)](https://hf.co/papers/2403.03507) significantly reduces memory usage when training large language models (LLMs). One of GaLores key benefits is *full-parameter* learning, unlike low-rank adaptation methods like [LoRA](https://hf.co/papers/2106.09685), which produces better model performance. [Gradient Low-Rank Projection (GaLore)](https://hf.co/papers/2403.03507) significantly reduces memory usage when training large language models (LLMs). One of GaLores key benefits is *full-parameter* learning, unlike low-rank adaptation methods like [LoRA](https://hf.co/papers/2106.09685), which produces better model performance.
Install the [GaLore](https://github.com/jiaweizzhao/GaLore) library, [TRL](https://hf.co/docs/trl/index), and [Datasets](https://hf.co/docs/datasets/index). Install the [GaLore](https://github.com/jiaweizzhao/GaLore) and [TRL](https://hf.co/docs/trl/index) libraries.
```bash ```bash
pip install galore-torch trl datasets pip install galore-torch trl
``` ```
Pick a GaLore optimizer (`"galore_adamw"`, `"galore_adafactor"`, `"galore_adamw_8bit`") and pass it to the `optim` parameter in [`TrainingArguments`]. Use the `optim_target_modules` parameter to specify which modules to adapt (can be a list of strings, regex, or a full path). Pick a GaLore optimizer (`"galore_adamw"`, `"galore_adafactor"`, `"galore_adamw_8bit`") and pass it to the `optim` parameter in [`trl.SFTConfig`]. Use the `optim_target_modules` parameter to specify which modules to adapt (can be a list of strings, regex, or a full path).
Extra parameters supported by GaLore, `rank`, `update_proj_gap`, and `scale`, should be passed to the `optim_args` parameter in [`TrainingArguments`]. Extra parameters supported by GaLore, `rank`, `update_proj_gap`, and `scale`, should be passed to the `optim_args` parameter in [`trl.SFTConfig`].
The example below enables GaLore with [`~trl.SFTTrainer`] that targets the `attn` and `mlp` layers with regex. The example below enables GaLore with [`~trl.SFTTrainer`] that targets the `attn` and `mlp` layers with regex.
@ -411,29 +411,22 @@ The example below enables GaLore with [`~trl.SFTTrainer`] that targets the `attn
<hfoption id="GaLore optimizer"> <hfoption id="GaLore optimizer">
```py ```py
import torch
import datasets import datasets
import trl from trl import SFTConfig, SFTTrainer
from transformers import TrainingArguments, AutoConfig, AutoTokenizer, AutoModelForCausalLM
train_dataset = datasets.load_dataset('imdb', split='train') train_dataset = datasets.load_dataset('imdb', split='train')
args = TrainingArguments( args = SFTConfig(
output_dir="./test-galore", output_dir="./test-galore",
max_steps=100, max_steps=100,
per_device_train_batch_size=2,
optim="galore_adamw", optim="galore_adamw",
optim_target_modules=[r".*.attn.*", r".*.mlp.*"], optim_target_modules=[r".*.attn.*", r".*.mlp.*"],
optim_args="rank=64, update_proj_gap=100, scale=0.10", optim_args="rank=64, update_proj_gap=100, scale=0.10",
gradient_checkpointing=True,
) )
config = AutoConfig.from_pretrained("google/gemma-2b") trainer = SFTTrainer(
tokenizer = AutoTokenizer.from_pretrained("google/gemma-2b") model="google/gemma-2b",
model = AutoModelForCausalLM.from_config("google/gemma-2b").to(0)
trainer = trl.SFTTrainer(
model=model,
args=args, args=args,
train_dataset=train_dataset, train_dataset=train_dataset,
dataset_text_field='text',
max_seq_length=512,
) )
trainer.train() trainer.train()
``` ```
@ -444,29 +437,22 @@ trainer.train()
Append `layerwise` to the optimizer name to enable layerwise optimization. For example, `"galore_adamw"` becomes `"galore_adamw_layerwise"`. This feature is still experimental and does not support Distributed Data Parallel (DDP). The code below can only be run on a [single GPU](https://github.com/jiaweizzhao/GaLore?tab=readme-ov-file#train-7b-model-with-a-single-gpu-with-24gb-memory). Other features like gradient clipping and DeepSpeed may not be available out of the box. Feel free to open an [issue](https://github.com/huggingface/transformers/issues) if you encounter any problems! Append `layerwise` to the optimizer name to enable layerwise optimization. For example, `"galore_adamw"` becomes `"galore_adamw_layerwise"`. This feature is still experimental and does not support Distributed Data Parallel (DDP). The code below can only be run on a [single GPU](https://github.com/jiaweizzhao/GaLore?tab=readme-ov-file#train-7b-model-with-a-single-gpu-with-24gb-memory). Other features like gradient clipping and DeepSpeed may not be available out of the box. Feel free to open an [issue](https://github.com/huggingface/transformers/issues) if you encounter any problems!
```py ```py
import torch
import datasets import datasets
import trl from trl import SFTConfig, SFTTrainer
from transformers import TrainingArguments, AutoConfig, AutoTokenizer, AutoModelForCausalLM
train_dataset = datasets.load_dataset('imdb', split='train') train_dataset = datasets.load_dataset('imdb', split='train')
args = TrainingArguments( args = SFTConfig(
output_dir="./test-galore", output_dir="./test-galore",
max_steps=100, max_steps=100,
per_device_train_batch_size=2,
optim="galore_adamw_layerwise", optim="galore_adamw_layerwise",
optim_target_modules=[r".*.attn.*", r".*.mlp.*"], optim_target_modules=[r".*.attn.*", r".*.mlp.*"],
optim_args="rank=64, update_proj_gap=100, scale=0.10", optim_args="rank=64, update_proj_gap=100, scale=0.10",
gradient_checkpointing=True,
) )
config = AutoConfig.from_pretrained("google/gemma-2b") trainer = SFTTrainer(
tokenizer = AutoTokenizer.from_pretrained("google/gemma-2b") model="google/gemma-2b",
model = AutoModelForCausalLM.from_config("google/gemma-2b").to(0)
trainer = trl.SFTTrainer(
model=model,
args=args, args=args,
train_dataset=train_dataset, train_dataset=train_dataset,
dataset_text_field='text',
max_seq_length=512,
) )
trainer.train() trainer.train()
``` ```

28
docs/source/ko/model_doc/mamba.md
View File

@ -58,34 +58,18 @@ print(tokenizer.batch_decode(out))
```python ```python
from datasets import load_dataset from datasets import load_dataset
from trl import SFTTrainer from trl import SFTConfig, SFTTrainer
from peft import LoraConfig from peft import LoraConfig
from transformers import AutoTokenizer, AutoModelForCausalLM, TrainingArguments
model_id = "state-spaces/mamba-130m-hf" model_id = "state-spaces/mamba-130m-hf"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)
dataset = load_dataset("Abirate/english_quotes", split="train") dataset = load_dataset("Abirate/english_quotes", split="train")
training_args = TrainingArguments( training_args = SFTConfig(dataset_text_field="quote")
output_dir="./results", lora_config = LoraConfig(target_modules=["x_proj", "embeddings", "in_proj", "out_proj"])
num_train_epochs=3,
per_device_train_batch_size=4,
logging_dir='./logs',
logging_steps=10,
learning_rate=2e-3
)
lora_config = LoraConfig(
r=8,
target_modules=["x_proj", "embeddings", "in_proj", "out_proj"],
task_type="CAUSAL_LM",
bias="none"
)
trainer = SFTTrainer( trainer = SFTTrainer(
model=model, model=model_id,
tokenizer=tokenizer,
args=training_args, args=training_args,
peft_config=lora_config,
train_dataset=dataset, train_dataset=dataset,
dataset_text_field="quote", peft_config=lora_config,
) )
trainer.train() trainer.train()
``` ```

35
docs/source/ko/model_doc/mamba2.md
View File

@ -57,40 +57,19 @@ print(tokenizer.batch_decode(out))
이곳은 미세조정을 위한 초안 스크립트입니다: 이곳은 미세조정을 위한 초안 스크립트입니다:
```python ```python
from trl import SFTTrainer from datasets import load_dataset
from peft import LoraConfig from peft import LoraConfig
from transformers import AutoTokenizer, Mamba2ForCausalLM, TrainingArguments from trl import SFTConfig, SFTTrainer
model_id = 'mistralai/Mamba-Codestral-7B-v0.1'
tokenizer = AutoTokenizer.from_pretrained(model_id, revision='refs/pr/9', from_slow=True, legacy=False)
tokenizer.pad_token = tokenizer.eos_token
tokenizer.padding_side = "left" #왼쪽 패딩으로 설정
model = Mamba2ForCausalLM.from_pretrained(model_id, revision='refs/pr/9') model_id = "mistralai/Mamba-Codestral-7B-v0.1"
dataset = load_dataset("Abirate/english_quotes", split="train") dataset = load_dataset("Abirate/english_quotes", split="train")
# CUDA 커널없이는, 배치크기 2가 80GB 장치를 하나 차지합니다. training_args = SFTConfig(dataset_text_field="quote", gradient_checkpointing=True, per_device_train_batch_size=4)
# 하지만 정확도는 감소합니다. lora_config = LoraConfig(target_modules=["x_proj", "embeddings", "in_proj", "out_proj"])
# 실험과 시도를 환영합니다!
training_args = TrainingArguments(
output_dir="./results",
num_train_epochs=3,
per_device_train_batch_size=2,
logging_dir='./logs',
logging_steps=10,
learning_rate=2e-3
)
lora_config = LoraConfig(
r=8,
target_modules=["embeddings", "in_proj", "out_proj"],
task_type="CAUSAL_LM",
bias="none"
)
trainer = SFTTrainer( trainer = SFTTrainer(
model=model, model=model_id,
tokenizer=tokenizer,
args=training_args, args=training_args,
peft_config=lora_config,
train_dataset=dataset, train_dataset=dataset,
dataset_text_field="quote", peft_config=lora_config,
) )
trainer.train() trainer.train()
``` ```

109
docs/source/ko/trainer.md
View File

@ -267,75 +267,45 @@ pip install galore-torch
그런 다음 `optim`에 `["galore_adamw", "galore_adafactor", "galore_adamw_8bit"]` 중 하나와 함께 `optim_target_modules`를 추가합니다. 이는 적용하려는 대상 모듈 이름에 해당하는 문자열, 정규 표현식 또는 전체 경로의 목록일 수 있습니다. 아래는 end-to-end 예제 스크립트입니다(필요한 경우 `pip install trl datasets`를 실행): 그런 다음 `optim`에 `["galore_adamw", "galore_adafactor", "galore_adamw_8bit"]` 중 하나와 함께 `optim_target_modules`를 추가합니다. 이는 적용하려는 대상 모듈 이름에 해당하는 문자열, 정규 표현식 또는 전체 경로의 목록일 수 있습니다. 아래는 end-to-end 예제 스크립트입니다(필요한 경우 `pip install trl datasets`를 실행):
```python ```python
import torch
import datasets import datasets
import trl from trl import SFTConfig, SFTTrainer
from transformers import TrainingArguments, AutoConfig, AutoTokenizer, AutoModelForCausalLM
train_dataset = datasets.load_dataset('imdb', split='train') train_dataset = datasets.load_dataset('imdb', split='train')
args = SFTConfig(
args = TrainingArguments(
output_dir="./test-galore", output_dir="./test-galore",
max_steps=100, max_steps=100,
per_device_train_batch_size=2,
optim="galore_adamw", optim="galore_adamw",
optim_target_modules=["attn", "mlp"] optim_target_modules=[r".*.attn.*", r".*.mlp.*"],
gradient_checkpointing=True,
) )
trainer = SFTTrainer(
model_id = "google/gemma-2b" model="google/gemma-2b",
config = AutoConfig.from_pretrained(model_id)
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_config(config).to(0)
trainer = trl.SFTTrainer(
model=model,
args=args, args=args,
train_dataset=train_dataset, train_dataset=train_dataset,
dataset_text_field='text',
max_seq_length=512,
) )
trainer.train() trainer.train()
``` ```
GaLore가 지원하는 추가 매개변수를 전달하려면 `optim_args`를 설정합니다. 예를 들어: GaLore가 지원하는 추가 매개변수를 전달하려면 `optim_args`를 설정합니다. 예를 들어:
```python ```python
import torch
import datasets import datasets
import trl from trl import SFTConfig, SFTTrainer
from transformers import TrainingArguments, AutoConfig, AutoTokenizer, AutoModelForCausalLM
train_dataset = datasets.load_dataset('imdb', split='train') train_dataset = datasets.load_dataset('imdb', split='train')
args = SFTConfig(
args = TrainingArguments(
output_dir="./test-galore", output_dir="./test-galore",
max_steps=100, max_steps=100,
per_device_train_batch_size=2,
optim="galore_adamw", optim="galore_adamw",
optim_target_modules=["attn", "mlp"], optim_target_modules=[r".*.attn.*", r".*.mlp.*"],
optim_args="rank=64, update_proj_gap=100, scale=0.10", optim_args="rank=64, update_proj_gap=100, scale=0.10",
gradient_checkpointing=True,
) )
trainer = SFTTrainer(
model_id = "google/gemma-2b" model="google/gemma-2b",
config = AutoConfig.from_pretrained(model_id)
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_config(config).to(0)
trainer = trl.SFTTrainer(
model=model,
args=args, args=args,
train_dataset=train_dataset, train_dataset=train_dataset,
dataset_text_field='text',
max_seq_length=512,
) )
trainer.train() trainer.train()
``` ```
@ -348,37 +318,22 @@ trainer.train()
다음과 같이 옵티마이저 이름에 `layerwise`를 추가하여 레이어별 최적화를 수행할 수도 있습니다: 다음과 같이 옵티마이저 이름에 `layerwise`를 추가하여 레이어별 최적화를 수행할 수도 있습니다:
```python ```python
import torch
import datasets import datasets
import trl from trl import SFTConfig, SFTTrainer
from transformers import TrainingArguments, AutoConfig, AutoTokenizer, AutoModelForCausalLM
train_dataset = datasets.load_dataset('imdb', split='train') train_dataset = datasets.load_dataset('imdb', split='train')
args = SFTConfig(
args = TrainingArguments(
output_dir="./test-galore", output_dir="./test-galore",
max_steps=100, max_steps=100,
per_device_train_batch_size=2,
optim="galore_adamw_layerwise", optim="galore_adamw_layerwise",
optim_target_modules=["attn", "mlp"] optim_target_modules=[r".*.attn.*", r".*.mlp.*"],
gradient_checkpointing=True,
) )
trainer = SFTTrainer(
model_id = "google/gemma-2b" model="google/gemma-2b",
config = AutoConfig.from_pretrained(model_id)
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_config(config).to(0)
trainer = trl.SFTTrainer(
model=model,
args=args, args=args,
train_dataset=train_dataset, train_dataset=train_dataset,
dataset_text_field='text',
max_seq_length=512,
) )
trainer.train() trainer.train()
``` ```
@ -398,39 +353,21 @@ LOMO 옵티마이저는 [제한된 자원으로 대형 언어 모델의 전체
다음은 IMDB 데이터셋에서 [google/gemma-2b](https://huggingface.co/google/gemma-2b)를 최대 정밀도로 미세 조정하는 간단한 스크립트입니다: 다음은 IMDB 데이터셋에서 [google/gemma-2b](https://huggingface.co/google/gemma-2b)를 최대 정밀도로 미세 조정하는 간단한 스크립트입니다:
```python ```python
import torch
import datasets import datasets
from transformers import TrainingArguments, AutoTokenizer, AutoModelForCausalLM from trl import SFTConfig, SFTTrainer
import trl
train_dataset = datasets.load_dataset('imdb', split='train') train_dataset = datasets.load_dataset('imdb', split='train')
args = SFTConfig(
args = TrainingArguments(
output_dir="./test-lomo", output_dir="./test-lomo",
max_steps=1000, max_steps=100,
per_device_train_batch_size=4,
optim="adalomo", optim="adalomo",
gradient_checkpointing=True, gradient_checkpointing=True,
logging_strategy="steps",
logging_steps=1,
learning_rate=2e-6,
save_strategy="no",
run_name="lomo-imdb",
) )
trainer = SFTTrainer(
model_id = "google/gemma-2b" model="google/gemma-2b",
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id).to(0)
trainer = trl.SFTTrainer(
model=model,
args=args, args=args,
train_dataset=train_dataset, train_dataset=train_dataset,
dataset_text_field='text',
max_seq_length=1024,
) )
trainer.train() trainer.train()
``` ```