Simplify and update trl examples

Author: Quentin Gallouédec

docs/source/ar/trainer.md

@@ -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`):
 
 ```python
-import torch
 import datasets
-import trl
-
-from transformers import TrainingArguments, AutoConfig, AutoTokenizer, AutoModelForCausalLM
+from trl import SFTConfig, SFTTrainer
 
 train_dataset = datasets.load_dataset('imdb', split='train')
-
-args = TrainingArguments(
-    output_dir="./test-galore"،
+args = SFTConfig(
+    output_dir="./test-galore",
     max_steps=100,
-    per_device_train_batch_size=2,
-    optim="galore_adamw"،
-    optim_target_modules=[r".*.attn.*"، r".*.mlp.*"]
+    optim="galore_adamw",
+    optim_target_modules=[r".*.attn.*", r".*.mlp.*"],
+    gradient_checkpointing=True,
 )
-
-model_id = "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, 
+trainer = SFTTrainer(
+    model="google/gemma-2b",
     args=args,
     train_dataset=train_dataset,
-    dataset_text_field='text',
-    max_seq_length=512,
 )
-
 trainer.train()
 ```
 
 لتمرير معامﻻت إضافية يدعمها  GaLore، يجب عليك تمرير `optim_args` بشكل صحيح، على سبيل المثال:
 
 ```python
-import torch
 import datasets
-import trl
-
-from transformers import TrainingArguments, AutoConfig, AutoTokenizer, AutoModelForCausalLM
+from trl import SFTConfig, SFTTrainer
 
 train_dataset = datasets.load_dataset('imdb', split='train')
-
-args = TrainingArguments(
+args = SFTConfig(
     output_dir="./test-galore",
     max_steps=100,
-    per_device_train_batch_size=2,
     optim="galore_adamw",
     optim_target_modules=[r".*.attn.*", r".*.mlp.*"],
     optim_args="rank=64, update_proj_gap=100, scale=0.10",
+    gradient_checkpointing=True,
 )
-
-model_id = "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, 
+trainer = SFTTrainer(
+    model="google/gemma-2b",
     args=args,
     train_dataset=train_dataset,
-    dataset_text_field='text',
-    max_seq_length=512,
 )
-
 trainer.train()
 ```
 يمكنك قراءة المزيد حول الطريقة في [المستودع الأصلي](https://github.com/jiaweizzhao/GaLore) أو [الورقة البحثية](https://arxiv.org/abs/2403.03507).
@@ -386,37 +356,23 @@ trainer.train()
 يمكنك أيضًا إجراء تحسين طبقة تلو الأخرى عن طريق إضافة `layerwise` إلى اسم المُحسِّن كما هو موضح أدناه:
 
 ```python
-import torch
 import datasets
-import trl
-
-from transformers import TrainingArguments، AutoConfig، AutoTokenizer، AutoModelForCausalLM
+from trl import SFTConfig, SFTTrainer
 
-train_dataset = datasets.load_dataset('imdb'، split='train')
-
-args = TrainingArguments(
-    output_dir="./test-galore"،
-    max_steps=100،
-    per_device_train_batch_size=2،
-    optim="galore_adamw_layerwise"،
-    optim_target_modules=[r".*.attn.*"، r".*.mlp.*"]
+train_dataset = datasets.load_dataset('imdb', split='train')
+args = SFTConfig(
+    output_dir="./test-galore",
+    max_steps=100,
+    optim="galore_adamw_layerwise",
+    optim_target_modules=[r".*.attn.*", r".*.mlp.*"],
+    optim_args="rank=64, update_proj_gap=100, scale=0.10",
+    gradient_checkpointing=True,
 )
-
-model_id = "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،
-    train_dataset=train_dataset،
-    dataset_text_field='text'،
-    max_seq_length=512،
+trainer = SFTTrainer(
+    model="google/gemma-2b",
+    args=args,
+    train_dataset=train_dataset,
 )
-
 trainer.train()
 ```
 
@@ -436,39 +392,21 @@ 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:
 
 ```python
-import torch
 import datasets
-from transformers import TrainingArguments، AutoTokenizer، AutoModelForCausalLM
-import trl
+from trl import SFTConfig, SFTTrainer
 
-train_dataset = datasets.load_dataset('imdb'، split='train')
-
-args = TrainingArguments(
-    output_dir="./test-lomo"،
-    max_steps=100،
-    per_device_train_batch_size=4،
-    optim="adalomo"،
-    gradient_checkpointing=True،
-    logging_strategy="steps"،
-    logging_steps=1،
-    learning_rate=2e-6،
-    save_strategy="no"،
-    run_name="lomo-imdb"،
+train_dataset = datasets.load_dataset('imdb', split='train')
+args = SFTConfig(
+    output_dir="./test-lomo",
+    max_steps=100,
+    optim="adalomo",
+    gradient_checkpointing=True,
 )
-
-model_id = "google/gemma-2b"
-
-tokenizer = AutoTokenizer.from_pretrained(model_id)
-model = AutoModelForCausalLM.from_pretrained(model_id، low_cpu_mem_usage=True).to(0)
-
-trainer = trl.SFTTrainer(
-    model=model،
-    args=args،
-    train_dataset=train_dataset،
-    dataset_text_field='text'،
-    max_seq_length=1024،
+trainer = SFTTrainer(
+    model="google/gemma-2b",
+    args=args,
+    train_dataset=train_dataset,
 )
-
 trainer.train()
 ```
 
@@ -524,39 +462,21 @@ 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:
 ```python
-import torch
 import datasets
-from transformers import TrainingArguments, AutoTokenizer, AutoModelForCausalLM
-import trl
+from trl import SFTConfig, SFTTrainer
 
 train_dataset = datasets.load_dataset('imdb', split='train')
-
-args = TrainingArguments(
-    output_dir="./test-schedulefree",
-    max_steps=1000,
-    per_device_train_batch_size=4,
+args = SFTConfig(
+    output_dir="./test-galore",
+    max_steps=100,
     optim="schedule_free_adamw",
     gradient_checkpointing=True,
-    logging_strategy="steps",
-    logging_steps=1,
-    learning_rate=2e-6,
-    save_strategy="no",
-    run_name="sfo-imdb",
 )
-
-model_id = "google/gemma-2b"
-
-tokenizer = AutoTokenizer.from_pretrained(model_id)
-model = AutoModelForCausalLM.from_pretrained(model_id, low_cpu_mem_usage=True).to(0)
-
-trainer = trl.SFTTrainer(
-    model=model, 
+trainer = SFTTrainer(
+    model="google/gemma-2b",
     args=args,
     train_dataset=train_dataset,
-    dataset_text_field='text',
-    max_seq_length=1024,
 )
-
 trainer.train()
 ```
 ## تسريع ومدرب

docs/source/en/model_doc/mamba.md

@@ -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.
 - The example below demonstrates how to fine-tune Mamba with [PEFT](https://huggingface.co/docs/peft).
 
-   ```py
-   from datasets import load_dataset
-   from trl import SFTTrainer
-   from peft import LoraConfig
-   from transformers import AutoTokenizer, AutoModelForCausalLM, TrainingArguments
-   
-   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")
-   training_args = TrainingArguments(
-       output_dir="./results",
-       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,
+  ```py
+  from datasets import load_dataset
+  from trl import SFTConfig, SFTTrainer
+  from peft import LoraConfig
+
+  model_id = "state-spaces/mamba-130m-hf"
+  dataset = load_dataset("Abirate/english_quotes", split="train")
+  training_args = SFTConfig(dataset_text_field="quote")
+  lora_config =  LoraConfig(target_modules=["x_proj", "embeddings", "in_proj", "out_proj"])
+  trainer = SFTTrainer(
+      model=model_id,
       args=training_args,
-       peft_config=lora_config,
-       train_dataset=dataset,
-       dataset_text_field="quote",
-   )
-   trainer.train()
+      train_dataset=dataset,
+      peft_config=lora_config,
+  )
+  trainer.train()
    ```
 
 ## MambaConfig

docs/source/en/model_doc/mamba2.md

@@ -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).
 
 ```python 
-from trl import SFTTrainer
+from datasets import load_dataset
 from peft import LoraConfig
-from transformers import AutoTokenizer, Mamba2ForCausalLM, TrainingArguments
-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
+from trl import SFTConfig, SFTTrainer
 
-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")
-# Without CUDA kernels, batch size of 2 occupies one 80GB device
-# but precision can be reduced.
-# 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"
-)
+training_args = SFTConfig(dataset_text_field="quote", gradient_checkpointing=True, per_device_train_batch_size=4)
+lora_config =  LoraConfig(target_modules=["x_proj", "embeddings", "in_proj", "out_proj"])
 trainer = SFTTrainer(
-    model=model,
-    tokenizer=tokenizer,
+    model=model_id,
     args=training_args,
-    peft_config=lora_config,
     train_dataset=dataset,
-    dataset_text_field="quote",
+    peft_config=lora_config,
 )
 trainer.train()
 ```