autorise la reprise d'unentrainement

Finetunning/finetunning.py

@@ -1,3 +1,4 @@
+import os
 import torch
 from datasets import load_dataset
 from transformers import (
@@ -11,15 +12,18 @@ from peft import (
     prepare_model_for_kbit_training,
 )
 from trl import SFTTrainer
-import os
+
+# ----------------------------
+# Environment safety (Windows)
+# ----------------------------
 os.environ["TORCHDYNAMO_DISABLE"] = "1"
 
 # ----------------------------
 # Model configuration
 # ----------------------------
-MODEL_NAME = "Qwen/Qwen2.5-14B-Instruct"
+MODEL_NAME = "Qwen/Qwen2.5-7B-Instruct"
 
-print("=== Starting fine-tuning script ===")
+print(f"=== Starting fine-tuning script {MODEL_NAME} ===")
 
 print(f"{80 * '_'}\n[1/7] Loading tokenizer...")
 tokenizer = AutoTokenizer.from_pretrained(
@@ -27,7 +31,7 @@ tokenizer = AutoTokenizer.from_pretrained(
     trust_remote_code=True
 )
 
-# Ensure padding token is defined
+# Ensure padding is defined
 tokenizer.pad_token = tokenizer.eos_token
 tokenizer.model_max_length = 1024
 
@@ -38,13 +42,19 @@ model = AutoModelForCausalLM.from_pretrained(
     MODEL_NAME,
     load_in_4bit=True,
     device_map="auto",
-    torch_dtype=torch.float16,  # OK for weights
+    torch_dtype=torch.float16,  # weights in fp16, gradients fp32
     trust_remote_code=True,
 )
 print("Model loaded.")
 
 print(f"{80 * '_'}\n[3/7] Preparing model for k-bit training...")
 model = prepare_model_for_kbit_training(model)
+
+# Fix future PyTorch checkpointing behavior
+model.gradient_checkpointing_enable(
+    gradient_checkpointing_kwargs={"use_reentrant": False}
+)
+
 print("Model prepared for k-bit training.")
 
 # ----------------------------
@@ -70,7 +80,7 @@ lora_config = LoraConfig(
 
 model = get_peft_model(model, lora_config)
 model.print_trainable_parameters()
-print("LoRA adapters attached to the model.")
+print("LoRA adapters attached.")
 
 # ----------------------------
 # Dataset loading
@@ -80,6 +90,7 @@ dataset = load_dataset(
     "json",
     data_files="traductions.json"
 )
+
 print(f"Dataset loaded with {len(dataset['train'])} samples.")
 
 print("Formatting dataset for Ukrainian → French translation...")
@@ -92,34 +103,32 @@ def format_prompt(example):
     )
     return {"text": prompt}
 
-dataset = dataset.map(format_prompt, remove_columns=dataset["train"].column_names)
+dataset = dataset.map(
+    format_prompt,
+    remove_columns=dataset["train"].column_names
+)
+
 print("Dataset formatting completed.")
 
 # ----------------------------
-# Training arguments
+# Training arguments (AMP OFF)
 # ----------------------------
 print(f"{80 * '_'}\n[6/7] Initializing training arguments...")
 training_args = TrainingArguments(
-    output_dir="./qwen-uk-fr-lora",
+    output_dir="./qwen2.5-7b-uk-fr-lora",
     per_device_train_batch_size=1,
     gradient_accumulation_steps=8,
     learning_rate=2e-4,
-    num_train_epochs=3,
-
+    num_train_epochs=2,  # 2 epochs usually enough for translation
     fp16=False,
     bf16=False,
-
     logging_steps=10,
     save_steps=500,
     save_total_limit=2,
-
-    # Use 32-bit optimizer
     optim="paged_adamw_32bit",
-
     report_to="none",
 )
 
-
 print("Training arguments ready.")
 
 # ----------------------------
@@ -138,15 +147,15 @@ print("Trainer initialized.")
 # Train
 # ----------------------------
 print(f"{80 * '_'}\n[7/7] Starting training...")
-trainer.train()
+trainer.train(resume_from_checkpoint=True)
 print("Training completed successfully.")
 
 # ----------------------------
 # Save LoRA adapter
 # ----------------------------
 print("Saving LoRA adapter and tokenizer...")
-trainer.model.save_pretrained("./qwen-uk-fr-lora")
-tokenizer.save_pretrained("./qwen-uk-fr-lora")
+trainer.model.save_pretrained("./qwen2.5-7b-uk-fr-lora")
+tokenizer.save_pretrained("./qwen2.5-7b-uk-fr-lora")
 
 print("=== Fine-tuning finished ===")
-print("LoRA adapter saved in ./qwen-uk-fr-lora")
+print("LoRA adapter saved in ./qwen2.5-7b-uk-fr-lora")