utilisation du GPU

Finetunning/finetunning.py

@@ -48,7 +48,11 @@ print(f"Max sequence length: {tokenizer.model_max_length}")
 # ----------------------------
 # [2/7] Quantization config (QLoRA)
 # ----------------------------
-print(f"{80 * '_'}\n[2/7] Configuring 4-bit quantization (BitsAndBytes)...")
+print(f"{80 * '_'}\n[2/7] Loading model in 4-bit mode (optimized QLoRA)...")
+
+assert torch.cuda.is_available(), "CUDA GPU not detected!"
+print(f"Using GPU: {torch.cuda.get_device_name(0)}")
+
 bnb_config = BitsAndBytesConfig(
     load_in_4bit=True,
     bnb_4bit_quant_type="nf4",
@@ -56,17 +60,16 @@ bnb_config = BitsAndBytesConfig(
     bnb_4bit_use_double_quant=True,
 )
 
-print("4-bit NF4 quantization configured.")
-
-print("Loading model...")
 model = AutoModelForCausalLM.from_pretrained(
     MODEL_NAME,
-    device_map="auto",
+    device_map="cuda",          # 🔥 SAFE
     quantization_config=bnb_config,
-    dtype=torch.float16,
+    low_cpu_mem_usage=True,
     trust_remote_code=True,
 )
-print("Model loaded successfully.")
+
+print("Model loaded successfully in 4-bit mode on GPU.")
+
 
 # ----------------------------
 # [3/7] Prepare model for k-bit training
@@ -119,8 +122,7 @@ print("Formatting dataset for Ukrainian → French translation...")
 
 def format_prompt(example):
     return {
-        "text": (
-            "<|user|>\n"
+        "text": ("<|user|>\n"
             "Translate the following Ukrainian text into French.\n"
             f"Ukrainian: {example['text']}\n"
             "<|assistant|>\n"
@@ -154,13 +156,13 @@ training_args = TrainingArguments(
     save_steps=500,
     save_total_limit=2,
     report_to="none",
+    dataloader_pin_memory=False,
 )
 
 print("Training arguments ready.")
 print(f"Output directory: {OUTPUT_DIR}")
 print(f"Epochs: {training_args.num_train_epochs}")
-print(
-    f"Effective batch size: "
+print(f"Effective batch size: "
     f"{training_args.per_device_train_batch_size * training_args.gradient_accumulation_steps}"
 )
 
@@ -171,7 +173,7 @@ print("Initializing SFTTrainer...")
 trainer = SFTTrainer(
     model=model,
     train_dataset=dataset["train"],
-    tokenizer=tokenizer,
+    processing_class=tokenizer,
     args=training_args,
 )
 print("Trainer initialized.")
@@ -181,12 +183,16 @@ print("Trainer initialized.")
 # ----------------------------
 print(f"{80 * '_'}\n[7/7] Starting training...")
 try:
-    trainer.train(resume_from_checkpoint=True)
+    train_output = trainer.train(resume_from_checkpoint=True)
 except Exception as e:
     print("No checkpoint found or resume failed, starting fresh training.")
     print(f"Reason: {e}")
-    trainer.train()
+    train_output = trainer.train()
 
+print("\n=== Training summary ===")
+print(f"Global steps: {train_output.global_step}")
+print(f"Training loss: {train_output.training_loss}")
+print(f"Metrics: {train_output.metrics}")
 print("Training completed successfully.")
 
 # ----------------------------