Update app.py

app.py

@@ -10,8 +10,8 @@ import spaces
 MODEL_NAME = "mistralai/Mistral-7B-v0.1"
 tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
 
-torch_dtype = torch.float32
-model = AutoModelForCausalLM.from_pretrained(MODEL_NAME, torch_dtype=torch_dtype)
+torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32
+model = AutoModelForCausalLM.from_pretrained(MODEL_NAME, torch_dtype=torch_dtype, device_map="auto")
 
 # === 2️⃣ LoRA AYARLARI ===
 lora_config = LoraConfig(
@@ -21,22 +21,26 @@ lora_config = LoraConfig(
     bias="none",
     target_modules=["q_proj", "v_proj"],
 )
-model = get_peft_model(model, lora_config)
+model = get_peft_model(model, lora_config).to("cuda" if torch.cuda.is_available() else "cpu")
 
 # === 3️⃣ VERİ SETİ ===
-dataset = load_dataset("oscar", "unshuffled_deduplicated_tr", trust_remote_code=True)
-subset = dataset["train"].shuffle(seed=42).select(range(10000))
-
 @spaces.GPU
-def tokenize_function(examples):
-    return tokenizer(examples["text"], truncation=True, max_length=512)
+def load_and_prepare_dataset():
+    dataset = load_dataset("oscar", "unshuffled_deduplicated_tr", trust_remote_code=True)
+    subset = dataset["train"].shuffle(seed=42).select(range(10000))
+    
+    def tokenize_function(examples):
+        return tokenizer(examples["text"], truncation=True, max_length=512)
+    
+    tokenized_datasets = subset.map(tokenize_function, batched=True)
+    return tokenized_datasets.train_test_split(test_size=0.1, seed=42)
 
-tokenized_datasets = subset.map(tokenize_function, batched=True)
+split_dataset = load_and_prepare_dataset()
 
 # === 4️⃣ EĞİTİM AYARLARI ===
 batch_size = 1  
 num_epochs = 1  
-max_steps = (len(tokenized_datasets) // batch_size) * num_epochs
+max_steps = (len(split_dataset["train"]) // batch_size) * num_epochs
 
 training_args = TrainingArguments(
     output_dir="./mistral_lora",
@@ -50,13 +54,14 @@ training_args = TrainingArguments(
     logging_dir="./logs",
     logging_steps=10,
     optim="adamw_torch",
-    no_cuda=True,  
+    fp16=torch.cuda.is_available(),  
 )
 
 trainer = Trainer(
     model=model,
     args=training_args,
-    train_dataset=tokenized_datasets,
+    train_dataset=split_dataset["train"],
+    eval_dataset=split_dataset["test"],
 )
 
 @spaces.GPU