Update app.py

app.py

@@ -2,78 +2,69 @@ import torch
 from transformers import AutoModelForCausalLM, AutoTokenizer, TrainingArguments, Trainer
 from peft import LoraConfig, get_peft_model
 from datasets import load_dataset
-import spaces
 import gradio as gr
+import os
+
 # === 1️⃣ MODEL VE TOKENIZER YÜKLEME ===
-MODEL_NAME = "mistralai/Mistral-7B-v0.1"  # Hugging Face model adı
+MODEL_NAME = "mistralai/Mistral-7B-v0.1"
 tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
 
 # === 2️⃣ CPU OPTİMİZASYONU ===
-
-
-
-
-zero = torch.Tensor([0]).cuda()
-print(zero.device) # <-- 'cpu' 🤔
-
-
-print(zero.device) # <-- 'cuda:0' 🤗
-    
-
-print(zero.device) # <-- 'cuda:0' 🤗
-
-
-
-torch_dtype = torch.float32  # float32 seçtik çünkü CPU'da bf16 genelde yok
-model = AutoModelForCausalLM.from_pretrained(MODEL_NAME, torch_dtype=torch_dtype)
+torch_dtype = torch.float32  # CPU için float32 en iyisi
+device = "cuda" if torch.cuda.is_available() else "cpu"
+model = AutoModelForCausalLM.from_pretrained(MODEL_NAME, torch_dtype=torch_dtype).to(device)
 
 # === 3️⃣ LoRA AYARLARI ===
 lora_config = LoraConfig(
-    r=8,              
-    lora_alpha=32,    
-    lora_dropout=0.1, 
+    r=8,
+    lora_alpha=32,
+    lora_dropout=0.1,
     bias="none",
-    target_modules=["q_proj", "v_proj"], 
+    target_modules=["q_proj", "v_proj"],
 )
 model = get_peft_model(model, lora_config)
 
-# === 4️⃣ VERİ SETİ ===
-dataset = load_dataset("oscar", "unshuffled_deduplicated_tr", trust_remote_code=True)  # 🔥 ÇÖZÜM: trust_remote_code=True
-train_data = dataset["train"].shuffle(seed=42).select(range(10000))  # Küçük subset
+# === 4️⃣ VERİ SETİ (OPTİMİZE) ===
+DATASET_PATH = "oscar_tr.parquet"
 
-# === 5️⃣ TOKENLEŞTİRME FONKSİYONU ===
+if os.path.exists(DATASET_PATH):
+    print("📂 Kaydedilmiş veri seti bulundu, yükleniyor...")
+    from datasets import Dataset
+    dataset = Dataset.from_parquet(DATASET_PATH)
+else:
+    print("🌍 Veri seti indiriliyor ve kaydediliyor...")
+    dataset = load_dataset("oscar", "unshuffled_deduplicated_tr", split="train")
+    dataset = dataset.shuffle(seed=42).select(range(10000))  # 10K veri ile sınırladık
+    dataset.to_parquet(DATASET_PATH)  # İlk çalışmada veriyi kaydediyoruz
+
+# === 5️⃣ TOKENLEŞTİRME (OPTİMİZE) ===
 def tokenize_function(examples):
     return tokenizer(examples["text"], truncation=True, max_length=512)
 
-tokenized_datasets = train_data.map(tokenize_function, batched=True)
+tokenized_datasets = dataset.map(tokenize_function, batched=True, num_proc=4)  # 🔥 Paralel işlem
 
 # === 6️⃣ EĞİTİM AYARLARI ===
 training_args = TrainingArguments(
     output_dir="./mistral_lora_cpu",
-    per_device_train_batch_size=1,  
-    gradient_accumulation_steps=16, 
-    learning_rate=5e-4,  
-    num_train_epochs=1,  
+    per_device_train_batch_size=1,
+    gradient_accumulation_steps=16,
+    learning_rate=5e-4,
+    num_train_epochs=1,
     save_steps=500,
     save_total_limit=2,
     logging_dir="./logs",
     logging_steps=10,
-    optim="adamw_torch",  # 🔥 ÇÖZÜM: bitsandbytes yerine adamw_torch
+    optim="adamw_torch",
+    dataloader_pin_memory=True,  # 🔥 GPU bellek optimizasyonu
 )
 
 # === 7️⃣ MODEL EĞİTİMİ ===
-@spaces.GPU
 def trainf():
-
-    v= Trainer(
-    
+    trainer = Trainer(
         model=model,
-    
         args=training_args,
-    
         train_dataset=tokenized_datasets,
-
     )
-    return v
+    trainer.train()
 
-trainf().train()
+trainf()