edits

app.py

@@ -10,13 +10,13 @@ from flask import Flask, request, jsonify
 from flask_cors import CORS
 import threading, torch, json
 from datasets import Dataset
-from transformers import AutoModelForCausalLM, AutoTokenizer, Trainer, TrainingArguments, BitsAndBytesConfig
+from transformers import AutoModelForCausalLM, AutoTokenizer, Trainer, TrainingArguments
 from peft import LoraConfig, get_peft_model, TaskType, PeftModel
 
 # ===============================
 # 2️⃣ إعداد المتغيرات
 # ===============================
-MODEL_NAME = "Qwen/Qwen2.5-7B-Instruct"   # ✅ موديل 7B
+MODEL_NAME = "Qwen/Qwen2.5-7B-Instruct"
 FLASK_PORT = 7860
 
 SAVE_DIR = "/tmp/lora_model"
@@ -33,30 +33,17 @@ device = "cuda" if torch.cuda.is_available() else "cpu"
 print("🔹 Using device:", device)
 
 # ===============================
-# 4️⃣ تحميل الموديل مع 4bit Quantization
+# 4️⃣ تحميل الموديل (بدون quantization)
 # ===============================
-print("🔄 Loading base model with 4-bit quantization...")
-
-bnb_config = BitsAndBytesConfig(
-    load_in_4bit=True,                 # ✅ تشغيل 4bit quantization
-    bnb_4bit_use_double_quant=True,
-    bnb_4bit_quant_type="nf4",
-    bnb_4bit_compute_dtype=torch.bfloat16
-)
-
+print("🔄 Loading base model...")
 tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
 
-try:
-    model = AutoModelForCausalLM.from_pretrained(
-        MODEL_NAME,
-        device_map="auto",
-        quantization_config=bnb_config,
-        trust_remote_code=True
-    )
-    print("✅ Base model loaded in 4-bit!")
-except Exception as e:
-    print("⚠️ Failed to load base model:", e)
-    raise e
+model = AutoModelForCausalLM.from_pretrained(
+    MODEL_NAME,
+    device_map="auto",
+    trust_remote_code=True
+)
+print("✅ Base model loaded!")
 
 # ===============================
 # 5️⃣ إعداد LoRA
@@ -79,7 +66,6 @@ if os.path.exists(SAVE_DIR):
         base_model = AutoModelForCausalLM.from_pretrained(
             MODEL_NAME,
             device_map="auto",
-            quantization_config=bnb_config,
             trust_remote_code=True
         )
         model = PeftModel.from_pretrained(base_model, SAVE_DIR)
@@ -121,7 +107,6 @@ def generate():
 
         user_name = user_info["name"]
 
-        # توليد نصوص من الموديل
         inputs = tokenizer(prompt, return_tensors="pt").to(device)
         outputs = model.generate(
             **inputs,
@@ -201,7 +186,6 @@ def train():
             else:
                 return jsonify({"error": "No file uploaded or JSON provided"}), 400
 
-        # الحفظ في dataset.json
         with open(DATASET_FILE, "w", encoding="utf-8") as f:
             json.dump(data, f, ensure_ascii=False, indent=2)
 

requirements.txt

@@ -1,4 +1,6 @@
 transformers
+accelerate
+bitsandbytes
 peft
 torch
 datasets