edits

app.py

@@ -6,17 +6,17 @@ os.environ["HF_HOME"] = "/tmp"
 os.environ["TRANSFORMERS_CACHE"] = "/tmp"
 os.environ["PEFT_HOME"] = "/tmp"
 
+from transformers import AutoModelForCausalLM, AutoTokenizer, Trainer, TrainingArguments
+from peft import LoraConfig, get_peft_model, TaskType, PeftModel
 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
-from peft import LoraConfig, get_peft_model, TaskType, PeftModel
 
 # ===============================
 # 2️⃣ إعداد المتغيرات
 # ===============================
-MODEL_NAME = "Qwen/Qwen2.5-7B-Instruct"
+MODEL_NAME = "Qwen/Qwen2.5-0.5B-Instruct"
 FLASK_PORT = 7860
 
 SAVE_DIR = "/tmp/lora_model"
@@ -33,29 +33,37 @@ device = "cuda" if torch.cuda.is_available() else "cpu"
 print("🔹 Using device:", device)
 
 # ===============================
-# 4️⃣ تحميل الموديل (بدون quantization)
+# 4️⃣ تحميل الموديل الأساسي بأمان
 # ===============================
 print("🔄 Loading base model...")
 tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
 
-model = AutoModelForCausalLM.from_pretrained(
-    MODEL_NAME,
-    device_map="auto",
-    trust_remote_code=True
-)
-print("✅ Base model loaded!")
+try:
+    model = AutoModelForCausalLM.from_pretrained(
+        MODEL_NAME,
+        device_map="auto" if device=="cuda" else {"": "cpu"},
+        torch_dtype=torch.float16 if device=="cuda" else torch.float32,
+        load_in_8bit=(device=="cuda"),
+        trust_remote_code=True
+    )
+    model.to(device)
+    print("✅ Base model loaded!")
+except Exception as e:
+    print("⚠️ Failed to load base model:", e)
+    raise e
 
 # ===============================
 # 5️⃣ إعداد LoRA
 # ===============================
 lora_config = LoraConfig(
     task_type=TaskType.CAUSAL_LM,
-    r=8,
-    lora_alpha=16,
-    lora_dropout=0.1,
+    r=2,
+    lora_alpha=8,
+    lora_dropout=0.05,
     target_modules=["q_proj", "v_proj"]
 )
 model = get_peft_model(model, lora_config)
+model.to(device)
 print("✅ LoRA model ready on device:", device)
 
 # ===============================
@@ -65,10 +73,13 @@ if os.path.exists(SAVE_DIR):
     try:
         base_model = AutoModelForCausalLM.from_pretrained(
             MODEL_NAME,
-            device_map="auto",
+            device_map="auto" if device=="cuda" else {"": "cpu"},
+            torch_dtype=torch.float16 if device=="cuda" else torch.float32,
+            load_in_8bit=(device=="cuda"),
             trust_remote_code=True
         )
         model = PeftModel.from_pretrained(base_model, SAVE_DIR)
+        model.to(device)
         print("✅ LoRA weights auto-loaded from disk!")
     except Exception as e:
         print("⚠️ Failed to auto-load LoRA weights:", e)
@@ -107,10 +118,34 @@ def generate():
 
         user_name = user_info["name"]
 
+        if current_dataset is None:
+            return jsonify({
+                "messages": [
+                    {"role": "assistant", "content": f"⚖️ أهلاً {user_name}! أنا زين، مساعدك الذكي. برجاء طرح سؤال قانوني."}
+                ]
+            })
+
+        # Search in dataset
+        for item in current_dataset:
+            lines = item["text"].split("\n")
+            for line in lines:
+                if line.lower().startswith("user:") and prompt in line:
+                    response_messages = []
+                    for l in lines:
+                        if ":" in l:
+                            role, content = l.split(":", 1)
+                            response_messages.append({
+                                "role": role.strip().lower(),
+                                "content": content.strip()
+                            })
+                    response_messages.insert(0, {"role": "assistant", "content": f"أهلاً {user_name} 👋"})
+                    return jsonify({"messages": response_messages})
+
+        # Fallback: generate
         inputs = tokenizer(prompt, return_tensors="pt").to(device)
         outputs = model.generate(
             **inputs,
-            max_new_tokens=200,
+            max_new_tokens=150,
             temperature=0.7,
             do_sample=True
         )
@@ -137,6 +172,7 @@ def load_dataset():
                 data = json.load(f)
             except:
                 data = []
+
         if data:
             formatted = []
             for item in data:
@@ -167,6 +203,7 @@ def train():
         epochs = 1
         batch_size = 1
 
+        # 1️⃣ JSON مباشر
         if request.is_json:
             body = request.get_json()
             if isinstance(body, list):
@@ -175,8 +212,11 @@ def train():
                 data = body["data"]
             else:
                 return jsonify({"error": "JSON format not recognized"}), 400
+
             epochs = int(body.get("epochs", 1)) if isinstance(body, dict) else 1
             batch_size = int(body.get("batch_size", 1)) if isinstance(body, dict) else 1
+
+        # 2️⃣ form-data ملف
         else:
             file = request.files.get("file")
             if file:
@@ -186,11 +226,35 @@ def train():
             else:
                 return jsonify({"error": "No file uploaded or JSON provided"}), 400
 
+        # دمج وإزالة التكرارات حسب محتوى الرسائل
+        old_data = []
+        if os.path.exists(DATASET_FILE):
+            with open(DATASET_FILE, "r", encoding="utf-8") as f:
+                try:
+                    old_data = json.load(f)
+                except:
+                    old_data = []
+
+        combined = old_data + data
+        unique_data, seen_texts = [], set()
+        for item in combined:
+            if "messages" in item:
+                text_key = "\n".join([f"{m['role'].strip().lower()}:{m['content'].strip()}" for m in item["messages"]])
+            elif "text" in item:
+                text_key = item["text"].strip()
+            else:
+                continue
+            if text_key not in seen_texts:
+                seen_texts.add(text_key)
+                unique_data.append(item)
+
+        # الحفظ
         with open(DATASET_FILE, "w", encoding="utf-8") as f:
-            json.dump(data, f, ensure_ascii=False, indent=2)
+            json.dump(unique_data, f, ensure_ascii=False, indent=2)
 
+        # تجهيز Dataset للتدريب
         formatted = []
-        for item in data:
+        for item in unique_data:
             if "messages" in item:
                 conversation = ""
                 for msg in item["messages"]:
@@ -202,7 +266,7 @@ def train():
         current_dataset = Dataset.from_list(formatted)
 
         def tokenize_function(examples):
-            return tokenizer(examples["text"], truncation=True, padding="max_length", max_length=512)
+            return tokenizer(examples["text"], truncation=True, padding="max_length", max_length=256)
 
         tokenized_dataset = current_dataset.map(tokenize_function, batched=True, remove_columns=["text"])
 
@@ -210,7 +274,7 @@ def train():
             output_dir=SAVE_DIR,
             per_device_train_batch_size=batch_size,
             num_train_epochs=epochs,
-            logging_steps=5,
+            logging_steps=2,
             save_strategy="epoch",
             fp16=torch.cuda.is_available(),
             report_to=None
@@ -218,6 +282,7 @@ def train():
 
         trainer = Trainer(model=model, args=training_args, train_dataset=tokenized_dataset)
 
+        # تشغيل التدريب في Thread منفصل
         def run_training():
             print("🔄 Training started...")
             trainer.train()
@@ -230,12 +295,13 @@ def train():
             "status": "training started",
             "epochs": epochs,
             "batch_size": batch_size,
-            "records": len(data)
+            "records": len(unique_data)
         })
 
     except Exception as e:
         return jsonify({"error": str(e)})
 
+
 # ===============================
 # 8️⃣ تشغيل Flask
 # ===============================
@@ -243,7 +309,7 @@ def train():
 def home():
     return jsonify({
         "status": "ok",
-        "message": "Qwen-7B LoRA API is running 🚀",
+        "message": "Graduation Project API is running 🚀",
         "endpoints": ["/generate", "/train"]
     })