updatE

app.py

@@ -2,8 +2,7 @@
 
 import os
 import torch
-import transformers
-from transformers import pipeline, AutoTokenizer, AutoModelForCausalLM
+from transformers import AutoTokenizer, AutoModelForCausalLM
 import gradio as gr
 import spaces
 
@@ -17,159 +16,127 @@ def load_system_prompt():
 
 DEFAULT_SYSTEM_PROMPT = load_system_prompt()
 
-model_path = "unsloth/gemma-3-4b-it-unsloth-bnb-4bit"
+model_path = "anaspro/iraqi-7b"
+
+# Jais chat prompts from documentation
+prompt_eng = """### Instruction:Your name is 'Jais', and you are named after Jebel Jais, the highest mountain in UAE. You were made by 'Inception' in the UAE. You are a helpful, respectful, and honest assistant. Always answer as helpfully as possible, while being safe. Complete the conversation between [|Human|] and [|AI|]:
+### Input: [|Human|] {Question}
+[|AI|]
+### Response :"""
+
+prompt_ar = """### Instruction:اسمك "جيس" وسميت على اسم جبل جيس اعلى جبل في الامارات. تم بنائك بواسطة Inception في الإمارات. أنت مساعد مفيد ومحترم وصادق. أجب دائمًا بأكبر قدر ممكن من المساعدة، مع الحفاظ على البقاء أمناً. أكمل المحادثة بين [|Human|] و[|AI|] :
+### Input:[|Human|] {Question}
+[|AI|]
+### Response :"""
 
 # إذا كان فيه HF_TOKEN في البيئة
 hf_token = os.getenv("HF_TOKEN")
 
-# Initialize model and tokenizer for ZeroGPU
-print("Loading model and tokenizer for ZeroGPU...")
+device = "cuda" if torch.cuda.is_available() else "cpu"
+
+tokenizer = AutoTokenizer.from_pretrained(model_path, token=hf_token)
+model = AutoModelForCausalLM.from_pretrained(model_path, device_map="auto", trust_remote_code=True, token=hf_token)
+
+if tokenizer.pad_token is None:
+    tokenizer.pad_token = tokenizer.eos_token
+
+def get_response(text, tokenizer=tokenizer, model=model):
+    """نفس الدالة من documentation مع تعديل لـ chat model"""
+    tokenized = tokenizer(text, return_tensors="pt")
+    input_ids, attention_mask = tokenized['input_ids'].to(device), tokenized['attention_mask'].to(device)
+    input_len = input_ids.shape[-1]
+    generate_ids = model.generate(
+        input_ids,
+        attention_mask=attention_mask,
+        top_p=0.9,
+        temperature=0.3,
+        max_length=2048,
+        min_length=input_len + 4,
+        repetition_penalty=1.2,
+        do_sample=True,
+        pad_token_id=tokenizer.pad_token_id
+    )
+    response = tokenizer.batch_decode(
+        generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=True
+    )[0]
+    response = response.split("### Response :")[-1].lstrip()
+    return response
 
-# Load tokenizer first
-tokenizer = AutoTokenizer.from_pretrained(
-    model_path,
-    token=hf_token,
-    trust_remote_code=True
-)
+def format_conversation_history(chat_history):
+    messages = []
+    for item in chat_history:
+        role = item["role"]
+        content = item["content"]
+        if isinstance(content, list):
+            content = content[0]["text"] if content and "text" in content[0] else str(content)
+        messages.append({"role": role, "content": content})
+    return messages
+
+def detect_language(text):
+    """Simple language detection - Arabic vs English"""
+    arabic_chars = sum(1 for char in text if '\u0600' <= char <= '\u06FF')
+    total_chars = len(text.replace(' ', ''))
+
+    if total_chars == 0:
+        return 'ar'  # default to Arabic
+
+    arabic_ratio = arabic_chars / total_chars
+    return 'ar' if arabic_ratio > 0.3 else 'en'
+
+@spaces.GPU()
+def generate_response(input_data, chat_history, max_new_tokens, temperature, top_p, top_k, repetition_penalty):
+    # Detect language of the current question
+    lang = detect_language(input_data)
+    prompt_template = prompt_ar if lang == 'ar' else prompt_eng
+
+    # Build conversation for Jais format
+    conversation_parts = []
+
+    # Add system prompt as part of the instruction (keep it short for Jais)
+    system_instruction = "اسمك \"أليكس\" وأنت مساعد خدمة العملاء في شركة TechSolutions. مهمتك مساعدة العملاء في حل مشاكلهم مع المنتجات والإجابة عن أسئلتهم حول الخدمات. كن ودوداً وصبوراً ومحترماً. أجب بالعربية أو الإنجليزية حسب تفضيل العميل. ابدأ بالتحية وكن مباشراً في الحلول."
+
+    # Add chat history
+    if chat_history:
+        for item in chat_history:
+            role = item["role"]
+            content = item["content"]
+            if isinstance(content, list):
+                content = content[0]["text"] if content and "text" in content[0] else str(content)
+
+            if role == "user":
+                conversation_parts.append(f"[|Human|] {content}")
+            elif role == "assistant":
+                conversation_parts.append(f"[|AI|] {content}")
+
+    # Add current user message
+    conversation_parts.append(f"[|Human|] {input_data}")
+    conversation_parts.append("[|AI|]")
+
+    # Join conversation
+    conversation = "\n".join(conversation_parts)
+
+    # Create full prompt using Jais format with our system prompt
+    full_prompt = f"### Instruction:{system_instruction}\n### Input:{conversation}\n### Response :"
 
-# For ZeroGPU, load model without specifying device_map
-# The @spaces.GPU() decorator will handle GPU allocation
-model = AutoModelForCausalLM.from_pretrained(
-    model_path,
-    token=hf_token,
-    trust_remote_code=True,
-    torch_dtype=torch.float16,  # Use float16 for ZeroGPU
-    low_cpu_mem_usage=True
-)
-
-print("Model loaded successfully!")
-
-
-def generate_with_pipeline(messages, max_new_tokens=256, temperature=0.7, top_p=0.9, top_k=50, repetition_penalty=1.0):
-    """Generate response using the model with messages format"""
-    # Gemma expects messages in format: [{"role": "user", "content": "..."}, {"role": "model", "content": "..."}]
-    # Convert 'assistant' to 'model' for Gemma
-    gemma_messages = []
-    
-    for msg in messages:
-        role = msg['role']
-        # Gemma uses 'model' instead of 'assistant'
-        if role == 'assistant':
-            role = 'model'
-        # Gemma doesn't use system role in the same way - prepend to first user message
-        if role == 'system':
-            continue  # We'll handle system prompt differently
-        gemma_messages.append({"role": role, "content": msg['content']})
-    
-    # If there's a system prompt, prepend it to the first user message
-    if messages and messages[0]['role'] == 'system' and gemma_messages:
-        system_content = messages[0]['content']
-        if gemma_messages[0]['role'] == 'user':
-            gemma_messages[0]['content'] = f"{system_content}\n\n{gemma_messages[0]['content']}"
-    
-    # Apply chat template
     try:
-        prompt = tokenizer.apply_chat_template(
-            gemma_messages,
-            tokenize=False,
-            add_generation_prompt=True
-        )
-    except Exception as template_error:
-        print(f"Template application error: {template_error}")
-        # Fallback: manually format messages for Gemma
-        prompt = ""
-        for msg in gemma_messages:
-            if msg['role'] == 'user':
-                prompt += f"<start_of_turn>user\n{msg['content']}<end_of_turn>\n"
-            elif msg['role'] == 'model':
-                prompt += f"<start_of_turn>model\n{msg['content']}<end_of_turn>\n"
-        prompt += "<start_of_turn>model\n"
-
-    # Debug: print final prompt
-    print(f"Final prompt preview: {prompt[:200]}...")
-
-    # Tokenize
-    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
-    
-    # Generate
-    with torch.no_grad():
-        outputs = model.generate(
-            **inputs,
-            max_new_tokens=max_new_tokens,
-            temperature=temperature,
-            top_p=top_p,
-            top_k=top_k,
-            repetition_penalty=repetition_penalty,
-            do_sample=True,
-            pad_token_id=tokenizer.pad_token_id,
-            eos_token_id=tokenizer.eos_token_id
-        )
-    
-    # Decode only the new tokens
-    response = tokenizer.decode(outputs[0][inputs['input_ids'].shape[1]:], skip_special_tokens=True)
-    return response
+        # استخدام دالة get_response من documentation
+        response = get_response(full_prompt)
 
+        # استخراج الرد الجديد فقط (بعد "### Response :")
+        if "### Response :" in response:
+            response = response.split("### Response :")[-1].strip()
 
-@spaces.GPU()  # This decorator handles GPU allocation for ZeroGPU
-def generate_response(message, history, max_new_tokens, temperature, top_p, top_k, repetition_penalty):
-    """
-    Generate response with full conversation history
-    
-    Args:
-        message: Current user message (dict with 'text' key when type="messages")
-        history: List of previous messages (already in correct format for type="messages")
-        max_new_tokens, temperature, top_p, top_k, repetition_penalty: Generation parameters
-    """
-    try:
-        # Build messages list - Gemma template expects alternating user/model
-        messages = []
-
-        # Add system message first (will be prepended to first user message)
-        messages.append({"role": "system", "content": DEFAULT_SYSTEM_PROMPT})
-
-        # Add conversation history
-        if history:
-            for msg in history:
-                if isinstance(msg, dict) and 'role' in msg and 'content' in msg:
-                    messages.append({"role": msg['role'], "content": msg['content']})
-
-        # Add current user message
-        if isinstance(message, dict):
-            current_message = message.get("text", "") or message.get("content", "")
-        else:
-            current_message = str(message)
-
-        messages.append({"role": "user", "content": current_message})
-        
-        # Debug: print messages structure
-        print(f"Messages sent to model: {len(messages)} messages")
-        for i, msg in enumerate(messages):
-            content_preview = msg['content'][:50] if len(msg['content']) > 50 else msg['content']
-            print(f"  Message {i}: role={msg['role']}, content_preview={content_preview}...")
-        
-        # Generate response
-        response = generate_with_pipeline(
-            messages,
-            max_new_tokens=max_new_tokens,
-            temperature=temperature,
-            top_p=top_p,
-            top_k=top_k,
-            repetition_penalty=repetition_penalty
-        )
-
-        if not response or response.strip() == "":
+        if not response:
             response = "أهلاً! أنا أليكس مساعد خدمة العملاء. كيف أقدر أساعدك اليوم؟"
 
-        return response
+        yield response
 
     except Exception as e:
         print(f"Error in generate_response: {e}")
         import traceback
         print(traceback.format_exc())
-        return "عذراً، حدث خطأ. يرجى المحاولة مرة أخرى."
-
+        yield "أهلاً! أنا أليكس مساعد خدمة العملاء. كيف أقدر أساعدك اليوم؟"
 
-# Create Gradio interface
 demo = gr.ChatInterface(
     fn=generate_response,
     additional_inputs=[
@@ -180,11 +147,11 @@ demo = gr.ChatInterface(
         gr.Slider(label="عقوبة التكرار", minimum=1.0, maximum=2.0, step=0.05, value=1.0)
     ],
     examples=[
-        ["النت عندي معطل من الصبح، تقدر تساعدني؟"],
-        ["عندي مشكلة بالاتصال بالواي فاي"],
-        ["شنو الباقات المتوفرة عندكم؟"],
-        ["كيف أعيد ضبط الجهاز؟"],
-        ["My device is not working properly"],
+        [{"text": "النت عندي معطل من الصبح، تقدر تساعدني؟"}],
+        [{"text": "عندي مشكلة بالاتصال بالواي فاي"}],
+        [{"text": "شنو الباقات المتوفرة عندكم؟"}],
+        [{"text": "كيف أعيد ضبط الجهاز؟"}],
+        [{"text": "My device is not working properly"}],
     ],
     cache_examples=False,
     type="messages",
@@ -196,8 +163,7 @@ demo = gr.ChatInterface(
 - 💬 لهجة محادثة طبيعية
 - 🔧 دعم فني واستكشاف الأخطاء
 - 📋 معلومات الخدمات والإرشاد
-- 🧠 **يتذكر المحادثة السابقة** - يمكنك الرجوع للمواضيع السابقة
-- 🎯 مدعوم بـ موديل Gemma-3-4B-IT
+- 🎯 مدعوم بـ موديل Unsloth Meta-Llama-3.1-8B-Instruct (مع تحسينات الأداء)
 
 احجي مع أليكس لحل مشاكلك التقنية، استفسر عن الخدمات، أو احصل على معلومات المنتجات.""",
     fill_height=True,
@@ -211,4 +177,4 @@ demo = gr.ChatInterface(
 )
 
 if __name__ == "__main__":
-    demo.launch()
+    demo.launch()