upadte

app.py

@@ -18,6 +18,63 @@ DEFAULT_SYSTEM_PROMPT = load_system_prompt()
 
 model_path = "inceptionai/jais-adapted-7b-chat"
 
+# Gemma-3 chat template for compatibility
+GEMMA_CHAT_TEMPLATE = "{{ bos_token }}{% if messages[0]['role'] == 'system' %}{% set system_message = messages[0]['content'] %}{% set messages = messages[1:] %}{% else %}{% set system_message = false %}{% endif %}{% for message in messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if (message['role'] == 'assistant') %}{% set role = 'model' %}{% else %}{% set role = message['role'] %}{% endif %}{% if loop.first and system_message %}{{ '<start_of_turn>' + role + '\n' + system_message + '\n\n' + message['content'] | trim + '<end_of_turn>\n' }}{% else %}{{ '<start_of_turn>' + role + '\n' + message['content'] | trim + '<end_of_turn>\n' }}{% endif %}{% endfor %}{% if add_generation_prompt %}{{ '<start_of_turn>model\n' }}{% endif %}"
+
+def apply_gemma_template(messages, add_generation_prompt=True):
+    """Apply Gemma-3 chat template for models based on Gemma-3"""
+    try:
+        # Try to use tokenizer's built-in template first
+        if hasattr(tokenizer, 'apply_chat_template'):
+            return tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=add_generation_prompt)
+
+        # Manual implementation based on the template
+        result = tokenizer.bos_token or ""
+
+        system_message = None
+        if messages and messages[0]['role'] == 'system':
+            system_message = messages[0]['content']
+            messages = messages[1:]
+
+        for i, message in enumerate(messages):
+            if message['role'] == 'assistant':
+                role = 'model'
+            else:
+                role = message['role']
+
+            result += f"<start_of_turn>{role}\n"
+
+            if i == 0 and system_message:
+                result += f"{system_message}\n\n"
+
+            if isinstance(message['content'], str):
+                result += message['content'].strip()
+            elif isinstance(message['content'], list):
+                for item in message['content']:
+                    if item.get('type') == 'text':
+                        result += item['text'].strip()
+
+            result += "<end_of_turn>\n"
+
+        if add_generation_prompt:
+            result += "<start_of_turn>model\n"
+
+        return result
+
+    except Exception as e:
+        print(f"Error in Gemma template: {e}")
+        # Fallback
+        prompt = ""
+        for msg in messages:
+            if msg["role"] == "system":
+                prompt += f"System: {msg['content']}\n"
+            elif msg["role"] == "user":
+                prompt += f"Human: {msg['content']}\n"
+            elif msg["role"] == "assistant":
+                prompt += f"Assistant: {msg['content']}\n"
+        prompt += "Assistant:"
+        return prompt
+
 # 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}
@@ -85,50 +142,41 @@ def detect_language(text):
 
 @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."
+    try:
+        # Build messages for Gemma format
+        messages = []
 
-    # 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 DEFAULT_SYSTEM_PROMPT:
+            messages.append({"role": "system", "content": DEFAULT_SYSTEM_PROMPT})
 
-            if role == "user":
-                conversation_parts.append(f"[|Human|] {content}")
-            elif role == "assistant":
-                conversation_parts.append(f"[|AI|] {content}")
+        # 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)
+                messages.append({"role": role, "content": content})
 
-    # Add current user message
-    conversation_parts.append(f"[|Human|] {input_data}")
-    conversation_parts.append("[|AI|]")
+        # Add current user input
+        messages.append({"role": "user", "content": input_data})
 
-    # Join conversation
-    conversation = "\n".join(conversation_parts)
+        # Use Gemma template for the model
+        prompt = apply_gemma_template(messages)
 
-    # Create full prompt using Jais format with our system prompt
-    full_prompt = f"### Instruction:{system_instruction}\n### Input:{conversation}\n### Response :"
+        print(f"Generated Gemma prompt: {prompt[:200]}...")  # Debug
 
-    try:
-        # استخدام دالة get_response من documentation
-        response = get_response(full_prompt)
+        # استخدام دالة get_response مع Gemma prompt
+        response = get_response(prompt)
 
-        # استخراج الرد الجديد فقط (بعد "### Response :")
-        if "### Response :" in response:
-            response = response.split("### Response :")[-1].strip()
+        # استخراج الرد الجديد فقط (بعد آخر <end_of_turn>)
+        if "<end_of_turn>" in response:
+            response = response.split("<end_of_turn>")[-1].strip()
 
         if not response:
             response = "أهلاً! أنا أليكس مساعد خدمة العملاء. كيف أقدر أساعدك اليوم؟"
 
+        print(f"Final response: {response[:100]}...")  # Debug
         yield response
 
     except Exception as e:

app2.py

@@ -1,7 +1,8 @@
+# -*- coding: utf-8 -*-
+
 import os
 import torch
-from transformers import AutoTokenizer, AutoModelForCausalLM, TextIteratorStreamer, pipeline
-from threading import Thread
+from transformers import AutoTokenizer, AutoModelForCausalLM
 import gradio as gr
 import spaces
 
@@ -15,74 +16,51 @@ def load_system_prompt():
 
 DEFAULT_SYSTEM_PROMPT = load_system_prompt()
 
-model_path = "unsloth/Meta-Llama-3.1-8B-Instruct-bnb-4bit"
+model_path = "inceptionai/jais-adapted-7b-chat"
+
+# 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")
 
-# استخدام ChatPipeline بدلاً من text-generation العادي
+device = "cuda" if torch.cuda.is_available() else "cpu"
+
 tokenizer = AutoTokenizer.from_pretrained(model_path, token=hf_token)
-model = AutoModelForCausalLM.from_pretrained(model_path, torch_dtype="auto", device_map="auto", token=hf_token)
-
-# إنشاء chat pipeline مخصص مع streaming
-def create_chat_pipeline(tokenizer, model):
-    """إنشاء pipeline مخصص للدردشة مع chat template و streaming"""
-    def chat_generate(messages, streamer=None, **kwargs):
-        # تحويل الرسائل للـ chat template
-        if hasattr(tokenizer, 'apply_chat_template') and tokenizer.chat_template is not None:
-            prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
-        else:
-            # Fallback للموديلات اللي ما عندها chat template
-            prompt = ""
-            for msg in messages:
-                if msg["role"] == "system":
-                    prompt += f"System: {msg['content']}\n"
-                elif msg["role"] == "user":
-                    prompt += f"Human: {msg['content']}\n"
-                elif msg["role"] == "assistant":
-                    prompt += f"Assistant: {msg['content']}\n"
-            prompt += "Assistant:"
-
-        inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
-
-        # توليد الرد مع streaming إذا كان مطلوب
-        if streamer:
-            generation_kwargs = {
-                **inputs,
-                "max_new_tokens": kwargs.get('max_new_tokens', 512),
-                "temperature": kwargs.get('temperature', 0.7),
-                "top_p": kwargs.get('top_p', 0.9),
-                "top_k": kwargs.get('top_k', 50),
-                "repetition_penalty": kwargs.get('repetition_penalty', 1.1),
-                "do_sample": True,
-                "pad_token_id": tokenizer.eos_token_id,
-                "streamer": streamer,
-            }
-
-            # نرجع الـ thread للتشغيل
-            return generation_kwargs
-        else:
-            # للتوليد العادي بدون streaming
-            with torch.no_grad():
-                outputs = model.generate(
-                    **inputs,
-                    max_new_tokens=kwargs.get('max_new_tokens', 512),
-                    temperature=kwargs.get('temperature', 0.7),
-                    top_p=kwargs.get('top_p', 0.9),
-                    top_k=kwargs.get('top_k', 50),
-                    repetition_penalty=kwargs.get('repetition_penalty', 1.1),
-                    do_sample=True,
-                    pad_token_id=tokenizer.eos_token_id,
-                    return_dict_in_generate=True,
-                    output_scores=False,
-                )
-
-            response = tokenizer.decode(outputs.sequences[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)
-            return [{"generated_text": response}]
-
-    return chat_generate
-
-pipe = create_chat_pipeline(tokenizer, model)
+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
 
 def format_conversation_history(chat_history):
     messages = []
@@ -94,39 +72,71 @@ def format_conversation_history(chat_history):
         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):
-    # حذف الهستوري كله - نبدأ محادثة جديدة دايماً
-    # chat_history = []  # تعطيل الهستوري
-
-    new_message = {"role": "user", "content": input_data}
-
-    # Build messages for Llama chat template
-    messages = [{"role": "system", "content": DEFAULT_SYSTEM_PROMPT}]
-    # لا نضيف chat_history القديم
-
-    # استخدام ChatPipeline المخصص مع streaming
-    streamer = TextIteratorStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)
-
-    generation_kwargs = pipe(
-        messages,
-        streamer=streamer,
-        max_new_tokens=max_new_tokens,
-        temperature=temperature,
-        top_p=top_p,
-        top_k=top_k,
-        repetition_penalty=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 = []
 
-    thread = Thread(target=model.generate, kwargs=generation_kwargs)
-    thread.start()
+    # 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 :"
+
+    try:
+        # 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()
+
+        if not response:
+            response = "أهلاً! أنا أليكس مساعد خدمة العملاء. كيف أقدر أساعدك اليوم؟"
 
-    # Stream the response
-    response = ""
-    for chunk in streamer:
-        response += chunk
         yield response
 
+    except Exception as e:
+        print(f"Error in generate_response: {e}")
+        import traceback
+        print(traceback.format_exc())
+        yield "أهلاً! أنا أليكس مساعد خدمة العملاء. كيف أقدر أساعدك اليوم؟"
+
 demo = gr.ChatInterface(
     fn=generate_response,
     additional_inputs=[
@@ -153,7 +163,7 @@ demo = gr.ChatInterface(
 - 💬 لهجة محادثة طبيعية
 - 🔧 دعم فني واستكشاف الأخطاء
 - 📋 معلومات الخدمات والإرشاد
-- 🎯 مدعوم بـ موديل anaspro العراقي (Llama 3.1 محسن للعربية العراقية)
+- 🎯 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,