Update main.py

main.py

@@ -7,33 +7,38 @@ import torch
 import os
 import asyncio
 
-# ✅ Set all cache directories to a writable location
+# Set cache directories
 cache_dir = "/tmp/hf_home"
 os.environ["HF_HOME"] = cache_dir
 os.environ["TRANSFORMERS_CACHE"] = cache_dir
 os.environ["HUGGINGFACE_HUB_CACHE"] = cache_dir
 
-# ✅ Create cache directory with proper permissions
+# Create cache directory with proper permissions
 os.makedirs(cache_dir, exist_ok=True)
-os.chmod(cache_dir, 0o777)  # Make writable by all
+os.chmod(cache_dir, 0o777)
 
-# ✅ Load model and tokenizer
+# Load model and tokenizer
 model_name = "Qwen/Qwen2.5-0.5B-Instruct"
-try:
-    tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True, cache_dir=cache_dir)
-    model = AutoModelForCausalLM.from_pretrained(model_name, trust_remote_code=True, cache_dir=cache_dir)
-except Exception as e:
-    print(f"Error loading model: {e}")
-    raise
+tokenizer = AutoTokenizer.from_pretrained(
+    model_name,
+    trust_remote_code=True,
+    cache_dir=cache_dir
+)
+model = AutoModelForCausalLM.from_pretrained(
+    model_name,
+    trust_remote_code=True,
+    cache_dir=cache_dir,
+    torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32
+)
 
-# ✅ Use CUDA if available
+# Set device
 device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
 model.to(device)
 
-# ✅ Initialize FastAPI
+# Initialize FastAPI
 app = FastAPI()
 
-# ✅ Enable CORS
+# Enable CORS
 app.add_middleware(
     CORSMiddleware,
     allow_origins=["*"],
@@ -42,36 +47,52 @@ app.add_middleware(
     allow_headers=["*"],
 )
 
-# ✅ Input data model
+# Input model
 class Question(BaseModel):
     question: str
 
-# ✅ Instructional system prompt
+# System prompt
 SYSTEM_PROMPT = "You are Orion, an intelligent AI assistant created by Abdullah Ali, a 13-year-old from Lahore. Respond kindly and wisely."
 
-# ✅ Streaming response generator
 async def generate_response_chunks(prompt: str):
-    qwen_prompt = (
-        f"<|im_start|>system\n{SYSTEM_PROMPT}<|im_end|>\n"
-        f"<|im_start|>user\n{prompt}<|im_end|>\n"
-        f"<|im_start|>assistant\n"
+    # Create the chat template
+    messages = [
+        {"role": "system", "content": SYSTEM_PROMPT},
+        {"role": "user", "content": prompt}
+    ]
+    
+    # Apply chat template
+    qwen_prompt = tokenizer.apply_chat_template(
+        messages,
+        tokenize=False,
+        add_generation_prompt=True
     )
+    
+    # Tokenize and generate
     inputs = tokenizer(qwen_prompt, return_tensors="pt").to(device)
     outputs = model.generate(
         **inputs,
-        max_new_tokens=256,
+        max_new_tokens=512,
         do_sample=True,
         temperature=0.7,
         top_p=0.9,
         pad_token_id=tokenizer.eos_token_id
     )
-    full_output = tokenizer.decode(outputs[0], skip_special_tokens=True)
-    reply = full_output.split("<|im_start|>assistant\n")[-1].strip()
-    for word in reply.split():
+    
+    # Decode and clean the output
+    full_output = tokenizer.decode(outputs[0], skip_special_tokens=False)
+    
+    # Extract only the assistant's response
+    response = full_output[len(qwen_prompt):].split(tokenizer.eos_token)[0].strip()
+    
+    # Stream the response
+    for word in response.split():
         yield word + " "
-        await asyncio.sleep(0.01)
+        await asyncio.sleep(0.05)
 
-# ✅ API route
 @app.post("/ask")
 async def ask(question: Question):
-    return StreamingResponse(generate_response_chunks(question.question), media_type="text/plain")
+    return StreamingResponse(
+        generate_response_chunks(question.question),
+        media_type="text/plain"
+    )