Update app.py

app.py

@@ -1,13 +1,14 @@
 import os
+import time
 from fastapi import FastAPI, Request
 from fastapi.middleware.cors import CORSMiddleware
-from fastapi.responses import HTMLResponse, JSONResponse
+from fastapi.responses import JSONResponse
+from huggingface_hub import hf_hub_download
 from llama_cpp import Llama
-import time
 
 app = FastAPI(title="Qwen 2.5 Coder OpenAI Server")
 
-# Enable CORS so your local desktop IDE tools can securely reach the Space
+# Enable CORS for local VS Code extension requests
 app.add_middleware(
     CORSMiddleware,
     allow_origins=["*"],
@@ -16,11 +17,19 @@ app.add_middleware(
     allow_headers=["*"],
 )
 
-print("Loading Qwen 2.5 Coder 7B GGUF into RAM...")
+# Safe Hugging Face runtime model download path
+MODEL_REPO = "Qwen/Qwen2.5-Coder-7B-Instruct-GGUF"
+MODEL_FILE = "qwen2.5-coder-7b-instruct-q4_k_m.gguf"
+
+print("Checking and downloading Qwen 2.5 Coder 7B GGUF model...")
+# Downloads dynamically on container boot time via the HF library (cached across restarts)
+model_path = hf_hub_download(repo_id=MODEL_REPO, filename=MODEL_FILE)
+
+print(f"Loading model into memory from: {model_path}")
 llm = Llama(
-    model_path="./qwen2.5-coder-7b-instruct-q4_k_m.gguf",
-    n_ctx=4096,      # Expanded slightly for code context, safely within 16GB
-    n_threads=2,     # Matches your 2 vCPU limit
+    model_path=model_path,
+    n_ctx=4096,      # Expanded safely within 16GB memory bounds
+    n_threads=2,     # Perfectly matches your 2 vCPU resource cap
     n_batch=512,
     verbose=False
 )
@@ -32,9 +41,8 @@ async def chat_completions(request: Request):
     messages = body.get("messages", [])
     temperature = body.get("temperature", 0.2)
     max_tokens = body.get("max_tokens", 1024)
-    stream = body.get("stream", False)
 
-    # 1. Format standard OpenAI/Claude message dicts into Qwen's ChatML structure
+    # Convert incoming standard OpenAI messages array to Qwen ChatML formatting
     formatted_prompt = ""
     for msg in messages:
         role = msg.get("role", "user")
@@ -42,13 +50,7 @@ async def chat_completions(request: Request):
         formatted_prompt += f"<|im_start|>{role}\n{content}<|im_end|>\n"
     formatted_prompt += "<|im_start|>assistant\n"
 
-    # 2. Block streaming requests to keep CPU resource usage stable 
-    if stream:
-        # High-performance streaming on 2 vCPUs can stall the event loop; 
-        # Forcing non-stream mode keeps the CPU connection robust
-        pass 
-
-    # 3. Generate completion
+    # Synchronous processing avoids lock stalls on CPU
     response = llm(
         formatted_prompt,
         max_tokens=max_tokens,
@@ -56,9 +58,8 @@ async def chat_completions(request: Request):
         temperature=temperature
     )
 
-    text_output = response["choices"][0]["text"]
+    text_output = response["choices"]["text"]
     
-    # 4. Construct a compliant OpenAI JSON response structure
     return JSONResponse(content={
         "id": f"chatcmpl-{int(time.time())}",
         "object": "chat.completion",
@@ -86,3 +87,4 @@ def index():
         "model": "qwen2.5-coder-7b-instruct-q4_k_m", 
         "endpoint": "/v1/chat/completions"
     }
+