Update app.py

app.py

@@ -17,22 +17,23 @@ app.add_middleware(
 )
 
 print("Downloading highly optimized VibeThinker-3B Q4_K_M GGUF model...")
-# Fixed typo: changed hyphen to dot before Q4_K_M to match the repository file system
 model_path = hf_hub_download(
     repo_id="prithivMLmods/VibeThinker-3B-GGUF", 
     filename="VibeThinker-3B.Q4_K_M.gguf"
 )
 
-print("Initializing high-speed llama.cpp execution runtime...")
+print("Initializing memory-optimized llama.cpp execution runtime...")
+# Hardened configurations specifically tuned to avoid OOM crashes on free shared CPU tiers
 llm = Llama(
     model_path=model_path, 
-    n_ctx=8192,  # Optimized context window to keep memory footprint light on CPU
-    n_threads=2  # Formatted to maximize dual-core free container allocation limits
+    n_ctx=4096,   # Shifting context window to 4K slashes active memory requirements in half
+    n_batch=32,   # Drop processing batches from 512 to 32 to eliminate peak memory spikes
+    n_threads=2   # Explicitly bounds CPU thread contention blocks
 )
 
 @app.get("/")
 def read_root():
-    return {"status": "online", "engine": "llama.cpp Optimized C++ Core"}
+    return {"status": "online", "engine": "llama.cpp Memory-Hardened Core"}
 
 @app.post("/v1/chat/completions")
 async def chat_completions(request: Request):
@@ -41,20 +42,23 @@ async def chat_completions(request: Request):
     
     response = llm.create_chat_completion(
         messages=messages,
-        temperature=0.7,
+        temperature=0.3,
         top_p=0.95,
         stream=True,
-        max_tokens=2048
+        max_tokens=1536
     )
     
     def stream_generator():
-        for chunk in response:
-            # Extract raw token changes
-            delta = chunk.get("choices", [{}])[0].get("delta", {})
-            if "content" in delta:
-                # Direct transparent formatting pass-through
-                yield f"data: {json.dumps(chunk)}\n\n"
-        yield "data: [DONE]\n\n"
+        try:
+            for chunk in response:
+                delta = chunk.get("choices", [{}])[0].get("delta", {})
+                if "content" in delta:
+                    yield f"data: {json.dumps(chunk)}\n\n"
+            yield "data: [DONE]\n\n"
+        except Exception as e:
+            # Catch silent disconnects or timeouts cleanly without crashing the Uvicorn thread
+            print(f"Streaming trace intercepted safely: {str(e)}")
+            return
 
     return StreamingResponse(stream_generator(), media_type="text/event-stream")