Update app.py

app.py

@@ -1,44 +1,48 @@
 from fastapi import FastAPI
 from pydantic import BaseModel
 from llama_cpp import Llama
+import llama_cpp
 
 app = FastAPI()
 
-# Download and initialize the model when the server starts
+# --- HIGHEST OPTIMIZATION FOR 1.5B ---
 llm = Llama.from_pretrained(
-    repo_id="Qwen/Qwen2.5-Coder-0.5B-Instruct-GGUF", # Make sure this says 0.5B!
-    filename="*q4_k_m.gguf", 
-    n_ctx=2048, 
-    n_threads=2, 
-    n_batch=512  
+    repo_id="Qwen/Qwen2.5-Coder-1.5B-Instruct-GGUF",
+    filename="*q4_k_m.gguf",
+    n_ctx=1024,
+    n_threads=2,             # Match your physical cores
+    n_batch=512,
+    flash_attn=True,
+    n_mlock=True,            # Keep model in RAM for consistent speed
+    type_k=llama_cpp.GGML_TYPE_Q4_0, # 4-bit KV Cache for faster processing
+    verbose=False
 )
 
 class EvalRequest(BaseModel):
     task_description: str
     python_code: str
 
-# --- ADDED HEALTH CHECK ROUTE HERE ---
 @app.get("/")
 async def health_check():
-    return {"status": "Online", "message": "AI Code Evaluator is running! Send POST requests to /evaluate"}
-# -------------------------------------
+    return {"status": "Online", "message": "Optimized 1.5B Evaluator Ready"}
 
 @app.post("/evaluate")
 async def evaluate_code(request: EvalRequest):
-    prompt = f"Task Context:\n{request.task_description}\n\nStudent Code:\n{request.python_code}"
+    # Minimalist prompt for faster processing
+    prompt = f"TASK: {request.task_description}\n\nCODE:\n{request.python_code}\n\nEVALUATE:"
     
+    system_prompt = 'You are a Python tutor. Output ONLY JSON: {"score": int, "feedback": str, "improvements": list}'
+
     response = llm.create_chat_completion(
         messages=[
-            {
-                "role": "system", 
-                "content": "You are a friendly Python grader. Output ONLY valid JSON."
-            },
+            {"role": "system", "content": system_prompt},
             {"role": "user", "content": prompt}
         ],
-        max_tokens=512, 
+        max_tokens=400,
         temperature=0.1,
-        # THIS IS THE MAGIC LINE:
+        repeat_penalty=1.1,
+        stop=["}"], # STOP IMMEDIATELY when JSON closes
         response_format={"type": "json_object"} 
     )
         
-    return {"evaluation": response['choices'][0]['message']['content']}
+    return {"evaluation": response['choices'][0]['message']['content']}