coder_model.py

import os
import logging
import asyncio
from huggingface_hub import hf_hub_download
from llama_cpp import Llama

logger = logging.getLogger("nexari.coder")
BASE_DIR = "./models/coder"
model = None

# === OPTIMIZED FOR 2 vCPU: Qwen 2.5 Coder 3B (Q4_K_M) ===
# TECHNIQUE: Q6 (Heavy) -> Q4 (Fast). 
# Quality drop is negligible, but speed boosts by ~40%.
REPO_ID = "Qwen/Qwen2.5-Coder-3B-Instruct-GGUF"
FILENAME = "qwen2.5-coder-3b-instruct-q4_k_m.gguf"

def load_model(local_dir: str = None):
    global model
    if not local_dir: local_dir = BASE_DIR
    try:
        os.makedirs(local_dir, exist_ok=True)
        path = os.path.join(local_dir, FILENAME)
        
        # Download (~1.7 GB instead of 2.8 GB)
        if not os.path.exists(path):
            logger.info(f"⬇️ Downloading Qwen 3B Coder (Fast Q4)...")
            hf_hub_download(repo_id=REPO_ID, filename=FILENAME, local_dir=local_dir)
        
        model = Llama(
            model_path=path, 
            n_ctx=8192, 
            n_threads=2, 
            n_batch=512, # Batch size increased to ingest prompts faster
            verbose=False
        )
        logger.info("✅ Coder Model Ready (Qwen 3B - Turbo Mode)")
        return model
    except Exception as e:
        logger.error(f"Coder Load Error: {e}")
        model = None

async def load_model_async():
    return await asyncio.to_thread(load_model)