"""Loads the fine-tuned GGUF via llama-cpp-python.
Pattern follows the HF ZeroGPU + small-talk reference:
  - hf_hub_download files at module import (warm cache)
  - instantiate Llama inside @spaces.GPU function on each request
"""
from huggingface_hub import hf_hub_download
from .config import MODEL_REPO, GGUF_FILE, MMPROJ_FILE

print(f"[model_loader] downloading {MODEL_REPO}/{GGUF_FILE} …")
MODEL_PATH = hf_hub_download(repo_id=MODEL_REPO, filename=GGUF_FILE)

try:
    MMPROJ_PATH = hf_hub_download(repo_id=MODEL_REPO, filename=MMPROJ_FILE)
    print(f"[model_loader] mmproj ready: {MMPROJ_PATH}")
except Exception as e:
    print(f"[model_loader] mmproj unavailable ({e}) — vision disabled")
    MMPROJ_PATH = None


def make_llm():
    """Create a fresh Llama inside a GPU context.
    The .gguf file is filesystem-cached, so this is fast after the first call."""
    from llama_cpp import Llama

    chat_handler = None
    if MMPROJ_PATH:
        try:
            from llama_cpp.llama_chat_format import MiniCPMv26ChatHandler
            chat_handler = MiniCPMv26ChatHandler(clip_model_path=MMPROJ_PATH, verbose=False)
        except Exception as e:
            print(f"[model_loader] vision chat handler failed: {e}")

    return Llama(
        model_path=MODEL_PATH,
        chat_handler=chat_handler,
        n_gpu_layers=-1,
        n_ctx=8192,
        flash_attn=True,
        verbose=False,
    )