llama_bootstrap.py

"""
llama-cpp-python bootstrap — MUST import before gradio or anything else.

2026-07-27: simplified after the great wheel hunt.
- PyPI = source-only (compiles 20+ min — the old "runs and runs" bug).
- abetlen CUDA indexes = abandoned (newest cu121 wheel is v0.2.59). Do NOT attempt.
- abetlen CPU index = current prebuilt wheels. That is the only fast path.
GPU speed returns via the Level-2 llama-server route (official llama.cpp CUDA
binaries + MTP flags) — separate spec, not this file.
"""

from __future__ import annotations

import os
import subprocess
import sys

CPU_WHEEL_INDEX = "https://abetlen.github.io/llama-cpp-python/whl/cpu"


def _pip(args: list) -> None:
    cmd = [sys.executable, "-m", "pip", *args]
    print(f"[LLAMA] $ {' '.join(cmd)}", flush=True)
    env = os.environ.copy()
    env["PIP_DISABLE_PIP_VERSION_CHECK"] = "1"
    env["PIP_ROOT_USER_ACTION"] = "ignore"
    subprocess.check_call(cmd, env=env)


def try_import_llama():
    import llama_cpp  # noqa: F401
    return llama_cpp


def bootstrap_llama_cpp():
    try:
        mod = try_import_llama()
        print("[LLAMA] Ready (prebuilt wheel already installed)", flush=True)
    except ImportError:
        print("[LLAMA] Not installed — installing CPU wheel (fast path)...", flush=True)
        _pip([
            "install", "--no-cache-dir", "llama-cpp-python==0.3.34",
            "--extra-index-url", CPU_WHEEL_INDEX,
        ])
        mod = try_import_llama()

    # CPU wheel => CPU inference. Be explicit so nothing pretends otherwise.
    os.environ.setdefault("DAVIDAU_N_GPU_LAYERS", "0")
    print("[LLAMA] CPU mode (n_gpu_layers=0) — GPU speed is the Level-2 llama-server job", flush=True)
    return mod


bootstrap_llama_cpp()