openhands
openhands
Add selective multimodal-inference manifest, downloader and VRAM planner
bb83d2a Download scripts/plan_inference.py from Frankenstein-Labs/NOEON: direct link, hf CLI and curl.
- Browser
- Download file 1.3 kB
-
https://huggingface.co/Frankenstein-Labs/NOEON/resolve/main/scripts/plan_inference.py
- Command line
-
hf download hf://Frankenstein-Labs/NOEON/scripts/plan_inference.py
-
curl -L -o plan_inference.py https://huggingface.co/Frankenstein-Labs/NOEON/resolve/main/scripts/plan_inference.py
1.3 kB
| #!/usr/bin/env python3 | |
| """Plan which NOEON download profile fits this machine's GPU memory. | |
| Reads ``weights/inference_manifest.json`` and reports, per profile, the exact | |
| weight bytes, the estimated bf16 VRAM need and an honest verdict. It never | |
| assumes a GPU is present: on a CPU-only box every profile reports | |
| ``cpu-only (no CUDA GPU detected)`` instead of a false "fits". | |
| Usage: | |
| python scripts/plan_inference.py | |
| """ | |
| from __future__ import annotations | |
| import json | |
| import sys | |
| from pathlib import Path | |
| sys.path.insert(0, str(Path(__file__).resolve().parent.parent)) | |
| from noeon.core.resources import detect, plan_profiles # noqa: E402 | |
| def main() -> int: | |
| env = detect() | |
| print("environment:") | |
| print(json.dumps(env, indent=2)) | |
| plan = plan_profiles(env) | |
| if "error" in plan: | |
| print("error:", plan["error"]) | |
| return 1 | |
| print("\ninference plan (bf16 estimates):") | |
| for name, v in plan.items(): | |
| print(f" {name:20s} {v['n_files']:3d} files " | |
| f"{v['weights_gb']:6.2f} GB weights " | |
| f"~{v['est_vram_gb_bf16']:5.1f} GB VRAM -> {v['verdict']}") | |
| print("\nnote: VRAM figures are estimates (weights x 1.25 for activations), " | |
| "not guarantees.") | |
| return 0 | |
| if __name__ == "__main__": | |
| raise SystemExit(main()) | |