Add selective multimodal-inference manifest, downloader and VRAM planner
openhands+1: openhands
bb83d2a
- scripts/build_inference_manifest.py: classify all 380 HF files into tiers
(core/noeon_own/alignment/loading/optional) + named download profiles, using
the Hub tree API only (no bytes transferred). LFS files carry SHA-256; regular
files carry a git blob SHA-1 (the two are not interchangeable).
- scripts/selective_download.py: fetch one profile via HTTP (Range resume +
per-file hash verification) or Git LFS; never the whole 43.5 GB.
- noeon/core/resources.py: plan_profiles() estimates per-profile bf16 VRAM.
- scripts/plan_inference.py + tests/test_selective_inference.py (15 tests).
Verified: metadata_only profile downloads 69/69 files with 0 hash errors; one
real LFS weight (VAE, 345 MB) downloads and verifies; 131 tests pass.
Co-authored-by: openhands <openhands@all-hands.dev>