SIGN IN SIGN UP

fix(auto-loader): Klein base-model 404 — use 50x-above/below, not 30x-below/50x

detect_gpu_variant() returned "50x"/"30x-below" for Klein-4B/9B, which
download_base_model substitutes into klein-*-series-{variant}-base-model.
Those directories do not exist on the published ModelScope/HF repos — every
series (Klein included) ships only *-50x-above-base-model and
*-50x-below-base-model. snapshot_download then fetched nothing and failed with
"Download completed but model_index.json not found".

Drop the Klein special-case so all series resolve to the real 50x-above
(SM>=120, FP4 text-encoder base) / 50x-below (else, INT4 base) dirs. The 3-tier
transformer + precision-config naming (50x/40x/30x-below) is unaffected: those
files really do use the 3-tier names and are selected per-file, independently
of which base-model is downloaded.

Verified with a dry run (repo file-tree listing only, zero downloads) that all
five series resolve to an existing <dir>/model_index.json for both tiers.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Y
yejianhui committed
4cbd632be27f1aa57782039cde75376f8e2d751c
Parent: 7423ea7