Pull requests / #277

#277 iq_pack: keep experts.bin only when its blobs match the GGUF

closed · @sergqwer · 0 Kommentare · Auf GitHub

Models & quants

Beschreibung

## Summary

`tools/iq_pack.py --experts-bin` kept an existing `experts.bin` whenever its size matched the layout. A different
checkpoint with the same geometry (a re-quantization, a fine-tune) has the same size, so it was kept and the engine
served the old experts beside the new model's dense weights, without a word.

Now the first and last blobs of the first, middle and last layers are compared with the GGUF before the file is
kept; a mismatch rewrites it. Six blobs are read, so an unchanged pack is still reused at once.

## Test

`test_experts_bin_is_kept_only_with_this_ggufs_blobs` (tools/test_iq_pack.py) packs two GGUFs of the same geometry
into one folder. It fails on main and passes here; the other 8 tests pass.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

https://claude.ai/code/session_01VZy1yKaDDiA8a7svdwaHio

Mehr auf der Site

Links zu Install, Modellen, Releases.