Skip to content

v0.7.2 — trellis PLE fix for Gemma-4 E2B/E4B

Choose a tag to compare

View all tags
@cnygaard cnygaard released this 25 Jul 23:31
v0.7.2
d179c85
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Patch release. One fix since v0.7.1:

--codebook trellis now works on Gemma-4 E2B/E4B (models with per-layer-input embeddings). The PLE embedding table routes to the shell codebook on BOTH sides — at quantize time (the trellis codebook forbids padded RHT, so PLE previously hit an assert) and at load time (E8RHTEmbedding only has a shell decode path, so trellis-quantized checkpoints previously failed with an attribute error).

Validated end-to-end on gemma-4-E2B and gemma-4-E2B-it at trellis-3INST 4 bpw: quantization completes, the checkpoint loads through the official AutoModelForMultimodalLM path, and sampled generation is coherent.

No kernel or format changes — checkpoints produced by 0.7.0/0.7.1 load unchanged.

Assets 2
Loading