Skip to content

index: tha-g2p-small-1.0-int4 client rung - #30

Merged
ronaldtse merged 1 commit into
mainfrom
release/tha-int4
Aug 23, 2026
Merged

index: tha-g2p-small-1.0-int4 client rung#30
ronaldtse merged 1 commit into
mainfrom
release/tha-int4

Conversation

@ronaldtse

Copy link
Copy Markdown
Contributor

Summary

  • additional precision rung on the tha-g2p-small-1.0 release: int4 (MatMulNBits blockwise) zip uploaded as a release asset; index entry registers it as tha-g2p-small-1.0-int4 (202MB vs int8 257MB)
  • parity 0.0734pp / 600 samples; quality cost ~0.17pp CER vs int8
  • exporter/schema int4 support lands separately on feat/int4-tier

Test plan

  • CI green
  • Model.load("tha-g2p-small-1.0-int4") resolves (Python/TS; Ruby will fail loudly on MatMulNBits — expected)

Same run-003 student, blockwise 4-bit graphs: parity cer_delta 0.0734pp
over 600 samples (int4 limit 3.0pp); CER vs gold 2.90pp vs int8's
2.73pp — 0.17pp quality for 55MB smaller. Client crystals only: the
Ruby gem's bundled ORT cannot execute MatMulNBits.
@ronaldtse
ronaldtse merged commit d3b6c62 into main Aug 23, 2026
10 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant