Skip to content

cortiq 0.8.26

Choose a tag to compare

@infosave2007 infosave2007 released this 21 Oct 13:54
· 312 commits to master since this release

[0.8.26] - 2026-10-21

Fixed

  • import-gguf: the native importer reads ggml type 41 (Q1_0, upstream
    since April 2026) — the 1-bit format of the Q1_0 models — and carries its
    weights into the q1 base unchanged: the source's per-128 block scale
    becomes four per-32 groups with the same scale, so no weight moves (the body
    grows from 1.125 to 1.5 bits per weight — the q1 base is per 32, the source
    per 128). A row that is not a multiple of 128 cannot be re-packed without
    changing a weight: the values are kept, the tensor goes to the requested
    profile, and the import says so.
  • import-gguf: ggml type 42 is decoded as upstream's Q2_0 (64-element
    blocks, the codes -1/0/+1/+2). The sparse-ternary format of the still-open
    upstream PR #22836 carried the same number and is now used only when the
    file's own data extent says so — before this, a real Q2_0 file was
    silently mis-decoded as sparse ternary.