Skip to content

tools/paradee: pin the GGUF that the current converter makes - #78

Merged
lalalune merged 1 commit into
mainfrom
grace/paradee-gguf-pin
Oct 14, 2026
Merged

lalalune merged 1 commit into
mainfrom
grace/paradee-gguf-pin

Conversation

@lalalune

Copy link
Copy Markdown
Member

Problem

The Paradee converter writes its own git blob SHA-1 into the GGUF metadata (paradee.converter.git_blob_sha1). The note in tools/paradee/pins.json says that the output changes when the converter changes.

Commit ad8cf2e (#69) changed the converter for the ty type check (one import and one type annotation). The blob changed, so the GGUF changed. But pins.json still pins the GGUF of the old converter.

Result at main (533d0d8):

Change

tools/paradee/pins.json: the gguf.f32.sha256 value is now the SHA-256 of the GGUF that the current converter makes. The size does not change (32072768 bytes).

Evidence

I compared the old GGUF (made at 3f131ab) and the new GGUF (made at 533d0d8) with gguf-py:

  • Both have 603 tensors. All tensor shapes, types and data are the same.
  • Only one metadata value is different: paradee.converter.git_blob_sha1, 08ed1650… (old) and b8396ffc… (new).

After this change, on macOS arm64 with the pinned Python environment:

python tools/paradee/convert_paradee_pt_to_gguf.py --model-dir <pinned Paradee-8M-v1.0 files> --outtype f32 --output p.gguf --check-output
wrote p.gguf: 603 tensors, 8005564 parameters, 32072768 bytes, sha256 357477651abb53b3fb1911a6e932026f3f1697fd772668326152e3ca04c9be27
output matches pins.json

A follow-up eliza PR moves the eliza gitlink to this commit and changes the three eliza copies of this hash (paradee-voice/inputs.json, host-tts-worker.ts, BionicSpeechInstrumentedTest.java).

🤖 Generated with Claude Code

The converter writes its own git blob SHA-1 into the GGUF metadata
(paradee.converter.git_blob_sha1). Commit ad8cf2e (#69) changed the
converter for the ty type check. The blob changed, so the output changed
too, but pins.json still pins the output of the old converter.

Because of this, convert_paradee_pt_to_gguf.py --check-output fails at
main, and the eliza paradee-voice build stops at the GGUF step.

The new output has the same 603 tensors with the same bytes. Only the
paradee.converter.git_blob_sha1 value is different (08ed165 to b8396ff).
The size does not change.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@lalalune
lalalune merged commit 866a6da into main Oct 14, 2026
2 checks passed
lalalune added a commit to elizaOS/eliza that referenced this pull request Oct 14, 2026
…makes (#36235)

The llama.cpp gitlink moved to 533d0d89 in #36212. At that commit the
Paradee converter writes a new git blob SHA-1 into the GGUF metadata, so
the GGUF has a new SHA-256. The fork pins.json, paradee-voice/inputs.json,
host-tts-worker.ts and BionicSpeechInstrumentedTest still named the old
hash. The paradee-voice build stopped at the GGUF step.

elizaOS/llama.cpp#78 pins the new GGUF in the fork. This change moves the
gitlink to that merge (866a6da2) and uses the same hash in the three
eliza copies. The tensors and the size do not change.

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant