Skip to content

v0.23.0rc3 — wasm 13bc10960bb4a4a8abe2fa846d6618360753892a6e49358edb297530eebc41f3

Pre-release
Pre-release

Choose a tag to compare

@github-actions github-actions released this 23 Oct 19:18
· 9 commits to main since this release

Overview

v0.23.0rc3 is the third release candidate of v0.23.0, with the fixes from walkthroughs of the v0.23.0rc2 release zip. A run_update loop no longer repeats a token, a downloaded prompt cache can be uploaded and continued with the new stamp_uploaded_prompt_cache, upload details no longer report removed files, and the model guides moved to docs/ and ship in the release zip.

Highlights

  • A downloaded prompt cache can be uploaded and continued: scripts/upload.py stamps it with stamp_uploaded_prompt_cache (#69)
  • A run_update loop no longer repeats the first token of the answer between the last ingest call and the first generate call (#68)
  • The model guides moved to docs/, indexed from README.md; the release zip ships them, the classifier example and the log monitor (#66)

Behaviour changes for apps

  • The ingest call that ends the prompt now returns the first answer token; join the output of every call, ingest included (#68)
  • A loop that ignores ingest outputs now drops that first token; per-call n_tokens_generated adds up to the answer's tokens (#68)
  • Re-sending a prompt to a cache that also holds a response no longer adds a duplicate KV cell, so it gives the same reply (#68)
  • uploaded_file_details and uploaded_prompt_cache_details report a file only while it exists with its uploaded size (#67)
  • upload_prompt_cache_chunk, copy_prompt_cache and remove_prompt_cache refuse names ending in .icppfmt or .logits (#69)
  • Models with recurrent or hybrid memory (LFM2.5) refuse to stamp an upload or to continue a cache without logits (#69)
  • -n 0 on a fully cached prompt counts it as cached and reports 0 decoded tokens, instead of a wrapped huge number (#69)
  • stamp_uploaded_prompt_cache accepts a stamp only for the loaded model and layout, a matching sha256 and a valid header (#69)
  • Generating on a classifier loaded without --embedding --pooling rank returns an Err instead of trapping (#65)

Candid API changes

  • New stamp_uploaded_prompt_cache : (StampPromptCacheInputRecord) -> (StatusCodeRecordResult), with { promptcache; stamp; filesha256 } (#69)

Upgrade notes

  • An upgrade keeps the files in stable memory and the access settings; call load_model and set_max_tokens again after it (#65)

Models and guides

  • Guides now live in docs/, under the same file names; links to the old top-level paths must be updated (#66)
  • README.md matches icp-cli 1.7.0; the classifier guide's longest text is exactly -ub tokens (192 at -ub 192) (#65)
  • README "Prompt Caching" explains how to save and restore a cache, and what happens on a mismatch (#69)

Scripts and tooling

  • scripts/download.py also fetches a cache's .icppfmt stamp; scripts/upload.py stamps it, or warns that it starts cold (#69)
  • The release zip ships docs/, images/, examples/classifier/ and scripts/monitor_logs.sh with monitor_logs.py (#66)
  • make docker-verify-wasm reads the deployed hash with icp canister status --public, so non-controllers can verify too (#65)

Changelog since v0.23.0rc2

  • 953603c Restore an uploaded prompt cache with stamp_uploaded_prompt_cache (#69)
  • f200fdd Decode the token that ends a call on -n, so a run_update loop no longer repeats it (#68)
  • 3c61a88 Model guides in docs/, indexed from README.md; the release zip ships the guides, the classifier example and the log monitor (#66)
  • 415312c Upload details only count while the uploaded file exists with its uploaded size (#67)
  • af308e8 Fixes from the v0.23.0rc2 README walkthroughs: no trap on an encoder loaded without rank pooling, and the READMEs match the release (#65)

Reproducible build

llama_cpp.wasm sha256 13bc10960bb4a4a8abe2fa846d6618360753892a6e49358edb297530eebc41f3
llama_cpp_canister commit 623ba718ebd17590a82b530f66d031493c2647c9
llama_cpp_onicai_fork commit 702142008552e33844f99dddefc409550166d175
built with Docker, linux/amd64, icpp-pro 7.0.0rc3
built on 2026-10-23

Reproduce this wasm yourself

git clone https://github.com/onicai/llama_cpp_canister.git
cd llama_cpp_canister
git checkout 623ba718ebd17590a82b530f66d031493c2647c9
make docker-build-base
make docker-build-wasm

The last command prints the sha256. It must equal
13bc10960bb4a4a8abe2fa846d6618360753892a6e49358edb297530eebc41f3.

No Mac and no local toolchain are needed: the llama.cpp fork commit,
icpp-pro, wasi-sdk, Rust and Binaryen are all pinned inside the image.

Check what a deployed canister is running

Anyone can read a deployed module hash, with no special rights, from the
public read-state path:

dfx canister --network ic info <canister-id>

It must print Module hash: 0x13bc10960bb4a4a8abe2fa846d6618360753892a6e49358edb297530eebc41f3.

If you control the canister you can instead build and compare in one step:

make docker-verify-wasm VERIFY_CANISTER=<canister-id>

(That target uses icp canister status, a controller-only management
call; it fails with IC0542 for anyone else.)