Skip to content

v0.23.0rc2 — wasm 70e87be4f835b6d848092ea75109468fc911973259080c4e143f86a49f02fcef

Pre-release
Pre-release

Choose a tag to compare

@github-actions github-actions released this 22 Oct 20:57
· 15 commits to main since this release

Overview

v0.23.0rc2 is the second release candidate of v0.23.0. It moves to icpp-pro 7.0.0rc3 and its new stable-memory file system, with an upload fix that the new file system needs, and it adds sequence classifiers: run_classify and run_classify_query, loaded with upstream llama.cpp's --embedding --pooling rank, with a DAO moderation classifier as a worked example.

Highlights

  • New run_classify (update) and run_classify_query (query) for sequence classifiers such as BERT and ModernBERT (#64)
  • Example: a DAO-channel moderation classifier (Ettin-17M, f32, 15 scores per message), published as onicai/dao-moderation-classifier (#64)
  • New stable-memory file system (icpp-pro 7.0.0rc3): a model takes about its file size, and LFM2.5-2.6B now loads (#62)

Candid API changes

  • New run_classify : (ClassifyInputRecord) -> (ClassifyResult) and the query run_classify_query with the same types (#64)
  • New records ClassifyInputRecord { text } and ClassifyOutputRecord { scores; labels; input_tokens; instructions } (#64)

Upgrade notes

  • An upgrade keeps the models and stored decision states that the old file system wrote (verified with Kev-0.8B) (#62)
  • Plan stable memory at about a model's file size: the 1.67 GB LFM2.5-2.6B takes 1.74 GB, down from 2.81 GB before (#62)
  • A later upload chunk no longer truncates the file; without that fix, the new file system kept only the last chunk (#62)
  • A corrupt uploaded_files_metadata.dat no longer makes the metadata reader allocate without bound (#62)

Models and guides

  • New README-classifier-models.md: load arguments, measurements, query vs update, limits; f32 is recommended over f16 (#64)
  • New README-classifier-training.md and examples/classifier/: the DAO moderation example step by step, up to the canister test (#64)
  • A classifier text is limited to 8192 bytes and the -ub tokens of the model's load; a longer one gets an Err (#64)
  • With a classifier loaded, new_chat, run_update, run_query and run_update_full return "use run_classify" (#64)
  • LFM2.5-2.6B guide: the current Hugging Face file loads with this release and answers a run_update_full call (#62)

icpp-pro

  • icpp-pro 7.0.0rc3, with the stable-fs 0.13.0 file system (ic-wasi-polyfill v0.13.0) (#62)

Scripts and tooling

  • scripts/upload.py and scripts/download.py stop with a clear error when the identity or the canister id cannot be found (#63)
  • For contributors, make test-llm-native runs the native tests under a memory guard that stops a run above 4 GB (#62)

Changelog since v0.23.0rc1

  • 42fcaa6 Add run_classify and run_classify_query for sequence classifiers, with upstream's --embedding --pooling rank (#64)
  • 52d19c8 Fail clearly on a missing identity or canister id, restore max_tokens after the Hindi tests, and add the upstream-llama.cpp-first rule (#63)
  • 2c5d04b Bump icpp-pro to 7.0.0rc3, keep written upload chunks, and run the native tests under a memory guard (#62)

Reproducible build

llama_cpp.wasm sha256 70e87be4f835b6d848092ea75109468fc911973259080c4e143f86a49f02fcef
llama_cpp_canister commit c072ae40e09552b7f28d6557af37ba9b9f57cb6b
llama_cpp_onicai_fork commit 702142008552e33844f99dddefc409550166d175
built with Docker, linux/amd64, icpp-pro 7.0.0rc3
built on 2026-10-22

Reproduce this wasm yourself

git clone https://github.com/onicai/llama_cpp_canister.git
cd llama_cpp_canister
git checkout c072ae40e09552b7f28d6557af37ba9b9f57cb6b
make docker-build-base
make docker-build-wasm

The last command prints the sha256. It must equal
70e87be4f835b6d848092ea75109468fc911973259080c4e143f86a49f02fcef.

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: 0x70e87be4f835b6d848092ea75109468fc911973259080c4e143f86a49f02fcef.

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.)