inclusionAI/AReno v0.10.0

v0.10.0

Repository: inclusionAI/AReno

Tag: v0.10.0

Published: 2026-10-23T13:59:30Z

Prerelease: no

Release notes:

Highlights

• Support MLX QLoRA on Apple Silicon, with expanded Ling LoRA targets and adapter export/reload validation (#601, #623).
• Export trained CUDA text models directly to GGUF and an Ollama package through Trainer.export_gguf() or areno train --export-gguf, without an intermediate checkpoint or external converter (#615, #625, #626).
• Add native CUDA LoRA and QLoRA for Qwen3.5 dense language backbones (#616).
• Add Bailing V3 MTP training loss and opt-in speculative decoding (#550).
• Add the JEV algorithm for joint decision-head and backbone-LoRA training, with typed decision serving on MLX (#618).

Features and performance

• Extend Ling MLX LoRA to attention gates and dense, shared-expert, and routed-expert MLP projections; preserve packed-expert adapters through PEFT export/import (#623).
• Add online MLX Q4 quantization and parallel gated-delta training support (#619).
• Support F16 and Q8_0 GGUF output for registered CUDA text architectures: Llama, OLMo2, Qwen3 dense/MoE, Qwen3.5 dense/MoE, Gemma4 dense/MoE, and Bailing V3. Export can merge supported native LoRA adapters and remains disabled by default (#625, #626).
• Complete the Phi-4 Multimodal training and inference flow, including image/audio alignment, checkpoint handling, validation, and audio optimization (#594, #597).
• Train a shared JEV decision head alongside backbone LoRA and serve typed choice, noul, and score decisions with --decision-checkpoint (#618).

Fixes

• Aggregate agentic policy loss by trajectory (#605).
• Remove adaptive mini-sequence tuning and reduce default chunk sizes (#614).
• Fix cancellation handling with ignore_eos=True and keep the cancellation fallback marker out of truncation stop tokens (#550).
• Fix MLX documentation heading markup so the strict Sphinx build passes.

Documentation and milestone

• Expand MLX fine-tuning documentation with quantization, adapter reload, tested configurations, and capacity limits (#623).
• Document Harbor harness and task adaptation, including the Pi workflow (#624).
• Add Phi-4 Multimodal usage and validation documentation (#597).

Published packages

• Python package: areno 0.10.0 on PyPI.
• CUDA installation: python -m pip install areno==0.10.0 --no-build-isolation after preparing CUDA-enabled PyTorch >= 2.6, the CUDA toolkit, and psutil.
• The attached areno-0.10.0.tar.gz is the same source distribution published on PyPI. Its SHA-256 is listed in the attached SHA256SUMS.

Docker images

• GHCR: ghcr.io/inclusionai/areno:v0.10.0 (latest also published from v0.10.0).
• Aliyun ACR: crpi-3kcpcptydwdbnr4i.cn-hangzhou.personal.cr.aliyuncs.com/areno/areno:v0.10.0 (latest also published from v0.10.0).
• Platform: Linux amd64, CUDA/PyTorch runtime.

docker pull ghcr.io/inclusionai/areno:v0.10.0
docker run --gpus all --rm -it ghcr.io/inclusionai/areno:v0.10.0 areno check

Versioned image digests:

• GHCR: sha256:bbe2e712c7e2c44b06fb2c69050a32ff976cbb55cf669319c48c642a65a97687.
• Aliyun ACR: sha256:a73b7483b60e4536bd7aaff7cee69d7d4aad717e588b2de3ef19c3c578ee6fff.