ReleaseInclusionAI (Ant Group)InclusionAI (Ant Group)published Oct 12, 2026seen 1w

inclusionAI/AReno v0.9.0

inclusionAI/AReno

Open original ↗

Captured source

source ↗
GH
GitHub/github.com/inclusionAI/AReno
inclusionAI/AReno v0.9.0
Source ↗
published Oct 12, 2026seen 1wcaptured 1whttp 200method plain

v0.9.0

Repository: inclusionAI/AReno

Tag: v0.9.0

Published: 2026-10-12T17:44:57Z

Prerelease: no

Release notes:

Highlights

  • Add Huawei Ascend NPU training, rollout, scoring, and OpenAI-compatible serving alongside CUDA and Apple Silicon MLX (#571).
  • Train native CUDA QLoRA adapters for Qwen3 dense and Bailing-MoE V3 with NF4 weights, double quantization, paged optimizers, and CUDA Graph rollout (#577).
  • Add Phi-4 Multimodal support for text, images, audio, and combined image/audio inputs (#521, #523, #525).
  • Integrate Modal training, serving, and job monitoring into the Dashboard, with editable plans, GPU/cost estimates, and live metrics and logs (#568).
  • Add a repository-repair training example using the pi coding agent and ModelScope SWE-smith tasks (#581).

Features

  • Add Ascend NPU backend selection, native operators, and support for SFT, DPO, GRPO, GSPO, and PPO (#571).
  • Add native NF4 quantization for frozen LoRA base weights and paged optimizer state management on CUDA (#577).
  • Add the Phi-4 Multimodal text adapter, SigLIP vision tower, Conformer audio encoder, modality projectors, and vision/speech LoRA weights (#521, #523, #525).
  • Add independent Qwen3.5 vision-tower and merger unfreezing controls (#559).
  • Add experimental Ling-3.0-tiny attention-only LoRA training and serving support on MLX (#584).
  • Add Dashboard-managed Modal workflows for datasets, training, deployment, credentials, pricing, and job monitoring (#568).
  • Add pi agent integration with exact per-turn trajectories, verifier rewards, and reproducible SWE-smith repository-repair dataset generation (#581).

Fixes

  • Preserve sampling seeds in CUDA rollout (#573).
  • Defer Torch imports so the public SDK can be used in MLX-only environments (#578).
  • Check MLX runtime readiness and display Metal runtime status correctly on Apple Silicon (#579, #583).

Performance and maintenance

  • Add model-independent mini-sequence execution for token-independent training blocks on CUDA with tensor parallel size 1 (#587).
  • Organize the quantized AdamW conceptual documentation (#553).
  • Add community meeting notes from September 4 and September 18, 2026 (#557, #572).
  • Require a single commit per pull request (#585).
  • Update the package version and documented container image tag to v0.9.0.

Upgrade notes

  • Ascend installation requires a CANN-compatible PyTorch and torch_npu environment. AReno detects torch_npu on Linux and selects the NPU dependency set; retain the versions matched to your CANN installation.
  • CUDA QLoRA currently supports Qwen3 dense and Bailing-MoE V3. Initialize native LoRA before quantizing the frozen base model.
  • Ling-3.0-tiny LoRA on MLX is experimental and uses attention-only targets. See the MLX guide for supported projections and compatible MLX-LM requirements.
  • The pi repository-repair example requires the pi CLI and a separate preinstalled task Python environment; see the pi example guide for setup.

Installation

For CUDA source builds, use an existing Linux environment with CUDA-enabled PyTorch and the CUDA toolkit:

python -m pip install psutil
python -m pip install areno==0.9.0 --no-build-isolation

The attached areno-0.9.0.tar.gz is the same source distribution published on PyPI; its SHA-256 is listed in SHA256SUMS.

Full Changelog: https://github.com/inclusionAI/AReno/compare/v0.0.8...v0.9.0