Ascend NPU

AReno provides an Ascend NPU backend for SFT, DPO, GRPO, GSPO and PPO, as well as rollout, scoring and OpenAI-compatible serving.

Installation

Use Linux/aarch64 with Ascend hardware and a CANN-compatible installation of PyTorch and torch_npu. The target stack is CANN 9.0.0, PyTorch 2.10.0+cpu and torch_npu 2.10.0.post2. Source the CANN toolkit’s set_env.sh before building, and make sure CMake and the CANN development tools are available.

From the AReno repository root:

python -m pip install -e . --no-build-isolation

The installer detects torch_npu and selects the NPU dependencies. Keep the PyTorch and torch_npu versions supplied with your CANN environment.

Training

Use the standard areno train command with --backend npu. Select the training method with --algo sft, dpo, grpo, gspo or ppo and use the corresponding dataset and reward configuration.

Set --world-size to the number of devices and --tp-size to the tensor parallel size. The CLI also selects NPU automatically on Linux when torch_npu is installed.

See CLI Reference for training options and Quickstart for the standard training workflow.

Serving

Start an OpenAI-compatible server with a local checkpoint:

areno serve --backend npu --model-path /path/to/local/checkpoint \
  --world-size 1 --tp-size 1 --port 8000

Use http://localhost:8000/v1 as the base URL for your OpenAI-compatible client. Add --attn-backend native to select native NPU attention for training or serving.