nanochat-ascend is a special version of nanochat designed to run on Ascend NPUs. It is built to help with training large language models (LLMs) using torch_npu and HCCL technology. This software is experimental and focuses on using Huawei Ascend hardware for AI model training.
The goal is to let you run AI chat and training tools smoothly on devices with Ascend NPUs. It offers basic chat features combined with training capabilities suited for researchers and developers who use Huawei’s AI chips.
Before downloading and running nanochat-ascend, make sure your PC meets these requirements:
- Windows 10 64-bit or newer version
- Huawei Ascend NPU device installed and properly set up
- At least 8 GB of RAM
- Minimum 10 GB free disk space
- Nvidia CUDA drivers NOT required (uses Ascend's torch_npu)
- Administrator rights to install software and drivers
nanochat-ascend works specifically with Huawei Ascend AI processors. It uses CANN (Compute Architecture for Neural Networks) and HCCL (Huawei Collective Communication Library) for distributed training.
You will need:
- Huawei Ascend NPU device access
- Ascend CANN software stack installed
- PyTorch framework with torch_npu extension enabled
- Your system configured for multi-core Ascend training node use
Follow these steps to install nanochat-ascend on your Windows machine:
-
Visit the download page
Click here to visit the releases page:
Download nanochat-ascend releases -
Choose the right installer
Look for the latest Windows installer file. It will have
.exeas the file extension. -
Download the installer
Click on the
.exefile link to download it to your PC. -
Run the installer
Double-click the downloaded file. Follow the instructions on the screen.
-
Allow permissions
If Windows asks for permission to run the app or install drivers, approve it.
-
Wait for installation to finish
The installer will copy files and set up the program.
-
Check your environment
After installation, verify that your Huawei Ascend NPU drivers and CANN software are properly installed.
-
Restart your PC
Restarting helps the system fully enable the Ascend software and drivers.
- Find the nanochat-ascend shortcut on your desktop or Start menu.
- Double-click to launch the program.
- The chat window will open. You can type your messages or commands here.
- To train or test your model, follow the on-screen prompts or refer to the usage guide below.
nanochat-ascend lets you chat with language models and also run training tasks on Ascend NPUs.
- Chat Mode: Type your message and receive responses generated by the AI model.
- Training Mode: Load your training data and start model training using torch_npu modules.
- Settings: Adjust device options and training configurations.
You can switch between modes via the menu bar in the application window.
Make sure your Windows PC is set to use the Ascend AI platform properly:
- Install the latest version of the Ascend CANN toolkit from Huawei’s official site.
- Configure environment variables such as
ASCEND_HOMEand include paths to CANN binaries in yourPATH. - Confirm torch_npu is installed with a compatible PyTorch version.
- Check that HCCL cluster settings are correctly set if using multiple Ascend devices.
Check the releases page regularly for updates:
Download nanochat-ascend releases
Updates may include bug fixes, new features, and improved hardware support.
To update:
- Download the new installer as described in the installation section.
- Run the new installer; it will overwrite existing files automatically.
- Restart your PC to apply changes.
- App doesn’t launch: Verify Ascend drivers and CANN toolkit are installed. Restart the system.
- Training errors: Check if your data format matches the expected input. Confirm torch_npu is set up correctly.
- NPU not detected: Ensure Ascend hardware is connected and visible in device manager.
- Slow performance: Close other heavy programs and check GPU/NPU usage via Huawei monitoring tools.
- Huawei official Ascend developer guides
- torch_npu and PyTorch documentation
- nanochat open-source repository for base code reference
ascend, cann, hccl, huawei-ascend, llm, llm-training, nanochat, npu, pytorch, torch-npu
