Skip to main content

Troubleshooting & FAQ

Common issues and solutions when working with AlphaApollo.

Network errors (ModelScope / Hugging Face download failures)​

Use a stable network + proxy settings first.

  • configure http_proxy, https_proxy, no_proxy
  • pre-download models to local cache, then run training/eval

OOM solutions​

Try these first:

  • reduce batch sizes / rollout group size / sequence lengths
  • if still OOM, use a smaller model or fewer concurrent workers

FlashAttention (flash-attn) error​

Recommended approach: install from a matching release wheel.

  • download a flash-attn wheel from the official release that matches your PyTorch + CUDA + Python
  • install it directly, for example:
pip install /path/to/flash_attn-<version>-<python>-<platform>.whl