PKU-Alignment

End-to-end remote VLA-Arena operations on GPU servers: collect user-provided host/path/repo/eval parameters, clone required repos, configure fixed HF mirror endpoint (https://hf-mirror.com) and centralized caches (HF_HOME/UV_CACHE_DIR/etc.), install uv, download Hugging Face models, patch train/eval YAMLs, run per-model uv environments for training/evaluation, split jobs across GPUs, and debug logs/results. Use when user asks for remote setup, model download, benchmark execution, or troubleshooting dependency/runtime failures.