haxiaofeng/patchvla-lehome-migration

PatchVLA / LeHome migration snapshot

This repository contains a Docker environment and runtime data snapshot exported from the owner's A6000 server for restoration on a PRO5000 server. It is an environment transfer bundle, not a Hugging Face model or a tabular dataset.

Contents

• —docker-images.tar.zst.part-*: a single Docker save archive containing lehome-lerobot:0.6.0-vla and openpi-pi05:ubuntu22, Zstandard-compressed and split into 4 GiB parts.
• —runtime-data.tar.zst.part-*: LeHome Assets and scripts, model checkpoint 29999, OpenPI auxiliary assets, and a source-code snapshot.
• —MANIFEST.json and SHA256SUMS: exact image identities and archive checksums.

The source uses Isaac Sim 5.1 and PyTorch 2.7.0+cu128 for simulation; the actor environment includes JAX/jaxlib 0.5.3. The target Blackwell GPU has NOT been validated by this upload job.

Download and restore on the PRO5000 HOST

Install huggingface_hub in a separate Python environment if needed. Docker and zstd must be available on the host.

mkdir -p /data0/patchvla-lehome-migration/download
hf download haxiaofeng/patchvla-lehome-migration --repo-type dataset --local-dir /data0/patchvla-lehome-migration/download
cd /data0/patchvla-lehome-migration/download
sha256sum -c SHA256SUMS
cat docker-images.tar.zst.part-* | zstd -d -c | docker load
mkdir -p /data0/patchvla-lehome-migration/runtime
cat runtime-data.tar.zst.part-* | zstd -d -c | tar -xf - -C /data0/patchvla-lehome-migration/runtime

Run these pipelines in Bash with set -o pipefail and stop on any failed checksum or command. Keep at least 150 GiB free until import and extraction finish; actual requirements depend on shared Docker layers.

Mount mapping

For the simulation container, mount runtime/lehome/Assets to /opt/lehome-challenge/Assets and runtime/lehome/scripts to /opt/lehome-challenge/scripts, both read-only. Use the interpreter /opt/lehome-challenge/.venv/bin/python.

For the policy container, mount runtime/code_snapshot to /app, runtime/checkpoint/29999 to /checkpoint, and runtime/openpi_assets to /openpi_assets, all read-only. Use /.venv/bin/python. The original services used host networking, 8 GiB shared memory, and one GPU. Clear image-inherited HTTP proxies when they do not apply to the target machine; disable or redirect old GPU caches.

The evaluation scripts observed in critic_stage_event_v3_20260916 were marked WITHDRAWN while this migration was prepared. Their diagnostic launcher and results are deliberately not offered as a valid default evaluation workflow. The source code snapshot came from that run's source directory; it is not a claim that the withdrawn evaluation protocol is valid. Select the current approved evaluation launcher before testing.

No host login credentials, HF token cache, SSH home directory, GPU caches, or evaluation results were selected for this bundle. Standard credential locations and common token formats were checked before upload; this is not a general security audit of all third-party image contents. Included third-party components remain subject to their respective licenses.