Railgun526's picture
Document native-A0/no-SDF reproduction and anonymous model download qualification
4c30652 verified
|
Raw History Blame Contribute Delete
7.29 kB
# Dependencies and installation
The active method uses the official NVIDIA Flash-CHORD JAX FlashSAC learner. A2 also requires the complete official Utonia backbone/checkpoint, including no-SDF and `no_utonia_features` runs; the latter still computes the encoder and masks features at fusion. A trained A2 policy additionally needs its matching `.estimator.pt`. Missing official components cannot be replaced with a toy model.
Source origins and licenses are in `third_party/sources.json`; assets and unpublished experimental files are listed in [data_assets.md](data_assets.md). Canonical native-A0 versus no-SDF-A2 reproduction commands are in [reproduction.md](reproduction.md). The selected engine remains `third_party/mujoco_warp_1771/mujoco_warp`, with the official, still unmerged PR 1771 witness correction. Latest upstream main is not an equivalent engine.
## Installation status
**The existing environment ran the experiments; the clean installation recipe below has not been executed or validated on a fresh machine.** It reconstructs the source layout and package separation used by `src/curiosity/runtime.py`, rather than claiming that `pip install -r requirements-runtime.txt` is a complete installer. Driver support, CUDA toolkit/extension compilation and shared-library compatibility must be checked on the target compute node before training. Do not modify the working experimental environment merely to test these instructions.
Version evidence is the retained metadata export [runtime-packages.json](runtime-packages.json), supplemented by installed sparse package metadata. Absolute paths in that export describe the original environment, not installation destinations on another machine:
| Component | Observed version / location |
|---|---|
| Python | 3.11; Linux, H200 experiment hardware |
| JAX / jaxlib / CUDA 13 plugin / PJRT | 0.10.2 in `.venv/flash-chord` |
| Flax / Optax / Distrax / safetensors | 0.12.7 / 0.2.8 / 0.1.9 / 0.8.0 |
| MuJoCo / MuJoCo Warp / Newton / Warp | 3.8.1 / 3.8.1 / 1.2.0 / 1.14.0 |
| NumPy | 2.4.6 in the official runtime |
| Torch | 2.6.0+cu124 in `.venv/lib/python3.11/site-packages` |
| FlashAttention | 2.7.4.post1 in `.venv/g1_packages` |
| torch-scatter | 2.1.2+pt26cu124 in `.venv/overlays/object_predictor_pkgs` |
| spconv / cumm | spconv-cu126 2.3.8 / cumm-cu126 0.8.2 in the same overlay; metadata queries for cu120 reported absent, not the installed variant |
| timm / addict / einops / huggingface-hub | 1.0.12 / 2.4.0 / 0.8.2 / 0.36.2 |
| NCCL loaded for Torch | nvidia-nccl-cu12 2.21.5 in `.venv/overlays/pkgs` |
| cuDNN loaded into the shared process | Official JAX environment's cuDNN 9; observed nvidia-cudnn-cu13 9.24.0.43 |
## Source and Python setup
Run installation, imports, compilation, tests and all simulation on an allocated compute node through tmux + srun. The commands below assume a **new checkout without an existing `.venv`**, Python 3.11, compiler/build tools and a CUDA toolkit compatible with the Torch extensions. Do not fake Slurm variables to bypass the guard.
```bash
# Restore the retained official source snapshots from vendor before installation.
bash tools/restore_official_sources.sh
# Keep the Torch dependency tree separate from the official JAX environment.
python3.11 -m venv .venv
python3.11 -m venv .venv/flash-chord
.venv/flash-chord/bin/python -m pip install -e 'third_party/video_to_data/robotic_grounding/flash_chord[train]'
.venv/flash-chord/bin/python -m pip install 'PyYAML==6.0.3' 'h5py==3.16.0'
.venv/flash-chord/bin/python -m pip install --no-deps -e .
# Torch uses its matching CUDA 12.4 wheel family. torchvision 0.21.0 is
# the matching distribution needed by timm, not a separately audited experiment pin.
.venv/bin/python -m pip install 'torch==2.6.0' 'torchvision==0.21.0' --index-url https://download.pytorch.org/whl/cu124
.venv/bin/python -m pip install 'timm==1.0.12' 'einops==0.8.2' 'huggingface-hub==0.36.2' 'safetensors==0.8.0' packaging ninja wheel
mkdir -p .venv/overlays/object_predictor_pkgs .venv/overlays/pkgs .venv/g1_packages
.venv/bin/python -m pip install --target .venv/overlays/object_predictor_pkgs \
'spconv-cu126==2.3.8' 'cumm-cu126==0.8.2' 'addict==2.4.0'
.venv/bin/python -m pip install --no-deps --target .venv/overlays/object_predictor_pkgs \
'torch-scatter==2.1.2+pt26cu124' -f https://data.pyg.org/whl/torch-2.6.0+cu124.html
.venv/bin/python -m pip install --no-build-isolation --no-deps --target .venv/g1_packages 'flash-attn==2.7.4.post1'
.venv/bin/python -m pip install --no-deps --target .venv/overlays/pkgs 'nvidia-nccl-cu12==2.21.5'
.venv/flash-chord/bin/python -m pip install 'nvidia-cudnn-cu13==9.24.0.43'
```
The commands are installable package entry points, **not a claim of a fully locked or successfully rebuilt environment**. FlashAttention may compile locally and needs a compatible toolkit/compiler. The spconv-cu126 and Torch-cu124 combination reflects the retained local stack; its acceptance by another GPU/driver must be established, not silently replaced with a smaller encoder. Transitive dependencies are not all pinned, and pip availability may change. If an exact wheel or source asset is unavailable, report the missing dependency.
Utonia is imported from the restored `third_party/Utonia` source by the provider; an unrelated pip package is unnecessary. Download `checkpoints/utonia/utonia.pth` and official task assets as described in [data_assets.md](data_assets.md). Do not apply Utonia's optional demo dependency `numpy==1.26.4` to this environment: Flash-CHORD uses NumPy 2.4.6.
`configure_dependencies()` appends `.venv/lib/python3.11/site-packages`, `.venv/g1_packages` and the object-predictor overlay to the official runtime. `preload_torch_libraries()` loads `libcudnn*.so.9` from `.venv/flash-chord/lib/python3.11/site-packages/nvidia/cudnn/lib`, then `.venv/overlays/pkgs/nvidia/nccl/lib/libnccl.so.2` with global visibility. Torch's separately installed cuDNN 9.1.0.70 is **not** the intended cuDNN for the shared JAX/Torch process. Do not copy both cuDNN distributions into the same package directory or assume `LD_LIBRARY_PATH` alone selects them: `prepare_backend()` clears that variable before re-exec.
The legacy cuDNN overlay path exported by `tools/run.sh` does not replace this explicit preload contract. Training/evaluation should continue through that wrapper; direct ad-hoc imports are not an equivalent runtime check. Complete full-shape smoke and checkpoint restoration before formal training, following [reproduction.md](reproduction.md).
## Browser and video dependencies
On the compute node, with Node/npm available:
```bash
npm install --no-save playwright@1.58.2
npx playwright install chromium
```
The renderers default to local `require('playwright')` and Playwright's installed Chromium. Optional `PLAYWRIGHT_MODULE` selects another module path; optional `PLAYWRIGHT_CHROMIUM_EXECUTABLE_PATH` selects an existing Chrome executable. `FFMPEG_BINARY` may select an existing ffmpeg; otherwise the project renderer uses the official environment's `imageio_ffmpeg` binary. A headless browser may still require host system libraries, which these npm commands do not provision. These rendering dependencies do not change the learner or produce new trajectories.