---
tags:
  - apptainer
  - environment
  - visual-language-tracking
  - pytorch
  - blackwell
---

# track_env — TrackingAnything training environment

Public environment-image release: **large-cu128-wandb-v1**, built 2026-10-03.
Repository: [jeffry77/track_env](https://huggingface.co/jeffry77/track_env).
Renamed from `jeffry77/tracking_anything_env`; the versioned artifact and its
SHA256 are unchanged.
This is an Apptainer SIF, **not model weights or a dataset**. The full
TrackingAnything large-training project must be obtained separately.

## Artifact

- File: `tracking-anything-large-cu128-wandb-v1.sif`
- Size: **10,113,572,864 bytes** (approximately 9.42 GiB)
- SHA256: `d1af2707df2bf5424c54fb4931b0a14424fdd37a26e3343f833fb05855b65784`
- Build/runtime tooling checked: Apptainer 1.4.5, Linux x86_64.

| Component | Version / behavior |
|---|---|
| Python | 3.10 |
| CUDA base / PyTorch CUDA | 12.8.1 / 12.8 |
| PyTorch / torchvision | 2.7.1+cu128 / 0.22.1+cu128 |
| Transformers / PEFT | 4.57.1 / 0.12.0 |
| Liger | 0.3.1 |
| Attention backend | SDPA; no external FlashAttention |
| W&B | 0.22.2; default mode online; credentials supplied at runtime |
| HTTP SOCKS support | requests 2.32.5 / PySocks 1.7.1 |

This preserves the original Blackwell environment and supplies a new versioned
image plus explicit two-node launch wrappers. It does not include project code,
model weights, datasets, login caches, or user credentials. Build dependencies
and distribution licenses remain those of their respective upstream packages;
this release does not relicense NVIDIA CUDA/cuDNN or other bundled components.

## Download and verify

No Hugging Face login is required to download this public environment.
Credentials for W&B, private models or restricted datasets remain separate;
do not put them in source files or the image.

```bash
hf download jeffry77/track_env --local-dir ./tracking-anything-env
cd tracking-anything-env
sha256sum -c tracking-anything-large-cu128-wandb-v1.sif.sha256

# CPU dependency check; does not initialize W&B or allocate a GPU.
apptainer exec --cleanenv --no-home tracking-anything-large-cu128-wandb-v1.sif \
  /opt/venv/bin/python /opt/tracking-anything-env/check_large_env.py
```

## Use with the full training project

```bash
# These must be shared, identical absolute paths on both nodes.
export PROJECT_DIR=/shared/TrackingAnything-large-training
export IMAGE=/shared/tracking-anything-env/tracking-anything-large-cu128-wandb-v1.sif
export DATA_DIR=/shared/datasets
export MODEL_DIR=/shared/models
export CONTAINER_CACHE_DIR=/shared/user-cache/tracking-anything
export PYTHON_BIN="$PROJECT_DIR/apptainer/python.sh"
chmod +x "$PYTHON_BIN"
export WANDB_MODE=online
# Supply WANDB_API_KEY securely, or use WANDB_NETRC_FILE as documented.

cd "$PROJECT_DIR"
GPUS=none "$PYTHON_BIN" scripts/preflight_large_training.py --require-wandb --deep-cache --require-2x8
sbatch --export=ALL --partition=YOUR_PRO6000_PARTITION scripts/train_2node_16gpu.slurm
```

Complete native-data preparation, initialization and training configuration
before preflight, including the training-mode Check0 report. Use the current
project's runtime wrappers so code-level asset/mount fixes stay current without
rebuilding this image. The image alone cannot supply these artifacts. See
[the detailed environment guide](apptainer/README.md) for same-path binds,
W&B offline mode, secrets handling and NCCL/RoCE settings.

## Verification scope

The derivative passed build-time CPU dependency checks. Post-build runtime
checks confirm the new profile, W&B online default and expected package versions.
The SIF checksum was also verified after transfer from the build machine.

**Not yet validated:** this derivative's full-model GPU training/inference,
RTX PRO 6000 multi-GPU kernels, 16-GPU/two-node NCCL or RoCE throughput, and
online W&B logging. Historical single-GPU synthetic checks on the base image
are not a substitute. Use an allocated target-cluster smoke before formal
training. No speed, convergence, or tracking-accuracy result is claimed here.

Apptainer preserves some inherited OCI labels (including Ubuntu/base-profile
labels). The release filename, this checksum, and
`/opt/tracking-anything-env/large-profile` identify this derivative; inherited
base labels alone do not identify the release.

## Rebuild

The included definition builds a derivative from the exact original baseline
SIF with SHA256 `9f854044eeebe8475e22363d94569f85c02eba40743a6feda27c655372a28101`.
That separate base image is required **only for rebuilding**, not for executing
this standalone SIF. Build scripts verify the base and refuse to overwrite
existing outputs. Credentials must never be embedded during rebuild or upload.