Blame for README.md

---
license: other
license_name: observed-training-environments-and-reproduction-recipes
license_link: https://huggingface.co/datasets/diffusion-gif/training-environments/blob/main/README.md
---
# Ego-Exo / robot training reproduction environments

This package contains **actual retained environment snapshots**, exact observed
package-version constraints, role-specific installation requirements, source
commit locks, portable training command generators and preflight checks. It is
not a claim that a clean GPU reinstall or an eight-GPU training/metric reproduction
was completed on this machine.

## Four separate profiles — do not mix them

| Profile | Python | Torch / CUDA | Transformers | Purpose |
|---|---|---|---|---|
| egoexo-generator | 3.11.15 | 2.9.1+cu130 / 13.0 | 4.54.0 | Wan/VACE HOI, caption Global/Spatial generation training |
| egoexo-captioner-phiseg | 3.12.14 | 2.10.0+cu128 / 12.8 | 4.57.6 | Qwen3-VL-8B/ms-swift4.4.2 captioner + PhiSeg |
| robot-sft-grpo | 3.12.13 | 2.11.0+cu130 / 13.0 | 5.9.0 | Original 27B mix6 SFT / verl0.9.1 + vLLM0.24.0 GRPO |
| robot-data-cpu | 3.12.13 | no Torch required | not required | data restoration, validation and CPU tests |

The 27B checkpoint nickname is Qwen3.8; its actual architecture is `qwen3_5`.
It requires the 5.x Transformers stack. Do not use the generator's Transformers
4.54 environment for that model or mix vLLM0.19/Torch2.10 with vLLM0.24/Torch2.11.

## Create a NEW environment

Choose/create the exact listed Python interpreter first (Conda explicit package
URLs with MD5 are supplied for the observed Conda profiles). Never install into
an existing training environment by mistake.

```bash
python tools/install_environment.py --profile robot-sft-grpo --python /path/to/python3.12.13 --prefix /path/to/new_env
# inspect the plan; then add --execute
/path/to/new_env/bin/python tools/verify_environment.py --profile robot-sft-grpo --gpu-basic-check
```

`requirements.observed-full.lock.txt` captures the retained inventory;
`requirements.core.in` selects training-role dependencies while using that inventory
as constraints. The CPU profile additionally provides a resolved hash-locked
install file. Full GPU clean installation has not been revalidated here. FlashAttention
is a post-Torch build step (`--flash-attn --cuda-home /path/to/cuda13`), not an
uncontrolled build-isolation install before Torch exists.

## CUDA and platform pitfalls actually found

- This host has driver580.126.20 / one SM12.0 GPU. Three GPU profiles passed a
  tiny BF16 CUDA tensor operation; that is not eight-GPU model training.
- The system toolkit here is nvcc12.8, even when Torch's wheel runtime is13.0.
  Use a **matching CUDA13 devel toolkit** for torch-cu130 source builds and
  FlashAttention/GDN kernels. A driver reporting CUDA13 is not an nvcc installation.
- Historical robot GRPO uses TP4 and eight GPUs; one-GPU runs are not equivalent.
- The retained Decord0.6.0 wheel has an incorrect internal cp36 tag although its
  ctypes library imports on Python3.11/3.12. Its original pip-check platform warning
  is recorded, not hidden. Use the official py3 Linux wheel; validate import/video
  decoding for your host before training. Do not classify that warning as a
  proven numerical-model failure or claim all fresh wheel installations passed.
- The uv-created robot env has no pip module. `uv pip check --python ENV/bin/python`
  passed all256 installed packages; `python -m pip check` alone is not a valid
  dependency-health test of that original no-pip environment.

## Get exact source / models / old training data

Use `SOURCE_LOCK.json` and checkout those commits, not moving main branches.
Some source repositories remain private: owner authorization is required.
For Ego-Exo use `code/core`, the supplied source overlay where needed, and the
public weight bundle's sharded-Wan compatibility/attachment tool. DiffSynth Wan
code is pinned; unrelated local Flux experiments/images are excluded.

Download the already published `TRAINING_DATA_INDEX.json`: it distinguishes
legacy mix6 training (16,174 train / 3,228 heldout), later joint data, and Any2Ego
original versus common filtered CSVs. Original training files/NPZ are protected
in the pinned private vault. Large Ego-Exo original RGB/mask media are NOT fully
backed up remotely and require original-source permission. No tokens are bundled.

## Portable training recipes (dry-run by default)

```bash
python tools/run_training_recipe.py --recipe robot-sft --code-root /your/trick_prediction --python /new_env/bin/python --data-root /your/legacy_data --model-root /your/base27b --output /your/new_run --gpus 8
python tools/run_training_recipe.py --recipe egoexo-spatial --code-root /your/exo-ego/code/core --python /new_env/bin/python --data-root /your/ego_data --model-root /your/any2ego_bundle --diffsynth-root /your/DiffSynth-Studio --output /your/new_run --gpus 8
# Only add --execute after data/model/version preflight. Outputs must be new.
```

These are safe, explicit entrypoint recipes rather than evidence of convergence.
The historical Spatial snapshot used teacher captions; switching to newly generated
GRPO captions is a different experiment. Old 27B checkpoints were not trained on
new fixed-EEF joint labels. SFT warm-start does not restore absent optimizer/RNG state.

## Verification boundaries

Recorded: package versions/ABI, CPU dependency hash-lock resolution, source revisions,
core imports, tiny GPU operations, standalone command dry-runs and syntax checks.
Not completed: fresh full GPU installation, eight-GPU distributed training, all
optional native source builds, model metrics, or the numerical identity of an
unrecorded original training-launch environment. See VALIDATION_REPORT.json.