docs/CHECKPOINTS_FLUX.md

# FLUX cloth-folding inference checkpoint

## Released files

```text
checkpoints/flux_fold_clothes_bf16/
  config.json
  manifest.json
  model.safetensors
```

This is the exported BF16 policy from step 14418 (model weights). No further export is needed. `model.safetensors` is 13,894,466,592 bytes, SHA256 `4378d2ffd9f90f6b2efc0767512c29c934cd9f78c1e66997b551b52ade4b1c45`.

The supplied export was downloaded and verified against OSS CRC64 and its original manifest. Tensor bytes are unchanged. Internal encoder paths were replaced with `./resources/flux/video_vae.safetensors` and `./resources/flux/text_encoder`; disabled training-cache fields `cached_video_latents` and `latent_cache_contract_sha256`, unsupported by the bundled `PolicyConfig`, were removed. The manifest now records the sanitized config hash. Keep config, manifest and model together.

## Load with the robot adapter

From the repository root, after installing the FLUX environment:

```bash
export FLUX_RESOURCES=/absolute/path/to/flux_resources
export UMI_INFER_DATA=/absolute/path/to/calibrated_cell
bash scripts/run_flux_server.sh
```

The launcher defaults to the bundled checkpoint directory and injects the VAE/text encoder from `FLUX_RESOURCES`, overriding the portable config paths. Required resources:

```text
flux_resources/
  video_vae.safetensors
  text_encoder/  # matching pretrained text encoder and processor
```

Encoders, cell extrinsics and vendor SDKs are separate requirements. For direct `FluxActionPolicy.from_pretrained`, either supply `video_vae` and `text_encoder` explicitly or provide the relative resource paths from the working directory.

The export fixes two cameras (`images.left_ego`, `images.right_ego`), a 256×512 side-by-side canvas, 20-D state/actions, 32-action chunks, 30 Hz and absolute actions in each chunk's anchor frame. The robot adapter handles 16-D poses, relative-state construction, normalization, anchor restoration and cell extrinsics. Keep the saved state/action quantiles unchanged.

## Sampling and validation

The export has null sampler, step-count, guidance and shift fields. The server sets Euler, 4 steps, shift 42, guidance 1.0, action guidance 1.0 and seed 0 by default. Adjust `FLUX_SAMPLER`, `FLUX_STEPS`, `FLUX_SHIFT`, `FLUX_GUIDANCE`, `FLUX_ACTION_GUIDANCE` and `FLUX_SEED` explicitly when a validated deployment recipe is available. The training timestep shift does not establish an optimal inference shift.

File integrity and configuration compatibility have been checked on CPU. Frozen encoders, CUDA inference and physical-robot cloth folding were not tested in this release. See [VALIDATION.md](VALIDATION.md).