---
library_name: pytorch
tags: ["robotics", "world-model", "video-generation", "causal-forcing", "oscar", "repa"]
---

# needed_checkpoints_robotics_repa

The three checkpoints needed to run the DMD + REPA stage (`release/stage_configs/3_dmd_merlin.yaml`, branch
`release/merlin-causal-forcing` of causal_control_cosmos_causalforcingpp).

| path | config key | what |
|---|---|---|
| `models/oscar_2b/oscar_2b_dcp2torch.pt` | `cosmos_ckpt` | OSCAR-2B backbone (real/fake score init) |
| `causal_cd_oscar_chunk3/t1000g/checkpoint_model_002000/model.pt` | `generator_ckpt` | causal CD generator, step 2000 (DMD init) |
| `OmniStream/` | `encoder_kwargs.align_encoder_path` | frozen REPA target encoder ([StreamFormer/OmniStream](https://huggingface.co/StreamFormer/OmniStream), MIT) |

```bash
hf download csusupergear/needed_checkpoints_robotics_repa --local-dir $HF_DIR
sha256sum -c $HF_DIR/SHA256SUMS
```

The layout matches `<HF_DIR>` in the release configs, so `release/tools/relocate_paths.py` fills `cosmos_ckpt` and
`generator_ckpt`; set `align_encoder_path` to `$HF_DIR/OmniStream`.