| --- |
| license: other |
| library_name: pytorch |
| tags: |
| - custom-code |
| - visual-navigation |
| - worldvln |
| - safetensors |
| --- |
| |
| # WorldVLN Backbone |
|
|
| This repository was exported from a WorldVLN training checkpoint into a Hugging Face friendly layout. |
| It is meant for direct folder upload: upload this whole directory as the root of a Hugging Face model repo. |
|
|
| ## Included Weights |
|
|
| - `gpt/`: standard sharded `safetensors` export of `trainer.gpt_fsdp` |
| - `vae/`: standard sharded `safetensors` export of `trainer.vae_local` |
| - `load_weights.py`: helper utilities for loading the two subfolders directly |
| - `export_manifest.json`: export provenance and metadata |
|
|
| ## Source Checkpoint |
|
|
| - Original checkpoint: `/manifold-obs/vln-uav/rluavflowcheckpoint_partialfreeze_stageb_only/train_run_pf_stageb_clipmix_gatemean_tok20480_vb1_ac4_iter1200_20260408_084520/ckpts/WorldVLN_backbone.pth` |
| - Architecture: `infinity_qwen8b` |
| - Epoch: `0` |
| - Iter: `1200` |
| - Global step: `1200` |
|
|
| ## File Layout |
|
|
| - `gpt/model.safetensors.index.json` |
| - `gpt/model-00001-of-xxxxx.safetensors` |
| - `vae/model.safetensors.index.json` |
| - `vae/model-00001-of-xxxxx.safetensors` |
|
|
| GPT shard count: `4` |
|
|
| VAE shard count: `1` |
|
|
| ## Direct Loading |
|
|
| This export is intentionally split into two model folders instead of one mixed training checkpoint. |
| Instantiate your GPT model and VAE model with this project's code, then load them separately. |
|
|
| ```python |
| from load_weights import load_worldvln_models |
| |
| load_worldvln_models( |
| repo_dir=".", |
| gpt_model=infinity_model, |
| vae_model=vae_model, |
| strict=False, |
| device="cpu", |
| ) |
| ``` |
|
|
| Or load raw state dicts only: |
|
|
| ```python |
| from load_weights import load_worldvln_state_dicts |
| |
| bundle = load_worldvln_state_dicts(".", device="cpu") |
| gpt_state_dict = bundle["gpt"] |
| vae_state_dict = bundle["vae"] |
| ``` |
|
|
| ## Notes |
|
|
| - This is a custom-code model export, not a generic `transformers.AutoModel.from_pretrained(...)` repo. |
| - The weights are in standard sharded `safetensors` format and do not require manual file concatenation. |
| - For inference in this codebase, point the GPT loader to `gpt/` and the VAE loader to `vae/`. |
|
|
