Document the Transformers checkpoint and its current limits

Transformers generation is supported from the transformers/ subfolder. Partial diffusion and inverse folding are not: they call the autoencoder's encoder, and rfd4-proteina-prep does not yet emit incident_bond_type_counts or ground_truth_element. Ligand and atomized inputs stop with an error rather than conditioning on unaligned geometry. Diffusers remains on the previous release.

README.md

@@ -33,7 +33,7 @@ the root are the *native* weights, loaded by `Proteina.from_pretrained`.
 | Call | Weights | Location | Status |
 |---|---|---|---|
 | `Proteina.from_pretrained("nvidia/RFD4-Proteina")` | Native | root of this repo | **Supported** |
-| `Rfd4ProteinaModel.from_pretrained("nvidia/RFD4-Proteina", subfolder="transformers")` | Transformers | `transformers/` here | Pending |
+| `Rfd4ProteinaModel.from_pretrained("nvidia/RFD4-Proteina", subfolder="transformers")` | Transformers | `transformers/` here | **Supported**, with limits below |
 | `RFD4ProteinaPipeline.from_pretrained("nvidia/RFD4-Proteina-Diffusers")` | Diffusers | [its own repo](https://huggingface.co/nvidia/RFD4-Proteina-Diffusers) | Pending; still serves the previous release |
 
 It is one model throughout — nothing is retrained. The Diffusers weights are the same tensors
@@ -45,11 +45,18 @@ Diffusers needs `model_index.json` at the root of whatever it is pointed at, so
 repository. Annotated-CIF preprocessing for both libraries is in
 [`nvidia/RFD4-Proteina-prep`](https://huggingface.co/nvidia/RFD4-Proteina-prep).
 
-> **Only the native path runs this release.** The autoencoder in these weights uses a 147-class
-> residue vocabulary and a 32-dimensional latent, where the previous release used 33 classes and
-> 16 dimensions. Both library integrations are still written against the old vocabulary and
-> cannot convert these weights yet. `nvidia/RFD4-Proteina-Diffusers` still holds the **previous**
-> release — following its examples gives you the previous model, not this one.
+> **Diffusers is not updated yet.** `nvidia/RFD4-Proteina-Diffusers` still holds the **previous**
+> release (33-class vocabulary, 16-dimensional latent) — following its examples gives you the
+> previous model, not this one.
+>
+> **The Transformers checkpoint supports generation only, for now.** Preparing an annotated CIF
+> uses [`rfd4-proteina-prep`](https://huggingface.co/nvidia/RFD4-Proteina-prep), which does not
+> yet emit two features the unified autoencoder's encoder needs
+> (`incident_bond_type_counts`, `ground_truth_element`). Plain generation only calls the decoder
+> and works; **partial diffusion and inverse folding call the encoder and will fail.** Ligand and
+> atomized inputs also stop with a clear error rather than conditioning on unaligned geometry,
+> because chirality alignment needs fields prep does not emit. The published prep release also
+> pins `atomworks==3.0.1` and is not yet updated for the 3.0.3 this release uses.
 
 Training data lives in the companion [dataset repository](https://huggingface.co/datasets/nvidia/RFD4-Proteina).
 
@@ -177,17 +184,35 @@ print(list(Path("runs/enzyme/inference_samples").rglob("*_generated.cif*")))
 
 `from_pretrained` loads in evaluation mode on CPU by default; `run_generation` uses the native CUDA inference workflow. A local exported bundle directory also works as the model argument.
 
-### Transformers and Diffusers
+### Transformers
 
-**Neither works with this revision yet.** Both integrations are written against the previous
-release's 33-class residue vocabulary and 16-dimensional latent; these weights use 147 classes
-and a 32-dimensional latent, and both conversion paths reject them rather than load them
-incorrectly. Annotated-CIF preprocessing for both lives in the shared
-[`rfd4-proteina-prep`](https://huggingface.co/nvidia/RFD4-Proteina-prep) package, which needs the
-same vocabulary update.
+```bash
+pip install transformers git+https://huggingface.co/nvidia/RFD4-Proteina-prep
+```
+
+```python
+import torch
+from rfd4_proteina_prep import prepare_cif, write_cif
+from transformers import Rfd4ProteinaModel
+
+model = Rfd4ProteinaModel.from_pretrained(
+    "nvidia/RFD4-Proteina", subfolder="transformers", dtype=torch.bfloat16, device_map="cuda:0"
+)
+features, atoms = prepare_cif("input.cif", seed=42, device=model.device)
+output = model.sample(features, generator=[torch.Generator("cuda:0").manual_seed(42)])
+write_cif(output, features, atoms, "design.cif")
+```
+
+`subfolder="transformers"` is required: a Transformers export writes `config.json` and
+`model.safetensors`, the names the native bundle already occupies at the root. The sampler
+settings (`sampler_args`, `recycle_depths`) are part of the config, not the weights, and this
+checkpoint records the recipe resolved from the native inference config — 60 steps, 185 trunk
+passes, matching `recycle_strategy=thinking_medium`. See the limits in the banner above:
+generation works; partial diffusion and inverse folding do not yet.
+
+### Diffusers
 
-Until that work lands, use the native path above. If you specifically need the Diffusers pipeline
-today, [`nvidia/RFD4-Proteina-Diffusers`](https://huggingface.co/nvidia/RFD4-Proteina-Diffusers)
+Not updated for this release. [`nvidia/RFD4-Proteina-Diffusers`](https://huggingface.co/nvidia/RFD4-Proteina-Diffusers)
 still serves the **previous** release pair — a different model from this one, with its own
 `provenance.json` recording which native revision it came from.