Add JEPA / V-JEPA 2 / LeWorldModel HF adapter

[jascal]

What

Adds full support for JEPA-family joint-embedding world models so users can hf download / hf convert them like any other family. Primary focus: V-JEPA 2 (facebook/vjepa2-*) and LeWorldModel (quentinll/lewm-*); also covers I-JEPA (facebook/ijepa_*).

n-orca hf convert facebook/vjepa2-vitl-fpc64-256 --out vjepa2.n.orca.md --mermaid vjepa2.mmd
n-orca verify vjepa2.n.orca.md          # VALID  params=325,967,488  depth=222
n-orca hf convert quentinll/lewm-pusht --out lewm-pusht.n.orca.md

New JepaAdapter (n_orca/hf/adapters/jepa.py)

Normalizes two unrelated config schemas into one encoder → predictor DAG:

• Flat HF transformers configs — V-JEPA 2 (vjepa2 / VJEPA2Model) and I-JEPA (ijepa / IJepaModel), with pred_* predictor fields.
• Nested Hydra configs — LeWorldModel (quentinll/lewm-*), which ship no model_type and are matched structurally (_target_ + nested encoder/predictor dicts).

The graph parses the ViT encoder, the joint-embedding predictor (distinct width/depth), a latent bridge between them, optional additive action conditioning, and an optional projector head. It emits two outputs — encoder_latents (the representation) and predicted_latents (the forecast) — that share the encoder's embedding space, so a single output_shape invariant enforces latent-dim consistency.

Mask tokens, the EMA stop-gradient target, SIGReg, and rotary/sincos positions are captured as verification rules + hyperparameters (mirroring how the LLaMA adapter treats RoPE/RMSNorm), keeping output round-trippable and compilable.

Supporting changes

• Ops — TubeletEmbed (Conv3d, video) and PatchEmbed (Conv2d, image) patch embedders with shape inference + param counting. Verifier param count matches PyTorch exactly (11,943,104 for LeWM).
• Blocks — pre_norm_decoder_block gains a prefix= arg so encoder/predictor stacks don't collide (default reproduces prior names byte-for-byte).
• CLI — n-orca hf download --include-processor also fetches preprocessor_config.json / video_preprocessor_config.json (V-JEPA ships the latter).
• Examples — examples/hf-generated/vjepa2.{n.orca.md,mmd} (V-JEPA 2 ViT-L, 326M params) and lewm-pusht.{n.orca.md,mmd} (action-conditioned). Verified byte-identical to live-Hub hf convert output.
• README — JEPA row in the family table; TubeletEmbed/PatchEmbed in the op table; download/convert examples.

Testing

140 passing (was 118). +22 tests: new ops, dispatch (incl. structural LeWM detection with no model_type), clean verify, parser round-trip, Mermaid branch rendering, and torch forward passes for video V-JEPA 2 (5D (B,C,T,H,W) clip), encoder-only I-JEPA, and action-conditioned LeWM. The exact target commands were run against the live Hub.

🤖 Generated with Claude Code