mirror of
https://github.com/hanzoai/ml.git
synced 2026-11-20 23:08:29 +00:00
ViT encoder as exposed by transformers.IJepaModel: no CLS token, learned
per-patch position embeddings, per-patch last_hidden_state after a final
LayerNorm, plus a mean-pooled convenience. Module tree mirrors the published
facebook/ijepa_vith14_1k layout so the checkpoint loads with no key remap.
Falsifiable parity gate vs transformers.IJepaModel (fp32 CPU):
tiny 2-layer fixture (committed, no network): per-patch max_abs 7.15e-7
real ViT-H/14 (#[ignore], 2.5GB ckpt): per-patch rel 7.83e-5, cosine 1.0
pooled rel 1.78e-6, cosine 1.0
Encoder only; the latent predictor is not in IJepaModel (follow-up).