Commit History
-
new: add model_v2.onnx, a re-export of ColModernVBERT for fastembed
Verified signatureChanges compared with model.onnx: - Uses colmodernvbert-base revision 17604b47. Upstream updated the vision tower after model.onnx was exported, so image embeddings shift slightly (mean per-token cosine 0.9998). Text embeddings are unchanged. - All-zero tiles are dropped before the vision encoder, as in transformers. model.onnx gave wrong embeddings for an image that came after a zero-padded image in the same batch. - The vision encoder is skipped when pixel_values has no non-zero tile, so text-only inputs don't run it. - Vision attention uses onnxruntime's com.microsoft MultiHeadAttention operator, which roughly halves CPU memory per image tile at the same speed.b2e872d
