Paraphrase similarity across 50+ languages, mpnet-sized. Text in, one 768-dimensional vector out, for search and retrieval that never leaves the device.

sentence-transformers stores it per model, and the shelf's seven embedding models do not agree. This one pools mean and does not normalise, read from 1_Pooling/config.json and modules.json rather than inferred from the family name. Getting it wrong does not throw; it returns vectors that look fine and rank wrong.

*Mac arm64, one 256-token sequence, fastest of five medians of ten — a reference point for relative cost, not a device number. The host shares its cores with other work, and a single median does not survive that: the same eager model here measured 19.6 ms and 182.8 ms twenty minutes apart. Contention only ever adds time, so the fastest repetition is the one that means something. Torch eager fp32, measured the same way, is 34.5 ms.

Cosine is measured against the model run in eager through its own pooling, over eight sentences. The last column is the one that decides: rank those eight against each other, and ask whether this build's score error is smaller than the gap between the document a query retrieves and the runner-up. Every shipped build keeps all eight top-1 results.

F.scaled_dot_product_attention does not survive export as one operation. The edge dialect lowers it through _safe_softmax, whose guard against a row with no unmasked key at all leaves 11 operations XNNPACK cannot take, in every attention block — scalar_tensor, where, mul.Scalar, logical_not, eq, full_like, any.dim. Each one cuts the subgraph in two.

The switch is attn_implementation="eager": transformers then builds the mask itself, as torch.finfo(dtype).min, instead of handing F.sdpa a boolean mask for PyTorch to fill with -inf.

The guard is emitted whether or not it can ever fire, and here it cannot: it triggers only on -inf, and this arm never produces one. So the two differ only about rows that have no unmasked key at all — sdpa zeroes them, this one gives them a uniform row — and those are padding rows, which the pooling discards and which every real query row masks out anyway. Measured with all but eight positions masked, as adversarial as this shape gets, the two graphs agree to 1.4e-07.

XNNPACK fp32 goes from 62.1% to 77.8% delegated.

embed_paraphrase_multilingual_mpnet_xnnpack_int8.pte is 855.5 MB against fp16's 555.2 MB. Dynamic int8 quantises the linear weights and leaves the token embedding table in fp32, and here that table is 768 MB of the 1110.0 MB model — 69%. The size a build comes out at is 0.5 + 1.5 x (table share) times the fp16 build; at 69% that is 1.54, so there was never a smaller file to be had.

It is withheld on the number that decides. Ranking the eight test sentences against each other, this build moves a pair score by at most 0.0053 while the closest fp32 decision — the gap between the document a query retrieves and the runner-up — is 0.0207. That is 26% of the room available, against a bar of 50%.

Correlation reads 0.999704 for this build, which no correlation gate would stop.

torch.export -> to_edge_transform_and_lower(partitioner) -> .pte (conversion scripts: executorch-models)