README.md

metadata

license: apache-2.0
tags:
  - executorch
  - xnnpack
  - pte
  - on-device
  - feature-extraction
  - sentence-similarity
base_model:
  - nomic-ai/nomic-embed-text-v1.5

nomic-embed-text-v1.5 — ExecuTorch

A BERT with rotary embeddings and SwiGLU, trained so that a prefix of the vector is still a usable vector. Text in, one 768-dimensional vector out, for search and retrieval that never leaves the device.

• Source: nomic-ai/nomic-embed-text-v1.5 — 12 layers, 768 dimensions, 30,528 vocabulary
• License: apache-2.0
• Input: input_ids and attention_mask, both [1, 256] int64
• Output: [1, 768], mean-pooled and not normalised inside the graph

The recipe is in the graph, and it was read off this repo

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.

The prefix is not in the graph

This model is trained with search_query: in front of the text and expects it at inference. That happens before tokenisation, so the .pte never sees it as anything but tokens — and leaving it out does not throw. It returns a plausible vector that retrieves worse.

Verification

build	file	size (MB)	Mac ms*	backend takes	worst cosine vs eager	retrieval budget
fp32	embed_nomic_embed_text_v15_xnnpack_fp32.pte	547.2	45.1	64.5%	1.000000	0%
fp16	embed_nomic_embed_text_v15_xnnpack_fp16.pte	273.8	97.0	58.6%	0.999999	7%
Core ML (fp16, iOS)	embed_nomic_embed_text_v15_coreml_all.pte	274.8	8.1	100.0%	0.999793	46%

*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 44.9 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.

The attention is eager, and that is the faster export

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 8 operations XNNPACK cannot take, in every attention block — where, mul.Scalar, logical_not, eq, full_like, any.dim. Each one cuts the subgraph in two.

This model does not answer to attn_implementation at all. Its attention lives in nomic-ai/nomic-bert-2048 and calls F.sdpa itself, so the setting goes nowhere — measured, 58.8% delegated either way. That code already carries an explicit-softmax arm, behind a module-level scaled_dot_product_attention it binds at import because the symbol existed; clearing that symbol is the switch. It already passes an additive finfo(dtype).min mask rather than a boolean one, which is why its guard costs eight operations a block here rather than eleven — there is no boolean mask to convert.

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 58.8% to 64.5% delegated.

Matryoshka: the vector truncates

This model is trained so that a prefix of the vector is still a usable vector — 768 down to 512, 256, 128 or 64 dimensions, trading accuracy for index size. The graph returns the full 768, because the dimension is the caller's choice, and the truncation recipe is three lines:

import torch.nn.functional as F
v = F.layer_norm(v, (v.shape[1],))     # before truncating, not after
v = v[:, :dim]                         # 512 / 256 / 128 / 64
v = F.normalize(v, p=2, dim=1)

The layer_norm is what makes the prefix usable, and it is easy to skip. At the full 768 it barely matters — measured on this shelf, going through the layer_norm changes the direction of the vector by a cosine of 0.999944 and leaves the test pair's score at 0.8233 either way. It earns its place only once you truncate.

The four prefixes are a real part of the model. search_document: for what goes in the index, search_query: for what is asked of it, plus classification: and clustering: . Unlike E5's symmetric mode, the two retrieval prefixes are not interchangeable.

The architecture is not stock BERT. Rotary embeddings and a SwiGLU MLP, with the modelling code in nomic-ai/nomic-bert-2048 rather than in transformers — loading it needs trust_remote_code=True and einops installed. None of that reaches the .pte, which is a plain graph once exported.

Not shipped: int8

embed_nomic_embed_text_v15_xnnpack_int8.pte is 208.0 MB — smaller than fp16's 273.8 MB, because the token embedding table is only 94 MB of the 547.2 MB model (17%), leaving most of the weight in linears for int8 to shrink.

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 0.0221 while the closest fp32 decision — the gap between the document a query retrieves and the runner-up — is 0.0077. That is 288% of the room available, against a bar of 50%.

Correlation reads 0.992203 for this build, which no correlation gate would stop.

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