mlboydaisuke/multilingual-e5-small-ExecuTorch

multilingual-e5-small — ExecuTorch

The smallest of the multilingual E5 family. Text in, one 384-dimensional vector out, for search and retrieval that never leaves the device.

• Source: intfloat/multilingual-e5-small — 12 layers, 384 dimensions, 250,037 vocabulary
• License: mit
• Input: input_ids and attention_mask, both [1, 256] int64
• Output: [1, 384], mean-pooled and L2-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 normalises, 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 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_multilingual_e5_small_xnnpack_fp32.pte	470.2	89.8	78.5%	1.000000	0%
fp16	embed_multilingual_e5_small_xnnpack_fp16.pte	235.2	41.8	67.6%	1.000000	1%
Core ML (fp16, iOS)	embed_multilingual_e5_small_coreml_all.pte	235.8	4.9	100.0%	0.999979	15%

*Mac arm64, median of 10, one 256-token sequence — a reference point for relative cost, not a device number. Torch eager fp32 on the same machine is 124.0 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

On the sdpa path transformers hands F.scaled_dot_product_attention a boolean mask, PyTorch fills the masked entries with -inf, and the edge dialect lowers the whole thing to _safe_softmax — whose guard against a row of all -inf costs six operations XNNPACK cannot take (scalar_tensor, where, logical_not, eq, full_like, any.dim), once per attention block, each one cutting the subgraph in two.

attn_implementation="eager" masks with torch.finfo(dtype).min instead — a large finite number — so there is no guard to lower. The two disagree only about rows that have no unmasked key at all: sdpa zeroes them, eager gives them a uniform row. Those are padding rows, which the pooling discards and which every real query row masks out, so the output does not move. 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 62.6% to 78.5% delegated.

Not shipped: int8

embed_multilingual_e5_small_xnnpack_int8.pte is 406.7 MB against fp16's 235.2 MB. Dynamic int8 quantises the linear weights and leaves the token embedding table in fp32, and here that table is 384 MB of the 470.2 MB model — 82%. The size a build comes out at is 0.5 + 1.5 x (table share) times the fp16 build; at 82% that is 1.73, 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 0.0046 while the closest fp32 decision — the gap between the document a query retrieves and the runner-up — is 0.0105. That is 44% of the room available, against a bar of 50%.

Correlation reads 0.999568 for this build, which no correlation gate would stop.

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