siglip2-base-patch16-384 text model — ONNX

Only the text model of upshift/siglip2-base-patch16-384-onnx, the ONNX export of google/siglip2-base-patch16-384 (revision f775b65, opset 18), for when you only need text embeddings, for example to embed search queries against image embeddings computed elsewhere. The files are identical to the ones in the full repo, which also has the vision model.

File	Precision	Size	Inputs → output
`text_model.onnx` + `text_model.onnx_data`	fp32	1.13 GB	`input_ids` int64 `(B, 64)` → `text_embeds` float32 `(B, 768)`
`text_model_fp16.onnx` + `text_model_fp16.onnx_data`	fp16	565 MB	same as fp32

B (batch) is dynamic. Embeddings are not normalised. config.json and the tokenizer files are copied from the original repo; export_meta.json holds logit_scale (already exponentiated) and logit_bias for scoring against image embeddings.

Parity

Compared with the PyTorch text model on onnxruntime (CPU):

Precision	Inputs	Min cosine	Max abs diff	Threshold	Result
fp32	6	1.0000000	1.0e-05	0.9999	passed
fp16	6	0.9999989	1.7e-02	0.999	passed

Inputs

Lowercase the text, tokenise with the included Gemma tokenizer.json, append <eos> and pad with <pad> (id 0) to exactly 64 tokens. The model pools the last token, so the padding is required. In Python, use Siglip2Tokenizer (transformers 5.x AutoTokenizer returns a GemmaTokenizer that does not lowercase).

Scoring against an image embedding from the vision model of upshift/siglip2-base-patch16-384-onnx: p = sigmoid(logit_scale · cos(image, text) + logit_bias), per label.

Usage with onnxruntime-web

import * as ort from 'onnxruntime-web/webgpu';

const base = 'https://huggingface.co/upshift/siglip2-base-patch16-384-text-onnx/resolve/main/';
const text = await ort.InferenceSession.create(base + 'text_model.onnx', {
  executionProviders: ['webgpu', 'wasm'],
  externalData: [{ path: 'text_model.onnx_data', data: base + 'text_model.onnx_data' }],
});
const { text_embeds } = await text.run({
  input_ids: new ort.Tensor('int64', inputIds, [1, 64]),
});