Repository navigation
Conversation
[High impact] The PR appears safe to merge. Reviews (12) · Last reviewed commit: "fix: keep zero scores of either sign in ..." · Reviewed by Greptile |
Dismissed because a newer commit was pushed; Greptile will re-review the current head.
This comment has been minimized.
This comment has been minimized.
…rrect results or errors
Dismissed because a newer commit was pushed; Greptile will re-review the current head.
Dismissed because a newer commit was pushed; Greptile will re-review the current head.
… config, and run equal head counts in one batch
Dismissed because a newer commit was pushed; Greptile will re-review the current head.
…wen3, and image edge cases
…serialize the reranker cache cleanup
… where HF_HOME caches models
… only a finished ONNX Runtime build
Dismissed because a newer commit was pushed; Greptile will re-review the current head.
… a pause in another download
…nd the Qwen3 defaults
Dismissed because a newer commit was pushed; Greptile will re-review the current head.
… the allocation probe of the resize
Dismissed because a newer commit was pushed; Greptile will re-review the current head.
Dismissed because a newer commit was pushed; Greptile will re-review the current head.
## [7.7.1](https://github.com/[secure]/fastembed-rs/compare/v7.7.0...v7.7.1) (2026-11-16) ### 🐛 Bug Fixes * harden edge cases across the ONNX and candle models, and repair CI ([#355](https://github.com/[secure]/fastembed-rs/issues/355)) ([d76d093](https://github.com/[secure]/fastembed-rs/commit/d76d093c54e8c04dd50347d010098dbef2981240)) ### 📝 Documentation * Removes AtlasCloud banner README.md ([#354](https://github.com/[secure]/fastembed-rs/issues/354)) ([a1e6023](https://github.com/[secure]/fastembed-rs/commit/a1e6023bc0d2d4b85a8a6101583972329d9adae4)) ### 🔁 Continuous Integration * release with semantic-release 26 and publish to crates.io with trusted publishing ([#356](https://github.com/[secure]/fastembed-rs/issues/356)) ([d7fb1db](https://github.com/[secure]/fastembed-rs/commit/d7fb1db6ea5640ff4069977ff0e7df415cea499b))
|
🎉 This PR is included in version 7.7.1 🎉 The release is available on:
Your semantic-release bot 📦🚀 |
ONNX models
TextEmbedding::embedpools each batch before it runs the next one, sobatch_sizelimits the memory in use.TextEmbedding::embedandtransformwith no texts return an empty result for dynamically quantized models. Abatch_sizeof 0 is still an error.EmbeddingGemma300MQis batched like the other EmbeddingGemma variants, because its output does not depend on the batch.TextEmbedding::get_quantization_modereturnsQuantizationMode::Staticfor it, andembedsplits its texts bybatch_size.TextEmbedding::list_supported_modelsreturns the models in table order, instead of a different order in each process.[sep]or[mask], is tokenized as ordinary text, as the model'stokenizer.jsondoes.max_lengthbelow the tokenizer's special-token count returns an error, instead of a panic or no truncation.TextRerankchecks the count for a pair.TextRerank::rerankaccepts[batch]logits, reads the first column of[batch, n]logits, and returns an error for other shapes, instead of a panic.TextRerank::rerankandsimilarity::top_krank aNaNscore last.top_kkeeps equal scores in corpus order, and0.0and-0.0count as equal.SparseTextEmbeddingcomputes the SPLADE weights withln_1p, so a small positive logit keeps a positive weight. Other weights move by 2e-7 at most.SparseTextEmbedding::embedandBgem3Embedding::embedreturnError::InvalidShapefor an output with no row for a text or token, instead of a panic.ImageEmbeddingreturns an error for an image with no pixels. A pipeline that resizes aborted the process, and a pipeline that only pads returned the vector of an empty image.imagedecoders. A 1 x 100,000 image is an example. A resizer set withwith_resizehas no limit.ImageEmbedding::embedandembed_byteshold one decoded image at a time, instead of the whole batch.ImageEmbedding::embeddecodes by the file extension first, and by the file content when that fails. A pipe is read once.ImageEmbedding::embed_preprocessedandembed_imageswith an empty batch return an empty list, instead of an error.preprocessor_config.jsonaccepts the processor namesCLIPFeatureExtractorandConvNextImageProcessor.preprocessor_config.jsonwithresample0, 1, or 2 resizes with the nearest, Lanczos, or bilinear filter that it names, instead of bicubic. The built-in models use bicubic, as before.ort-load-dynamic, a missing ONNX Runtime library returns an error, instead of a panic.ExternalInitializerFileis exported from the crate root.MxbaiEmbedLargeV1andMxbaiEmbedLargeV1Qspell the vendor name Mixedbread.Hugging Face Hub
LockAcquisitionerror after five seconds (Bug(?) when running test suite #203). The load also waits through a pause of up to one minute in that download. With 16 loads at once on an empty cache, 7.7.0 returned the error to nine or ten of them.HF_HOMEset, the token thathuggingface-cli loginsaved inHF_HOMEis used.HF_ENDPOINTmeans the default endpoint, and a trailing slash is removed.model.safetensorsdownload reports its own cause, next to the cause for the index file.model.safetensors.index.jsonthat points outside the repository returns an error.candle models
BertTextEmbeddingembeds RoBERTa, XLM-RoBERTa, and CamemBERT checkpoints with the position ids of that architecture. It used BERT position ids, which gave wrong vectors.BertTextEmbeddingloads checkpoints that name the LayerNorm tensorsgammaandbeta, such asgoogle-bert/bert-base-uncased.BertTextEmbeddingreports a missing tensor under the name prefix that the file uses.BertTextEmbeddingcapsmax_lengthat the longest text that the position table holds, so long texts no longer fail.NomicV2MoeTextEmbeddingreturns unit-length vectors withDType::F16andDType::BF16weights. The norms were up to 1.09.NomicV2MoeTextEmbedding::newwith a tokenizer that has no padding pads each batch, instead of rows that mix the tokens of two texts.nn.GELU()in the reference model does.rope_thetaand the MRoPE settings underrope_parameters.Qwen/Qwen3-VL-Embedding-8Bloads. Itstext_confighas notie_word_embeddings.embed_paddedruns one padded row, instead of an error. This path has a test that only the macOS CI job runs.Qwen3TextEmbedding::newsets a tokenizer that pads to pad on the left. A tokenizer that padded on the right gave wrong vectors.Qwen3TextEmbedding::embed_paddedpads the rows itself for a tokenizer with no padding, and returns an error for rows of unequal length.Qwen3VLEmbedding::embed_imagesruns withDType::BF16weights on the CPU, instead of an error.Qwen3VLEmbeddingaccepts amax_lengthup tousize::MAX, instead of an overflow panic in a debug build.max_lengthbelow the special-token count.Qwen3VLEmbedding::embed_imagesdecodes by the file content when the extension fails.Tests
BertTextEmbeddingon a RoBERTa checkpoint with transformers, and theClipVitB32image vectors with fastembed for Python.HF_HOMEwhen it is set.test_rerankdeletes its models only on CI, and waits while another test reads them.Docs
query:andpassage:prefixes belong to other models.Qvariants are half precision, lists thecandle-coreandortdependency lines, and uses the Qwen3 prompt format of the model card.HF_HOMEis the model cache directory itself.Qwen3TextEmbeddingembeds a text as given, and thatQwen3VLEmbedding::embed_textsadds the prompt of the model card.image-modelsfeature forImageEmbedding, and docs.rs builds thebert-textfeature.CI
bert-textandort-load-dynamicon Linux, and a new test row runs the BERT tests.