Repository navigation
fastembed 7.7.1: reliability release
This release hardens every model family. It has no new API and no API change. Models that gave correct vectors in 7.7.0 give the same vectors.
Highlights
🚦 Parallel model loads work on an empty cache
Start as many loads of one model as you want. One of them downloads, and the others wait and then read the cache. In 7.7.0, the others got a LockAcquisition error after five seconds (#203).
| 16 loads at once, empty cache | 7.7.0 | This release |
|---|---|---|
| Loads that succeed | 6 or 7 | 16 |
A load also waits through a pause of up to one minute in the download.
🧠 BertTextEmbedding gives correct vectors for RoBERTa models
RoBERTa, XLM-RoBERTa, and CamemBERT checkpoints now get the position ids of their architecture. In 7.7.0 they loaded and returned wrong vectors. The vectors now match transformers within 1.3e-6.
google-bert/bert-base-uncased and other checkpoints with gamma and beta LayerNorm names now load.
📏 Half-precision Nomic vectors have unit length
NomicV2MoeTextEmbedding with DType::F16 or DType::BF16 weights returned vectors with norms up to 1.09. The norms are now 1.
🖼️ Images fail safely
- An image with no pixels returns an error.
- A resize that needs a buffer above 512 MiB returns an error. In 7.7.0, a small file with an extreme aspect ratio could end the process. A resizer set with
with_resizehas no limit. ImageEmbedding::embedreads an image whose file extension names the wrong format.embedandembed_byteshold one decoded image at a time, instead of the whole batch.
🧮 TextEmbedding::embed uses less memory
Each batch is pooled before the next one runs, so batch_size now limits the memory in use.
More fixes
ONNX models
- Errors replace panics for: a
max_lengthbelow the special-token count, a model output with no tokens, reranker logits of an unexpected shape, and a missing ONNX Runtime library withort-load-dynamic. TextRerank::rerankandsimilarity::top_krank aNaNscore last.top_kkeeps equal scores in corpus order, and0.0and-0.0count as equal.SparseTextEmbeddingkeeps a positive weight for a small positive logit.TextEmbedding::list_supported_modelsreturns the models in table order.- A lower-case form of a special token in a text, such as
[sep], is tokenized as ordinary text. EmbeddingGemma300MQis batched bybatch_size, like the other EmbeddingGemma variants.- A user-defined
preprocessor_config.jsongets the resize filter that itsresamplevalue names. ExternalInitializerFileis exported from the crate root.
Qwen3 and Qwen3-VL
Qwen/Qwen3-VL-Embedding-8Bloads, and so do configs written by transformers 5.Qwen3VLEmbedding::embed_imagesruns withDType::BF16weights on the CPU.- The Qwen3-VL patch merger uses exact GELU, as the reference model does.
Qwen3TextEmbedding::newgives correct vectors for a tokenizer that pads on the right.- The prompt of an image at the token limit is no longer cut.
Hugging Face Hub
- A model loads in a process with no home directory.
- With
HF_HOMEset, the token saved byhuggingface-cli loginis used. - An empty
HF_ENDPOINTmeans the default endpoint. - Download and parse errors name their cause.
Behavior to know before you upgrade
- RoBERTa-family vectors change. If you stored vectors from
BertTextEmbeddingwith a RoBERTa, XLM-RoBERTa, or CamemBERT checkpoint, embed those texts again. - Half-precision Nomic vectors change length. Their direction stays the same within rounding.
- The default resizer has a 512 MiB limit for each buffer. An image must be about 150,000 pixels wide, or have an aspect ratio above about 3,500 to 1, to reach it.
get_quantization_modereturnsStaticforEmbeddingGemma300MQ.
CI/CD
- Releases publish to crates.io with trusted publishing.
Full Changelog: v7.7.0...v7.7.1