nurysso/eulix is a project that builds structures and semantic indices in order to handle code on the scale of millions of lines locally. The October 25, 2026 commit d19a730 moved the embedder's batch size computation code into a shared base class and fixed a Torch pipeline crash.
Batch size is the number of sentences fed into the model and processed at once. This repository groups sentences by length interval (bucket) and processes them in groups, reducing the batch size as the interval gets longer so that attention memory usage does not grow. The computation linearly reduced the batch size for intervals exceeding a reference length of 512 tokens and set the minimum to 1. This code existed only inside the full embedder wrapper (ONNX), but this commit moved it into the newly created eulix-embed/embedders/mixins.py as BucketBatchMixin so that the Torch-side wrapper inherits it as well. The EmbeddingGenerator protocol of the shared pipeline specified a batch_size_for item to align the contract of the two implementations, and _batch_size_for was changed to call this method instead of a fixed batch_size. The misspelled _bacth_size_for method that remained in the two pipeline files was deleted.
The commit message states that this change fixes the Torch pipeline crash. The scale of the change is 20 lines added and 26 lines deleted across 7 files, combining a behavioral fix with the cleanup of duplicated code. However, the comparative figures for the impact of the batch reduction rule on accuracy or throughput are not included in this commit. Depending on how often long documents are processed, it also cannot be known from this commit alone whether the 512-token reference setting is appropriate, nor when the license and release schedule will be.