jacketsj

Touch screen pans and zooms on canvas by using drag and pinch

1

Useful for devices with separate active touch pens and touch screens.

Hello! Nice benchmarking repo. [ExpANN](https://github.com/jacketsj/expANN/) has been my personal dev repo for in-memory ANN experiments for quite some time now. One of the engines had enough performance to dominate all...

feat: cuda/cuvs acceleration for PQ training/assignment

2

blocked by #2816 Benchmarking ongoing on wikipedia-40M early tests indicate that cuvs pq params might need some better tuning Hardware for testing: - n1-standard-16 - nvidia T4

We currently have a pytorch-based k-means implementation for computing IVF centroids. This PR accelerates it with cuVS. This uses a tradeoff of faster iterations/less score improvement per iteration. By default,...

lance.torch.data.LanceDataset + torch.utils.data.DataLoader is slower than to_batches + frombuffer

3

I was writing some gpu code that iterates through an entire dataset, and I found that the torch integration was quite slow. I wrote up a simple to_batches-based implementation of...

Row IDs in the torch integration seem to be broken on main at the moment. Here's a repro: ```py import lance import pyarrow as pa import pyarrow.compute as pc import...

Specifically, it seems vectors respect the device type, while other types do not. I don't believe this is intended. Repro (requires cuda): ```py import torch import lance.torch.data import pyarrow as...

feat: "deterministic "parallel builds for HNSW

1

This draft PR is still WIP - there is a bug somewhere in my code. If anyone has any ideas, please help. The bug is quite robust to a bunch...

The following sequence of operations causes the shuffler to crash: 1. Create a table consisting of 1000 random 64-dimensional vectors 2. Index that table with ivfpq 3. Append 1000 64-dimensional...