Make ML Training
4× Faster
A next-generation data file format that eliminates the data loading bottleneck. Drop-in replacement for PyTorch DataLoader. One line to integrate.
4×
Faster End-to-End Training
20×
Faster I/O (ImageNet)
26×
Faster I/O (Flickr30k)
42.6
dB PSNR (Near-Lossless)
Why Kuattree?
Your GPU is idle 40-80% of training time waiting for data. We fix that.
Drop-In Replacement
Replace your DataLoader import with one line. Your training loop, augmentations, and hyperparameters stay exactly the same.
Near-Lossless Compression
0.977 SSIM, 42.6 dB PSNR. Quaternion-structured coding preserves perceptual quality. Minimal RAM footprint.
Cross-Modal Search
Semantic search across compressed datasets. Query by text, find matching images — without decompressing anything.
Single-File Datasets
Pack an entire dataset into one .kt file. No more scattered JPEGs, no filesystem overhead, no slow random reads.
Streaming Decode
Progressive decoding feeds batches to GPU as data streams from disk. Zero idle time between batches.
Rust + Python
Core engine in Rust for zero-copy performance. Python bindings via PyO3. pip install and go.
One-Line Integration
Swap your import. Your training code stays exactly the same. Kuattree handles compression, streaming, and GPU prefetching behind one familiar interface.
View on GitHub →# Before from torch.utils.data import DataLoader # After — one line change from kuattree import DataLoader # Everything else stays the same loader = DataLoader( dataset="imagenet.kt", batch_size=256, num_workers=8 ) for batch in loader: outputs = model(batch) loss.backward() optimizer.step()
Benchmarks
Tested across standard ML datasets. Numbers speak for themselves.
Stop Waiting for Data
Your GPU is already fast enough. Your data loading isn't. Fix it in one line.