streamlz

GPU LZ77 codec with CUDA and Vulkan backends. Decodes 2-3x faster than nvCOMP at equal or better compression ratio.

pip install streamlz

import streamlz

c = streamlz.Compressor()            # CUDA if present, else Vulkan
frame = c.compress(data, level=5)
back  = c.decompress(frame)
assert back == data

One Compressor handles both directions. Seven levels: L1 for speed, L5 for density, L6 adds a word-replacing transform for text, and L7 adds a price-aware parse on top of L6 for the smallest frames (several times slower to compress, about 2% slower to decode).

Backends

Both produce byte-identical frames, so a frame written by one decodes on the other.

streamlz.available_backends()          # e.g. ['cuda', 'vulkan']
c = streamlz.Compressor(backend="vulkan")

from streamlz import cuda, vulkan      # or import one directly

Neither needs a CUDA Toolkit or Vulkan SDK installed - the shared libraries ship in the wheel and only the driver is required at runtime. There is no CPU fallback: no GPU means no codec.

Device-to-device

Callers already holding GPU memory can skip the host round trip entirely, which is where the headline throughput lives:

c.compress_d2d(d_src, src_size, d_dst, dst_capacity, level=5)
c.decompress_d2d(d_frame, frame_size, d_dst, dst_capacity)

Notes

The first run on a new machine compiles the embedded GPU kernels for your card and the driver caches the result - about 140 ms cached against roughly 3.5 seconds uncached. In a container or CI runner where that cache cannot persist, point CUDA_CACHE_PATH at a writable directory that survives between runs.

Apache 2.0. Source, benchmarks and the wire-format spec: https://github.com/StreamLZ/Compressor-GPU