pyshmem

pyshmem is a Python library for low-latency, cross-process exchange of fixed shape NumPy arrays and CUDA-backed PyTorch tensors. CPU and GPU streams share the same small create / open / write / read API.

A pyshmem stream is a capacity-one latest-value exchange, not a queue. Each write replaces the previous payload; readers get a consistent snapshot and can inspect missed-publication counters when producers run faster than consumers. It supports Linux and macOS; CUDA IPC is Linux-only. Windows is unsupported.

Documentation · API reference · Source · Issues · Changelog

Quick start

Create and publish from one process:

import numpy as np
import pyshmem

writer = pyshmem.create("frames", shape=(480, 640), dtype=np.float32)
writer.write(np.ones((480, 640), dtype=np.float32))

Attach and read from another:

import pyshmem

reader = pyshmem.open("frames")
frame = reader.read()                    # latest consistent snapshot
next_frame = reader.read_new(timeout=1) # wait for the next publication
print(reader.missed_writes)

reader.close()

Use the same API for CUDA by creating with gpu_device="cuda:0"; reads return a CUDA torch.Tensor when attached to the GPU. Destroy persistent streams with writer.unlink() when they are no longer needed.

See the quick start, usage guide, and GPU guide for lifecycle, locking, asyncio, CPU mirrors, failure recovery, and CLI examples.

Installation

CPU support:

pip install pyshmem

CUDA support (installs the PyTorch dependency):

pip install "pyshmem[gpu]"

See the installation guide for supported Python versions, platform requirements, development setup, and installation verification.

Performance

The repository includes both single-process microbenchmarks and a calibrated, spawned-process request/acknowledgement benchmark with an unsafe raw shared memory lower-bound comparison.

On the primary Linux development machine (Python 3.12, NumPy 2.2.6, PyTorch 2.10, RTX 5090), the spawned-process 64 KiB benchmark measured:

The raw baseline omits pyshmem's locking, metadata validation, discovery, and consistent snapshots. The GPU row is a separate spawned process mapping the producer's CUDA tensor over torch IPC and reading a consistent device snapshot each round trip. Results are machine-specific; run the harness on the target deployment host (add --gpu for the CUDA baseline):

python benchmarks/benchmark_ipc.py \
  --payload-bytes 65536 --minimum-seconds 1 --repeats 5 --gpu

See the benchmark documentation and the versioned result for methodology, CPU/GPU measurements, and limitations.

License and contact

pyshmem is licensed under GPL-3.0-only. Applications that distribute pyshmem or derivative work should evaluate the GPL's obligations.

Use GitHub issues for bugs and feature requests. See CONTRIBUTING.md for development, SUPPORT.md for compatibility and support scope, and SECURITY.md for private vulnerability reporting.