openai/blocksparse

Health v2 · maintenance only

Flags: no_releases archived

Adoption not part of the score

1068 stars · 198 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-30, confidence not recorded

A Python package of TensorFlow ops and custom CUDA GPU kernels for block-sparse matrix multiplication and convolution, plus related ops like edge bias and sparse weight/layer norm. Developed by OpenAI to accelerate sparse neural network training on Nvidia GPUs.

Use cases

• speed up training of sparse neural networks on gpu
• block-sparse matrix multiplication in tensorflow
• train large language models with sparse weights
• block-sparse convolution kernels for cuda
• reduce gpu compute cost of dense matmul via sparsity

When to choose

• you use TensorFlow 1.x on Linux with an Nvidia Pascal/Maxwell GPU and want block-sparse ops
• you need fast custom CUDA kernels for sparse matmul or convolution
• you're reproducing OpenAI's block-sparse GPU kernel research

When to avoid

• you use modern frameworks like PyTorch or TensorFlow 2.x
• you need cross-platform or Windows/macOS support
• you want a maintained library - it targets old CUDA/TensorFlow versions and sees little activity

Facets

library · maturity maintenance

machine-learning gpu-computing deep-learning machine-learning deep-learning gpu-computing python cuda-kernels block-sparse sparse-matmul tensorflow-ops sparse-convolution linux gpu

Member repositories

For agents

markdown · JSON · MCP: product_card(name="openai/blocksparse")