facebookresearch/TensorComprehensions

Health v2 · maintenance only

Flags: archived

Adoption not part of the score

1767 stars · 211 forks observed · 2026-08-28

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

Tensor Comprehensions is a C++ library with a Python API that provides a domain-specific language for expressing machine learning workloads and automatically synthesizes high-performance GPU kernels via JIT compilation. It integrates with PyTorch and Caffe2 and includes an autotuner that searches for optimal kernel mapping options.

Use cases

• generate optimized GPU kernels for tensor operations automatically
• JIT-compile custom ML kernels for specific tensor sizes
• autotune tensordot or convolution kernels on CUDA GPUs
• express ML math in a compact DSL instead of writing CUDA by hand
• integrate custom kernels into PyTorch or Caffe2 models

When to choose

• you need high-performance custom tensor kernels without hand-writing CUDA
• you want JIT compilation and autotuning for specific tensor shapes
• you work in PyTorch or Caffe2 and need framework-agnostic kernel synthesis

When to avoid

• you need actively maintained tooling - the project has seen little recent development
• you target CPUs only or non-CUDA hardware
• you want a general-purpose deep learning framework rather than a kernel compiler

Facets

library · maturity maintenance

compiler machine-learning gpu-computing benchmarking machine-learning deep-learning gpu-computing compilers cpp python jit-compilation kernel-synthesis autotuning halide pytorch-integration caffe2 domain-specific-language cuda gpu linux macos

Member repositories

For agents

markdown · JSON · MCP: product_card(name="facebookresearch/TensorComprehensions")