pytorch/FBGEMM

Health v2 · maintenance only

Flags: no_license

Adoption not part of the score

1584 stars · 777 forks observed · 2026-08-28

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

FBGEMM is a collection of highly optimized low-precision matrix multiplication and convolution kernels for server-side deep learning inference, published as FBGEMM (CPU), FBGEMM_GPU, and FBGEMM-GenAI packages. It serves as the backend for PyTorch's quantized x86 CPU operators and provides GPU operators for recommendation systems and generative AI workloads.

Use cases

• run low-precision quantized inference on x86 CPUs
• speed up matrix multiplication for small batch sizes
• quantize models with row-wise or outlier-aware quantization
• train and serve recommendation system models on GPUs
• use FP8 quantization for generative AI inference
• accelerate PyTorch quantized operators

When to choose

• you need high-performance low-precision GEMM on x86 server CPUs
• you are deploying PyTorch quantized models
• you build recommendation systems with embedding operations on GPU
• you need FP8 kernels for GenAI inference

When to avoid

• you need a high-level training framework rather than kernels
• you target ARM or other non-x86 CPUs for the CPU library
• you want plug-and-play quantization without writing C++/CUDA integration

Facets

library · maturity active

machine-learning deep-learning llm-inference gpu-computing benchmarking deep-learning machine-learning large-language-models gpu-computing performance cpp python quantization matrix-multiplication inference-kernels recommendation-systems low-precision pytorch fp8 simd gpu linux cuda

Member repositories

For agents

markdown · JSON · MCP: product_card(name="pytorch/FBGEMM")