pytorch/fbgemm is a repository that provides optimized kernels for deep learning applications. The latest release is v1.8.0, but the provided materials do not list what changed in that version. The repository's primary language is C++. FBGEMM covers low-precision matrix multiplication and convolution for server-side inference, and quantized GEMM and fused operations are also in scope.
The repository consists of three packages: FBGEMM, FBGEMM_GPU and FBGEMM_GPU GenAI. FBGEMM is a matrix multiplication and convolution library for server inference. FBGEMM_GPU is a PyTorch GPU operator library for training and inference in recommendation systems, while FBGEMM_GPU GenAI is a PyTorch GPU operator library for generative AI applications. Low-precision operations mean computation that lowers the precision used for representation. The GenAI examples given include FP8 row-wise quantization and collective communication. The scope of this repository is therefore not limited to a single inference path; it covers server-side matrix operations together with GPU operators for recommendation systems and generative AI.
However, the provided materials alone do not make it possible to confirm the changes in v1.8.0 or the performance figures for each operation. The license entry is also NOASSERTION, so no specific license name can be confirmed. The scope of the packages is visible, but that information has to be checked separately before judging performance or usage conditions.