All Tools
F
Dev ToolsFreeOpen Source
FBGEMM
Low-precision GEMM and convolution kernels for server inference
BSD-3-Clause
ABOUT
Server inference at INT8 and mixed precision needs kernels that stay accurate at small batches. FBGEMM supplies quantized GEMM, convolution, and fused ops so PyTorch recommendation and GenAI stacks avoid rewriting SIMD kernels for every model.
INTEGRATION GUIDE
1. Accelerate quantized recommendation models on CPU servers
2. Fuse low-precision GEMM with bandwidth-bound ops in PyTorch
3. Train and serve recsys operators via FBGEMM-GPU
4. Run FP8 and collective kernels for generative inference
TAGS
c++inferencequantizationgemmpytorchrecommendationperformance