flashrt/fp8-gemm

FlashRT native CUDA FP8 GEMV/GEMM kernels for low-latency transformer and diffuser linear layers.

This package targets Blackwell sm_120a FP8 MMA.

Functions

• fp8_linear_bf16(input, weight, alpha=1.0, out=None, variant=0)
• fp8_linear_residual_bf16(input, weight, residual, alpha=1.0, variant=0)
• select_fp8_linear_tile(m, n, k, variant=0)

See the repository README for shape contracts, validation status, and examples.