W8A8 Block FP8 Kernel Configurations

This directory contains optimized kernel configurations for the W8A8 block FP8 matrix multiplication kernel.

Configuration File Format

Configuration files are named using the following pattern:

N={N},K={K},device_name={DEVICE_NAME},dtype=fp8_w8a8,block_shape=[{BLOCK_N},{BLOCK_K}].json

Where:

• N: Output dimension (number of columns in weight matrix)
• K: Input dimension (number of columns in activation matrix)
• DEVICE_NAME: GPU device name with spaces replaced by underscores (e.g., NVIDIA_H100_80GB_HBM3)
• BLOCK_N, BLOCK_K: Block quantization granularity (typically [128,128])