bf16_to_fp8

Function bf16_to_fp8 

Source
pub fn bf16_to_fp8(
    gpu: &dyn GpuBackend,
    kernel: KernelHandle,
    src: DevicePtr,
    dst: DevicePtr,
    total_elements: u32,
    stream: u64,
) -> Result<()>
Expand description

Convert BF16 activations to FP8 E4M3 for FP8×FP8 GEMM.

Grid: (ceil(total_elements/2 / 256), 1, 1) Block: (256, 1, 1)