pub fn bf16_absmax(
gpu: &dyn GpuBackend,
kernel: KernelHandle,
data: DevicePtr,
out_max: DevicePtr,
n_elems: u32,
stream: u64,
) -> Result<()>Expand description
Compute max absolute value of a BF16 buffer into a device-side f32.
Used for FP8 KV cache online scale calibration: accumulates max |K| and max |V| during warmup tokens. The output f32 is updated via atomicMax, so the caller must initialize it to 0.0 before the first call.
Kernel: bf16_absmax(data, out_max, n_elems)
Grid: (ceil(n_elems / (256*2)), 1, 1) Block: (256, 1, 1)