For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /max/get-started.md).

Mojo function

convert_e4m3fn_to_e4m3fnuz

def convert_e4m3fn_to_e4m3fnuz(input_buffer: TileTensor[DType.float8_e4m3fn, address_space=input_buffer.address_space, linear_idx_type=input_buffer.linear_idx_type, element_size=input_buffer.element_size], output_buffer: TileTensor[DType.float8_e4m3fnuz, address_space=output_buffer.address_space, linear_idx_type=output_buffer.linear_idx_type, element_size=output_buffer.element_size], context: DeviceContext)

Convert E4M3FN weights to E4M3FNUZ format for AMD GPU compatibility.

This conversion handles the key differences between E4M3FN and E4M3FNUZ:

The bit pattern 10000000 (-128) represents zero in E4M3FN but NaN in E4M3FNUZ

Args:

input_buffer (TileTensor[DType.float8_e4m3fn, address_space=input_buffer.address_space, linear_idx_type=input_buffer.linear_idx_type, element_size=input_buffer.element_size]): Input tensor in E4M3FN format.
output_buffer (TileTensor[DType.float8_e4m3fnuz, address_space=output_buffer.address_space, linear_idx_type=output_buffer.linear_idx_type, element_size=output_buffer.element_size]): Output tensor to store E4M3FNUZ format.
context (DeviceContext): Device context for kernel execution.