For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /get-started.md).
Mojo struct
Struct_grouped_quantize_nvfp4_row_scaled
struct Struct_grouped_quantize_nvfp4_row_scaled
Registers the mo.grouped.quantize.nvfp4.row_scaled graph op with the graph compiler.
Quantizes grouped BF16 rows to NVFP4 with a global scale per row, taken
from the row's amax inside the kernel, and returns the scales for the
grouped matmul's a_row_scales. The result is bit-identical to the
chain of a row amax, two BF16 roundings and
mo.grouped.quantize.dynamic.block.scaled with a tensor-wise scale of 1.
apply_relu2 quantizes relu(x) * relu(x) in place of x, with
relu from nn.activations.
Implemented traitsโ
Methodsโ
executeโ
static def execute[out_dtype: DType, scales_type: DType, in_dtype: DType, //, scales_rank: Int, apply_relu2: Int, target: StringSpan[ImmStaticOrigin]](output: ManagedTensorSlice[IOSpec[_, _].Output, static_spec=output.static_spec], scales: ManagedTensorSlice[IOSpec[_, _].Output, static_spec=scales.static_spec], row_scales: ManagedTensorSlice[IOSpec[_, _].Output, static_spec=row_scales.static_spec], input: ManagedTensorSlice[IOSpec[_, _].Input, static_spec=input.static_spec], row_offsets: ManagedTensorSlice[IOSpec[_, _].Input, static_spec=row_offsets.static_spec], scales_offsets: ManagedTensorSlice[IOSpec[_, _].Input, static_spec=scales_offsets.static_spec], context: DeviceContextBase)