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

quantize_tensor_dynamic_scaled_fp8

def quantize_tensor_dynamic_scaled_fp8[out_dtype: DType, in_dtype: DType, scales_dtype: DType, InputFnType: def[width: Int, alignment: Int](row: Int, col: Int) -> SIMD[in_dtype, width] & ImplicitlyCopyable & RegisterPassable, //, group_size_or_per_token: Int, num_cols: Int, pdl_level: PDLLevel = PDLLevel.ON](input_fn: InputFnType, scaled_output: TileTensor[out_dtype, Storage=scaled_output.Storage, address_space=scaled_output.address_space, linear_idx_type=scaled_output.linear_idx_type], scales: TileTensor[scales_dtype, Storage=scales.Storage, address_space=scales.address_space, linear_idx_type=scales.linear_idx_type], scale_ub: Float32, ctx: DeviceContext, num_rows: Int) where (eq InputFnType.in_dtype, in_dtype)

TileTensor primary implementation of dynamic scaled FP8 quantization.

Parameters:

out_dtype (DType): FP8 dtype of the quantized output; must be float8_e4m3fn.
in_dtype (DType): Dtype of the input values loaded by input_fn.
scales_dtype (DType): Dtype of the per-group scale factors written to scales; one of bfloat16, float16, or float32.
InputFnType (def[width: Int, alignment: Int](row: Int, col: Int) -> SIMD[in_dtype, width] & ImplicitlyCopyable & RegisterPassable): Compile-time callable type that loads SIMD[in_dtype, width] tiles for a given (row, col) (inferred).
group_size_or_per_token (Int): Number of columns per quantization group; -1 selects per-tensor scaling over all num_cols.
num_cols (Int): Number of columns in the input; the hidden dimension size.
pdl_level (PDLLevel): Programmatic dependent launch level for the kernels (defaults to PDLLevel.ON).

Args:

input_fn (InputFnType): Callable that loads input tiles for a given (row, col) pair.
scaled_output (TileTensor[out_dtype, Storage=scaled_output.Storage, address_space=scaled_output.address_space, linear_idx_type=scaled_output.linear_idx_type]): Rank-2 output TileTensor of FP8-quantized values.
scales (TileTensor[scales_dtype, Storage=scales.Storage, address_space=scales.address_space, linear_idx_type=scales.linear_idx_type]): Rank-2 output TileTensor of per-group scale factors.
scale_ub (Float32): Upper bound for the dynamic scale factor.
ctx (DeviceContext): Device context used to enqueue the kernels.
num_rows (Int): Number of rows to quantize; 0 skips the launch.