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

tpool_patch_merger

def tpool_patch_merger[dtype: DType, output_layout: TensorLayout, x_layout: TensorLayout, bounds_layout: TensorLayout](output: TileTensor[dtype, output_layout, MutAnyOrigin], x: TileTensor[dtype, x_layout, ImmutAnyOrigin], bounds: TileTensor[DType.int64, bounds_layout, ImmutAnyOrigin], kH: Int, kW: Int, max_h: Int, max_w: Int, ctx: DeviceContext)

Temporal pooling patch merger entry point.

Parameters:

dtype (DType): Element type of the input and output tensors.
output_layout (TensorLayout): Memory layout of the output tensor.
x_layout (TensorLayout): Memory layout of the input tensor.
bounds_layout (TensorLayout): Memory layout of the bounds tensor.

Args:

output (TileTensor[dtype, output_layout, MutAnyOrigin]): Contiguous output tensor [total_output_patches, D].
x (TileTensor[dtype, x_layout, ImmutAnyOrigin]): Input tensor [n_tokens, D].
bounds (TileTensor[DType.int64, bounds_layout, ImmutAnyOrigin]): Grid dimensions tensor [n_vids, 3] with (T, H, W) per video.
kH (Int): Merge kernel height.
kW (Int): Merge kernel width.
max_h (Int): Maximum H across all videos (for grid sizing).
max_w (Int): Maximum W across all videos (for grid sizing).
ctx (DeviceContext): Device context.