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

scatter_set_constant

def scatter_set_constant[data_type: DType, index_type: DType, //, target: StringSlice[ImmStaticOrigin]](data: TileTensor[data_type, Storage=data.Storage, address_space=data.address_space, linear_idx_type=data.linear_idx_type], indices: TileTensor[index_type, Storage=indices.Storage, address_space=indices.address_space, linear_idx_type=indices.linear_idx_type], fill_value: Scalar[data_type], ctx: DeviceContext)

Scatter the fill_value into the data at the specified indices.

Example: Suppose we have a 3x3 matrix data initialized to zeros:

data = [[0, 0, 0], [0, 0, 0], [0, 0, 0]]

And indices is a 2D tensor with shape [2, 2]:

indices = [[0, 1], [2, 0]]

If fill_value is 5, after calling scatter_set_constant, data will be:

data = [[0, 5, 0], [0, 0, 0], [5, 0, 0]]

Parameters:

data_type (DType): Element type of data and fill_value.
index_type (DType): Integral element type of the indices tensor.
target (StringSlice[ImmStaticOrigin]): Target backend to execute on, such as "cpu" or "cuda".

Args:

data (TileTensor[data_type, Storage=data.Storage, address_space=data.address_space, linear_idx_type=data.linear_idx_type]): The data to scatter the updates into.
indices (TileTensor[index_type, Storage=indices.Storage, address_space=indices.address_space, linear_idx_type=indices.linear_idx_type]): The indices to scatter the updates into.
fill_value (Scalar[data_type]): The value to fill the data with.
ctx (DeviceContext): The device context.