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_nd_generator

def scatter_nd_generator[output_type: DType, indices_type: DType, //, oob_index_strategy: ScatterOobIndexStrategy = ScatterOobIndexStrategy.UNDEFINED, target: StringSlice[ImmStaticOrigin] = StringSlice("cpu"), reduce_fn: OptionalReg[def[dtype: DType, width: SIMDLength](SIMD[dtype, width], SIMD[dtype, width]) thin -> SIMD[dtype, width]] = None, *, _trace_description: StringSlice[ImmStaticOrigin] = StringSlice("scatter_nd")](data: TileTensor[output_type, Storage=data.Storage, linear_idx_type=data.linear_idx_type], indices: TileTensor[indices_type, Storage=indices.Storage, linear_idx_type=indices.linear_idx_type], updates: TileTensor[output_type, Storage=updates.Storage, linear_idx_type=updates.linear_idx_type], output: TileTensor[output_type, Storage=output.Storage, linear_idx_type=output.linear_idx_type], context: DeviceContext)

Implements ONNX ScatterND operation as defined in https://github.com/onnx/onnx/blob/main/docs/Operators.md#ScatterND.

Parameters:

output_type (DType): Type of data, updates, and output tensors.
indices_type (DType): Type of the indices tensor.
oob_index_strategy (ScatterOobIndexStrategy): Strategy to handle out of bounds indices.
target (StringSlice[ImmStaticOrigin]): Target cpu or cuda.
reduce_fn (OptionalReg[def[dtype: DType, width: SIMDLength](SIMD[dtype, width], SIMD[dtype, width]) thin -> SIMD[dtype, width]]): Reduction function to apply: none (default), add, mul, max, min. When set, every update is folded in atomically, in unspecified order — the atomic runs on all updates, not only detected duplicates, since duplicate index vectors can only be known at runtime. Without a reduce_fn, duplicates leave an unspecified winner instead.
_trace_description (StringSlice[ImmStaticOrigin]): A description of the function, used for profiling and tracing.

Args:

data (TileTensor[output_type, Storage=data.Storage, linear_idx_type=data.linear_idx_type]): Tensor of rank data_rank >= 1.
indices (TileTensor[indices_type, Storage=indices.Storage, linear_idx_type=indices.linear_idx_type]): Tensor of rank indices_rank containing indices for the scatter operation.
updates (TileTensor[output_type, Storage=updates.Storage, linear_idx_type=updates.linear_idx_type]): Tensor containing values to update output tensor based on indices tensor.
output (TileTensor[output_type, Storage=output.Storage, linear_idx_type=output.linear_idx_type]): Tensor of rank data_rank, shaped the same as data tensor.
context (DeviceContext): Pointer to DeviceContext.