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

gather_reduce

def gather_reduce[dtype: DType, gather_axis: Int, reduce_axis: Int, simd_width: Int, reduce_fn: def[dtype: DType, width: SIMDLength](SIMD[dtype, width], SIMD[dtype, width]) thin -> SIMD[dtype, width]](output: TileTensor[dtype, Storage=output.Storage, address_space=output.address_space, linear_idx_type=output.linear_idx_type], input: TileTensor[dtype, Storage=input.Storage, address_space=input.address_space, linear_idx_type=input.linear_idx_type], indices: TileTensor[DType.int32, Storage=indices.Storage, address_space=indices.address_space, linear_idx_type=indices.linear_idx_type], reduce_init: Scalar[dtype], ctx: Optional[DeviceContext] = None)

Computes output[i, j, k] = input[indices[i, j], k] and simultaneously reduces the output across axis 1 to produce output[i, k].

The motivating use-case for this is multi-hot embeddings in recommender models. This provides similar functionality to Torch's EmbeddingBag layer. In that context, i is the batch dimension, j is the multi-hot dimension, and k is the embedding dimension.

Parameters:

dtype (DType): Element type of input, output, and reduce_init.
gather_axis (Int): Axis of input to gather along (must be 0).
reduce_axis (Int): Axis of the gathered result to reduce across (must be 1).
simd_width (Int): SIMD vector width used for tiling the embedding dimension.
reduce_fn (def[dtype: DType, width: SIMDLength](SIMD[dtype, width], SIMD[dtype, width]) thin -> SIMD[dtype, width]): Binary reduction function combining accumulated and gathered values into the output.

Args:

output (TileTensor[dtype, Storage=output.Storage, address_space=output.address_space, linear_idx_type=output.linear_idx_type]): Output tensor holding the reduced embeddings, shape [batch, embedding_dim].
input (TileTensor[dtype, Storage=input.Storage, address_space=input.address_space, linear_idx_type=input.linear_idx_type]): Source embedding table, shape [num_embeddings, embedding_dim].
indices (TileTensor[DType.int32, Storage=indices.Storage, address_space=indices.address_space, linear_idx_type=indices.linear_idx_type]): Multi-hot indices, shape [batch, multi_hot], int32.
reduce_init (Scalar[dtype]): Initial accumulator value for the reduction.
ctx (Optional[DeviceContext]): Optional device context for parallel execution (defaults to None).