For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /get-started.md).

Mojo trait

Reducer

Parallel scalar reducer over the closed set of associative ops.

Reduces a scalar across all participants in the current parallel scope; every participant receives the reduced value (like MPI_Allreduce).

Concrete impls provide hardware-fast sum / max / min over a closed set of dtypes, plus a generic fallback that combines any ReduceOp via its join (slower; dtype-agnostic).

Implemented traits

AnyType

Required methods

sum

def sum[dtype: DType](self, val: Scalar[dtype]) -> Scalar[dtype]

Returns the sum of val across all participants.

Parameters:

• ​dtype (DType): The scalar dtype.

Args:

• ​val (Scalar[dtype]): The per-participant value.

Returns:

Scalar[dtype]: The reduced sum, broadcast to every participant.

max

def max[dtype: DType](self, val: Scalar[dtype]) -> Scalar[dtype]

Returns the maximum of val across all participants.

Parameters:

• ​dtype (DType): The scalar dtype.

Args:

• ​val (Scalar[dtype]): The per-participant value.

Returns:

Scalar[dtype]: The reduced maximum, broadcast to every participant.

min

def min[dtype: DType](self, val: Scalar[dtype]) -> Scalar[dtype]

Returns the minimum of val across all participants.

Parameters:

• ​dtype (DType): The scalar dtype.

Args:

• ​val (Scalar[dtype]): The per-participant value.

Returns:

Scalar[dtype]: The reduced minimum, broadcast to every participant.

generic

def generic[S: ReduceOp](self, mut state: S)

Combines state across all participants using only join — the multi-field path for monoids whose state isn't decomposable into scalar sum/max/min (Welford, ArgMax, ArgMin, …).

Impls pick by state size: small states (a few uint32 words) use a register-only warp-shuffle butterfly, larger states a shmem tree — both valid since ReduceOp is TrivialRegisterPassable. Monoids that can decompose (ReduceSum, ReduceMax, OnlineLogSumExp, …) instead call reducer.sum/max/min directly, which compile to one hardware redux.sync per field and are faster.

Parameters:

• ​S (ReduceOp): The monoid type being combined.

Args:

• ​state (S): Per-participant state; on return, holds the combined value on every participant.