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

broadcast

def broadcast[dtype: DType, in_layout: TensorLayout, in_origin: Origin[mut=in_origin.mut], //, ngpus: Int, pdl_level: PDLLevel = PDLLevel(), use_multimem: Bool = False](input_tensor: TileTensor[dtype, in_layout, in_origin], output_tensor: TileTensor[dtype, in_layout], rank_sigs: InlineArray[UnsafePointer[Signal, MutAnyOrigin], 8], ctx: DeviceContext, root: Int, _max_num_blocks: Optional[Int] = None)

Broadcast data from root GPU to all participating GPUs.

Parameters:

dtype (DType): Data type of the tensor elements.
in_layout (TensorLayout): Layout of the input TileTensor.
in_origin (Origin[mut=in_origin.mut]): Origin of the input TileTensor.
ngpus (Int): Number of GPUs participating in the broadcast.
pdl_level (PDLLevel): Controls PDL behavior for P2P kernels.
use_multimem (Bool): Whether to use multimem mode for improved performance.

Args:

input_tensor (TileTensor[dtype, in_layout, in_origin]): Input tensor from root GPU as a TileTensor.
output_tensor (TileTensor[dtype, in_layout]): Output tensor for THIS GPU as a TileTensor.
rank_sigs (InlineArray[UnsafePointer[Signal, MutAnyOrigin], 8]): Per-GPU Signal pointers.
ctx (DeviceContext): Device context for THIS GPU.
root (Int): Root GPU rank (source of broadcast data).
_max_num_blocks (Optional[Int]): Optional grid limit.