class CuGraphGATConv(in_channels: int, out_channels: int, heads: int = 1, concat: bool = True, negative_slope: float = 0.2, bias: bool = True)[source]

Bases: CuGraphModule

The graph attentional operator from the “Graph Attention Networks” paper.

CuGraphGATConv is an optimized version of GATConv based on the cugraph-ops package that fuses message passing computation for accelerated execution and lower memory footprint.

forward(x: Tensor, edge_index: EdgeIndex, max_num_neighbors: Optional[int] = None) Tensor[source]

Runs the forward pass of the module.

  • x (torch.Tensor) – The node features.

  • edge_index (EdgeIndex) – The edge indices.

  • max_num_neighbors (int, optional) – The maximum number of neighbors of a target node. It is only effective when operating in a bipartite graph. When not given, the value will be computed on-the-fly, leading to slightly worse performance. (default: None)


Resets all learnable parameters of the module.