flashinfer.comm.vllm_register_graph_buffers

flashinfer.comm.vllm_register_graph_buffers(fa: int, handles: List[List[int]], offsets: List[List[int]]) → None

Register graph-capture buffers across the all-reduce world.

Parameters:

• fa (int) – Handle returned by init_custom_ar().

• handles (list[list[int]]) – Per-rank IPC handles published via get_graph_buffer_ipc_meta().

• offsets (list[list[int]]) – Per-rank offsets matching handles.