flashinfer.moe_ep.MoEEpDispatchResult

class flashinfer.moe_ep.MoEEpDispatchResult(hidden_states: torch.Tensor, topk_ids: torch.Tensor, topk_weights: torch.Tensor | None, tokens_per_rank: int, hidden_states_scale: torch.Tensor | None = None, eplb_gathered_stats: torch.Tensor | None = None)

Bases: object

Rank-major receive buffers produced by MoEEpCommunication.dispatch().

All tensors have ep_size * tokens_per_rank rows (see the module docstring for the row contract). They may be views into backend-owned buffers that stay valid until the matching combine().