flashinfer.comm.trtllm_moe_finalize_allreduce_fusion

flashinfer.comm.trtllm_moe_finalize_allreduce_fusion(allreduce_in: Tensor, residual_in: Tensor, norm_weight: Tensor, expanded_idx_to_permuted_idx: Tensor, norm_out: Tensor, residual_out: Tensor, workspace_ptrs: Tensor, launch_with_pdl: bool, world_rank: int, world_size: int, eps: float, shared_expert_output: Tensor | None, expert_scale_factor: Tensor | None) → None

参数: - allreduce_in: 输入张量。 [token_num, top_k, hidden_dim] - residual_in: 残差输入张量。 [token_num, hidden_dim] - norm_weight: 归一化权重张量。 [hidden_dim] - expanded_idx_to_permuted_idx: 扩展索引到排列索引张量。 [token_num, top_k] - norm_out: 归一化输出张量。 [token_num, hidden_dim] - residual_out: 残差输出张量。 [token_num, hidden_dim] - workspace_ptrs: 工作区指针。 - launch_with_pdl: 是否使用 pdl 启动。 - world_rank: 当前进程的 rank。 - world_size: 进程组的大小。 - eps: epsilon 值。 - shared_expert_output: 共享专家输出张量。 [token_num, hidden_dim] - expert_scale_factor: 专家缩放因子张量。 [token_num, top_k]