vllm.model_executor.layers.fused_moe.fused_globalize_align_block_size
¶
Functions:
-
fused_globalize_align_block_size–Fuse the DeepEP-v2 INDEXED decode metadata chain into one launch.
fused_globalize_align_block_size(recv_topk_idx, psum_recv_per_rank, rank_expert_offset, global_num_experts, local_num_experts, block_size, return_scatter_idx=False)
¶
Fuse the DeepEP-v2 INDEXED decode metadata chain into one launch.
Replaces _globalize_recv_topk_idx + moe_align_block_size +
count_and_sort_expert_tokens. recv_topk_idx holds LOCAL expert ids
and is globalized in place; valid-token count is read on-device from
psum_recv_per_rank[-1] (CUDA-graph safe).
Returns the globalized recv_topk_idx and the moe_align_block_size
outputs (sorted_ids, expert_ids, num_tokens_post_pad). When requested,
the final output is a masked identity scatter index for valid routes.