flashinfer.fused_moe.TrtllmGenRoutingResult

class flashinfer.fused_moe.TrtllmGenRoutingResult(topk_ids: Tensor, topk_weights: Tensor, total_num_padded_tokens: Tensor, expanded_idx_to_permuted_idx: Tensor, permuted_idx_to_token_idx: Tensor, cta_idx_xy_to_batch_idx: Tensor, cta_idx_xy_to_mn_limit: Tensor, num_non_exiting_ctas: Tensor)

Outputs of the trtllm-gen MoE routing stage.

topk_ids/topk_weights are the per-token expert selection. The remaining fields are the permutation/bookkeeping tensors the fused MoE kernels consume; entries beyond the actual padded count (see total_num_padded_tokens) are undefined.

In from-logits mode the routing kernels never emit expert ids directly — expert identity lives in the permuted layout. topk_ids is therefore reconstructed on the torch side: permuted slot p belongs to CTA tile p // tile_tokens_dim, and cta_idx_xy_to_batch_idx maps each CTA tile to its expert.

__init__()

Methods

__init__()
count(value, /)	Return number of occurrences of value.
index(value[, start, stop])	Return first index of value.

Attributes

cta_idx_xy_to_batch_idx	Alias for field number 5
cta_idx_xy_to_mn_limit	Alias for field number 6
expanded_idx_to_permuted_idx	Alias for field number 3
num_non_exiting_ctas	Alias for field number 7
permuted_idx_to_token_idx	Alias for field number 4
topk_ids	Alias for field number 0
topk_weights	Alias for field number 1
total_num_padded_tokens	Alias for field number 2