flashinfer.fused_moe.AlphaMoERoutePlan

class flashinfer.fused_moe.AlphaMoERoutePlan(topk_weights: torch.Tensor, topk_ids: torch.Tensor, sorted_token_ids: torch.Tensor, expert_ids: torch.Tensor, num_tokens_post_padded: torch.Tensor, expert_counts: torch.Tensor, expert_offsets: torch.Tensor, expert_scatter_offsets: torch.Tensor)

Device-resident routing outputs and reusable private workspace.

The first five fields are the route-plan ABI consumed by AlphaMoE compute kernels. expert_counts, expert_offsets, and expert_scatter_offsets are implementation-owned workspace retained in the tuple so repeated calls can avoid allocation. Consumers must use only the prefix of sorted_token_ids named by num_tokens_post_padded[0] and the corresponding prefix of expert_ids; reading that device scalar on the host is not required to launch a compatible compute kernel.

__init__()

Methods

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

Attributes

expert_counts	Alias for field number 5
expert_ids	Alias for field number 3
expert_offsets	Alias for field number 6
expert_scatter_offsets	Alias for field number 7
max_padded_pairs	Allocated number of padded token/route entries.
max_route_blocks	Allocated number of route blocks.
num_tokens_post_padded	Alias for field number 4
sorted_token_ids	Alias for field number 2
topk_ids	Alias for field number 1
topk_weights	Alias for field number 0