flashinfer.fused_moe.ExpertConfig

class flashinfer.fused_moe.ExpertConfig(intermediate_size: int, local_expert_offset: int = 0, local_num_experts: int | None = None, num_fused_shared_experts: int = 0)

Expert geometry.

Parameters:

• intermediate_size (int) – Hidden dimension of the expert FFN (the N in gemm1’s MxK → MxN).

• local_expert_offset (int) – Start index for expert-parallel sharding.

• local_num_experts (int or None) – Number of experts on this rank. None → num_experts at runtime.

• num_fused_shared_experts (int) – Number of shared experts run for every token. Their rows follow the routed experts, so weights have E + S rows while routing fields remain routed-only. Cross-field constraints are checked by MoEConfig.

__init__(intermediate_size: int, local_expert_offset: int = 0, local_num_experts: int | None = None, num_fused_shared_experts: int = 0) → None

Methods

__init__(intermediate_size[, ...])

Attributes

local_expert_offset
local_num_experts
num_fused_shared_experts
intermediate_size