SharedMoEWeights Struct Reference

Contains the globally shared, read-only weights for the Factorized MoE network. More...

#include <MoECacheModel.hpp>

Public Member Functions
void	init_architecture (int branchConvLayers)

Public Attributes
std::array< Branch, 12 >	branches {}
std::array< float,(size_t) 12 *NET_MIXER_OUT *NET_BRANCH_DIM >	mixerWBr {}
std::array< float,(size_t) NET_BYPASS *NET_MIXER_OUT >	mixerWBp {}
std::array< float, NET_MIXER_OUT >	mixerB {}
std::array< float,(size_t) NET_MIXER_OUT *NET_GLOBALS >	globalW {}
std::array< float, NET_MIXER_OUT >	globalB {}
std::array< float,(size_t) NET_EXPERTS *NET_GLOBALS >	gateW {}
std::array< float, NET_EXPERTS >	gateB {}
std::array< Expert, NET_EXPERTS >	experts {}

Static Public Attributes
static constexpr int	bd = NET_BRANCH_DIM
static constexpr int	nf = NET_MIXER_OUT
static constexpr int	nBypass = NET_BYPASS
static constexpr int	nGlobals = NET_GLOBALS
static constexpr int	nExperts = NET_EXPERTS
static constexpr int	ebo = NET_EXPERT_BOTTLENECK
static constexpr int	eh = NET_EXPERT_HIDDEN

Detailed Description

Contains the globally shared, read-only weights for the Factorized MoE network.

This struct stores the pre-trained weights for all convolution branches, the mixer layer, the router gate, and the individual expert networks. In the engine, these weights are loaded once into memory and shared across all search threads.

Note

Includes forward(), a full-batch non-incremental evaluation function strictly used by benchmark tools. The engine does not use forward().

Member Function Documentation

init_architecture()

void SharedMoEWeights::init_architecture ( int branchConvLayers )

inline

Here is the call graph for this function:

Here is the caller graph for this function:

Member Data Documentation

bd

int SharedMoEWeights::bd = NET_BRANCH_DIM

staticconstexpr

branches

std::array<Branch, 12> SharedMoEWeights::branches {}

ebo

int SharedMoEWeights::ebo = NET_EXPERT_BOTTLENECK

staticconstexpr

eh

int SharedMoEWeights::eh = NET_EXPERT_HIDDEN

staticconstexpr

experts

std::array<Expert, NET_EXPERTS> SharedMoEWeights::experts {}

gateB

std::array<float, NET_EXPERTS> SharedMoEWeights::gateB {}

gateW

std::array< float, (size_t)NET_EXPERTS * NET_GLOBALS> SharedMoEWeights::gateW {}

globalB

std::array<float, NET_MIXER_OUT> SharedMoEWeights::globalB {}

globalW

std::array<float, (size_t)NET_MIXER_OUT * NET_GLOBALS> SharedMoEWeights::globalW {}

mixerB

std::array<float, NET_MIXER_OUT> SharedMoEWeights::mixerB {}

mixerWBp

std::array<float, (size_t)NET_BYPASS * NET_MIXER_OUT> SharedMoEWeights::mixerWBp {}

mixerWBr

std::array<float, (size_t)12 * NET_MIXER_OUT * NET_BRANCH_DIM> SharedMoEWeights::mixerWBr {}

nBypass

int SharedMoEWeights::nBypass = NET_BYPASS

staticconstexpr

nExperts

int SharedMoEWeights::nExperts = NET_EXPERTS

staticconstexpr

nf

int SharedMoEWeights::nf = NET_MIXER_OUT

staticconstexpr

nGlobals

int SharedMoEWeights::nGlobals = NET_GLOBALS

staticconstexpr

---

The documentation for this struct was generated from the following file:

• engine/src/eval/MoECacheModel.hpp