MoeFlatmmKernelArgs< ScaleM, ScaleN, ExpertBias > Struct Template Reference

MoeFlatmmKernelArgs&lt; ScaleM, ScaleN, ExpertBias &gt; Struct Template Reference#

Composable Kernel: ck_tile::MoeFlatmmKernel< TilePartitioner_, FlatmmPipeline_, EpiloguePipeline_, kind, FusedActivation >::MoeFlatmmKernelArgs< ScaleM, ScaleN, ExpertBias > Struct Template Reference
Public Attributes | List of all members
ck_tile::MoeFlatmmKernel< TilePartitioner_, FlatmmPipeline_, EpiloguePipeline_, kind, FusedActivation >::MoeFlatmmKernelArgs< ScaleM, ScaleN, ExpertBias > Struct Template Reference

#include <moe_flatmm_kernel.hpp>

Public Attributes

const ck_tile::index_t * p_sorted_token_ids
const ck_tile::index_t * p_sorted_expert_ids
const ck_tile::index_t * p_max_token_id
const void * p_sorted_expert_weights
const void * a_ptr
const void * b_ptr
void * e_ptr
ck_tile::index_t NumTokens
ck_tile::index_t TopK
ck_tile::index_t M
ck_tile::index_t N
ck_tile::index_t K
ck_tile::index_t stride_A
ck_tile::index_t stride_B
ck_tile::index_t stride_C
ck_tile::index_t k_batch
ck_tile::index_t n_padded_zeros
ck_tile::index_t k_padded_zeros
ScaleM scale_m
ScaleN scale_n
ExpertBias exp_bias

Member Data Documentation

◆ a_ptr

template<typename TilePartitioner_, typename FlatmmPipeline_, typename EpiloguePipeline_, MoeFlatmmKind kind, typename FusedActivation = moe::MoeSilu>
template<class ScaleM = FlatmmScalePointer<-1>, class ScaleN = FlatmmScalePointer<-1>, class ExpertBias = FlatmmScalePointer<-1>>
const void* ck_tile::MoeFlatmmKernel< TilePartitioner_, FlatmmPipeline_, EpiloguePipeline_, kind, FusedActivation >::MoeFlatmmKernelArgs< ScaleM, ScaleN, ExpertBias >::a_ptr

◆ b_ptr

template<typename TilePartitioner_, typename FlatmmPipeline_, typename EpiloguePipeline_, MoeFlatmmKind kind, typename FusedActivation = moe::MoeSilu>
template<class ScaleM = FlatmmScalePointer<-1>, class ScaleN = FlatmmScalePointer<-1>, class ExpertBias = FlatmmScalePointer<-1>>
const void* ck_tile::MoeFlatmmKernel< TilePartitioner_, FlatmmPipeline_, EpiloguePipeline_, kind, FusedActivation >::MoeFlatmmKernelArgs< ScaleM, ScaleN, ExpertBias >::b_ptr

◆ e_ptr

template<typename TilePartitioner_, typename FlatmmPipeline_, typename EpiloguePipeline_, MoeFlatmmKind kind, typename FusedActivation = moe::MoeSilu>
template<class ScaleM = FlatmmScalePointer<-1>, class ScaleN = FlatmmScalePointer<-1>, class ExpertBias = FlatmmScalePointer<-1>>
void* ck_tile::MoeFlatmmKernel< TilePartitioner_, FlatmmPipeline_, EpiloguePipeline_, kind, FusedActivation >::MoeFlatmmKernelArgs< ScaleM, ScaleN, ExpertBias >::e_ptr

◆ exp_bias

template<typename TilePartitioner_, typename FlatmmPipeline_, typename EpiloguePipeline_, MoeFlatmmKind kind, typename FusedActivation = moe::MoeSilu>
template<class ScaleM = FlatmmScalePointer<-1>, class ScaleN = FlatmmScalePointer<-1>, class ExpertBias = FlatmmScalePointer<-1>>
ExpertBias ck_tile::MoeFlatmmKernel< TilePartitioner_, FlatmmPipeline_, EpiloguePipeline_, kind, FusedActivation >::MoeFlatmmKernelArgs< ScaleM, ScaleN, ExpertBias >::exp_bias

◆ K

template<typename TilePartitioner_, typename FlatmmPipeline_, typename EpiloguePipeline_, MoeFlatmmKind kind, typename FusedActivation = moe::MoeSilu>
template<class ScaleM = FlatmmScalePointer<-1>, class ScaleN = FlatmmScalePointer<-1>, class ExpertBias = FlatmmScalePointer<-1>>
ck_tile::index_t ck_tile::MoeFlatmmKernel< TilePartitioner_, FlatmmPipeline_, EpiloguePipeline_, kind, FusedActivation >::MoeFlatmmKernelArgs< ScaleM, ScaleN, ExpertBias >::K

◆ k_batch

template<typename TilePartitioner_, typename FlatmmPipeline_, typename EpiloguePipeline_, MoeFlatmmKind kind, typename FusedActivation = moe::MoeSilu>
template<class ScaleM = FlatmmScalePointer<-1>, class ScaleN = FlatmmScalePointer<-1>, class ExpertBias = FlatmmScalePointer<-1>>
ck_tile::index_t ck_tile::MoeFlatmmKernel< TilePartitioner_, FlatmmPipeline_, EpiloguePipeline_, kind, FusedActivation >::MoeFlatmmKernelArgs< ScaleM, ScaleN, ExpertBias >::k_batch

◆ k_padded_zeros

template<typename TilePartitioner_, typename FlatmmPipeline_, typename EpiloguePipeline_, MoeFlatmmKind kind, typename FusedActivation = moe::MoeSilu>
template<class ScaleM = FlatmmScalePointer<-1>, class ScaleN = FlatmmScalePointer<-1>, class ExpertBias = FlatmmScalePointer<-1>>
ck_tile::index_t ck_tile::MoeFlatmmKernel< TilePartitioner_, FlatmmPipeline_, EpiloguePipeline_, kind, FusedActivation >::MoeFlatmmKernelArgs< ScaleM, ScaleN, ExpertBias >::k_padded_zeros

◆ M

template<typename TilePartitioner_, typename FlatmmPipeline_, typename EpiloguePipeline_, MoeFlatmmKind kind, typename FusedActivation = moe::MoeSilu>
template<class ScaleM = FlatmmScalePointer<-1>, class ScaleN = FlatmmScalePointer<-1>, class ExpertBias = FlatmmScalePointer<-1>>
ck_tile::index_t ck_tile::MoeFlatmmKernel< TilePartitioner_, FlatmmPipeline_, EpiloguePipeline_, kind, FusedActivation >::MoeFlatmmKernelArgs< ScaleM, ScaleN, ExpertBias >::M

◆ N

template<typename TilePartitioner_, typename FlatmmPipeline_, typename EpiloguePipeline_, MoeFlatmmKind kind, typename FusedActivation = moe::MoeSilu>
template<class ScaleM = FlatmmScalePointer<-1>, class ScaleN = FlatmmScalePointer<-1>, class ExpertBias = FlatmmScalePointer<-1>>
ck_tile::index_t ck_tile::MoeFlatmmKernel< TilePartitioner_, FlatmmPipeline_, EpiloguePipeline_, kind, FusedActivation >::MoeFlatmmKernelArgs< ScaleM, ScaleN, ExpertBias >::N

◆ n_padded_zeros

template<typename TilePartitioner_, typename FlatmmPipeline_, typename EpiloguePipeline_, MoeFlatmmKind kind, typename FusedActivation = moe::MoeSilu>
template<class ScaleM = FlatmmScalePointer<-1>, class ScaleN = FlatmmScalePointer<-1>, class ExpertBias = FlatmmScalePointer<-1>>
ck_tile::index_t ck_tile::MoeFlatmmKernel< TilePartitioner_, FlatmmPipeline_, EpiloguePipeline_, kind, FusedActivation >::MoeFlatmmKernelArgs< ScaleM, ScaleN, ExpertBias >::n_padded_zeros

◆ NumTokens

template<typename TilePartitioner_, typename FlatmmPipeline_, typename EpiloguePipeline_, MoeFlatmmKind kind, typename FusedActivation = moe::MoeSilu>
template<class ScaleM = FlatmmScalePointer<-1>, class ScaleN = FlatmmScalePointer<-1>, class ExpertBias = FlatmmScalePointer<-1>>
ck_tile::index_t ck_tile::MoeFlatmmKernel< TilePartitioner_, FlatmmPipeline_, EpiloguePipeline_, kind, FusedActivation >::MoeFlatmmKernelArgs< ScaleM, ScaleN, ExpertBias >::NumTokens

◆ p_max_token_id

template<typename TilePartitioner_, typename FlatmmPipeline_, typename EpiloguePipeline_, MoeFlatmmKind kind, typename FusedActivation = moe::MoeSilu>
template<class ScaleM = FlatmmScalePointer<-1>, class ScaleN = FlatmmScalePointer<-1>, class ExpertBias = FlatmmScalePointer<-1>>
const ck_tile::index_t* ck_tile::MoeFlatmmKernel< TilePartitioner_, FlatmmPipeline_, EpiloguePipeline_, kind, FusedActivation >::MoeFlatmmKernelArgs< ScaleM, ScaleN, ExpertBias >::p_max_token_id

◆ p_sorted_expert_ids

template<typename TilePartitioner_, typename FlatmmPipeline_, typename EpiloguePipeline_, MoeFlatmmKind kind, typename FusedActivation = moe::MoeSilu>
template<class ScaleM = FlatmmScalePointer<-1>, class ScaleN = FlatmmScalePointer<-1>, class ExpertBias = FlatmmScalePointer<-1>>
const ck_tile::index_t* ck_tile::MoeFlatmmKernel< TilePartitioner_, FlatmmPipeline_, EpiloguePipeline_, kind, FusedActivation >::MoeFlatmmKernelArgs< ScaleM, ScaleN, ExpertBias >::p_sorted_expert_ids

◆ p_sorted_expert_weights

template<typename TilePartitioner_, typename FlatmmPipeline_, typename EpiloguePipeline_, MoeFlatmmKind kind, typename FusedActivation = moe::MoeSilu>
template<class ScaleM = FlatmmScalePointer<-1>, class ScaleN = FlatmmScalePointer<-1>, class ExpertBias = FlatmmScalePointer<-1>>
const void* ck_tile::MoeFlatmmKernel< TilePartitioner_, FlatmmPipeline_, EpiloguePipeline_, kind, FusedActivation >::MoeFlatmmKernelArgs< ScaleM, ScaleN, ExpertBias >::p_sorted_expert_weights

◆ p_sorted_token_ids

template<typename TilePartitioner_, typename FlatmmPipeline_, typename EpiloguePipeline_, MoeFlatmmKind kind, typename FusedActivation = moe::MoeSilu>
template<class ScaleM = FlatmmScalePointer<-1>, class ScaleN = FlatmmScalePointer<-1>, class ExpertBias = FlatmmScalePointer<-1>>
const ck_tile::index_t* ck_tile::MoeFlatmmKernel< TilePartitioner_, FlatmmPipeline_, EpiloguePipeline_, kind, FusedActivation >::MoeFlatmmKernelArgs< ScaleM, ScaleN, ExpertBias >::p_sorted_token_ids

◆ scale_m

template<typename TilePartitioner_, typename FlatmmPipeline_, typename EpiloguePipeline_, MoeFlatmmKind kind, typename FusedActivation = moe::MoeSilu>
template<class ScaleM = FlatmmScalePointer<-1>, class ScaleN = FlatmmScalePointer<-1>, class ExpertBias = FlatmmScalePointer<-1>>
ScaleM ck_tile::MoeFlatmmKernel< TilePartitioner_, FlatmmPipeline_, EpiloguePipeline_, kind, FusedActivation >::MoeFlatmmKernelArgs< ScaleM, ScaleN, ExpertBias >::scale_m

◆ scale_n

template<typename TilePartitioner_, typename FlatmmPipeline_, typename EpiloguePipeline_, MoeFlatmmKind kind, typename FusedActivation = moe::MoeSilu>
template<class ScaleM = FlatmmScalePointer<-1>, class ScaleN = FlatmmScalePointer<-1>, class ExpertBias = FlatmmScalePointer<-1>>
ScaleN ck_tile::MoeFlatmmKernel< TilePartitioner_, FlatmmPipeline_, EpiloguePipeline_, kind, FusedActivation >::MoeFlatmmKernelArgs< ScaleM, ScaleN, ExpertBias >::scale_n

◆ stride_A

template<typename TilePartitioner_, typename FlatmmPipeline_, typename EpiloguePipeline_, MoeFlatmmKind kind, typename FusedActivation = moe::MoeSilu>
template<class ScaleM = FlatmmScalePointer<-1>, class ScaleN = FlatmmScalePointer<-1>, class ExpertBias = FlatmmScalePointer<-1>>
ck_tile::index_t ck_tile::MoeFlatmmKernel< TilePartitioner_, FlatmmPipeline_, EpiloguePipeline_, kind, FusedActivation >::MoeFlatmmKernelArgs< ScaleM, ScaleN, ExpertBias >::stride_A

◆ stride_B

template<typename TilePartitioner_, typename FlatmmPipeline_, typename EpiloguePipeline_, MoeFlatmmKind kind, typename FusedActivation = moe::MoeSilu>
template<class ScaleM = FlatmmScalePointer<-1>, class ScaleN = FlatmmScalePointer<-1>, class ExpertBias = FlatmmScalePointer<-1>>
ck_tile::index_t ck_tile::MoeFlatmmKernel< TilePartitioner_, FlatmmPipeline_, EpiloguePipeline_, kind, FusedActivation >::MoeFlatmmKernelArgs< ScaleM, ScaleN, ExpertBias >::stride_B

◆ stride_C

template<typename TilePartitioner_, typename FlatmmPipeline_, typename EpiloguePipeline_, MoeFlatmmKind kind, typename FusedActivation = moe::MoeSilu>
template<class ScaleM = FlatmmScalePointer<-1>, class ScaleN = FlatmmScalePointer<-1>, class ExpertBias = FlatmmScalePointer<-1>>
ck_tile::index_t ck_tile::MoeFlatmmKernel< TilePartitioner_, FlatmmPipeline_, EpiloguePipeline_, kind, FusedActivation >::MoeFlatmmKernelArgs< ScaleM, ScaleN, ExpertBias >::stride_C

◆ TopK

template<typename TilePartitioner_, typename FlatmmPipeline_, typename EpiloguePipeline_, MoeFlatmmKind kind, typename FusedActivation = moe::MoeSilu>
template<class ScaleM = FlatmmScalePointer<-1>, class ScaleN = FlatmmScalePointer<-1>, class ExpertBias = FlatmmScalePointer<-1>>
ck_tile::index_t ck_tile::MoeFlatmmKernel< TilePartitioner_, FlatmmPipeline_, EpiloguePipeline_, kind, FusedActivation >::MoeFlatmmKernelArgs< ScaleM, ScaleN, ExpertBias >::TopK
The documentation for this struct was generated from the following file: