kernels-community/flash-mla

1#pragma once2 3////////////////////////////////////////////////////////////////////////////////////////////////////4 5struct Flash_fwd_mla_params {6    using index_t = int64_t;7 8    int b, seqlen_q, d, d_v;9    int h, h_h_k_ratio, ngroups;10    bool is_causal;11    float scale_softmax, scale_softmax_log2;12    int *__restrict__ cu_seqlens_k;13 14    void *__restrict__ q_ptr;15    void *__restrict__ k_ptr;16    void *__restrict__ v_ptr;17    void *__restrict__ o_ptr;18    void *__restrict__ softmax_lse_ptr;19 20    index_t q_batch_stride;21    index_t k_batch_stride;22    index_t v_batch_stride;23    index_t o_batch_stride;24    index_t q_row_stride;25    index_t k_row_stride;26    index_t v_row_stride;27    index_t o_row_stride;28    index_t q_head_stride;29    index_t k_head_stride;30    index_t v_head_stride;31    index_t o_head_stride;32 33    int *__restrict__ block_table;34    index_t block_table_batch_stride;35    int page_block_size;36 37    int *__restrict__ tile_scheduler_metadata_ptr;38    int num_sm_parts;39    int *__restrict__ num_splits_ptr;40 41    void *__restrict__ softmax_lseaccum_ptr;42    void *__restrict__ oaccum_ptr;43};44 45static constexpr int TileSchedulerMetaDataSize = 8;46// [begin_idx, begin_seqlen, end_idx, end_seqlen, begin_n_split_idx, _, _, _]47 48////////////////////////////////////////////////////////////////////////////////////////////////////49 50template<typename T, int Headdim>51void run_mha_fwd_splitkv_mla(Flash_fwd_mla_params &params, cudaStream_t stream);52 53struct Mla_metadata_params {54    int *__restrict__ seqlens_k_ptr;55    int *__restrict__ tile_scheduler_metadata_ptr;56    int *__restrict__ num_splits_ptr;57    int batch_size;58    int block_size_n;59    int fixed_overhead_num_blocks;60    int num_sm_parts;61};62 63void get_mla_metadata_func(Mla_metadata_params &params, cudaStream_t stream);64