Hugging Face's logo

kernels-community
/
flash-mla

kernel
flash-mla
deepseek
kernel-builder
Model card Files Community
flash-mla
File size: 1,806 Bytes 
1f83cde
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
 
#pragma once

////////////////////////////////////////////////////////////////////////////////////////////////////

struct Flash_fwd_mla_params {
    using index_t = int64_t;

    int b, seqlen_q, d, d_v;
    int h, h_h_k_ratio, ngroups;
    bool is_causal;
    float scale_softmax, scale_softmax_log2;
    int *__restrict__ cu_seqlens_k;

    void *__restrict__ q_ptr;
    void *__restrict__ k_ptr;
    void *__restrict__ v_ptr;
    void *__restrict__ o_ptr;
    void *__restrict__ softmax_lse_ptr;

    index_t q_batch_stride;
    index_t k_batch_stride;
    index_t v_batch_stride;
    index_t o_batch_stride;
    index_t q_row_stride;
    index_t k_row_stride;
    index_t v_row_stride;
    index_t o_row_stride;
    index_t q_head_stride;
    index_t k_head_stride;
    index_t v_head_stride;
    index_t o_head_stride;

    int *__restrict__ block_table;
    index_t block_table_batch_stride;
    int page_block_size;

    int *__restrict__ tile_scheduler_metadata_ptr;
    int num_sm_parts;
    int *__restrict__ num_splits_ptr;

    void *__restrict__ softmax_lseaccum_ptr;
    void *__restrict__ oaccum_ptr;
};

static constexpr int TileSchedulerMetaDataSize = 8;
// [begin_idx, begin_seqlen, end_idx, end_seqlen, begin_n_split_idx, _, _, _]

////////////////////////////////////////////////////////////////////////////////////////////////////

template<typename T, int Headdim>
void run_mha_fwd_splitkv_mla(Flash_fwd_mla_params &params, cudaStream_t stream);

struct Mla_metadata_params {
    int *__restrict__ seqlens_k_ptr;
    int *__restrict__ tile_scheduler_metadata_ptr;
    int *__restrict__ num_splits_ptr;
    int batch_size;
    int block_size_n;
    int fixed_overhead_num_blocks;
    int num_sm_parts;
};

void get_mla_metadata_func(Mla_metadata_params &params, cudaStream_t stream);