Skip to content

GitLab

Switch branch/tag
Find file Normal viewHistoryPermalink
flash_c_api.h 8.31 KB
Newer Older
zhangshao's avatar
first commit  
zhangshao committed
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22 
#include <fstream>
#include <cstdlib>
#include <vector>
#include "numeric_types.h"
#include "config.h"
#include "static_switch.h"
#include "flash_singleton.h"
#include "flash_memory.h"


// ====================================================================================================================================
//                                                               FWD
// ====================================================================================================================================
void run_mha_fwd(Flash_fwd_params &params, hipStream_t stream, bool force_split_kernel=false) {
#if defined(BUILD_FA_FWD)
    const char* fa_debug = std::getenv("FA_DEBUG");
    if (fa_debug != nullptr) {
        if (std::strcmp(fa_debug, "5") == 0) return;
        else if (std::strcmp(fa_debug, "C") == 0) {PRINT_PARAMS}; // for c interface debug
    }
    if (params.seqused_k != nullptr) {
        // Prefix prefill attention
hly's avatar
Import latest aicc hipcc fp8 pa snapshot.  
hly committed
23
24
25
26
27
28
29
30
31 
        if (params.is_e4m3) {
            FP16_SWITCH(!params.is_bf16, [&] {
                if (params.d == 128 and params.d_value == 128) {
                    run_fp8_mha_fwd_prefix_prefill_<elem_type, 128, 128>(params, stream);
                } else {
                    assert(false && "FP8 prefix prefill only supports head_dim=128");
                }
            });
        } else if (!params.is_int8){
zhangshao's avatar
first commit  
zhangshao committed
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
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192 
            FP16_SWITCH(!params.is_bf16, [&] {
                if (params.d == 128 and params.d_value == 128) {
                    run_mha_fwd_prefix_prefill_<elem_type, 128, 128>(params, stream);
                } else if (params.d == 192 and params.d_value == 128) {
                    run_mha_fwd_prefix_prefill_<elem_type, 192, 128>(params, stream);
                } else if (params.d == 192 and params.d_value == 192) {
                    run_mha_fwd_prefix_prefill_<elem_type, 192, 192>(params, stream);
                } else if (params.d == 256 and params.d_value == 256) {
                    run_mha_fwd_prefix_prefill_<elem_type, 256, 256>(params, stream); // used in gemma2-9b
                }
            });
        } else {
            FP16_SWITCH(!params.is_bf16, [&] {
                if (params.d == 128 and params.d_value == 128) {
                    run_int8_mha_fwd_prefix_prefill_<elem_type, 128, 128>(params, stream);
                } else if (params.d == 192 and params.d_value == 128) {
                    run_int8_mha_fwd_prefix_prefill_<elem_type, 192, 128>(params, stream);
                } else if (params.d == 192 and params.d_value == 192) {
                    run_int8_mha_fwd_prefix_prefill_<elem_type, 192, 192>(params, stream);
                }
            });
        }
    }
    else if (params.attn_mask != nullptr) {
        // Broadcastable mask attention like torch.sdpa
        FP16_SWITCH(!params.is_bf16, [&] {
            if (params.d == 128) {
                run_mha_fwd_attn_mask_<elem_type, 128, 128>(params, stream);
            }
        });
    }
    else if (params.padding_mask != nullptr) {
        // Encoder attention, bert. etc.
        FP16_SWITCH(!params.is_bf16, [&] {
            if (params.d == 64) {
                run_mha_fwd_padding_mask_<elem_type, 64, 64>(params, stream);
            } else if (params.d == 128) {
                run_mha_fwd_padding_mask_<elem_type, 128, 128>(params, stream);
            }
        });
    }
    else {
        // Decoder-only attention
        FP16_SWITCH(!params.is_bf16, [&] {
        #if defined(HEADDIM_128_ONLY)
            run_mha_fwd_<elem_type, 128, 128>(params, stream);
        #elif defined(HEADDIM_192_128_ONLY)
            run_mha_fwd_<elem_type, 192, 128>(params, stream);
        #else
            ALL_HEADDIM_SWITCH(params.d, params.d_value, [&] {
                run_mha_fwd_<elem_type, kHeadDimQ, kHeadDimV>(params, stream);
            });
        #endif
        });
    }
#endif
}

void (*run_mha_fwd_c)(Flash_fwd_params&, hipStream_t, bool) = run_mha_fwd;



// ====================================================================================================================================
//                                                               BWD
// ====================================================================================================================================
void run_mha_bwd(Flash_bwd_params &params, hipStream_t stream, const bool configure=false) {
#if defined(BUILD_FA_BWD)
    const char* fa_debug = std::getenv("FA_DEBUG");
    if (fa_debug != nullptr) {
        if (std::strcmp(fa_debug, "5") == 0) return;
        else { printFlashBwdParams(params); };
    }
    ElementType_SWITCH(params.is_bf16, params.is_e4m3, [&] {
        #if defined(HEADDIM_128_ONLY)
            run_mha_bwd_<elem_type, 128, 128>(params, stream, configure);
        #elif defined(HEADDIM_192_128_ONLY)
            run_mha_bwd_<elem_type, 192, 128>(params, stream, configure);
        #else
            HEADDIM_SWITCH(params.d, params.d_value, [&] {
                run_mha_bwd_<elem_type, kHeadDimQ, kHeadDimV>(params, stream, configure);
            });
        #endif
    });
#endif
}



// ====================================================================================================================================
//                                                              PA
// ====================================================================================================================================
void run_mha_fwd_kvcache(Flash_fwd_params &params, hipStream_t stream, bool force_split_kernel=false) {
#if defined(BUILD_FA_KVCACHE)
    const char* fa_debug = std::getenv("FA_DEBUG");
    if (fa_debug != nullptr) {
        if (std::strcmp(fa_debug, "5") == 0) return;
        else if (std::strcmp(fa_debug, "C") == 0) {PRINT_PARAMS}; // for c interface debug
    }
    FP16_SWITCH(!params.is_bf16, [&] {
    #ifdef HEADDIM_128_ONLY
        run_mha_fwd_splitkv_dispatch<elem_type, 128, 128>(params, stream);
    #elif defined(HEADDIM_192_128_ONLY)
        if (params.d == 192 and params.d_value == 128)
            run_mha_fwd_splitkv_dispatch<elem_type, 192, 128>(params, stream);
        else if (params.d == 576 and params.d_value == 512)
            run_mha_fwd_splitkv_dispatch<elem_type, 576, 512>(params, stream);
    #else
        PA_HEADDIM_SWITCH(params.d, params.d_value, [&] {
            run_mha_fwd_splitkv_dispatch<elem_type, kHeadDimQ, kHeadDimV>(params, stream);
        });
    #endif
    });
#endif
}

void run_int8_fwd_kvcache(Flash_fwd_params &params, hipStream_t stream, bool force_split_kernel=false) {
#if defined(BUILD_FA_KVCACHE)
    const char* fa_debug = std::getenv("FA_DEBUG");
    if (fa_debug != nullptr) {
        if (std::strcmp(fa_debug, "5") == 0) return;
        else if (std::strcmp(fa_debug, "C") == 0) {PRINT_PARAMS}; // for c interface debug
    }
    FP16_SWITCH(!params.is_bf16, [&] {
        if (params.d != 128 or params.d_value != 128){
            printf("int8 pa only support headdim=128!\n");
            assert(params.d == 128 and params.d_value == 128);
        }
        run_int8_fwd_splitkv_dispatch<elem_type, 128, 128>(params, stream);
    });
#endif
}


// ====================================================================================================================================
//                                                             FlashMLA
// ====================================================================================================================================
void run_fwd_flashmla(Flash_fwd_mla_params &params, hipStream_t stream, bool force_split_kernel=false) {
#if defined(BUILD_FLASHMLA)
    const char* fa_debug = std::getenv("FA_DEBUG");
    if (fa_debug != nullptr) {
        if (std::strcmp(fa_debug, "5") == 0) return;
        else if (std::strcmp(fa_debug, "C") == 0) {PRINT_MLA_PARAMS}; // for c interface debug
    }
    FP16_SWITCH(!params.is_bf16, [&] {
        run_mla_fwd_splitkv_dispatch<elem_type, 576, 512>(params, stream);
    });
#endif
}


void run_fwd_prefix_prefill_mla(Flash_fwd_mla_params &params, hipStream_t stream) {
#if defined(BUILD_FA_FWD)
    const char* fa_debug = std::getenv("FA_DEBUG");
    if (fa_debug != nullptr) {
        if (std::strcmp(fa_debug, "5") == 0) return;
        else if (std::strcmp(fa_debug, "C") == 0) {PRINT_MLA_PARAMS}; // for c interface debug
    }
    FP16_SWITCH(!params.is_bf16, [&] {
        run_mla_fwd_prefix_prefill_dispatch_<elem_type, 576, 512>(params, stream);
    });
#endif
hly's avatar
Import latest aicc hipcc fp8 pa snapshot.  
hly committed
193 
}