File size: 13,134 Bytes
61b850a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
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
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
#include "common.cuh"
#include "fattn-common.cuh"
#include "fattn-tile-f32.cuh"

#define FATTN_KQ_STRIDE_TILE_F32 32

template<int D, int ncols, int nwarps, int parallel_blocks, bool use_logit_softcap> // D == head size
#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
__launch_bounds__(nwarps*WARP_SIZE, 1)
#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
static __global__ void flash_attn_tile_ext_f32(
        const char * __restrict__ Q,
        const char * __restrict__ K,
        const char * __restrict__ V,
        const char * __restrict__ mask,
        float      * __restrict__ dst,
        float2     * __restrict__ dst_meta,
        const float scale,
        const float max_bias,
        const float m0,
        const float m1,
        const uint32_t n_head_log2,
        const float logit_softcap,
        const int ne00,
        const int ne01,
        const int ne02,
        const int ne03,
        const int ne10,
        const int ne11,
        const int ne12,
        const int ne13,
        const int ne31,
        const int nb31,
        const int nb01,
        const int nb02,
        const int nb03,
        const int nb11,
        const int nb12,
        const int nb13,
        const int nb21,
        const int nb22,
        const int nb23,
        const int ne0,
        const int ne1,
        const int ne2,
        const int ne3) {
#ifndef FLASH_ATTN_AVAILABLE
    NO_DEVICE_CODE;
    return;
#endif // FLASH_ATTN_AVAILABLE

    // Skip unused kernel variants for faster compilation:
#ifdef FP16_MMA_AVAILABLE
    NO_DEVICE_CODE;
    return;
#endif // FP16_MMA_AVAILABLE
    if (use_logit_softcap && !(D == 128 || D == 256)) {
        NO_DEVICE_CODE;
        return;
    }

    // In this kernel Q, K, V are matrices while i, j, k are matrix indices.

    const int ic0 = (blockIdx.x / parallel_blocks) * ncols; // Index of the Q/QKV column to work on.
    const int ip  =  blockIdx.x % parallel_blocks; // Index in group of blocks running for the same column in parallel.

    const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
    const float2 * Q_f2  = (const float2 *) (Q    + nb02* blockIdx.y              + nb01*ic0);
    const half2  * K_h2  = (const half2  *) (K    + nb12*(blockIdx.y / gqa_ratio));
    const half2  * V_h2  = (const half2  *) (V    + nb12*(blockIdx.y / gqa_ratio)); // K and V have same shape
    const half   * maskh = (const half   *)  mask + ne11*ic0;

    const int stride_KV2 = nb11 / sizeof(half2);

    const float slope = get_alibi_slope(max_bias, blockIdx.y, n_head_log2, m0, m1);

    static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64.");

    __shared__ float KQ[ncols*FATTN_KQ_STRIDE_TILE_F32];

    __shared__ float KV_tmp[FATTN_KQ_STRIDE_TILE_F32][D + 1]; // Pad D to avoid memory bank conflicts.
    float2 * KV_tmp2 = (float2 *) KV_tmp;

    float kqmax[ncols/nwarps];
#pragma unroll
    for (int j0 = 0; j0 < ncols; j0 += nwarps) {
        kqmax[j0/nwarps] = -FLT_MAX/2.0f;
    }
    float kqsum[ncols/nwarps] = {0.0f};

    float2 VKQ[ncols/nwarps][(D/2)/WARP_SIZE] = {{{0.0f, 0.0f}}};

    // Convert Q to half2 and store in registers:
    __shared__ float Q_f[ncols][D];
#pragma unroll
    for (int j0 = 0; j0 < ncols; j0 += nwarps) {
        const int j = j0 + threadIdx.y;

#pragma unroll
        for (int i0 = 0; i0 < D; i0 += 2*WARP_SIZE) {
            float2 tmp = ic0 + j < ne01 ? Q_f2[j*(nb01/sizeof(float2)) + i0/2 + threadIdx.x] : make_float2(0.0f, 0.0f);
            Q_f[j][i0 + 0*WARP_SIZE + threadIdx.x] = tmp.x * scale;
            Q_f[j][i0 + 1*WARP_SIZE + threadIdx.x] = tmp.y * scale;
        }
    }

    __syncthreads();

    const int k_start = parallel_blocks == 1 ? 0 : ip*FATTN_KQ_STRIDE_TILE_F32;
    for (int k_VKQ_0 = k_start; k_VKQ_0 < ne11; k_VKQ_0 += parallel_blocks*FATTN_KQ_STRIDE_TILE_F32) {
        // Calculate KQ tile and keep track of new maximum KQ values:

        float kqmax_new[ncols/nwarps];
#pragma unroll
        for (int j = 0; j < ncols/nwarps; ++j) {
            kqmax_new[j] = kqmax[j];
        }

#pragma unroll
        for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE_TILE_F32; i_KQ_0 += nwarps) {
            const int i_KQ = i_KQ_0 + threadIdx.y;

#pragma unroll
            for (int k_KQ_0 = 0; k_KQ_0 < D; k_KQ_0 += 2*WARP_SIZE) {
                const half2 tmp = K_h2[(k_VKQ_0 + i_KQ)*stride_KV2 + k_KQ_0/2 + threadIdx.x];
                KV_tmp[i_KQ][k_KQ_0 + 0*WARP_SIZE + threadIdx.x] =  __low2float(tmp);
                KV_tmp[i_KQ][k_KQ_0 + 1*WARP_SIZE + threadIdx.x] = __high2float(tmp);
            }
        }

        __syncthreads();

        float sum[FATTN_KQ_STRIDE_TILE_F32/WARP_SIZE][ncols/nwarps] = {{0.0f}};

#pragma unroll
        for (int k_KQ = 0; k_KQ < D; ++k_KQ) {
            float K_k[FATTN_KQ_STRIDE_TILE_F32/WARP_SIZE];
            float Q_k[ncols/nwarps];

#pragma unroll
            for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE_TILE_F32; i_KQ_0 += WARP_SIZE) {
                const int i_KQ = i_KQ_0 + threadIdx.x;

                K_k[i_KQ_0/WARP_SIZE] = KV_tmp[i_KQ][k_KQ];
            }
#pragma unroll
            for (int j_KQ_0 = 0; j_KQ_0 < ncols; j_KQ_0 += nwarps) {
                const int j_KQ = j_KQ_0 + threadIdx.y;

                Q_k[j_KQ_0/nwarps] = Q_f[j_KQ][k_KQ];
            }

#pragma unroll
            for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE_TILE_F32; i_KQ_0 += WARP_SIZE) {
#pragma unroll
                for (int j_KQ_0 = 0; j_KQ_0 < ncols; j_KQ_0 += nwarps) {
                    sum[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps] += K_k[i_KQ_0/WARP_SIZE] * Q_k[j_KQ_0/nwarps];
                }
            }
        }

#pragma unroll
        for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE_TILE_F32; i_KQ_0 += WARP_SIZE) {
            const int i_KQ = i_KQ_0 + threadIdx.x;

#pragma unroll
            for (int j_KQ_0 = 0; j_KQ_0 < ncols; j_KQ_0 += nwarps) {
                const int j_KQ = j_KQ_0 + threadIdx.y;

                if (use_logit_softcap) {
                    sum[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps] = logit_softcap * tanhf(sum[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps]);
                }

                sum[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps] += mask ? slope*__half2float(maskh[j_KQ*ne11 + k_VKQ_0 + i_KQ]) : 0.0f;

                kqmax_new[j_KQ_0/nwarps] = fmaxf(kqmax_new[j_KQ_0/nwarps], sum[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps]);

                KQ[j_KQ*FATTN_KQ_STRIDE_TILE_F32 + i_KQ] = sum[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps];
            }
        }

        __syncthreads();

#pragma unroll
        for (int j0 = 0; j0 < ncols; j0 += nwarps) {
            const int j = j0 + threadIdx.y;

            kqmax_new[j0/nwarps] = warp_reduce_max(kqmax_new[j0/nwarps]);
            const float KQ_max_scale = expf(kqmax[j0/nwarps] - kqmax_new[j0/nwarps]);
            kqmax[j0/nwarps] = kqmax_new[j0/nwarps];

            float kqsum_add = 0.0f;
#pragma unroll
            for (int i0 = 0; i0 < FATTN_KQ_STRIDE_TILE_F32; i0 += WARP_SIZE) {
                const int i = i0 + threadIdx.x;

                const float diff = KQ[j*FATTN_KQ_STRIDE_TILE_F32 + i] - kqmax[j0/nwarps];
                const float val = expf(diff);
                kqsum_add += val;
                KQ[j*FATTN_KQ_STRIDE_TILE_F32 + i] = val;
            }
            kqsum[j0/nwarps] = kqsum[j0/nwarps]*KQ_max_scale + kqsum_add;

#pragma unroll
            for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
                VKQ[j0/nwarps][i0/WARP_SIZE].x *= KQ_max_scale;
                VKQ[j0/nwarps][i0/WARP_SIZE].y *= KQ_max_scale;
            }
        }

        __syncthreads();

#pragma unroll
        for (int k0 = 0; k0 < FATTN_KQ_STRIDE_TILE_F32; k0 += nwarps) {
            const int k = k0 + threadIdx.y;

#pragma unroll
            for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
                const int i = i0 + threadIdx.x;

                KV_tmp2[k*(D/2) + i].x =  __low2float(V_h2[(k_VKQ_0 + k)*stride_KV2 + i]);
                KV_tmp2[k*(D/2) + i].y = __high2float(V_h2[(k_VKQ_0 + k)*stride_KV2 + i]);
            }
        }

        __syncthreads();

#pragma unroll
        for (int k = 0; k < FATTN_KQ_STRIDE_TILE_F32; ++k) {
            float2 V_k[(D/2)/WARP_SIZE];
            float  KQ_k[ncols/nwarps];

#pragma unroll
            for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
                const int i = i0 + threadIdx.x;

                V_k[i0/WARP_SIZE] = KV_tmp2[k*(D/2) + i];
            }
#pragma unroll
            for (int j0 = 0; j0 < ncols; j0 += nwarps) {
                const int j = j0 + threadIdx.y;

                KQ_k[j0/nwarps] = KQ[j*FATTN_KQ_STRIDE_TILE_F32 + k];
            }

#pragma unroll
            for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
#pragma unroll
                for (int j0 = 0; j0 < ncols; j0 += nwarps) {
                    VKQ[j0/nwarps][i0/WARP_SIZE].x += V_k[i0/WARP_SIZE].x*KQ_k[j0/nwarps];
                    VKQ[j0/nwarps][i0/WARP_SIZE].y += V_k[i0/WARP_SIZE].y*KQ_k[j0/nwarps];
                }
            }
        }

        __syncthreads();
    }

#pragma unroll
    for (int j_VKQ_0 = 0; j_VKQ_0 < ncols; j_VKQ_0 += nwarps) {
        const int j_VKQ = j_VKQ_0 + threadIdx.y;

        if (ic0 + j_VKQ >= ne01) {
            return;
        }

        float kqsum_j = kqsum[j_VKQ_0/nwarps];
        kqsum_j = warp_reduce_sum(kqsum_j);

#pragma unroll
        for (int i00 = 0; i00 < D; i00 += 2*WARP_SIZE) {
            const int i0 = i00 + 2*threadIdx.x;

            float2 dst_val = VKQ[j_VKQ_0/nwarps][i0/(2*WARP_SIZE)];
            if (parallel_blocks == 1) {
                dst_val.x /= kqsum_j;
                dst_val.y /= kqsum_j;
            }
            const int j_dst = (ic0 + j_VKQ)*parallel_blocks + ip;
            dst[j_dst*D*gridDim.y + D*blockIdx.y + i0 + 0] = dst_val.x;
            dst[j_dst*D*gridDim.y + D*blockIdx.y + i0 + 1] = dst_val.y;
        }

        if (parallel_blocks != 1 && threadIdx.x == 0) {
            dst_meta[(ic0 + j_VKQ)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = make_float2(kqmax[j_VKQ_0/nwarps], kqsum_j);
        }
    }
}

template <int cols_per_block, int parallel_blocks, bool use_logit_softcap>
void launch_fattn_tile_f32_64_128(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
    const ggml_tensor * Q = dst->src[0];
    switch (Q->ne[0]) {
        case  64: {
            constexpr int    D             = 64;
            constexpr int    nwarps        = 8;
            constexpr size_t nbytes_shared = 0;
            fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f32<D, cols_per_block, nwarps, parallel_blocks, use_logit_softcap>;
            launch_fattn<D, cols_per_block, parallel_blocks, -1>(ctx, dst, fattn_kernel, nwarps, nbytes_shared, true, true);
        } break;
        case 128: {
            constexpr int    D             = 128;
            constexpr int    nwarps        = 8;
            constexpr size_t nbytes_shared = 0;
            fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f32<D, cols_per_block, nwarps, parallel_blocks, use_logit_softcap>;
            launch_fattn<D, cols_per_block, parallel_blocks, -1>(ctx, dst, fattn_kernel, nwarps, nbytes_shared, true, true);
        } break;
        default: {
            GGML_ABORT("FlashAttention without tensor cores only supports head sizes 64 and 128.");
        } break;
    }
}

void ggml_cuda_flash_attn_ext_tile_f32(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
    const ggml_tensor * KQV = dst;
    const ggml_tensor * Q = dst->src[0];

    float logit_softcap;
    memcpy(&logit_softcap, (const float *) KQV->op_params + 2, sizeof(float));

    if (Q->ne[1] <= 16) {
        constexpr int cols_per_block = 16;
        constexpr int parallel_blocks = 4;
        if (logit_softcap == 0.0f) {
            constexpr bool use_logit_softcap = false;
            launch_fattn_tile_f32_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
        } else {
            constexpr bool use_logit_softcap = true;
            launch_fattn_tile_f32_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
        }
        return;
    }

    if (Q->ne[1] <= 32) {
        constexpr int cols_per_block = 32;
        constexpr int parallel_blocks = 4;
        if (logit_softcap == 0.0f) {
            constexpr bool use_logit_softcap = false;
            launch_fattn_tile_f32_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
        } else {
            constexpr bool use_logit_softcap = true;
            launch_fattn_tile_f32_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
        }
        return;
    }

    constexpr int cols_per_block = 32;
    constexpr int parallel_blocks = 1;
    if (logit_softcap == 0.0f) {
        constexpr bool use_logit_softcap = false;
        launch_fattn_tile_f32_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
    } else {
        constexpr bool use_logit_softcap = true;
        launch_fattn_tile_f32_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
    }
}