llama.cpp/ggml-cuda/fattn-vec-f16.cu

#include "common.cuh"
#include "fattn-common.cuh"
#include "fattn-vec-f16.cuh"

template<int D, int ncols, int parallel_blocks> // D == head size
#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
__launch_bounds__(D, 1)
#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
static __global__ void flash_attn_vec_ext_f16(
        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 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 ne0,
        const int ne1,
        const int ne2,
        const int ne3) {
#if FP16_AVAILABLE
    //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 half   * V_h   = (const half   *) (V    + nb12*(blockIdx.y / gqa_ratio)); // K and V have same shape
    const half   * maskh = (const half   *)  mask + ne11*ic0;

    const int stride_KV  = nb11 / sizeof(half);
    const int stride_KV2 = nb11 / sizeof(half2);

    const float slopef = get_alibi_slope(max_bias, blockIdx.y, n_head_log2, m0, m1);
    const half  slopeh = __float2half(slopef);

    static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64.");
    constexpr int nwarps = D / WARP_SIZE;
    const int tid = WARP_SIZE*threadIdx.y + threadIdx.x;
    __builtin_assume(tid < D);

    __shared__ half KQ[ncols*D];
#pragma unroll
    for (int j = 0; j < ncols; ++j) {
        KQ[j*D + tid] = -HALF_MAX_HALF;
    }
    half2 * KQ2 = (half2 *) KQ;

    half kqmax[ncols];
#pragma unroll
    for (int j = 0; j < ncols; ++j) {
        kqmax[j] = -HALF_MAX_HALF;
    }
    half kqsum[ncols] = {0.0f};

    __shared__ half kqmax_shared[ncols][WARP_SIZE];
    __shared__ half kqsum_shared[ncols][WARP_SIZE];
#pragma unroll
    for (int j = 0; j < ncols; ++j) {
        if (threadIdx.y == 0) {
            kqmax_shared[j][threadIdx.x] = -HALF_MAX_HALF;
            kqsum_shared[j][threadIdx.x] = 0.0f;
        }
    }
    __syncthreads();

    // Convert Q to half2 and store in registers:
    half2 Q_h2[ncols][D/(2*WARP_SIZE)];
#pragma unroll
    for (int j = 0; j < ncols; ++j) {
#pragma unroll
        for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
            const int i = i0 + threadIdx.x;

            const float2 tmp = ncols <= 2 || ic0 + j < ne01 ? Q_f2[j*(nb01/sizeof(float2)) + i] : make_float2(0.0f, 0.0f);
            Q_h2[j][i0/WARP_SIZE] = make_half2(scale, scale) * make_half2(tmp.x, tmp.y);
        }
    }

    half2 VKQ[ncols] = {{0.0f, 0.0f}};

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

        // For unknown reasons using a half array of size 1 for kqmax_new causes a performance regression,
        // see https://github.com/ggerganov/llama.cpp/pull/7061 .
        // Therefore this variable is defined twice but only used once (so that the compiler can optimize out the unused variable).
        half kqmax_new = kqmax[0];
        half kqmax_new_arr[ncols];
#pragma unroll
        for (int j = 0; j < ncols; ++j) {
            kqmax_new_arr[j] = kqmax[j];
        }

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

            if ((i_KQ_0 + nwarps > D && i_KQ >= D) || (FATTN_KQ_STRIDE % D != 0 && k_VKQ_0 + i_KQ >= ne11)) {
                break;
            }

            half2 sum2[ncols] = {{0.0f, 0.0f}};
#pragma unroll
            for (int k_KQ_0 = 0; k_KQ_0 < D/2; k_KQ_0 += WARP_SIZE) {
                const int k_KQ = k_KQ_0 + threadIdx.x;

                const half2 K_ik = K_h2[(k_VKQ_0 + i_KQ)*stride_KV2 + k_KQ];
#pragma unroll
                for (int j = 0; j < ncols; ++j) {
                    sum2[j] += K_ik * Q_h2[j][k_KQ_0/WARP_SIZE];
                }
            }

#pragma unroll
            for (int j = 0; j < ncols; ++j) {
                sum2[j] = warp_reduce_sum(sum2[j]);
                half sum = __low2half(sum2[j]) + __high2half(sum2[j]);
                sum += mask ? slopeh*maskh[j*ne11 + k_VKQ_0 + i_KQ] : __float2half(0.0f);

                if (ncols == 1) {
                    kqmax_new        = ggml_cuda_hmax(kqmax_new,        sum);
                } else {
                    kqmax_new_arr[j] = ggml_cuda_hmax(kqmax_new_arr[j], sum);
                }

                if (threadIdx.x == 0) {
                    KQ[j*D + i_KQ] = sum;
                }
            }
        }

#pragma unroll
        for (int j = 0; j < ncols; ++j) {
            half kqmax_new_j = ncols == 1 ? kqmax_new : kqmax_new_arr[j];

            kqmax_new_j = warp_reduce_max(kqmax_new_j);
            if (threadIdx.x == 0) {
                kqmax_shared[j][threadIdx.y] = kqmax_new_j;
            }
        }

        __syncthreads();

#pragma unroll
        for (int j = 0; j < ncols; ++j) {
            half kqmax_new_j = kqmax_shared[j][threadIdx.x];
            kqmax_new_j = warp_reduce_max(kqmax_new_j);

            const half KQ_max_scale = hexp(kqmax[j] - kqmax_new_j);
            kqmax[j] = kqmax_new_j;

            const half val = hexp(KQ[j*D + tid] - kqmax[j]);
            kqsum[j] = kqsum[j]*KQ_max_scale + val;
            KQ[j*D + tid] = val;

            VKQ[j] *= __half2half2(KQ_max_scale);
        }

        __syncthreads();

#pragma unroll
        for (int k0 = 0; k0 < D; k0 += 2) {
            if (FATTN_KQ_STRIDE % D != 0 && k_VKQ_0 + k0 >= ne11) {
                break;
            }

            half2 V_k;
            reinterpret_cast<half&>(V_k.x) = V_h[(k_VKQ_0 + k0 + 0)*stride_KV + tid];
            reinterpret_cast<half&>(V_k.y) = V_h[(k_VKQ_0 + k0 + 1)*stride_KV + tid];
#pragma unroll
            for (int j = 0; j < ncols; ++j) {
                VKQ[j] += V_k*KQ2[j*(D/2) + k0/2];
            }
        }

        __syncthreads();
    }

#pragma unroll
    for (int j = 0; j < ncols; ++j) {
        kqsum[j] = warp_reduce_sum(kqsum[j]);
        if (threadIdx.x == 0) {
            kqsum_shared[j][threadIdx.y] = kqsum[j];
        }
    }

    __syncthreads();

#pragma unroll
    for (int j_VKQ = 0; j_VKQ < ncols; ++j_VKQ) {
        if (ncols > 2 && ic0 + j_VKQ >= ne01) {
            break;
        }

        kqsum[j_VKQ] = kqsum_shared[j_VKQ][threadIdx.x];
        kqsum[j_VKQ] = warp_reduce_sum(kqsum[j_VKQ]);

        half dst_val = (__low2half(VKQ[j_VKQ]) + __high2half(VKQ[j_VKQ]));
        if (parallel_blocks == 1) {
            dst_val /= kqsum[j_VKQ];
        }
        const int j_dst = (ic0 + j_VKQ)*parallel_blocks + ip;
        dst[j_dst*D*gridDim.y + D*blockIdx.y + tid] = dst_val;
    }

    if (parallel_blocks != 1 && tid < ncols && (ncols <= 2 || ic0 + tid < ne01)) {
        dst_meta[(ic0 + tid)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = make_float2(kqmax[tid], kqsum[tid]);
    }
#else
   NO_DEVICE_CODE;
#endif // FP16_AVAILABLE
}

void ggml_cuda_flash_attn_ext_vec_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
    ggml_tensor * KQV = dst;
    ggml_tensor * Q   = dst->src[0];

    const int32_t precision = KQV->op_params[2];
    GGML_ASSERT(precision == GGML_PREC_DEFAULT);

    constexpr int cols_per_block  = 1;
    constexpr int parallel_blocks = 4;
    switch (Q->ne[0]) {
        case  64: {
            constexpr int      D = 64;
            constexpr int nwarps = D/WARP_SIZE;
            fattn_kernel_t fattn_kernel = flash_attn_vec_ext_f16<D, cols_per_block, parallel_blocks>;
            launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block);
        } break;
        case 128: {
            constexpr int      D = 128;
            constexpr int nwarps = D/WARP_SIZE;
            fattn_kernel_t fattn_kernel = flash_attn_vec_ext_f16<D, cols_per_block, parallel_blocks>;
            launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block);
        } break;
        case 256: {
            constexpr int      D = 256;
            constexpr int nwarps = D/WARP_SIZE;
            fattn_kernel_t fattn_kernel = flash_attn_vec_ext_f16<D, cols_per_block, parallel_blocks>;
            launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block);
        } break;
        default:
            GGML_ASSERT(false);
            break;
    }
}

template <int cols_per_block, int parallel_blocks>
void launch_fattn_vec_f16_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 = D/WARP_SIZE;
            fattn_kernel_t fattn_kernel = flash_attn_vec_ext_f16<D, cols_per_block, parallel_blocks>;
            launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block);
        } break;
        case 128: {
            constexpr int      D = 128;
            constexpr int nwarps = D/WARP_SIZE;
            fattn_kernel_t fattn_kernel = flash_attn_vec_ext_f16<D, cols_per_block, parallel_blocks>;
            launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block);
        } break;
        default: {
            GGML_ASSERT(false && "FlashAttention without tensor cores only supports head sizes 64 and 128.");
        } break;
    }
}

void ggml_cuda_flash_attn_ext_vec_f16_no_mma(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
    const ggml_tensor * KQV = dst;
    const ggml_tensor * Q   = dst->src[0];

    const int32_t precision = KQV->op_params[2];
    GGML_ASSERT(precision == GGML_PREC_DEFAULT);

    if (Q->ne[1] == 1) {
        ggml_cuda_flash_attn_ext_vec_f16(ctx, dst);
        return;
    }

    if (Q->ne[1] == 2) {
        constexpr int cols_per_block  = 2;
        constexpr int parallel_blocks = 4;
        launch_fattn_vec_f16_64_128<cols_per_block, parallel_blocks>(ctx, dst);
        return;
    }

    if (Q->ne[1] <= 4) {
        constexpr int cols_per_block  = 4;
        constexpr int parallel_blocks = 4;
        launch_fattn_vec_f16_64_128<cols_per_block, parallel_blocks>(ctx, dst);
        return;
    }

    if (Q->ne[1] <= 8) {
        constexpr int cols_per_block  = 8;
        constexpr int parallel_blocks = 4;
        launch_fattn_vec_f16_64_128<cols_per_block, parallel_blocks>(ctx, dst);
        return;
    }

    constexpr int cols_per_block  = 8;
    constexpr int parallel_blocks = 1;
    launch_fattn_vec_f16_64_128<cols_per_block, parallel_blocks>(ctx, dst);
}