llama.cpp/ggml-cuda.cu

#include <cstddef>
#include <cstdint>
#include <limits>
#include <stdint.h>
#include <stdio.h>
#include <atomic>
#include <assert.h>

#include <cuda_runtime.h>
#include <cublas_v2.h>
#include <cuda_fp16.h>

#include "ggml-cuda.h"
#include "ggml.h"

#if defined(_MSC_VER)
#pragma warning(disable: 4244 4267) // possible loss of data
#endif

static_assert(sizeof(half) == sizeof(ggml_fp16_t), "wrong fp16 size");

#define CUDA_CHECK(err)                                                                 \
    do {                                                                                \
        cudaError_t err_ = (err);                                                       \
        if (err_ != cudaSuccess) {                                                      \
            fprintf(stderr, "CUDA error %d at %s:%d: %s\n", err_, __FILE__, __LINE__,   \
                cudaGetErrorString(err_));                                              \
            exit(1);                                                                    \
        }                                                                               \
    } while (0)

#if CUDART_VERSION >= 12000
#define CUBLAS_CHECK(err)                                                               \
    do {                                                                                \
        cublasStatus_t err_ = (err);                                                    \
        if (err_ != CUBLAS_STATUS_SUCCESS) {                                            \
            fprintf(stderr, "\ncuBLAS error %d at %s:%d: %s\n",                         \
                    err_, __FILE__, __LINE__, cublasGetStatusString(err_));             \
            exit(1);                                                                    \
        }                                                                               \
    } while (0)
#else
#define CUBLAS_CHECK(err)                                                               \
    do {                                                                                \
        cublasStatus_t err_ = (err);                                                    \
        if (err_ != CUBLAS_STATUS_SUCCESS) {                                            \
            fprintf(stderr, "\ncuBLAS error %d at %s:%d\n", err_, __FILE__, __LINE__);  \
            exit(1);                                                                    \
        }                                                                               \
    } while (0)
#endif // CUDART_VERSION >= 11

#ifdef GGML_CUDA_DMMV_F16
typedef half dfloat; // dequantize float
typedef half2 dfloat2;
#else
typedef float dfloat; // dequantize float
typedef float2 dfloat2;
#endif //GGML_CUDA_DMMV_F16

typedef void (*dequantize_kernel_t)(const void * vx, const int ib, const int iqs, dfloat2 & v);
typedef void (*to_fp32_cuda_t)(const void * x, float * y, int k, cudaStream_t stream);
typedef void (*dot_kernel_k_t)(const void * vx, const int ib, const int iqs, const float * y, float & v);
typedef void (*cpy_kernel_t)(const char * cx, char * cdst);
typedef void (*ggml_cuda_func_t)(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst);
typedef void (*ggml_cuda_op_t)(
    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, char * src0_ddq_i, float * src0_ddf_i,
    float * src1_ddf_i, float * dst_ddf_i, int64_t i02, int64_t i01_low, int64_t i01_high, int i1,
    cudaStream_t & cudaStream_main);

// QK = number of values after dequantization
// QR = QK / number of values before dequantization

#define QK4_0 32
#define QR4_0 2
typedef struct {
    half    d;              // delta
    uint8_t qs[QK4_0 / 2];  // nibbles / quants
} block_q4_0;
static_assert(sizeof(block_q4_0) == sizeof(ggml_fp16_t) + QK4_0 / 2, "wrong q4_0 block size/padding");

#define QK4_1 32
#define QR4_1 2
typedef struct {
    half    d;              // delta
    half    m;              // min
    uint8_t qs[QK4_1 / 2];  // nibbles / quants
} block_q4_1;
static_assert(sizeof(block_q4_1) == sizeof(ggml_fp16_t) * 2 + QK4_1 / 2, "wrong q4_1 block size/padding");

#define QK5_0 32
#define QR5_0 2
typedef struct {
    half d;                 // delta
    uint8_t qh[4];          // 5-th bit of quants
    uint8_t qs[QK5_0 / 2];  // nibbles / quants
} block_q5_0;
static_assert(sizeof(block_q5_0) == sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_0 / 2, "wrong q5_0 block size/padding");

#define QK5_1 32
#define QR5_1 2
typedef struct {
    half d;                 // delta
    half m;                 // min
    uint8_t qh[4];          // 5-th bit of quants
    uint8_t qs[QK5_1 / 2];  // nibbles / quants
} block_q5_1;
static_assert(sizeof(block_q5_1) == 2 * sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_1 / 2, "wrong q5_1 block size/padding");

#define QK8_0 32
#define QR8_0 1
typedef struct {
    half    d;              // delta
    int8_t  qs[QK8_0];      // quants
} block_q8_0;
static_assert(sizeof(block_q8_0) == sizeof(ggml_fp16_t) + QK8_0, "wrong q8_0 block size/padding");

//================================= k-quants

#ifdef GGML_QKK_64
#define QK_K 64
#define K_SCALE_SIZE 4
#else
#define QK_K 256
#define K_SCALE_SIZE 12
#endif

typedef struct {
    uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
    uint8_t qs[QK_K/4];      // quants
    half d;                  // super-block scale for quantized scales
    half dmin;               // super-block scale for quantized mins
} block_q2_K;
static_assert(sizeof(block_q2_K) == 2*sizeof(ggml_fp16_t) + QK_K/16 + QK_K/4, "wrong q2_K block size/padding");

typedef struct {
    uint8_t hmask[QK_K/8];     // quants - high bit
    uint8_t qs[QK_K/4];        // quants - low 2 bits
#ifdef GGML_QKK_64
    uint8_t scales[2]; // scales, quantized with 8 bits
#else
    uint8_t scales[K_SCALE_SIZE]; // scales, quantized with 6 bits
#endif
    half d;             // super-block scale
} block_q3_K;
//static_assert(sizeof(block_q3_K) == sizeof(ggml_fp16_t) + QK_K / 4 + QK_K / 8 + K_SCALE_SIZE, "wrong q3_K block size/padding");

#ifdef GGML_QKK_64
typedef struct {
    half    d[2];              // super-block scales/mins
    uint8_t scales[2];         // 4-bit block scales/mins
    uint8_t qs[QK_K/2];        // 4--bit quants
} block_q4_K;
static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_fp16_t) + QK_K/2 + 2, "wrong q4_K block size/padding");
#else
typedef struct {
    half d;                    // super-block scale for quantized scales
    half dmin;                 // super-block scale for quantized mins
    uint8_t scales[3*QK_K/64]; // scales, quantized with 6 bits
    uint8_t qs[QK_K/2];        // 4--bit quants
} block_q4_K;
static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_fp16_t) + 3*QK_K/64 + QK_K/2, "wrong q4_K block size/padding");
#endif

#ifdef GGML_QKK_64
typedef struct {
    half d;                  // super-block scale
    int8_t scales[QK_K/16];  // block scales
    uint8_t qh[QK_K/8];      // quants, high bit
    uint8_t qs[QK_K/2];      // quants, low 4 bits
} block_q5_K;
static_assert(sizeof(block_q5_K) == sizeof(ggml_fp16_t) + QK_K/2 + QK_K/8 + QK_K/16, "wrong q5_K block size/padding");
#else
typedef struct {
    half d;               // super-block scale for quantized scales
    half dmin;            // super-block scale for quantized mins
    uint8_t scales[K_SCALE_SIZE];   // scales and mins, quantized with 6 bits
    uint8_t qh[QK_K/8];          // quants, high bit
    uint8_t qs[QK_K/2];          // quants, low 4 bits
} block_q5_K;
static_assert(sizeof(block_q5_K) == 2*sizeof(ggml_fp16_t) + K_SCALE_SIZE + QK_K/2 + QK_K/8, "wrong q5_K block size/padding");
#endif

typedef struct {
    uint8_t ql[QK_K/2];   // quants, lower 4 bits
    uint8_t qh[QK_K/4];   // quants, upper 2 bits
    int8_t  scales[QK_K/16]; // scales
    half    d;         // delta
} block_q6_K;
static_assert(sizeof(block_q6_K) == sizeof(ggml_fp16_t) + 13*QK_K/16, "wrong q6_K block size/padding");

#define WARP_SIZE 32

#define CUDA_ADD_BLOCK_SIZE 256
#define CUDA_MUL_BLOCK_SIZE 256
#define CUDA_SILU_BLOCK_SIZE 256
#define CUDA_CPY_BLOCK_SIZE 32
#define CUDA_SCALE_BLOCK_SIZE 256
#define CUDA_ROPE_BLOCK_SIZE 256
#define CUDA_DIAG_MASK_INF_BLOCK_SIZE 32
#define CUDA_DEQUANTIZE_BLOCK_SIZE 256

// dmmv = dequantize_mul_mat_vec
#ifndef GGML_CUDA_DMMV_X
#define GGML_CUDA_DMMV_X 32
#endif
#ifndef GGML_CUDA_DMMV_Y
#define GGML_CUDA_DMMV_Y 1
#endif

#ifndef K_QUANTS_PER_ITERATION
#define K_QUANTS_PER_ITERATION 2
#else
static_assert(K_QUANTS_PER_ITERATION == 1 || K_QUANTS_PER_ITERATION == 2, "K_QUANTS_PER_ITERATION must be 1 or 2");
#endif

static __global__ void add_f32(const float * x, const float * y, float * dst, const int k) {
    const int i = blockDim.x*blockIdx.x + threadIdx.x;

    if (i >= k) {
        return;
    }
    dst[i] = x[i] + y[i];
}

static __global__ void add_f16_f32_f16(const half * x, const float * y, half * dst, const int k) {
    const int i = blockDim.x*blockIdx.x + threadIdx.x;

    if (i >= k) {
        return;
    }
    dst[i] = __hadd(x[i], __float2half(y[i]));
}

static __global__ void mul_f32(const float * x, const float * y, float * dst, const int kx, const int ky) {
    const int i = blockDim.x*blockIdx.x + threadIdx.x;

    if (i >= kx) {
        return;
    }
    dst[i] = x[i] * y[i%ky];
}

static __global__ void silu_f32(const float * x, float * dst, const int k) {
    const int i = blockDim.x*blockIdx.x + threadIdx.x;

    if (i >= k) {
        return;
    }
    dst[i] = x[i] / (1.0f + expf(-x[i]));
}

static __global__ void rms_norm_f32(const float * x, float * dst, const int ncols) {
    const int row = blockIdx.x*blockDim.y + threadIdx.y;
    const int tid = threadIdx.x;

    const float eps = 1e-6;

    float tmp = 0.0f; // partial sum for thread in warp

    for (int i = 0; i < ncols; i += WARP_SIZE) {
        const int col = i + tid;
        const float xi = x[row*ncols + col];
        tmp += xi * xi;
    }

    // sum up partial sums
    __syncthreads();
#pragma unroll
    for (int mask = 16; mask > 0; mask >>= 1) {
        tmp += __shfl_xor_sync(0xffffffff, tmp, mask, 32);
    }

    const float mean = tmp / ncols;
    const float scale = 1.0f / sqrtf(mean + eps);

    for (int i = 0; i < ncols; i += WARP_SIZE) {
        const int col = i + tid;
        dst[row*ncols + col] = scale * x[row*ncols + col];
    }
}

static __device__ __forceinline__ void dequantize_q4_0(const void * vx, const int ib, const int iqs, dfloat2 & v){
    const block_q4_0 * x = (const block_q4_0 *) vx;

    const dfloat d = x[ib].d;

    const int vui = x[ib].qs[iqs];

    v.x = vui & 0xF;
    v.y = vui >> 4;

#ifdef GGML_CUDA_DMMV_F16
    v = __hsub2(v, {8.0f, 8.0f});
    v = __hmul2(v, {d, d});
#else
    v.x = (v.x - 8.0f) * d;
    v.y = (v.y - 8.0f) * d;
#endif // GGML_CUDA_DMMV_F16
}

static __device__ __forceinline__ void dequantize_q4_1(const void * vx, const int ib, const int iqs, dfloat2 & v){
    const block_q4_1 * x = (const block_q4_1 *) vx;

    const dfloat d = x[ib].d;
    const dfloat m = x[ib].m;

    const int vui = x[ib].qs[iqs];

    v.x = vui & 0xF;
    v.y = vui >> 4;

#ifdef GGML_CUDA_DMMV_F16
    v = __hmul2(v, {d, d});
    v = __hadd2(v, {m, m});
#else
    v.x = (v.x * d) + m;
    v.y = (v.y * d) + m;
#endif // GGML_CUDA_DMMV_F16
}

static __device__ __forceinline__ void dequantize_q5_0(const void * vx, const int ib, const int iqs, dfloat2 & v){
    const block_q5_0 * x = (const block_q5_0 *) vx;

    const dfloat d = x[ib].d;

    uint32_t qh;
    memcpy(&qh, x[ib].qh, sizeof(qh));

    const int xh_0 = ((qh >> (iqs +  0)) << 4) & 0x10;
    const int xh_1 = ((qh >> (iqs + 12))     ) & 0x10;

    v.x = ((x[ib].qs[iqs] & 0xf) | xh_0);
    v.y = ((x[ib].qs[iqs] >>  4) | xh_1);

#ifdef GGML_CUDA_DMMV_F16
    v = __hsub2(v, {16.0f, 16.0f});
    v = __hmul2(v, {d, d});
#else
    v.x = (v.x - 16.0f) * d;
    v.y = (v.y - 16.0f) * d;
#endif // GGML_CUDA_DMMV_F16
}

static __device__ __forceinline__ void dequantize_q5_1(const void * vx, const int ib, const int iqs, dfloat2 & v){
    const block_q5_1 * x = (const block_q5_1 *) vx;

    const dfloat d = x[ib].d;
    const dfloat m = x[ib].m;

    uint32_t qh;
    memcpy(&qh, x[ib].qh, sizeof(qh));

    const int xh_0 = ((qh >> (iqs +  0)) << 4) & 0x10;
    const int xh_1 = ((qh >> (iqs + 12))     ) & 0x10;

    v.x = ((x[ib].qs[iqs] & 0xf) | xh_0);
    v.y = ((x[ib].qs[iqs] >>  4) | xh_1);

#ifdef GGML_CUDA_DMMV_F16
    v = __hmul2(v, {d, d});
    v = __hadd2(v, {m, m});
#else
    v.x = (v.x * d) + m;
    v.y = (v.y * d) + m;
#endif // GGML_CUDA_DMMV_F16
}

static __device__ __forceinline__ void dequantize_q8_0(const void * vx, const int ib, const int iqs, dfloat2 & v){
    const block_q8_0 * x = (const block_q8_0 *) vx;

    const dfloat d = x[ib].d;

    v.x = x[ib].qs[iqs + 0];
    v.y = x[ib].qs[iqs + 1];

#ifdef GGML_CUDA_DMMV_F16
    v = __hmul2(v, {d, d});
#else
    v.x *= d;
    v.y *= d;
#endif // GGML_CUDA_DMMV_F16
}

//================================== k-quants

static __global__ void dequantize_block_q2_K(const void * vx, float * yy) {

    const int i   = blockIdx.x;
    const block_q2_K * x = (const block_q2_K *) vx;

    const int tid = threadIdx.x;
#if QK_K == 256
    const int n   = tid/32;
    const int l   = tid - 32*n;
    const int is  = 8*n + l/16;

    const uint8_t q = x[i].qs[32*n + l];
    float * y = yy + i*QK_K + 128*n;

    float dall = x[i].d;
    float dmin = x[i].dmin;
    y[l+ 0] = dall * (x[i].scales[is+0] & 0xF) * ((q >> 0) & 3) - dmin * (x[i].scales[is+0] >> 4);
    y[l+32] = dall * (x[i].scales[is+2] & 0xF) * ((q >> 2) & 3) - dmin * (x[i].scales[is+2] >> 4);
    y[l+64] = dall * (x[i].scales[is+4] & 0xF) * ((q >> 4) & 3) - dmin * (x[i].scales[is+4] >> 4);
    y[l+96] = dall * (x[i].scales[is+6] & 0xF) * ((q >> 6) & 3) - dmin * (x[i].scales[is+6] >> 4);
#else
    const int is = tid/16;  // 0 or 1
    const int il = tid%16;  // 0...15
    const uint8_t q = x[i].qs[il] >> (2*is);
    float * y = yy + i*QK_K + 16*is + il;
    float dall = x[i].d;
    float dmin = x[i].dmin;
    y[ 0] = dall * (x[i].scales[is+0] & 0xF) * ((q >> 0) & 3) - dmin * (x[i].scales[is+0] >> 4);
    y[32] = dall * (x[i].scales[is+2] & 0xF) * ((q >> 4) & 3) - dmin * (x[i].scales[is+2] >> 4);
#endif

}

static __global__ void dequantize_block_q3_K(const void * vx, float * yy) {

    const int i = blockIdx.x;
    const block_q3_K * x = (const block_q3_K *) vx;

#if QK_K == 256
    const int r = threadIdx.x/4;
    const int tid = r/2;
    const int is0 = r%2;
    const int l0 = 16*is0 + 4*(threadIdx.x%4);
    const int n = tid / 4;
    const int j = tid - 4*n;

    uint8_t m = 1 << (4*n + j);
    int is = 8*n + 2*j + is0;
    int shift = 2*j;

    int8_t us = is <  4 ? (x[i].scales[is-0] & 0xF) | (((x[i].scales[is+8] >> 0) & 3) << 4) :
                is <  8 ? (x[i].scales[is-0] & 0xF) | (((x[i].scales[is+4] >> 2) & 3) << 4) :
                is < 12 ? (x[i].scales[is-8] >>  4) | (((x[i].scales[is+0] >> 4) & 3) << 4) :
                          (x[i].scales[is-8] >>  4) | (((x[i].scales[is-4] >> 6) & 3) << 4);
    float d_all = x[i].d;
    float dl = d_all * (us - 32);

    float * y = yy + i*QK_K + 128*n + 32*j;
    const uint8_t * q = x[i].qs + 32*n;
    const uint8_t * hm = x[i].hmask;

    for (int l = l0; l < l0+4; ++l) y[l] = dl * ((int8_t)((q[l] >> shift) & 3) - ((hm[l] & m) ? 0 : 4));
#else
    const int tid = threadIdx.x;
    const int is  = tid/16;  // 0 or 1
    const int il  = tid%16;  // 0...15
    const int im  = il/8;    // 0...1
    const int in  = il%8;    // 0...7

    float * y = yy + i*QK_K + 16*is + il;

    const uint8_t q = x[i].qs[il] >> (2*is);
    const uint8_t h = x[i].hmask[in] >> (2*is + im);
    const float   d = (float)x[i].d;

    if (is == 0) {
        y[ 0] = d * ((x[i].scales[0] & 0xF) - 8) * ((int8_t)((q >> 0) & 3) - ((h >> 0) & 1 ? 0 : 4));
        y[32] = d * ((x[i].scales[1] & 0xF) - 8) * ((int8_t)((q >> 4) & 3) - ((h >> 4) & 1 ? 0 : 4));
    } else {
        y[ 0] = d * ((x[i].scales[0] >>  4) - 8) * ((int8_t)((q >> 0) & 3) - ((h >> 0) & 1 ? 0 : 4));
        y[32] = d * ((x[i].scales[1] >>  4) - 8) * ((int8_t)((q >> 4) & 3) - ((h >> 4) & 1 ? 0 : 4));
    }
#endif

}

#if QK_K == 256
static inline __device__ void get_scale_min_k4(int j, const uint8_t * q, uint8_t & d, uint8_t & m) {
    if (j < 4) {
        d = q[j] & 63; m = q[j + 4] & 63;
    } else {
        d = (q[j+4] & 0xF) | ((q[j-4] >> 6) << 4);
        m = (q[j+4] >>  4) | ((q[j-0] >> 6) << 4);
    }
}
#endif

static __global__ void dequantize_block_q4_K(const void * vx, float * yy) {
    const block_q4_K * x = (const block_q4_K *) vx;

    const int i = blockIdx.x;

#if QK_K == 256
    // assume 32 threads
    const int tid = threadIdx.x;
    const int il  = tid/8;
    const int ir  = tid%8;
    const int is  = 2*il;
    const int n   = 4;

    float * y = yy + i*QK_K + 64*il + n*ir;

    const float dall = x[i].d;
    const float dmin = x[i].dmin;

    const uint8_t * q = x[i].qs + 32*il + n*ir;

    uint8_t sc, m;
    get_scale_min_k4(is + 0, x[i].scales, sc, m);
    const float d1 = dall * sc; const float m1 = dmin * m;
    get_scale_min_k4(is + 1, x[i].scales, sc, m);
    const float d2 = dall * sc; const float m2 = dmin * m;
    for (int l = 0; l < n; ++l) {
        y[l + 0] = d1 * (q[l] & 0xF) - m1;
        y[l +32] = d2 * (q[l] >>  4) - m2;
    }
#else
    const int tid = threadIdx.x;
    const uint8_t * q = x[i].qs;
    float * y = yy + i*QK_K;
    const float d = (float)x[i].d[0];
    const float m = (float)x[i].d[1];
    y[tid+ 0] = d * (x[i].scales[0] & 0xF) * (q[tid] & 0xF) - m * (x[i].scales[0] >> 4);
    y[tid+32] = d * (x[i].scales[1] & 0xF) * (q[tid] >>  4) - m * (x[i].scales[1] >> 4);
#endif
}

static __global__ void dequantize_block_q5_K(const void * vx, float * yy) {
    const block_q5_K * x = (const block_q5_K *) vx;

    const int i = blockIdx.x;

#if QK_K == 256
    // assume 64 threads - this is very slightly better than the one below
    const int tid = threadIdx.x;
    const int il  = tid/16;   // il is in 0...3
    const int ir  = tid%16;   // ir is in 0...15
    const int is  = 2*il;     // is is in 0...6

    float * y = yy + i*QK_K + 64*il + 2*ir;

    const float dall = x[i].d;
    const float dmin = x[i].dmin;

    const uint8_t * ql = x[i].qs + 32*il + 2*ir;
    const uint8_t * qh = x[i].qh + 2*ir;

    uint8_t sc, m;
    get_scale_min_k4(is + 0, x[i].scales, sc, m);
    const float d1 = dall * sc; const float m1 = dmin * m;
    get_scale_min_k4(is + 1, x[i].scales, sc, m);
    const float d2 = dall * sc; const float m2 = dmin * m;

    uint8_t   hm  = 1 << (2*il);
    y[ 0] = d1 * ((ql[ 0] & 0xF) + (qh[ 0] & hm ? 16 : 0)) - m1;
    y[ 1] = d1 * ((ql[ 1] & 0xF) + (qh[ 1] & hm ? 16 : 0)) - m1;
    hm <<= 1;
    y[32] = d2 * ((ql[ 0] >>  4) + (qh[ 0] & hm ? 16 : 0)) - m2;
    y[33] = d2 * ((ql[ 1] >>  4) + (qh[ 1] & hm ? 16 : 0)) - m2;
#else
    const int tid = threadIdx.x;
    const uint8_t q = x[i].qs[tid];
    const int im = tid/8;  // 0...3
    const int in = tid%8;  // 0...7
    const int is = tid/16; // 0 or 1
    const uint8_t h = x[i].qh[in] >> im;
    const float d = x[i].d;
    float * y = yy + i*QK_K + tid;
    y[ 0] = d * x[i].scales[is+0] * ((q & 0xF) - ((h >> 0) & 1 ? 0 : 16));
    y[32] = d * x[i].scales[is+2] * ((q >>  4) - ((h >> 4) & 1 ? 0 : 16));
#endif
}

static __global__ void dequantize_block_q6_K(const void * vx, float * yy) {
    const block_q6_K * x = (const block_q6_K *) vx;

    const int i = blockIdx.x;
#if QK_K == 256

    // assume 64 threads - this is very slightly better than the one below
    const int tid = threadIdx.x;
    const int ip  = tid/32;   // ip is 0 or 1
    const int il  = tid - 32*ip; // 0...32
    const int is  = 8*ip + il/16;

    float * y = yy + i*QK_K + 128*ip + il;

    const float d = x[i].d;

    const uint8_t * ql = x[i].ql + 64*ip + il;
    const uint8_t   qh = x[i].qh[32*ip + il];
    const int8_t  * sc = x[i].scales + is;

    y[ 0] = d * sc[0] * ((int8_t)((ql[ 0] & 0xF) | (((qh >> 0) & 3) << 4)) - 32);
    y[32] = d * sc[2] * ((int8_t)((ql[32] & 0xF) | (((qh >> 2) & 3) << 4)) - 32);
    y[64] = d * sc[4] * ((int8_t)((ql[ 0]  >> 4) | (((qh >> 4) & 3) << 4)) - 32);
    y[96] = d * sc[6] * ((int8_t)((ql[32]  >> 4) | (((qh >> 6) & 3) << 4)) - 32);
#else

    // assume 32 threads
    const int tid = threadIdx.x;
    const int ip  = tid/16;         // 0 or 1
    const int il  = tid - 16*ip;    // 0...15

    float * y = yy + i*QK_K + 16*ip + il;

    const float d = x[i].d;

    const uint8_t   ql = x[i].ql[16*ip + il];
    const uint8_t   qh = x[i].qh[il] >> (2*ip);
    const int8_t  * sc = x[i].scales;

    y[ 0] = d * sc[ip+0] * ((int8_t)((ql & 0xF) | (((qh >> 0) & 3) << 4)) - 32);
    y[32] = d * sc[ip+2] * ((int8_t)((ql  >> 4) | (((qh >> 4) & 3) << 4)) - 32);
#endif
}

static __global__ void dequantize_mul_mat_vec_q2_k(const void * vx, const float * yy, float * dst, const int ncols, int nrows) {

    static_assert(16%K_QUANTS_PER_ITERATION == 0, "16 must be divisible by K_QUANTS_PER_ITERATION");

    const int row = blockIdx.y*blockDim.y + threadIdx.y;
    if (row > nrows) return;

    const int num_blocks_per_row = ncols / QK_K;
    const int ib0 = row*num_blocks_per_row;

    const block_q2_K * x = (const block_q2_K *)vx + ib0;

    float tmp = 0; // partial sum for thread in warp

#if QK_K == 256
    const int tid = threadIdx.x/K_QUANTS_PER_ITERATION;  // 0...31 or 0...15
    const int ix  = threadIdx.x%K_QUANTS_PER_ITERATION;  // 0 or 0,1

    const int step = 16/K_QUANTS_PER_ITERATION;

    const int im = tid/step;                             // 0 or 1. 0 computes 0..., 1 computes 128...
    const int in = tid - step*im;                        // 0...15 or 0...7

    const int l0 = K_QUANTS_PER_ITERATION*in;            // 0...15 or 0...14 in steps of 2
    const int q_offset = 32*im + l0;
    const int s_offset = 8*im;
    const int y_offset = 128*im + l0;

    uint32_t aux[4];
    const uint8_t * d = (const uint8_t *)aux;
    const uint8_t * m = (const uint8_t *)(aux + 2);

    for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {

        const float   * y = yy + i * QK_K + y_offset;
        const uint8_t * q = x[i].qs + q_offset;

        const float dall = x[i].d;
        const float dmin = x[i].dmin;

        const uint32_t * a = (const uint32_t *)(x[i].scales + s_offset);
        aux[0] = a[0] & 0x0f0f0f0f;
        aux[1] = a[1] & 0x0f0f0f0f;
        aux[2] = (a[0] >> 4) & 0x0f0f0f0f;
        aux[3] = (a[1] >> 4) & 0x0f0f0f0f;

        float sum1 = 0, sum2 = 0;
        for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
            sum1 += y[l+ 0] * d[0] * ((q[l+ 0] >> 0) & 3)
                  + y[l+32] * d[2] * ((q[l+ 0] >> 2) & 3)
                  + y[l+64] * d[4] * ((q[l+ 0] >> 4) & 3)
                  + y[l+96] * d[6] * ((q[l+ 0] >> 6) & 3)
                  + y[l+16] * d[1] * ((q[l+16] >> 0) & 3)
                  + y[l+48] * d[3] * ((q[l+16] >> 2) & 3)
                  + y[l+80] * d[5] * ((q[l+16] >> 4) & 3)
                  +y[l+112] * d[7] * ((q[l+16] >> 6) & 3);
            sum2 += y[l+ 0] * m[0] + y[l+32] * m[2] + y[l+64] * m[4] + y[ l+96] * m[6]
                  + y[l+16] * m[1] + y[l+48] * m[3] + y[l+80] * m[5] + y[l+112] * m[7];

        }
        tmp += dall * sum1 - dmin * sum2;

    }
#else
    const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION);  // 0...15 or 0...7
    const int ix  = threadIdx.x%(2*K_QUANTS_PER_ITERATION);  // 0....1 or 0...3
    const int offset = tid * K_QUANTS_PER_ITERATION;

    uint32_t uaux[2];
    const uint8_t * d = (const uint8_t *)uaux;

    for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) {

        const float   * y = yy + i * QK_K + offset;
        const uint8_t * q = x[i].qs + offset;
        const uint32_t * s = (const uint32_t *)x[i].scales;

        uaux[0] = s[0] & 0x0f0f0f0f;
        uaux[1] = (s[0] >> 4) & 0x0f0f0f0f;

        const half2 * dh = (const half2 *)&x[i].d;

        const float2 dall = __half22float2(dh[0]);

        float sum1 = 0, sum2 = 0;
        for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
            const uint8_t ql = q[l];
            sum1 += y[l+ 0] * d[0] * ((ql >> 0) & 3)
                  + y[l+16] * d[1] * ((ql >> 2) & 3)
                  + y[l+32] * d[2] * ((ql >> 4) & 3)
                  + y[l+48] * d[3] * ((ql >> 6) & 3);
            sum2 += y[l+0] * d[4] + y[l+16] * d[5] + y[l+32] * d[6] + y[l+48] * d[7];
        }
        tmp += dall.x * sum1 - dall.y * sum2;
    }
#endif

    // sum up partial sums and write back result
    __syncthreads();
#pragma unroll
    for (int mask = 16; mask > 0; mask >>= 1) {
        tmp += __shfl_xor_sync(0xffffffff, tmp, mask, 32);
    }

    if (threadIdx.x == 0) {
        dst[row] = tmp;
    }
}

static __global__ void dequantize_mul_mat_vec_q3_k(const void * vx, const float * yy, float * dst, const int ncols, int nrows) {

    const int row = blockIdx.y*blockDim.y + threadIdx.y;
    if (row > nrows) return;

    const int num_blocks_per_row = ncols / QK_K;
    const int ib0 = row*num_blocks_per_row;

    const block_q3_K * x = (const block_q3_K *)vx + ib0;

    float tmp = 0; // partial sum for thread in warp

#if QK_K == 256

    const uint16_t kmask1 = 0x0303;
    const uint16_t kmask2 = 0x0f0f;

    const int tid = threadIdx.x/K_QUANTS_PER_ITERATION;  // 0...31 or 0...16
    const int ix  = threadIdx.x%K_QUANTS_PER_ITERATION;  // 0 or 0,1

    const int n  = K_QUANTS_PER_ITERATION;               // iterations in the inner loop
    const int step = 16/K_QUANTS_PER_ITERATION;
    const int im = tid/step;                             // 0 or 1. 0 computes 0..., 1 computes 128...
    const int in = tid - step*im;                        // 0....15 or 0...7

    const uint8_t m = 1 << (4*im);

    const int l0 = n*in;                                 // 0...15 or 0...14 in steps of 2
    const int q_offset =  32*im + l0;
    const int y_offset = 128*im + l0;

    uint16_t utmp[4];
    const int8_t * s = (const int8_t *)utmp;

    const uint16_t s_shift = 4*im;

    for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {

        const float   * y  = yy + i * QK_K + y_offset;
        const uint8_t * q = x[i].qs + q_offset;
        const uint8_t * h = x[i].hmask + l0;

        const uint16_t * a = (const uint16_t *)x[i].scales;
        utmp[0] = ((a[0] >> s_shift) & kmask2) | (((a[4] >> (s_shift + 0)) & kmask1) << 4);
        utmp[1] = ((a[1] >> s_shift) & kmask2) | (((a[5] >> (s_shift + 0)) & kmask1) << 4);
        utmp[2] = ((a[2] >> s_shift) & kmask2) | (((a[4] >> (s_shift + 2)) & kmask1) << 4);
        utmp[3] = ((a[3] >> s_shift) & kmask2) | (((a[5] >> (s_shift + 2)) & kmask1) << 4);

        const float d = x[i].d;

        float sum = 0;
        for (int l = 0; l < n; ++l) {
            sum += y[l+ 0] * (s[0] - 32) * (((q[l] >> 0) & 3) - (h[l] & (m << 0) ? 0 : 4))
                 + y[l+32] * (s[2] - 32) * (((q[l] >> 2) & 3) - (h[l] & (m << 1) ? 0 : 4))
                 + y[l+64] * (s[4] - 32) * (((q[l] >> 4) & 3) - (h[l] & (m << 2) ? 0 : 4))
                 + y[l+96] * (s[6] - 32) * (((q[l] >> 6) & 3) - (h[l] & (m << 3) ? 0 : 4));
            sum += y[l+16] * (s[1] - 32) * (((q[l+16] >> 0) & 3) - (h[l+16] & (m << 0) ? 0 : 4))
                 + y[l+48] * (s[3] - 32) * (((q[l+16] >> 2) & 3) - (h[l+16] & (m << 1) ? 0 : 4))
                 + y[l+80] * (s[5] - 32) * (((q[l+16] >> 4) & 3) - (h[l+16] & (m << 2) ? 0 : 4))
                + y[l+112] * (s[7] - 32) * (((q[l+16] >> 6) & 3) - (h[l+16] & (m << 3) ? 0 : 4));
        }
        tmp += d * sum;

    }
#else

    const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION);  // 0...15 or 0...7
    const int ix  = threadIdx.x%(2*K_QUANTS_PER_ITERATION);  // 0....1 or 0...3
    const int offset = tid * K_QUANTS_PER_ITERATION;         // 0...15 or 0...14
    const int in = offset/8;                                 // 0 or 1
    const int im = offset%8;                                 // 0...7

    for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) {

        const float   * y = yy + i * QK_K + offset;
        const uint8_t * q = x[i].qs + offset;
        const uint8_t * s = x[i].scales;

        const float dall = (float)x[i].d;

        float sum = 0;
        for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
            const uint8_t hl = x[i].hmask[im+l] >> in;
            const uint8_t ql = q[l];
            sum += y[l+ 0] * dall * ((s[0] & 0xF) - 8) * ((int8_t)((ql >> 0) & 3) - ((hl >> 0) & 1 ? 0 : 4))
                 + y[l+16] * dall * ((s[0] >>  4) - 8) * ((int8_t)((ql >> 2) & 3) - ((hl >> 2) & 1 ? 0 : 4))
                 + y[l+32] * dall * ((s[1] & 0xF) - 8) * ((int8_t)((ql >> 4) & 3) - ((hl >> 4) & 1 ? 0 : 4))
                 + y[l+48] * dall * ((s[1] >>  4) - 8) * ((int8_t)((ql >> 6) & 3) - ((hl >> 6) & 1 ? 0 : 4));
        }
        tmp += sum;
    }
#endif

    // sum up partial sums and write back result
    __syncthreads();
#pragma unroll
    for (int mask = 16; mask > 0; mask >>= 1) {
        tmp += __shfl_xor_sync(0xffffffff, tmp, mask, 32);
    }

    if (threadIdx.x == 0) {
        dst[row] = tmp;
    }
}

static __global__ void dequantize_mul_mat_vec_q4_k(const void * vx, const float * yy, float * dst, const int ncols, int nrows) {

    const int row = blockIdx.y*blockDim.y + threadIdx.y;
    if (row > nrows) return;
    const int num_blocks_per_row = ncols / QK_K;
    const int ib0 = row*num_blocks_per_row;

    const block_q4_K * x = (const block_q4_K *)vx + ib0;

#if QK_K == 256
    const uint16_t kmask1 = 0x3f3f;
    const uint16_t kmask2 = 0x0f0f;
    const uint16_t kmask3 = 0xc0c0;

    const int tid = threadIdx.x/K_QUANTS_PER_ITERATION;  // 0...31 or 0...16
    const int ix  = threadIdx.x%K_QUANTS_PER_ITERATION;  // 0 or 0,1

    const int step = 8/K_QUANTS_PER_ITERATION;           // 8 or 4

    const int il  = tid/step;                            // 0...3
    const int ir  = tid - step*il;                       // 0...7 or 0...3
    const int n   = 2 * K_QUANTS_PER_ITERATION;          // 2 or 4

    const int im = il/2;  // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224
    const int in = il%2;

    const int l0 = n*(2*ir + in);
    const int q_offset = 32*im + l0;
    const int y_offset = 64*im + l0;

    uint16_t aux[4];
    const uint8_t * sc = (const uint8_t *)aux;

    float tmp = 0; // partial sum for thread in warp

    for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {

        const uint8_t * q1 = x[i].qs + q_offset;
        const uint8_t * q2 = q1 + 64;
        const float   * y1 = yy + i*QK_K + y_offset;
        const float   * y2 = y1 + 128;

        const float dall = x[i].d;
        const float dmin = x[i].dmin;

        const uint16_t * a = (const uint16_t *)x[i].scales;
        aux[0] = a[im+0] & kmask1;
        aux[1] = a[im+2] & kmask1;
        aux[2] = ((a[im+4] >> 0) & kmask2) | ((a[im+0] & kmask3) >> 2);
        aux[3] = ((a[im+4] >> 4) & kmask2) | ((a[im+2] & kmask3) >> 2);

        float4 s = {0.f, 0.f, 0.f, 0.f};
        float smin = 0;
        for (int l = 0; l < n; ++l) {
            s.x += y1[l] * (q1[l] & 0xF); s.y += y1[l+32] * (q1[l] >> 4);
            s.z += y2[l] * (q2[l] & 0xF); s.w += y2[l+32] * (q2[l] >> 4);
            smin += y1[l] * sc[2] + y1[l+32] * sc[3] + y2[l] * sc[6] + y2[l+32] * sc[7];
        }
        tmp += dall * (s.x * sc[0] + s.y * sc[1] + s.z * sc[4] + s.w * sc[5]) - dmin * smin;

    }
#else
    const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION);  // 0...15
    const int ix  = threadIdx.x%(2*K_QUANTS_PER_ITERATION);

    const int step = tid * K_QUANTS_PER_ITERATION;

    uint16_t aux16[2];
    const uint8_t * s = (const uint8_t *)aux16;

    float tmp = 0;

    for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) {
        const uint8_t * q = x[i].qs + step;
        const float   * y = yy + i*QK_K + step;
        const uint16_t * a = (const uint16_t *)x[i].scales;
        aux16[0] = a[0] & 0x0f0f;
        aux16[1] = (a[0] >> 4) & 0x0f0f;
        const float d = (float)x[i].d[0];
        const float m = (float)x[i].d[1];
        float sum = 0.f;
        for (int j = 0; j < K_QUANTS_PER_ITERATION; ++j) {
            sum += y[j+ 0] * (d * s[0] * (q[j+ 0] & 0xF) - m * s[2])
                 + y[j+16] * (d * s[0] * (q[j+16] & 0xF) - m * s[2])
                 + y[j+32] * (d * s[1] * (q[j+ 0] >>  4) - m * s[3])
                 + y[j+48] * (d * s[1] * (q[j+16] >>  4) - m * s[3]);
        }
        tmp += sum;
    }

#endif

    // sum up partial sums and write back result
    __syncthreads();
#pragma unroll
    for (int mask = 16; mask > 0; mask >>= 1) {
        tmp += __shfl_xor_sync(0xffffffff, tmp, mask, 32);
    }

    if (tid == 0) {
        dst[row] = tmp;
    }
}

static __global__ void dequantize_mul_mat_vec_q5_k(const void * vx, const float * yy, float * dst, const int ncols) {

    const int row = blockIdx.x;
    const int num_blocks_per_row = ncols / QK_K;
    const int ib0 = row*num_blocks_per_row;

    const block_q5_K * x = (const block_q5_K *)vx + ib0;

    float tmp = 0; // partial sum for thread in warp

#if QK_K == 256
    const uint16_t kmask1 = 0x3f3f;
    const uint16_t kmask2 = 0x0f0f;
    const uint16_t kmask3 = 0xc0c0;

    const int tid = threadIdx.x/2;  // 0...15
    const int ix  = threadIdx.x%2;

    const int il  = tid/4;     // 0...3
    const int ir  = tid - 4*il;// 0...3
    const int n   = 2;

    const int im = il/2;  // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224
    const int in = il%2;

    const int l0 = n*(2*ir + in);
    const int q_offset = 32*im + l0;
    const int y_offset = 64*im + l0;

    const uint8_t hm1  = 1 << (2*im);
    const uint8_t hm2  = hm1 << 4;

    uint16_t aux[4];
    const uint8_t * sc = (const uint8_t *)aux;

    for (int i = ix; i < num_blocks_per_row; i += 2) {

        const uint8_t * ql1 = x[i].qs + q_offset;
        const uint8_t * ql2 = ql1 + 64;
        const uint8_t * qh  = x[i].qh + l0;
        const float   * y1  = yy + i*QK_K + y_offset;
        const float   * y2  = y1 + 128;

        const float dall = x[i].d;
        const float dmin = x[i].dmin;

        const uint16_t * a = (const uint16_t *)x[i].scales;
        aux[0] = a[im+0] & kmask1;
        aux[1] = a[im+2] & kmask1;
        aux[2] = ((a[im+4] >> 0) & kmask2) | ((a[im+0] & kmask3) >> 2);
        aux[3] = ((a[im+4] >> 4) & kmask2) | ((a[im+2] & kmask3) >> 2);

        float4 sum = {0.f, 0.f, 0.f, 0.f};
        float smin = 0;
        for (int l = 0; l < n; ++l) {
            sum.x += y1[l+ 0] * ((ql1[l+ 0] & 0xF) + (qh[l+ 0] & (hm1 << 0) ? 16 : 0))
                   + y1[l+16] * ((ql1[l+16] & 0xF) + (qh[l+16] & (hm1 << 0) ? 16 : 0));
            sum.y += y1[l+32] * ((ql1[l+ 0] >>  4) + (qh[l+ 0] & (hm1 << 1) ? 16 : 0))
                   + y1[l+48] * ((ql1[l+16] >>  4) + (qh[l+16] & (hm1 << 1) ? 16 : 0));
            sum.z += y2[l+ 0] * ((ql2[l+ 0] & 0xF) + (qh[l+ 0] & (hm2 << 0) ? 16 : 0))
                   + y2[l+16] * ((ql2[l+16] & 0xF) + (qh[l+16] & (hm2 << 0) ? 16 : 0));
            sum.w += y2[l+32] * ((ql2[l+ 0] >>  4) + (qh[l+ 0] & (hm2 << 1) ? 16 : 0))
                   + y2[l+48] * ((ql2[l+16] >>  4) + (qh[l+16] & (hm2 << 1) ? 16 : 0));
            smin += (y1[l] + y1[l+16]) * sc[2] + (y1[l+32] + y1[l+48]) * sc[3]
                  + (y2[l] + y2[l+16]) * sc[6] + (y2[l+32] + y2[l+48]) * sc[7];
        }
        tmp += dall * (sum.x * sc[0] + sum.y * sc[1] + sum.z * sc[4] + sum.w * sc[5]) - dmin * smin;
    }

#else
    const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION);  // 0...15
    const int ix  = threadIdx.x%(2*K_QUANTS_PER_ITERATION);
    const int step = tid * K_QUANTS_PER_ITERATION;
    const int im = step/8;
    const int in = step%8;

    for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) {
        const uint8_t * q = x[i].qs + step;
        const int8_t  * s = x[i].scales;
        const float   * y = yy + i*QK_K + step;
        const float     d = x[i].d;
        float sum = 0.f;
        for (int j = 0; j < K_QUANTS_PER_ITERATION; ++j) {
            const uint8_t h = x[i].qh[in+j] >> im;
            sum += y[j+ 0] * d * s[0] * ((q[j+ 0] & 0xF) - ((h >> 0) & 1 ? 0 : 16))
                 + y[j+16] * d * s[1] * ((q[j+16] & 0xF) - ((h >> 2) & 1 ? 0 : 16))
                 + y[j+32] * d * s[2] * ((q[j+ 0] >>  4) - ((h >> 4) & 1 ? 0 : 16))
                 + y[j+48] * d * s[3] * ((q[j+16] >>  4) - ((h >> 6) & 1 ? 0 : 16));
        }
        tmp += sum;
    }
#endif

    // sum up partial sums and write back result
    __syncthreads();
#pragma unroll
    for (int mask = 16; mask > 0; mask >>= 1) {
        tmp += __shfl_xor_sync(0xffffffff, tmp, mask, 32);
    }

    if (threadIdx.x == 0) {
        dst[row] = tmp;
    }
}

static __global__ void dequantize_mul_mat_vec_q6_k(const void * vx, const float * yy, float * dst, const int ncols, int nrows) {

    static_assert(16%K_QUANTS_PER_ITERATION == 0, "16 must be divisible by K_QUANTS_PER_ITERATION");

    const int row = blockIdx.y*blockDim.y + threadIdx.y;
    if (row > nrows) return;

    const int num_blocks_per_row = ncols / QK_K;
    const int ib0 = row*num_blocks_per_row;

    const block_q6_K * x = (const block_q6_K *)vx + ib0;

#if QK_K == 256

    const int tid = threadIdx.x/K_QUANTS_PER_ITERATION;  // 0...31 or 0...16
    const int ix  = threadIdx.x%K_QUANTS_PER_ITERATION;  // 0 or 0, 1

    const int step = 16/K_QUANTS_PER_ITERATION;          // 16 or 8

    const int im = tid/step;                             // 0 or 1. 0 computes 0..., 1 computes 128...
    const int in = tid - step*im;                        // 0...15 or 0...7

#if K_QUANTS_PER_ITERATION == 1
    const int l0 = K_QUANTS_PER_ITERATION*in;            // 0...15
    const int is = 0;
#else
    const int l0 = 4 * in;                               // 0, 4, 8, ..., 28
    const int is = in / 4;
#endif
    const int ql_offset = 64*im + l0;
    const int qh_offset = 32*im + l0;
    const int s_offset  =  8*im + is;
    const int y_offset = 128*im + l0;

    float tmp = 0; // partial sum for thread in warp

    for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {

        const float   * y  = yy + i * QK_K + y_offset;
        const uint8_t * ql = x[i].ql + ql_offset;
        const uint8_t * qh = x[i].qh + qh_offset;
        const int8_t  * s  = x[i].scales + s_offset;

        const float d = x[i].d;

#if K_QUANTS_PER_ITERATION == 1
        float sum = y[ 0] * s[0] * d * ((int8_t)((ql[ 0] & 0xF) | ((qh[ 0] & 0x03) << 4)) - 32)
                  + y[16] * s[1] * d * ((int8_t)((ql[16] & 0xF) | ((qh[16] & 0x03) << 4)) - 32)
                  + y[32] * s[2] * d * ((int8_t)((ql[32] & 0xF) | ((qh[ 0] & 0x0c) << 2)) - 32)
                  + y[48] * s[3] * d * ((int8_t)((ql[48] & 0xF) | ((qh[16] & 0x0c) << 2)) - 32)
                  + y[64] * s[4] * d * ((int8_t)((ql[ 0]  >> 4) | ((qh[ 0] & 0x30) >> 0)) - 32)
                  + y[80] * s[5] * d * ((int8_t)((ql[16]  >> 4) | ((qh[16] & 0x30) >> 0)) - 32)
                  + y[96] * s[6] * d * ((int8_t)((ql[32]  >> 4) | ((qh[ 0] & 0xc0) >> 2)) - 32)
                  +y[112] * s[7] * d * ((int8_t)((ql[48]  >> 4) | ((qh[16] & 0xc0) >> 2)) - 32);
        tmp += sum;
#else
        float sum = 0;
        for (int l = 0; l < 4; ++l) {
            sum += y[l+ 0] * s[0] * d * ((int8_t)((ql[l+ 0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32)
                 + y[l+32] * s[2] * d * ((int8_t)((ql[l+32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32)
                 + y[l+64] * s[4] * d * ((int8_t)((ql[l+ 0]  >> 4) | (((qh[l] >> 4) & 3) << 4)) - 32)
                 + y[l+96] * s[6] * d * ((int8_t)((ql[l+32]  >> 4) | (((qh[l] >> 6) & 3) << 4)) - 32);
        }
        tmp += sum;
#endif

    }

#else

    const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION);  // 0...7
    const int ix  = threadIdx.x%(2*K_QUANTS_PER_ITERATION);  // 0...3

    const int step = tid * K_QUANTS_PER_ITERATION;

    float tmp = 0; // partial sum for thread in warp

    for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) {

        const float   * y  = yy + i * QK_K + step;
        const uint8_t * ql = x[i].ql + step;
        const uint8_t * qh = x[i].qh + step;
        const int8_t  * s  = x[i].scales;

        const float d = x[i+0].d;

        float sum = 0;
        for (int j = 0; j < K_QUANTS_PER_ITERATION; ++j) {
            sum += y[j+ 0] * s[0] * d * ((int8_t)((ql[j+ 0] & 0xF) | ((qh[j] & 0x03) << 4)) - 32)
                 + y[j+16] * s[1] * d * ((int8_t)((ql[j+16] & 0xF) | ((qh[j] & 0x0c) << 2)) - 32)
                 + y[j+32] * s[2] * d * ((int8_t)((ql[j+ 0] >>  4) | ((qh[j] & 0x30) >> 0)) - 32)
                 + y[j+48] * s[3] * d * ((int8_t)((ql[j+16] >>  4) | ((qh[j] & 0xc0) >> 2)) - 32);
        }
        tmp += sum;

    }

#endif

    // sum up partial sums and write back result
    __syncthreads();
#pragma unroll
    for (int mask = 16; mask > 0; mask >>= 1) {
        tmp += __shfl_xor_sync(0xffffffff, tmp, mask, 32);
    }

    if (tid == 0) {
        dst[row] = tmp;
    }
}

static __device__ void convert_f16(const void * vx, const int ib, const int iqs, dfloat2 & v){
    const half * x = (const half *) vx;

    // automatic half -> float type cast if dfloat == float
    v.x = x[ib + iqs + 0];
    v.y = x[ib + iqs + 1];
}

template <int qk, int qr, dequantize_kernel_t dequantize_kernel>
static __global__ void dequantize_block(const void * vx, float * y, const int k) {
    const int i = blockDim.x*blockIdx.x + 2*threadIdx.x;

    if (i >= k) {
        return;
    }

    const int ib = i/qk; // block index
    const int iqs = (i%qk)/qr; // quant index
    const int iybs = i - i%qk; // y block start index
    const int y_offset = qr == 1 ? 1 : qk/2;

    // dequantize
    dfloat2 v;
    dequantize_kernel(vx, ib, iqs, v);

    y[iybs + iqs + 0]        = v.x;
    y[iybs + iqs + y_offset] = v.y;
}

template <int qk, int qr, dequantize_kernel_t dequantize_kernel>
static __global__ void dequantize_mul_mat_vec(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows) {
    // qk = quantized weights per x block
    // qr = number of quantized weights per data value in x block
    const int row = blockIdx.y*blockDim.y + threadIdx.y;

    if (row >= nrows) {
        return;
    }

    const int tid = threadIdx.x;

    const int iter_stride = 2*GGML_CUDA_DMMV_X;
    const int vals_per_iter = iter_stride / WARP_SIZE; // num quantized vals per thread and i iter
    const int y_offset = qr == 1 ? 1 : qk/2;

// partial sum for each thread
#ifdef GGML_CUDA_DMMV_F16
    half2 tmp = {0.0f, 0.0f}; // two sums for f16 to take advantage of half2 intrinsics
#else
    float tmp = 0.0f;
#endif // GGML_CUDA_DMMV_F16

    for (int i = 0; i < ncols; i += iter_stride) {
        const int col = i + vals_per_iter*tid;
        const int ib = (row*ncols + col)/qk; // x block index
        const int iqs = (col%qk)/qr; // x quant index
        const int iybs = col - col%qk; // y block start index

// processing >2 values per i iter is faster for fast GPUs
#pragma unroll
        for (int j = 0; j < vals_per_iter; j += 2) {
            // process 2 vals per j iter

            // dequantize
            // for qr = 2 the iqs needs to increase by 1 per j iter because 2 weights per data val
            dfloat2 v;
            dequantize_kernel(vx, ib, iqs + j/qr, v);

            // matrix multiplication
            // for qr = 2 the y index needs to increase by 1 per j iter because of y_offset = qk/2
#ifdef GGML_CUDA_DMMV_F16
            tmp += __hmul2(v, {
                y[iybs + iqs + j/qr + 0],
                y[iybs + iqs + j/qr + y_offset]
            });
#else
            tmp += v.x * y[iybs + iqs + j/qr + 0];
            tmp += v.y * y[iybs + iqs + j/qr + y_offset];
#endif // GGML_CUDA_DMMV_F16
        }
    }

    // sum up partial sums and write back result
    __syncthreads();
#pragma unroll
    for (int mask = 16; mask > 0; mask >>= 1) {
        tmp += __shfl_xor_sync(0xffffffff, tmp, mask, 32);
    }

    if (tid == 0) {
#ifdef GGML_CUDA_DMMV_F16
        dst[row] = tmp.x + tmp.y;
#else
        dst[row] = tmp;
#endif // GGML_CUDA_DMMV_F16
    }
}

static __global__ void mul_mat_p021_f16_f32(const void * vx, const float * y, float * dst, const int ncols_x, const int nrows_x, const int nchannels_x) {
    const half * x = (half *) vx;

    const int row_x = blockDim.y*blockIdx.y + threadIdx.y;
    const int channel = blockDim.z*blockIdx.z + threadIdx.z;

    const int nrows_y = ncols_x;
    const int nrows_dst = nrows_x;
    const int row_dst = row_x;

    float tmp = 0.0f;

    for (int col_x0 = 0; col_x0 < ncols_x; col_x0 += blockDim.x) {
        const int col_x = col_x0 + threadIdx.x;

        if (col_x >= ncols_x) {
            break;
        }

        // x is transposed and permuted
        const int ix = row_x*nchannels_x*ncols_x + channel*ncols_x + col_x;
        const float xi = __half2float(x[ix]);

        const int row_y = col_x;


        // y is not transposed but permuted
        const int iy = channel*nrows_y + row_y;

        tmp += xi * y[iy];
    }

    // dst is not transposed and not permuted
    const int idst = channel*nrows_dst + row_dst;

    // sum up partial sums and write back result
    __syncthreads();
#pragma unroll
    for (int mask = 16; mask > 0; mask >>= 1) {
        tmp += __shfl_xor_sync(0xffffffff, tmp, mask, 32);
    }

    if (threadIdx.x == 0) {
        dst[idst] = tmp;
    }
}

static __global__ void mul_mat_vec_nc_f16_f32( // nc == non-contiguous
    const void * vx, const float * y, float * dst, const int ncols_x, const int nrows_x,
    const int row_stride_x, const int nchannels_x, const int channel_stride_x) {

    const half * x = (half *) vx;

    const int row_x = blockDim.y*blockIdx.y + threadIdx.y;
    const int channel = blockDim.z*blockIdx.z + threadIdx.z;

    const int nrows_y = ncols_x;
    const int nrows_dst = nrows_x;
    const int row_dst = row_x;

    const int idst = channel*nrows_dst + row_dst;

    float tmp = 0.0f;

    for (int col_x0 = 0; col_x0 < ncols_x; col_x0 += blockDim.x) {
        const int col_x = col_x0 + threadIdx.x;

        if (col_x >= ncols_x) {
            break;
        }

        const int ix = channel*channel_stride_x + row_x*row_stride_x + col_x;
        const float xi = __half2float(x[ix]);

        const int row_y = col_x;

        const int iy = channel*nrows_y + row_y;

        tmp += xi * y[iy];
    }

    // sum up partial sums and write back result
    __syncthreads();
#pragma unroll
    for (int mask = 16; mask > 0; mask >>= 1) {
        tmp += __shfl_xor_sync(0xffffffff, tmp, mask, 32);
    }

    if (threadIdx.x == 0) {
        dst[idst] = tmp;
    }
}

static __device__ void cpy_1_f32_f32(const char * cxi, char * cdsti) {
    const float * xi = (float *) cxi;
    float * dsti = (float *) cdsti;

    *dsti = *xi;
}

static __device__ void cpy_1_f32_f16(const char * cxi, char * cdsti) {
    const float * xi = (float *) cxi;
    half * dsti = (half *) cdsti;

    *dsti = __float2half(*xi);
}

template <cpy_kernel_t cpy_1>
static __global__ void cpy_f32_f16(const char * cx, char * cdst, const int ne,
                                   const int ne00, const int ne01, const int nb00, const int nb01, const int nb02,
                                   const int ne10, const int ne11, const int nb10, const int nb11, const int nb12) {
    const int i = blockDim.x*blockIdx.x + threadIdx.x;

    if (i >= ne) {
        return;
    }

    // determine indices i02/i12, i01/i11, i00/i10 as a function of index i of flattened tensor
    // then combine those indices with the corresponding byte offsets to get the total offsets
    const int i02 = i / (ne00*ne01);
    const int i01 = (i - i02*ne01*ne00) / ne00;
    const int i00 = i - i02*ne01*ne00 - i01*ne00;
    const int x_offset = i00*nb00 + i01*nb01 + i02*nb02;

    const int i12 = i / (ne10*ne11);
    const int i11 = (i - i12*ne10*ne11) / ne10;
    const int i10 = i - i12*ne10*ne11 - i11*ne10;
    const int dst_offset = i10*nb10 + i11*nb11 + i12*nb12;

    cpy_1(cx + x_offset, cdst + dst_offset);
}

// rope == RoPE == rotary positional embedding
static __global__ void rope_f32(const float * x, float * dst, const int ncols, const float p, const float theta_scale) {
    const int col = 2*(blockDim.x*blockIdx.x + threadIdx.x);

    if (col >= ncols) {
        return;
    }

    const int row = blockDim.y*blockIdx.y + threadIdx.y;
    const int i = row*ncols + col;

    const float theta = p*powf(theta_scale, col/2);
    const float sin_theta = sinf(theta);
    const float cos_theta = cosf(theta);

    const float x0 = x[i + 0];
    const float x1 = x[i + 1];

    dst[i + 0] = x0*cos_theta - x1*sin_theta;
    dst[i + 1] = x0*sin_theta + x1*cos_theta;
}

static __global__ void diag_mask_inf_f32(const float * x, float * dst, const int ncols, const int rows_per_channel, const int n_past) {
    const int col = blockDim.x*blockIdx.x + threadIdx.x;
    const int row = blockDim.y*blockIdx.y + threadIdx.y;

    if (col >= ncols) {
        return;
    }

    const int i = row*ncols + col;
    // dst[i] = col > n_past + row ? -INFINITY : x[i];
    dst[i] = x[i] - (col > n_past + row % rows_per_channel) * INT_MAX; // equivalent within rounding error but slightly faster on GPU
}

// the CUDA soft max implementation differs from the CPU implementation
// instead of doubles floats are used
// values are also not normalized to the maximum value by subtracting it in the exponential function
// theoretically these changes could cause problems with rounding error and arithmetic overflow but for LLaMa it seems to be fine
static __global__ void soft_max_f32(const float * x, float * dst, const int ncols) {
    const int row = blockDim.y*blockIdx.y + threadIdx.y;
    const int block_size = blockDim.x;
    const int tid = threadIdx.x;

    float tmp = 0.0;

    for (int block_start = 0; block_start < ncols; block_start += block_size) {
        const int col = block_start + tid;

        if (col >= ncols) {
            break;
        }

        const int i = row*ncols + col;
        const float val = expf(x[i]);
        tmp += val;
        dst[i] = val;
    }

    // sum up partial sums
    __syncthreads();
#pragma unroll
    for (int mask = 16; mask > 0; mask >>= 1) {
        tmp += __shfl_xor_sync(0xffffffff, tmp, mask, 32);
    }

    for (int block_start = 0; block_start < ncols; block_start += block_size) {
        const int col = block_start + tid;

        if (col >= ncols) {
            break;
        }

        const int i = row*ncols + col;
        dst[i] /= tmp;
    }
}

static __global__ void scale_f32(const float * x, float * dst, const float scale, const int k) {
    const int i = blockDim.x*blockIdx.x + threadIdx.x;

    if (i >= k) {
        return;
    }

    dst[i] = scale * x[i];
}

static void add_f32_cuda(const float * x, const float * y, float * dst, const int k, cudaStream_t stream) {
    const int num_blocks = (k + CUDA_ADD_BLOCK_SIZE - 1) / CUDA_ADD_BLOCK_SIZE;
    add_f32<<<num_blocks, CUDA_ADD_BLOCK_SIZE, 0, stream>>>(x, y, dst, k);
}

static void add_f16_f32_f16_cuda(const half * x, const float * y, half * dst, const int k, cudaStream_t stream) {
    const int num_blocks = (k + CUDA_ADD_BLOCK_SIZE - 1) / CUDA_ADD_BLOCK_SIZE;
    add_f16_f32_f16<<<num_blocks, CUDA_ADD_BLOCK_SIZE, 0, stream>>>(x, y, dst, k);
}

static void mul_f32_cuda(const float * x, const float * y, float * dst, const int kx, const int ky, cudaStream_t stream) {
    const int num_blocks = (kx + CUDA_MUL_BLOCK_SIZE - 1) / CUDA_MUL_BLOCK_SIZE;
    mul_f32<<<num_blocks, CUDA_MUL_BLOCK_SIZE, 0, stream>>>(x, y, dst, kx, ky);
}

static void silu_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
    const int num_blocks = (k + CUDA_SILU_BLOCK_SIZE - 1) / CUDA_SILU_BLOCK_SIZE;
    silu_f32<<<num_blocks, CUDA_SILU_BLOCK_SIZE, 0, stream>>>(x, dst, k);
}

static void rms_norm_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
    GGML_ASSERT(ncols % WARP_SIZE == 0);
    const dim3 block_dims(WARP_SIZE, 1, 1);
    rms_norm_f32<<<nrows, block_dims, 0, stream>>>(x, dst, ncols);
}

static void dequantize_row_q4_0_cuda(const void * vx, float * y, const int k, cudaStream_t stream) {
    const int num_blocks = (k + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE;
    dequantize_block<QK4_0, QR4_0, dequantize_q4_0><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>(vx, y, k);
}

static void dequantize_row_q4_1_cuda(const void * vx, float * y, const int k, cudaStream_t stream) {
    const int num_blocks = (k + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE;
    dequantize_block<QK4_1, QR4_1, dequantize_q4_1><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>(vx, y, k);
}

static void dequantize_row_q5_0_cuda(const void * vx, float * y, const int k, cudaStream_t stream) {
    const int num_blocks = (k + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE;
    dequantize_block<QK5_0, QR5_0, dequantize_q5_0><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>(vx, y, k);
}

static void dequantize_row_q5_1_cuda(const void * vx, float * y, const int k, cudaStream_t stream) {
    const int num_blocks = (k + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE;
    dequantize_block<QK5_1, QR5_1, dequantize_q5_1><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>(vx, y, k);
}

static void dequantize_row_q8_0_cuda(const void * vx, float * y, const int k, cudaStream_t stream) {
    const int num_blocks = (k + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE;
    dequantize_block<QK8_0, QR8_0, dequantize_q8_0><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>(vx, y, k);
}

static void dequantize_row_q2_K_cuda(const void * vx, float * y, const int k, cudaStream_t stream) {
    const int nb = k / QK_K;
#if QK_K == 256
    dequantize_block_q2_K<<<nb, 64, 0, stream>>>(vx, y);
#else
    dequantize_block_q2_K<<<nb, 32, 0, stream>>>(vx, y);
#endif
}

static void dequantize_row_q3_K_cuda(const void * vx, float * y, const int k, cudaStream_t stream) {
    const int nb = k / QK_K;
#if QK_K == 256
    dequantize_block_q3_K<<<nb, 64, 0, stream>>>(vx, y);
#else
    dequantize_block_q3_K<<<nb, 32, 0, stream>>>(vx, y);
#endif
}

static void dequantize_row_q4_K_cuda(const void * vx, float * y, const int k, cudaStream_t stream) {
    const int nb = k / QK_K;
    dequantize_block_q4_K<<<nb, 32, 0, stream>>>(vx, y);
}

static void dequantize_row_q5_K_cuda(const void * vx, float * y, const int k, cudaStream_t stream) {
    const int nb = k / QK_K;
#if QK_K == 256
    dequantize_block_q5_K<<<nb, 64, 0, stream>>>(vx, y);
#else
    dequantize_block_q5_K<<<nb, 32, 0, stream>>>(vx, y);
#endif
}

static void dequantize_row_q6_K_cuda(const void * vx, float * y, const int k, cudaStream_t stream) {
    const int nb = k / QK_K;
#if QK_K == 256
    dequantize_block_q6_K<<<nb, 64, 0, stream>>>(vx, y);
#else
    dequantize_block_q6_K<<<nb, 32, 0, stream>>>(vx, y);
#endif
}

static void dequantize_mul_mat_vec_q4_0_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
    GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
    const int block_num_y = (nrows + GGML_CUDA_DMMV_Y - 1) / GGML_CUDA_DMMV_Y;
    const dim3 block_nums(1, block_num_y, 1);
    const dim3 block_dims(WARP_SIZE, GGML_CUDA_DMMV_Y, 1);
    dequantize_mul_mat_vec<QK4_0, QR4_0, dequantize_q4_0>
        <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
}

static void dequantize_mul_mat_vec_q4_1_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
    GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
    const int block_num_y = (nrows + GGML_CUDA_DMMV_Y - 1) / GGML_CUDA_DMMV_Y;
    const dim3 block_nums(1, block_num_y, 1);
    const dim3 block_dims(WARP_SIZE, GGML_CUDA_DMMV_Y, 1);
    dequantize_mul_mat_vec<QK4_1, QR4_1, dequantize_q4_1>
        <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
}

static void dequantize_mul_mat_vec_q5_0_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
    GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
    const int block_num_y = (nrows + GGML_CUDA_DMMV_Y - 1) / GGML_CUDA_DMMV_Y;
    const dim3 block_nums(1, block_num_y, 1);
    const dim3 block_dims(WARP_SIZE, GGML_CUDA_DMMV_Y, 1);
    dequantize_mul_mat_vec<QK5_0, QR5_0, dequantize_q5_0>
        <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
}

static void dequantize_mul_mat_vec_q5_1_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
    GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
    const int block_num_y = (nrows + GGML_CUDA_DMMV_Y - 1) / GGML_CUDA_DMMV_Y;
    const dim3 block_nums(1, block_num_y, 1);
    const dim3 block_dims(WARP_SIZE, GGML_CUDA_DMMV_Y, 1);
    dequantize_mul_mat_vec<QK5_1, QR5_1, dequantize_q5_1>
        <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
}

static void dequantize_mul_mat_vec_q8_0_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
    GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
    const int block_num_y = (nrows + GGML_CUDA_DMMV_Y - 1) / GGML_CUDA_DMMV_Y;
    const dim3 block_nums(1, block_num_y, 1);
    const dim3 block_dims(WARP_SIZE, GGML_CUDA_DMMV_Y, 1);
    dequantize_mul_mat_vec<QK8_0, QR8_0, dequantize_q8_0>
        <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
}

static void dequantize_mul_mat_vec_q2_K_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
    GGML_ASSERT(ncols % QK_K == 0);
    const int ny = 2; // very slightly faster than 1 even when K_QUANTS_PER_ITERATION = 2
    const int block_num_y = (nrows + ny - 1) / ny;
    const dim3 block_nums(1, block_num_y, 1);
    const dim3 block_dims(32, ny, 1);
    dequantize_mul_mat_vec_q2_k<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
}

static void dequantize_mul_mat_vec_q3_K_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
    GGML_ASSERT(ncols % QK_K == 0);
    const int ny = 2 / K_QUANTS_PER_ITERATION;
    const int block_num_y = (nrows + ny - 1) / ny;
    const dim3 block_nums(1, block_num_y, 1);
    const dim3 block_dims(32, ny, 1);
    dequantize_mul_mat_vec_q3_k<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
}

static void dequantize_mul_mat_vec_q4_K_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
    GGML_ASSERT(ncols % QK_K == 0);
    const int ny = 2 / K_QUANTS_PER_ITERATION;
    const int block_num_y = (nrows + ny - 1) / ny;
    const dim3 block_nums(1, block_num_y, 1);
    const dim3 block_dims(32, ny, 1);
    dequantize_mul_mat_vec_q4_k<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
}

static void dequantize_mul_mat_vec_q5_K_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
    GGML_ASSERT(ncols % QK_K == 0);
    const dim3 block_dims(32, 1, 1);
    dequantize_mul_mat_vec_q5_k<<<nrows, block_dims, 0, stream>>>(vx, y, dst, ncols);
}

static void dequantize_mul_mat_vec_q6_K_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
    GGML_ASSERT(ncols % QK_K == 0);
    const int ny = 2 / K_QUANTS_PER_ITERATION;
    const int block_num_y = (nrows + ny - 1) / ny;
    const dim3 block_nums(1, block_num_y, 1);
    const dim3 block_dims(32, ny, 1);
    dequantize_mul_mat_vec_q6_k<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
}

static void convert_fp16_to_fp32_cuda(const void * vx, float * y, const int k, cudaStream_t stream) {
    const int num_blocks = (k + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE;
    dequantize_block<1, 1, convert_f16><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>(vx, y, k);
}

static void convert_mul_mat_vec_f16_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
    GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
    const int block_num_y = (nrows + GGML_CUDA_DMMV_Y - 1) / GGML_CUDA_DMMV_Y;
    const dim3 block_nums(1, block_num_y, 1);
    const dim3 block_dims(WARP_SIZE, GGML_CUDA_DMMV_Y, 1);
    dequantize_mul_mat_vec<1, 1, convert_f16>
        <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
}

static to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type) {
    switch (type) {
        case GGML_TYPE_Q4_0:
            return dequantize_row_q4_0_cuda;
        case GGML_TYPE_Q4_1:
            return dequantize_row_q4_1_cuda;
        case GGML_TYPE_Q5_0:
            return dequantize_row_q5_0_cuda;
        case GGML_TYPE_Q5_1:
            return dequantize_row_q5_1_cuda;
        case GGML_TYPE_Q8_0:
            return dequantize_row_q8_0_cuda;
        case GGML_TYPE_Q2_K:
            return dequantize_row_q2_K_cuda;
        case GGML_TYPE_Q3_K:
            return dequantize_row_q3_K_cuda;
        case GGML_TYPE_Q4_K:
            return dequantize_row_q4_K_cuda;
        case GGML_TYPE_Q5_K:
            return dequantize_row_q5_K_cuda;
        case GGML_TYPE_Q6_K:
            return dequantize_row_q6_K_cuda;
        case GGML_TYPE_F16:
            return convert_fp16_to_fp32_cuda;
        default:
            return nullptr;
    }
}

static void ggml_mul_mat_p021_f16_f32_cuda(const void * vx, const float * y, float * dst, const int ncols_x, const int nrows_x, const int nchannels_x, cudaStream_t stream) {
    const dim3 block_nums(1, nrows_x, nchannels_x);
    const dim3 block_dims(WARP_SIZE, 1, 1);
    mul_mat_p021_f16_f32<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols_x, nrows_x, nchannels_x);
}

static void ggml_mul_mat_vec_nc_f16_f32_cuda(
    const void * vx, const float * y, float * dst, const int ncols_x, const int nrows_x, const int row_stride_x,
    const int nchannels_x, const int channel_stride_x, cudaStream_t stream) {

    const dim3 block_nums(1, nrows_x, nchannels_x);
    const dim3 block_dims(WARP_SIZE, 1, 1);
    mul_mat_vec_nc_f16_f32<<<block_nums, block_dims, 0, stream>>>
        (vx, y, dst, ncols_x, nrows_x, row_stride_x, nchannels_x, channel_stride_x);
}

static void ggml_cpy_f32_f32_cuda(
    const char * cx, char * cdst, const int ne,
    const int ne00, const int ne01, const int nb00, const int nb01, const int nb02,
    const int ne10, const int ne11, const int nb10, const int nb11, const int nb12, cudaStream_t stream) {

    const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
    cpy_f32_f16<cpy_1_f32_f32><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
        (cx, cdst, ne, ne00, ne01, nb00, nb01, nb02, ne10, ne11, nb10, nb11, nb12);
}

static void ggml_cpy_f32_f16_cuda(
    const char * cx, char * cdst, const int ne,
    const int ne00, const int ne01, const int nb00, const int nb01, const int nb02,
    const int ne10, const int ne11, const int nb10, const int nb11, const int nb12, cudaStream_t stream) {

    const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
    cpy_f32_f16<cpy_1_f32_f16><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
        (cx, cdst, ne, ne00, ne01, nb00, nb01, nb02, ne10, ne11, nb10, nb11, nb12);
}

static void scale_f32_cuda(const float * x, float * dst, const float scale, const int k, cudaStream_t stream) {
    const int num_blocks = (k + CUDA_SCALE_BLOCK_SIZE - 1) / CUDA_SCALE_BLOCK_SIZE;
    scale_f32<<<num_blocks, CUDA_SCALE_BLOCK_SIZE, 0, stream>>>(x, dst, scale, k);
}

static void rope_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, const float p, const float theta_scale, cudaStream_t stream) {
    GGML_ASSERT(nrows % 2 == 0);
    const dim3 block_dims(2*CUDA_ROPE_BLOCK_SIZE, 1, 1);
    const int num_blocks_x = (ncols + 2*CUDA_ROPE_BLOCK_SIZE - 1) / (2*CUDA_ROPE_BLOCK_SIZE);
    const dim3 block_nums(num_blocks_x, nrows, 1);
    rope_f32<<<block_nums, block_dims, 0, stream>>>(x, dst, ncols, p, theta_scale);
}

static void diag_mask_inf_f32_cuda(const float * x, float * dst, const int ncols_x, const int nrows_x, const int rows_per_channel, const int n_past, cudaStream_t stream) {
    const dim3 block_dims(CUDA_DIAG_MASK_INF_BLOCK_SIZE, 1, 1);
    const int block_num_x = (ncols_x + CUDA_DIAG_MASK_INF_BLOCK_SIZE - 1) / CUDA_DIAG_MASK_INF_BLOCK_SIZE;
    const dim3 block_nums(block_num_x, nrows_x, 1);
    diag_mask_inf_f32<<<block_nums, block_dims, 0, stream>>>(x, dst, ncols_x, rows_per_channel, n_past);
}

static void soft_max_f32_cuda(const float * x, float * dst, const int ncols_x, const int nrows_x, cudaStream_t stream) {
    const dim3 block_dims(WARP_SIZE, 1, 1);
    const dim3 block_nums(1, nrows_x, 1);
    soft_max_f32<<<block_nums, block_dims, 0, stream>>>(x, dst, ncols_x);
}

// buffer pool for cuda
#define MAX_CUDA_BUFFERS 256

struct scoped_spin_lock {
    std::atomic_flag& lock;
    scoped_spin_lock(std::atomic_flag& lock) : lock(lock) {
        while (lock.test_and_set(std::memory_order_acquire)) {
            ; // spin
        }
    }
    ~scoped_spin_lock() {
        lock.clear(std::memory_order_release);
    }
    scoped_spin_lock(const scoped_spin_lock&) = delete;
    scoped_spin_lock& operator=(const scoped_spin_lock&) = delete;
};

struct cuda_buffer {
    void * ptr = nullptr;
    size_t size = 0;
};

static cuda_buffer g_cuda_buffer_pool[GGML_CUDA_MAX_DEVICES][MAX_CUDA_BUFFERS];
static std::atomic_flag g_cuda_pool_lock = ATOMIC_FLAG_INIT;

static void * ggml_cuda_pool_malloc(size_t size, size_t * actual_size) {
    scoped_spin_lock lock(g_cuda_pool_lock);
    int id;
    CUDA_CHECK(cudaGetDevice(&id));

    for (int i = 0; i < MAX_CUDA_BUFFERS; ++i) {
        cuda_buffer& b = g_cuda_buffer_pool[id][i];
        if (b.size >= size && b.ptr != nullptr) {
            void * ptr = b.ptr;
            *actual_size = b.size;
            b.ptr = nullptr;
            b.size = 0;
            return ptr;
        }
    }
    void * ptr;
    CUDA_CHECK(cudaMalloc((void **) &ptr, size));
    *actual_size = size;
    return ptr;
}

static void ggml_cuda_pool_free(void * ptr, size_t size) {
    scoped_spin_lock lock(g_cuda_pool_lock);
    int id;
    CUDA_CHECK(cudaGetDevice(&id));

    for (int i = 0; i < MAX_CUDA_BUFFERS; ++i) {
        cuda_buffer& b = g_cuda_buffer_pool[id][i];
        if (b.ptr == nullptr) {
            b.ptr = ptr;
            b.size = size;
            return;
        }
    }
    fprintf(stderr, "WARNING: cuda buffer pool full, increase MAX_CUDA_BUFFERS\n");
    CUDA_CHECK(cudaFree(ptr));
}


static void * g_scratch_buffer = nullptr;
static size_t g_scratch_size = 1024*1024*1024; // 1 GB by default
static size_t g_scratch_offset = 0;

static int g_device_count = -1;
static int g_main_device = 0;
static float g_tensor_split[GGML_CUDA_MAX_DEVICES] = {0};

static cublasHandle_t g_cublas_handles[GGML_CUDA_MAX_DEVICES] = {nullptr};

static cudaStream_t g_cudaStreams_main[GGML_CUDA_MAX_DEVICES] = { nullptr };

void ggml_init_cublas() {
    static bool initialized = false;

    if (!initialized) {
        CUDA_CHECK(cudaGetDeviceCount(&g_device_count));
        GGML_ASSERT(g_device_count <= GGML_CUDA_MAX_DEVICES);
        int64_t total_vram = 0;
        fprintf(stderr, "%s: found %d CUDA devices:\n", __func__, g_device_count);
        for (int id = 0; id < g_device_count; ++id) {
            cudaDeviceProp prop;
            CUDA_CHECK(cudaGetDeviceProperties(&prop, id));
            fprintf(stderr, "  Device %d: %s\n", id, prop.name);
            g_tensor_split[id] = total_vram;
            total_vram += prop.totalGlobalMem;
        }
        for (int id = 0; id < g_device_count; ++id) {
            g_tensor_split[id] /= total_vram;
        }

        for (int id = 0; id < g_device_count; ++id) {
            CUDA_CHECK(cudaSetDevice(id));

            // create main stream
            CUDA_CHECK(cudaStreamCreateWithFlags(&g_cudaStreams_main[id], cudaStreamNonBlocking));

            // create cublas handle
            CUBLAS_CHECK(cublasCreate(&g_cublas_handles[id]));
            CUBLAS_CHECK(cublasSetMathMode(g_cublas_handles[id], CUBLAS_TF32_TENSOR_OP_MATH));
        }

        // configure logging to stdout
        // CUBLAS_CHECK(cublasLoggerConfigure(1, 1, 0, nullptr));

        initialized = true;
    }
}

void ggml_cuda_set_tensor_split(const float * tensor_split) {
    bool all_zero = true;
    for (int i = 0; i < g_device_count; ++i) {
        if (tensor_split[i] != 0.0f) {
            all_zero = false;
            break;
        }
    }
    if (all_zero) {
        return;
    }
    float split_sum = 0.0f;
    for (int i = 0; i < g_device_count; ++i) {
        g_tensor_split[i] = split_sum;
        split_sum += tensor_split[i];
    }
    for (int i = 0; i < g_device_count; ++i) {
        g_tensor_split[i] /= split_sum;
    }
}

void * ggml_cuda_host_malloc(size_t size) {
    if (getenv("GGML_CUDA_NO_PINNED") != nullptr) {
        return nullptr;
    }

    void * ptr = nullptr;
    cudaError_t err = cudaMallocHost((void **) &ptr, size);
    if (err != cudaSuccess) {
        // The allocation error can be bypassed. A null ptr will assigned out of this function.
        // This can fixed the OOM error in WSL.
        cudaGetLastError();
        fprintf(stderr, "WARNING: failed to allocate %.2f MB of pinned memory: %s\n",
            size/1024.0/1024.0, cudaGetErrorString(err));
        return nullptr;
    }

    return ptr;
}

void ggml_cuda_host_free(void * ptr) {
    CUDA_CHECK(cudaFreeHost(ptr));
}

static cudaError_t ggml_cuda_cpy_tensor_2d(
    void * dst, const struct ggml_tensor * src, int64_t i3, int64_t i2, int64_t i1_low, int64_t i1_high, cudaStream_t stream) {

    cudaMemcpyKind kind;
    char * src_ptr;
    if (src->backend == GGML_BACKEND_CPU) {
        kind = cudaMemcpyHostToDevice;
        src_ptr = (char *) src->data;
    } else if (src->backend == GGML_BACKEND_GPU) {
        kind = cudaMemcpyDeviceToDevice;
        struct ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) src->extra;
        int id;
        CUDA_CHECK(cudaGetDevice(&id));
        src_ptr = (char *) extra->data_device[id];
    } else {
        GGML_ASSERT(false);
    }
    char * dst_ptr = (char *) dst;

    const int64_t ne0 = src->ne[0];
    const int64_t nb0 = src->nb[0];
    const int64_t nb1 = src->nb[1];
    const int64_t nb2 = src->nb[2];
    const int64_t nb3 = src->nb[3];
    const enum ggml_type type = src->type;
    const int64_t ts = ggml_type_size(type);
    const int64_t bs = ggml_blck_size(type);
    int64_t i1_diff = i1_high - i1_low;

    const char * x = src_ptr + i1_low*nb1 + i2*nb2 + i3*nb3;
    if (nb0 == ts && nb1 == ts*ne0/bs) {
        return cudaMemcpyAsync(dst_ptr, x, i1_diff*nb1, kind, stream);
    } else if (nb0 == ts) {
        return cudaMemcpy2DAsync(dst_ptr, ts*ne0/bs, x, nb1, ts*ne0/bs, i1_diff, kind, stream);
    } else {
        for (int64_t i1 = 0; i1 < i1_diff; i1++) {
            const void * rx = (const void *) ((const char *) x + i1*nb1);
            void * rd = (void *) (dst_ptr + i1*ts*ne0/bs);
            // pretend the row is a matrix with cols=1
            cudaError_t r = cudaMemcpy2DAsync(rd, ts/bs, rx, nb0, ts/bs, ne0, kind, stream);
            if (r != cudaSuccess) return r;
        }
        return cudaSuccess;
    }
}

inline void ggml_cuda_op_add(
    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, char * src0_ddq_i,
    float * src0_ddf_i, float * src1_ddf_i, float * dst_ddf_i, int64_t i02, int64_t i01_low, int64_t i01_high, int i1,
    cudaStream_t & cudaStream_main){

    GGML_ASSERT(src0_ddq_i != nullptr || src0_ddf_i != nullptr);
    GGML_ASSERT(src1_ddf_i != nullptr);
    GGML_ASSERT(dst_ddf_i != nullptr);

    const int64_t ne0 = src0->ne[0];
    const int64_t i01_diff = i01_high - i01_low;

    // compute
    if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
        add_f32_cuda(src0_ddf_i, src1_ddf_i, dst_ddf_i, ne0*i01_diff, cudaStream_main);
    } else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) {
        add_f16_f32_f16_cuda((half *) src0_ddq_i, src1_ddf_i, (half *) dst_ddf_i, ne0*i01_diff, cudaStream_main);
    } else {
        GGML_ASSERT(false);
    }
    CUDA_CHECK(cudaGetLastError());

    (void) src1;
    (void) dst;
    (void) src0_ddq_i;
    (void) i02;
    (void) i1;
}

inline void ggml_cuda_op_mul(
    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, char * src0_ddq_i,
    float * src0_ddf_i, float * src1_ddf_i, float * dst_ddf_i, int64_t i02, int64_t i01_low, int64_t i01_high, int i1,
    cudaStream_t & cudaStream_main){

    GGML_ASSERT(src0_ddf_i != nullptr);
    GGML_ASSERT(src1_ddf_i != nullptr);
    GGML_ASSERT(dst_ddf_i != nullptr);

    const int64_t ne00 = src0->ne[0];

    const int64_t ne10 = src1->ne[0];
    const int64_t ne11 = src1->ne[1];

    for (int64_t i01 = i01_low; i01 < i01_high; i01++) {
        const int64_t i11 = i1*ne11 + i01%ne11; // broadcast src1 across src0

        float * src0_ddf_i01 = src0_ddf_i + i01*ne00;
        float * src1_ddf_i01 = src1_ddf_i + i11*ne10;
        float * dst_ddf_i01 = dst_ddf_i + i01*ne00;

        // compute
        mul_f32_cuda(src0_ddf_i01, src1_ddf_i01, dst_ddf_i01, ne00, ne10, cudaStream_main);
        CUDA_CHECK(cudaGetLastError());
    }

    (void) dst;
    (void) src0_ddq_i;
    (void) i02;
}

inline void ggml_cuda_op_silu(
    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, char * src0_ddq_i,
    float * src0_ddf_i, float * src1_ddf_i, float * dst_ddf_i, int64_t i02, int64_t i01_low, int64_t i01_high, int i1,
    cudaStream_t & cudaStream_main){

    GGML_ASSERT(src0_ddf_i != nullptr);
    GGML_ASSERT(dst_ddf_i != nullptr);

    const int64_t ne00 = src0->ne[0];
    const int64_t i01_diff = i01_high - i01_low;

    // compute
    silu_f32_cuda(src0_ddf_i, dst_ddf_i, ne00*i01_diff, cudaStream_main);
    CUDA_CHECK(cudaGetLastError());

    (void) src1;
    (void) dst;
    (void) src0_ddq_i;
    (void) src1_ddf_i;
    (void) i02;
    (void) i1;
}

inline void ggml_cuda_op_rms_norm(
    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, char * src0_ddq_i,
    float * src0_ddf_i, float * src1_ddf_i, float * dst_ddf_i, int64_t i02, int64_t i01_low, int64_t i01_high, int i1,
    cudaStream_t & cudaStream_main){

    GGML_ASSERT(src0_ddf_i != nullptr);
    GGML_ASSERT(dst_ddf_i != nullptr);

    const int64_t ne00 = src0->ne[0];
    const int64_t i01_diff = i01_high - i01_low;

    // compute
    rms_norm_f32_cuda(src0_ddf_i, dst_ddf_i, ne00, i01_diff, cudaStream_main);
    CUDA_CHECK(cudaGetLastError());

    (void) src1;
    (void) dst;
    (void) src0_ddq_i;
    (void) src1_ddf_i;
    (void) i02;
    (void) i1;
}

inline void ggml_cuda_op_dequantize_mul_mat_vec(
    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, char * src0_ddq_i,
    float * src0_ddf_i, float * src1_ddf_i, float * dst_ddf_i, int64_t i02, int64_t i01_low, int64_t i01_high, int i1,
    cudaStream_t & cudaStream_main){

    GGML_ASSERT(src0_ddq_i != nullptr);
    GGML_ASSERT(src1_ddf_i != nullptr);
    GGML_ASSERT(dst_ddf_i != nullptr);

    const int64_t ne00 = src0->ne[0];
    const int64_t nrows = i01_high - i01_low;

// on some GPUs it is faster to convert src1 to half and to use half precision intrinsics
#ifdef GGML_CUDA_DMMV_F16
    size_t ash;
    dfloat * src1_dfloat = nullptr; // dfloat == half

    bool src1_convert_f16 = src0->type == GGML_TYPE_Q4_0 || src0->type == GGML_TYPE_Q4_1 ||
        src0->type == GGML_TYPE_Q5_0 || src0->type == GGML_TYPE_Q5_1 ||
        src0->type == GGML_TYPE_Q8_0 || src0->type == GGML_TYPE_F16;

    if (src1_convert_f16) {
        src1_dfloat = (half *) ggml_cuda_pool_malloc(ne00*sizeof(half), &ash);
        ggml_cpy_f32_f16_cuda((char *) src1_ddf_i, (char *) src1_dfloat, ne00,
                                ne00, 1, sizeof(float), 0, 0,
                                ne00, 1, sizeof(half),  0, 0, cudaStream_main);
    }
#else
    dfloat * src1_dfloat = src1_ddf_i; // dfloat == float, no conversion
#endif // GGML_CUDA_DMMV_F16

    switch (src0->type) {
        case GGML_TYPE_Q4_0:
            dequantize_mul_mat_vec_q4_0_cuda(src0_ddq_i, src1_dfloat, dst_ddf_i, ne00, nrows, cudaStream_main);
            break;
        case GGML_TYPE_Q4_1:
            dequantize_mul_mat_vec_q4_1_cuda(src0_ddq_i, src1_dfloat, dst_ddf_i, ne00, nrows, cudaStream_main);
            break;
        case GGML_TYPE_Q5_0:
            dequantize_mul_mat_vec_q5_0_cuda(src0_ddq_i, src1_dfloat, dst_ddf_i, ne00, nrows, cudaStream_main);
            break;
        case GGML_TYPE_Q5_1:
            dequantize_mul_mat_vec_q5_1_cuda(src0_ddq_i, src1_dfloat, dst_ddf_i, ne00, nrows, cudaStream_main);
            break;
        case GGML_TYPE_Q8_0:
            dequantize_mul_mat_vec_q8_0_cuda(src0_ddq_i, src1_dfloat, dst_ddf_i, ne00, nrows, cudaStream_main);
            break;
        case GGML_TYPE_Q2_K:
            dequantize_mul_mat_vec_q2_K_cuda(src0_ddq_i, src1_ddf_i, dst_ddf_i, ne00, nrows, cudaStream_main);
            break;
        case GGML_TYPE_Q3_K:
            dequantize_mul_mat_vec_q3_K_cuda(src0_ddq_i, src1_ddf_i, dst_ddf_i, ne00, nrows, cudaStream_main);
            break;
        case GGML_TYPE_Q4_K:
            dequantize_mul_mat_vec_q4_K_cuda(src0_ddq_i, src1_ddf_i, dst_ddf_i, ne00, nrows, cudaStream_main);
            break;
        case GGML_TYPE_Q5_K:
            dequantize_mul_mat_vec_q5_K_cuda(src0_ddq_i, src1_ddf_i, dst_ddf_i, ne00, nrows, cudaStream_main);
            break;
        case GGML_TYPE_Q6_K:
            dequantize_mul_mat_vec_q6_K_cuda(src0_ddq_i, src1_ddf_i, dst_ddf_i, ne00, nrows, cudaStream_main);
            break;
        case GGML_TYPE_F16:
            convert_mul_mat_vec_f16_cuda(src0_ddq_i, src1_dfloat, dst_ddf_i, ne00, nrows, cudaStream_main);
            break;
        default:
            GGML_ASSERT(false);
            break;
    }
    CUDA_CHECK(cudaGetLastError());

#ifdef GGML_CUDA_DMMV_F16
    if (src1_convert_f16) {
        ggml_cuda_pool_free(src1_dfloat, ash);
    }
#endif // GGML_CUDA_DMMV_F16

    (void) src1;
    (void) dst;
    (void) src0_ddf_i;
    (void) i02;
    (void) i1;
}

inline void ggml_cuda_op_mul_mat_cublas(
    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, char * src0_ddq_i,
    float * src0_ddf_i, float * src1_ddf_i, float * dst_ddf_i, int64_t i02, int64_t i01_low, int64_t i01_high, int i1,
    cudaStream_t & cudaStream_main){

    GGML_ASSERT(src0_ddf_i != nullptr);
    GGML_ASSERT(src1_ddf_i != nullptr);
    GGML_ASSERT(dst_ddf_i != nullptr);

    const float alpha = 1.0f;
    const float beta = 0.0f;

    const int64_t ne00 = src0->ne[0];

    const int64_t ne10 = src1->ne[0];
    const int64_t ne11 = src1->ne[1];

    const int64_t ne0 = dst->ne[0];
    const int64_t i01_diff = i01_high - i01_low;

    int id;
    CUDA_CHECK(cudaGetDevice(&id));

    // the main device has a larger memory buffer to hold the results from all GPUs
    // ldc == nrows of the matrix that cuBLAS writes into
    int ldc = dst->backend == GGML_BACKEND_GPU && id == g_main_device ? ne0 : i01_diff;

    CUBLAS_CHECK(cublasSetStream(g_cublas_handles[id], cudaStream_main));
    CUBLAS_CHECK(
        cublasSgemm(g_cublas_handles[id], CUBLAS_OP_T, CUBLAS_OP_N,
                i01_diff, ne11, ne10,
                &alpha, src0_ddf_i, ne00,
                        src1_ddf_i, ne10,
                &beta,  dst_ddf_i,  ldc));

    (void) dst;
    (void) src0_ddq_i;
    (void) i02;
    (void) i1;
}

inline void ggml_cuda_op_rope(
    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, char * src0_ddq_i,
    float * src0_ddf_i, float * src1_ddf_i, float * dst_ddf_i, int64_t i02, int64_t i01_low, int64_t i01_high, int i1,
    cudaStream_t & cudaStream_main){

    GGML_ASSERT(src0_ddf_i != nullptr);
    GGML_ASSERT(dst_ddf_i != nullptr);

    const int64_t ne00 = src0->ne[0];
    const int64_t i01_diff = i01_high - i01_low;

    const int n_past = ((int32_t *) src1->data)[0];
    const int n_dims = ((int32_t *) src1->data)[1];
    const int mode   = ((int32_t *) src1->data)[2];
    GGML_ASSERT(mode == 0);

    const float theta_scale = powf(10000.0, -2.0f/n_dims);
    const float p = ((mode & 1) == 0 ? n_past + i02 : i02);

    // compute
    rope_f32_cuda(src0_ddf_i, dst_ddf_i, ne00, i01_diff, p, theta_scale, cudaStream_main);
    CUDA_CHECK(cudaGetLastError());

    (void) dst;
    (void) src0_ddq_i;
    (void) src1_ddf_i;
    (void) i1;
}

inline void ggml_cuda_op_diag_mask_inf(
    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, char * src0_ddq_i,
    float * src0_ddf_i, float * src1_ddf_i, float * dst_ddf_i, int64_t i02, int64_t i01_low, int64_t i01_high, int i1,
    cudaStream_t & cudaStream_main){

    GGML_ASSERT(src0_ddf_i != nullptr);
    GGML_ASSERT(dst_ddf_i != nullptr);

    const int64_t ne00 = src0->ne[0];
    const int64_t ne01 = src0->ne[1];
    const int64_t i01_diff = i01_high - i01_low;

    const int n_past = ((int32_t *) src1->data)[0];

    // compute
    diag_mask_inf_f32_cuda(src0_ddf_i, dst_ddf_i, ne00, i01_diff, ne01, n_past, cudaStream_main);
    CUDA_CHECK(cudaGetLastError());

    (void) dst;
    (void) src0_ddq_i;
    (void) src1_ddf_i;
    (void) i02;
    (void) i1;
}

inline void ggml_cuda_op_soft_max(
    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, char * src0_ddq_i,
    float * src0_ddf_i, float * src1_ddf_i, float * dst_ddf_i, int64_t i02, int64_t i01_low, int64_t i01_high, int i1,
    cudaStream_t & cudaStream_main){

    GGML_ASSERT(src0_ddf_i != nullptr);
    GGML_ASSERT(dst_ddf_i != nullptr);

    const int64_t ne00 = src0->ne[0];
    const int64_t i01_diff = i01_high - i01_low;

    // compute
    soft_max_f32_cuda(src0_ddf_i, dst_ddf_i, ne00, i01_diff, cudaStream_main);
    CUDA_CHECK(cudaGetLastError());

    (void) src1;
    (void) dst;
    (void) src0_ddq_i;
    (void) src1_ddf_i;
    (void) i02;
    (void) i1;
}

inline void ggml_cuda_op_scale(
    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, char * src0_ddq_i,
    float * src0_ddf_i, float * src1_ddf_i, float * dst_ddf_i, int64_t i02, int64_t i01_low, int64_t i01_high, int i1,
    cudaStream_t & cudaStream_main){

    GGML_ASSERT(src0_ddf_i != nullptr);
    GGML_ASSERT(dst_ddf_i != nullptr);

    const float scale = ((float *) src1->data)[0];

    const int64_t ne00 = src0->ne[0];
    const int64_t i01_diff = i01_high - i01_low;

    // compute
    scale_f32_cuda(src0_ddf_i, dst_ddf_i, scale, ne00*i01_diff, cudaStream_main);
    CUDA_CHECK(cudaGetLastError());

    (void) src1;
    (void) dst;
    (void) src0_ddq_i;
    (void) src1_ddf_i;
    (void) i02;
    (void) i1;
}

static void ggml_cuda_op(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
                         ggml_cuda_op_t op, bool src0_needs_f32, bool flatten_rows) {
    const int64_t ne00 = src0->ne[0];
    const int64_t ne01 = src0->ne[1];
    const int64_t ne02 = src0->ne[2];
    const int64_t ne03 = src0->ne[3];
    const int64_t nrows0 = ggml_nrows(src0);

    const bool use_src1 = src1 != nullptr;
    const int64_t ne10 = use_src1 ? src1->ne[0] : 1;
    const int64_t ne11 = use_src1 ? src1->ne[1] : 1;
    const int64_t ne12 = use_src1 ? src1->ne[2] : 1;
    const int64_t ne13 = use_src1 ? src1->ne[3] : 1;

    const int64_t ne0 = dst->ne[0];
    const int64_t ne1 = dst->ne[1];

    const int nb2  = dst->nb[2];
    const int nb3  = dst->nb[3];

    GGML_ASSERT(dst->backend != GGML_BACKEND_GPU_SPLIT);
    GGML_ASSERT(!use_src1 || src1->backend != GGML_BACKEND_GPU_SPLIT);

    // strides for iteration over dims 3 and 2
    const int64_t num_iters = flatten_rows ? 1 : ne02 * ne03;
    const int64_t stride_mod = flatten_rows ? ne02 * ne03 : 1;
    const int64_t src0_stride = ne00 * ne01 * stride_mod;
    const int64_t src1_stride = ne10 * ne11 * stride_mod;
    const int64_t dst_stride = ne0 * ne1 * stride_mod;

    const size_t src0_ts = ggml_type_size(src0->type);
    const size_t src0_bs = ggml_blck_size(src0->type);

    struct ggml_tensor_extra_gpu * src0_extra =            (ggml_tensor_extra_gpu *) src0->extra;
    struct ggml_tensor_extra_gpu * src1_extra = use_src1 ? (ggml_tensor_extra_gpu *) src1->extra : nullptr;
    struct ggml_tensor_extra_gpu * dst_extra  =            (ggml_tensor_extra_gpu *) dst->extra;

    const bool src0_on_device = src0->backend == GGML_BACKEND_GPU || src0->backend == GGML_BACKEND_GPU_SPLIT;
    const bool src0_is_contiguous = ggml_is_contiguous(src0);
    const bool src0_is_f32 = src0->type == GGML_TYPE_F32;

    const bool src1_is_contiguous = use_src1 && ggml_is_contiguous(src1);
    const bool src1_stays_on_host = use_src1 && (
        dst->op == GGML_OP_SCALE || dst->op == GGML_OP_DIAG_MASK_INF || dst->op == GGML_OP_ROPE);

    const bool split = src0->backend == GGML_BACKEND_GPU_SPLIT;

    const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(src0->type);

    // dd = data device
    char  * src0_ddq[GGML_CUDA_MAX_DEVICES] = {nullptr}; // quantized
    float * src0_ddf[GGML_CUDA_MAX_DEVICES] = {nullptr}; // float
    float * src1_ddf[GGML_CUDA_MAX_DEVICES] = {nullptr};
    float *  dst_ddf[GGML_CUDA_MAX_DEVICES] = {nullptr};

    // asq = actual size quantized, asf = actual size float
    size_t src0_asq[GGML_CUDA_MAX_DEVICES] = {0};
    size_t src0_asf[GGML_CUDA_MAX_DEVICES] = {0};
    size_t src1_asf[GGML_CUDA_MAX_DEVICES] = {0};
    size_t  dst_asf[GGML_CUDA_MAX_DEVICES] = {0};

    // if multiple GPUs are used they need to wait for the main GPU to finish
    if (split && g_device_count > 1) {
        CUDA_CHECK(cudaSetDevice(g_main_device));
        CUDA_CHECK(cudaDeviceSynchronize());
    }

    for (int id = 0; id < g_device_count; ++id) {
        if (!split && id != g_main_device) {
            continue;
        }

        const bool src1_on_device = use_src1 && src1->backend == GGML_BACKEND_GPU && id == g_main_device;
        const bool dst_on_device = dst->backend == GGML_BACKEND_GPU && id == g_main_device;

        int64_t row_low, row_high;
        if (split) {
            row_low = id == 0 ? 0 : nrows0*g_tensor_split[id];
            row_high = id == g_device_count - 1 ? nrows0 : nrows0*g_tensor_split[id + 1];
        } else {
            row_low = 0;
            row_high = nrows0;
        }
        if (row_low == row_high) {
            continue;
        }

        int64_t row_diff = row_high - row_low;

        cudaSetDevice(id);

        if (src0_on_device && src0_is_contiguous) {
            if (src0_is_f32) {
                src0_ddf[id] = (float *) src0_extra->data_device[id];
            } else {
                src0_ddq[id] = (char *) src0_extra->data_device[id];
            }
        } else {
            if (src0_is_f32) {
                src0_ddf[id] = (float *) ggml_cuda_pool_malloc(row_diff*ne00 * sizeof(float), &src0_asf[id]);
            } else {
                src0_ddq[id] = (char *) ggml_cuda_pool_malloc(row_diff*ne00 * src0_ts/src0_bs, &src0_asq[id]);
            }
        }

        if (src0_needs_f32 && !src0_is_f32) {
            src0_ddf[id] = (float *) ggml_cuda_pool_malloc(row_diff*ne00 * sizeof(float), &src0_asf[id]);
        }

        if (use_src1 && !src1_stays_on_host) {
            if (src1_on_device && src1_is_contiguous) {
                src1_ddf[id] = (float *) src1_extra->data_device[id];
            } else {
                src1_ddf[id] = (float *) ggml_cuda_pool_malloc(num_iters*src1_stride * sizeof(float), &src1_asf[id]);
            }
        }
        if (dst_on_device) {
            dst_ddf[id] = (float *) dst_extra->data_device[id];
        } else {
            size_t size_dst_ddf = split ? row_diff*ne1 * sizeof(float) : num_iters*dst_stride * sizeof(float);
            dst_ddf[id] = (float *) ggml_cuda_pool_malloc(size_dst_ddf, &dst_asf[id]);
        }

        const int64_t i03_max = flatten_rows ? 1 : ne03;
        const int64_t i02_max = flatten_rows ? 1 : ne02;
        const int64_t rows_per_iter = flatten_rows ? nrows0 : ne01;

        for (int64_t i03 = 0; i03 < i03_max; i03++) {
            const int64_t i13 = i03 % ne13;
            for (int64_t i02 = 0; i02 < i02_max; i02++) {
                const int64_t i12 = i02 % ne12;

                const int64_t i0 = i03*ne02 + i02;

                // i0 values that contain the lower/upper rows for a split tensor when using multiple GPUs
                const int64_t i0_offset_low = row_low/rows_per_iter;
                const int64_t i0_offset_high = row_high/rows_per_iter;

                int64_t i01_low = 0;
                int64_t i01_high = rows_per_iter;
                if (split) {
                    if (i0 < i0_offset_low || i0 > i0_offset_high) {
                        continue;
                    }
                    if (i0 == i0_offset_low) {
                        i01_low = row_low % rows_per_iter;
                    }
                    if (i0 == i0_offset_high) {
                        i01_high = row_high % rows_per_iter;
                    }
                }

                // There is possibly a bug in the Windows nvcc compiler regarding instruction reordering or optimizing out local variables.
                // Removing the first assert or changing the order of the arguments causes the second assert to fail.
                // Removing both asserts results in i01_high becoming 0 which in turn results in garbage output.
                // The root cause seems to be a problem with i0_offset_high becoming 0 when it should always be >0 (for single GPU).
                GGML_ASSERT(i01_low == 0 || g_device_count > 1);
                GGML_ASSERT(i01_high == rows_per_iter || g_device_count > 1);

                const int64_t i01_diff = i01_high - i01_low;
                if (i01_diff == 0) {
                    continue;
                }
                const int64_t i11 = i13*ne12 + i12;

                cudaStream_t cudaStream_main = g_cudaStreams_main[id];

                // for split tensors the data begins at i0 == i0_offset_low
                char  * src0_ddq_i = src0_ddq[id] + (i0 - i0_offset_low)*src0_stride*src0_ts/src0_bs;
                float * src0_ddf_i = src0_ddf[id] + (i0 - i0_offset_low)*src0_stride;
                float * src1_ddf_i = src1_ddf[id] + i11*src1_stride;
                float * dst_ddf_i  =  dst_ddf[id] + (i0 - i0_offset_low)*dst_stride;

                // for split tensors the data pointer needs to be rounded down
                // to the bin edge for i03, i02 bins beyond the first
                if (i0 - i0_offset_low > 0) {
                    GGML_ASSERT(!flatten_rows);
                    src0_ddq_i -= (row_low % ne01)*ne00 * src0_ts/src0_bs;
                    src0_ddf_i -= (row_low % ne01)*ne00;
                    dst_ddf_i  -= (row_low % ne0)*ne1;
                }

                // the main device memory buffer can be on VRAM scratch, with space for all partial results
                // in that case an offset on dst_ddf_i is needed
                if (dst->backend == GGML_BACKEND_GPU && id == g_main_device) {
                    dst_ddf_i += i01_low; // offset is 0 if no tensor split
                }

                // copy src0, src1 to device if necessary
                if (use_src1 && !src1_stays_on_host) {
                    if (src1->backend == GGML_BACKEND_CPU) {
                        GGML_ASSERT(!flatten_rows || nrows0 == ggml_nrows(src1));
                        int64_t nrows1 = flatten_rows ? nrows0 : ne11;
                        CUDA_CHECK(ggml_cuda_cpy_tensor_2d(src1_ddf_i, src1, i03, i02, 0, nrows1, cudaStream_main));
                    } else if (src1->backend == GGML_BACKEND_GPU && src1_is_contiguous) {
                        if (id != g_main_device) {
                            GGML_ASSERT(!flatten_rows);
                            float * src1_ddf_i_source = (float *) src1_extra->data_device[g_main_device];
                            src1_ddf_i_source += i11*src1_stride;
                            CUDA_CHECK(cudaMemcpyAsync(src1_ddf_i, src1_ddf_i_source, src1_stride*sizeof(float),
                                                    cudaMemcpyDeviceToDevice, cudaStream_main));
                        }
                    } else if (src1_on_device && !src1_is_contiguous) {
                        GGML_ASSERT(!split);
                        CUDA_CHECK(ggml_cuda_cpy_tensor_2d(src1_ddf_i, src1, i03, i02, 0, ne11, cudaStream_main));
                    } else {
                        GGML_ASSERT(false);
                    }
                }

                if (!src0_on_device || !src0_is_contiguous) {
                    if (src0_is_f32) {
                        CUDA_CHECK(ggml_cuda_cpy_tensor_2d(src0_ddf_i, src0, i03, i02, i01_low, i01_high, cudaStream_main));
                    } else {
                        CUDA_CHECK(ggml_cuda_cpy_tensor_2d(src0_ddq_i, src0, i03, i02, i01_low, i01_high, cudaStream_main));
                    }
                }

                // convert src0 to f32 if it is necessary for the ggml_cuda_op
                if (src0_needs_f32 && !src0_is_f32) {
                    to_fp32_cuda(src0_ddq_i, src0_ddf_i, i01_diff*ne00, cudaStream_main);
                    CUDA_CHECK(cudaGetLastError());
                }

                // do the computation
                op(src0, src1, dst, src0_ddq_i, src0_ddf_i, src1_ddf_i, dst_ddf_i, i02, i01_low, i01_high, i11, cudaStream_main);

                // copy dst to host or other device if necessary
                if (!dst_on_device) {
                    void * dst_off_device;
                    cudaMemcpyKind kind;
                    if (dst->backend == GGML_BACKEND_CPU) {
                        dst_off_device = dst->data;
                        kind = cudaMemcpyDeviceToHost;
                    } else if (dst->backend == GGML_BACKEND_GPU) {
                        dst_off_device = dst_extra->data_device[g_main_device];
                        kind = cudaMemcpyDeviceToDevice;
                    } else {
                        GGML_ASSERT(false);
                    }
                    if (split) {
                        // src0 = weight matrix is saved as a transposed matrix for better memory layout.
                        // dst is NOT transposed.
                        // The outputs of cuBLAS matrix matrix multiplications can therefore NOT simply be concatenated for >1 GPU.
                        // Instead they need to be copied to the correct slice in ne0 = dst row index.
                        // If dst is a vector with ne0 == 1 then you don't have to do this but it still produces correct results.
                        for (int64_t j = 0; j < ne1; ++j) {
                            float * dhf_dst_i = (float *) ((char *) dst_off_device + (j*ne0 + i01_low)*sizeof(float) + i02*nb2 + i03*nb3);
                            CUDA_CHECK(cudaMemcpyAsync(dhf_dst_i, dst_ddf_i + j*i01_diff, i01_diff*sizeof(float), kind, cudaStream_main));
                        }
                    } else {
                        float * dhf_dst_i = (float *) ((char *) dst_off_device + i02*nb2 + i03*nb3);
                        CUDA_CHECK(cudaMemcpyAsync(dhf_dst_i, dst_ddf_i, dst_stride*sizeof(float), kind, cudaStream_main));
                    }
                }
            }
        }
    }

    // wait until each device is finished, then free their buffers
    for (int id = 0; id < g_device_count; ++id) {
        if (src0_asq[id] == 0 && src0_asf[id] == 0 && src1_asf[id] == 0 && dst_asf[id] == 0) {
            continue;
        }

        CUDA_CHECK(cudaSetDevice(id));
        CUDA_CHECK(cudaDeviceSynchronize());

        if (src0_asq[id] > 0) {
            ggml_cuda_pool_free(src0_ddq[id], src0_asq[id]);
        }
        if (src0_asf[id] > 0) {
            ggml_cuda_pool_free(src0_ddf[id], src0_asf[id]);
        }
        if (src1_asf[id] > 0) {
            ggml_cuda_pool_free(src1_ddf[id], src1_asf[id]);
        }
        if (dst_asf[id] > 0) {
            ggml_cuda_pool_free(dst_ddf[id], dst_asf[id]);
        }
    }
}

void ggml_cuda_add(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
    // ggml_cuda_add permits f16 dst even though this could in theory cause problems with the pointer arithmetic in ggml_cuda_op.
    // Due to flatten_rows == true this does in practice not make a difference however.
    // Better solution would be nice but right now that would require disproportionate changes.
    GGML_ASSERT(
        (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16) &&
        src1->type == GGML_TYPE_F32 &&
        (dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16));
    ggml_cuda_op(src0, src1, dst, ggml_cuda_op_add, false, true);
}

void ggml_cuda_mul(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
    GGML_ASSERT(src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32);
    ggml_cuda_op(src0, src1, dst, ggml_cuda_op_mul, true, false); // TODO ggml_cuda_op needs modification for flatten
}

void ggml_cuda_silu(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
    GGML_ASSERT(src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32);
    ggml_cuda_op(src0, src1, dst, ggml_cuda_op_silu, true, true);
}

void ggml_cuda_rms_norm(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
    GGML_ASSERT(src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32);
    ggml_cuda_op(src0, src1, dst, ggml_cuda_op_rms_norm, true, true);
}

bool ggml_cuda_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) {
    const int64_t ne10 = src1->ne[0];

    const int64_t ne0 = dst->ne[0];
    const int64_t ne1 = dst->ne[1];

    // TODO: find the optimal values for these
    if ((src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) &&
        src1->type == GGML_TYPE_F32 &&
        dst->type == GGML_TYPE_F32 &&
        (ne0 >= 32 && ne1 >= 32 && ne10 >= 32)) {
        return true;
    }

    return false;
}

void ggml_cuda_mul_mat_vec_p021(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst){
    GGML_ASSERT(ggml_is_permuted(src0) && ggml_is_permuted(src1));
    GGML_ASSERT(src0->backend != GGML_BACKEND_GPU_SPLIT);
    GGML_ASSERT(src0->nb[0] <= src0->nb[1] && src0->nb[2] <= src0->nb[3]); // 0213 permutation
    GGML_ASSERT(src1->nb[0] <= src1->nb[1] && src1->nb[2] <= src1->nb[3]); // 0213 permutation
    GGML_ASSERT(src0->type == GGML_TYPE_F16);
    GGML_ASSERT(src1->type == GGML_TYPE_F32);

    const int64_t ne00 = src0->ne[0];
    const int64_t ne01 = src0->ne[1];
    const int64_t ne02 = src0->ne[2];

    CUDA_CHECK(cudaSetDevice(g_main_device));
    cudaStream_t cudaStream_main = g_cudaStreams_main[g_main_device];

    struct ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra;
    void * src0_ddq = src0_extra->data_device[g_main_device];

    struct ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu *) src1->extra;
    float * src1_ddf = (float *) src1_extra->data_device[g_main_device];

    struct ggml_tensor_extra_gpu * dst_extra = (ggml_tensor_extra_gpu *) dst->extra;
    float * dst_ddf = (float *) dst_extra->data_device[g_main_device];

    ggml_mul_mat_p021_f16_f32_cuda(src0_ddq, src1_ddf, dst_ddf, ne00, ne01, ne02, cudaStream_main);
}

void ggml_cuda_mul_mat_vec_nc(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst){
    GGML_ASSERT(!ggml_is_contiguous(src0) && ggml_is_contiguous(src1));
    GGML_ASSERT(!ggml_is_permuted(src0));
    GGML_ASSERT(src0->backend != GGML_BACKEND_GPU_SPLIT);
    GGML_ASSERT(src0->type == GGML_TYPE_F16);
    GGML_ASSERT(src1->type == GGML_TYPE_F32);

    const int64_t ne00 = src0->ne[0];
    const int64_t ne01 = src0->ne[1];
    const int64_t ne02 = src0->ne[2];

    const int64_t nb01 = src0->nb[1];
    const int64_t nb02 = src0->nb[2];

    CUDA_CHECK(cudaSetDevice(g_main_device));
    cudaStream_t cudaStream_main = g_cudaStreams_main[g_main_device];

    struct ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra;
    void * src0_ddq = src0_extra->data_device[g_main_device];

    struct ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu *) src1->extra;
    float * src1_ddf = (float *) src1_extra->data_device[g_main_device];

    struct ggml_tensor_extra_gpu * dst_extra = (ggml_tensor_extra_gpu *) dst->extra;
    float * dst_ddf = (float *) dst_extra->data_device[g_main_device];

    const int row_stride_x = nb01 / sizeof(half);
    const int channel_stride_x = nb02 / sizeof(half);

    ggml_mul_mat_vec_nc_f16_f32_cuda(src0_ddq, src1_ddf, dst_ddf, ne00, ne01, row_stride_x, ne02, channel_stride_x, cudaStream_main);
}

void ggml_cuda_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
    bool all_on_device = (src0->backend == GGML_BACKEND_GPU || src0->backend == GGML_BACKEND_GPU_SPLIT) &&
        src1->backend == GGML_BACKEND_GPU && dst->backend == GGML_BACKEND_GPU;

    if (all_on_device && ggml_is_permuted(src0) && ggml_is_permuted(src1) && src1->ne[1] == 1) {
        ggml_cuda_mul_mat_vec_p021(src0, src1, dst);
    } else if (all_on_device && !ggml_is_contiguous(src0) && ggml_is_contiguous(src1) && src1->ne[1] == 1) {
        ggml_cuda_mul_mat_vec_nc(src0, src1, dst);
    }else if (src0->type == GGML_TYPE_F32) {
        ggml_cuda_op(src0, src1, dst, ggml_cuda_op_mul_mat_cublas, true, false);
    } else if (ggml_is_quantized(src0->type) || src0->type == GGML_TYPE_F16) {
        if (src1->ne[1] == 1 && src0->ne[0] % GGML_CUDA_DMMV_X == 0 && src0->ne[1] % GGML_CUDA_DMMV_Y == 0) {
            ggml_cuda_op(src0, src1, dst, ggml_cuda_op_dequantize_mul_mat_vec, false, false);
        } else {
            ggml_cuda_op(src0, src1, dst, ggml_cuda_op_mul_mat_cublas, true, false);
        }
    } else {
        GGML_ASSERT(false);
    }
}

void ggml_cuda_scale(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
    GGML_ASSERT(src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32);
    ggml_cuda_op(src0, src1, dst, ggml_cuda_op_scale, true, true);
}

void ggml_cuda_cpy(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
    const int64_t ne = ggml_nelements(src0);
    GGML_ASSERT(ne == ggml_nelements(src1));

    GGML_ASSERT(src0->backend == GGML_BACKEND_GPU);
    GGML_ASSERT(src1->backend == GGML_BACKEND_GPU);

    GGML_ASSERT(ggml_nbytes(src0) <= INT_MAX);
    GGML_ASSERT(ggml_nbytes(src1) <= INT_MAX);

    const int64_t ne00 = src0->ne[0];
    const int64_t ne01 = src0->ne[1];
    GGML_ASSERT(src0->ne[3] == 1);

    const int64_t nb00 = src0->nb[0];
    const int64_t nb01 = src0->nb[1];
    const int64_t nb02 = src0->nb[2];

    const int64_t ne10 = src1->ne[0];
    const int64_t ne11 = src1->ne[1];
    GGML_ASSERT(src1->ne[3] == 1);

    const int64_t nb10 = src1->nb[0];
    const int64_t nb11 = src1->nb[1];
    const int64_t nb12 = src1->nb[2];

    CUDA_CHECK(cudaSetDevice(g_main_device));
    cudaStream_t cudaStream_main = g_cudaStreams_main[g_main_device];

    const struct ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra;
    const struct ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu *) src1->extra;

    char * src0_ddc = (char *) src0_extra->data_device[g_main_device];
    char * src1_ddc = (char *) src1_extra->data_device[g_main_device];

    if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) {
        ggml_cpy_f32_f32_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, nb00, nb01, nb02,
                              ne10, ne11, nb10, nb11, nb12, cudaStream_main);
    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F16) {
        ggml_cpy_f32_f16_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, nb00, nb01, nb02,
                              ne10, ne11, nb10, nb11, nb12, cudaStream_main);
    } else {
        GGML_ASSERT(false);
    }

    (void) dst;
}

void ggml_cuda_diag_mask_inf(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
    GGML_ASSERT(src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32);
    ggml_cuda_op(src0, src1, dst, ggml_cuda_op_diag_mask_inf, true, true);
}

void ggml_cuda_soft_max(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
    GGML_ASSERT(src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32);
    ggml_cuda_op(src0, src1, dst, ggml_cuda_op_soft_max, true, true);
}

void ggml_cuda_rope(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
    GGML_ASSERT(src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32);
    ggml_cuda_op(src0, src1, dst, ggml_cuda_op_rope, true, false); // FIXME flatten changes results
}

void ggml_cuda_nop(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
    (void) src0;
    (void) src1;
    (void) dst;
}

void ggml_cuda_transform_tensor(void * data, struct ggml_tensor * tensor) {
    int nrows = ggml_nrows(tensor);
    const size_t nb1 = tensor->nb[1];
    ggml_backend backend = tensor->backend;
    struct ggml_tensor_extra_gpu * extra = new struct ggml_tensor_extra_gpu;
    memset(extra, 0, sizeof(*extra));

    for (int id = 0; id < g_device_count; ++id) {
        if (backend == GGML_BACKEND_GPU && id != g_main_device) {
            continue;
        }

        cudaSetDevice(id);

        int row_low, row_high;
        if (backend == GGML_BACKEND_GPU) {
            row_low = 0;
            row_high = nrows;
        } else if (backend == GGML_BACKEND_GPU_SPLIT) {
            row_low = id == 0 ? 0 : nrows*g_tensor_split[id];
            row_high = id == g_device_count - 1 ? nrows : nrows*g_tensor_split[id + 1];
        } else {
            GGML_ASSERT(false);
        }
        if (row_low == row_high) {
            continue;
        }

        int64_t nrows_split = row_high - row_low;

        const size_t offset_split = row_low*nb1;
        const size_t size = ggml_nbytes_split(tensor, nrows_split);

        void * buf;
        CUDA_CHECK(cudaMalloc(&buf, size));
        void * buf_host = (char*)data + offset_split;

        cudaMemcpy(buf, buf_host, size, cudaMemcpyHostToDevice);

        extra->data_device[id] = buf;
    }

    tensor->extra = extra;
}

void ggml_cuda_free_data(struct ggml_tensor * tensor) {
    if (tensor->backend != GGML_BACKEND_GPU && tensor->backend != GGML_BACKEND_GPU_SPLIT) {
        return;
    }

    ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) tensor->extra;

    for (int id = 0; id < g_device_count; ++id) {
        if (extra->data_device[id] == nullptr) {
            continue;
        }

        CUDA_CHECK(cudaSetDevice(id));
        CUDA_CHECK(cudaFree(extra->data_device[id]));
    }

    delete extra;
}

void ggml_cuda_assign_buffers_impl(struct ggml_tensor * tensor, bool scratch, bool force_inplace) {
    if (scratch && g_scratch_size == 0) {
        return;
    }

    // recursively assign CUDA buffers until a compute tensor is found
    if (tensor->src0 != nullptr && tensor->src0->backend == GGML_BACKEND_CPU) {
        const ggml_op src0_op = tensor->src0->op;
        if (src0_op == GGML_OP_RESHAPE || src0_op == GGML_OP_TRANSPOSE || src0_op == GGML_OP_VIEW) {
            ggml_cuda_assign_buffers_impl(tensor->src0, scratch, force_inplace);
        }
    }
    if (tensor->op == GGML_OP_CPY && tensor->src1->backend == GGML_BACKEND_CPU) {
        ggml_cuda_assign_buffers_impl(tensor->src1, scratch, force_inplace);
    }

    tensor->backend = GGML_BACKEND_GPU;
    struct ggml_tensor_extra_gpu * extra = new ggml_tensor_extra_gpu;
    memset(extra, 0, sizeof(*extra));

    const bool inplace = (tensor->src0 != nullptr && tensor->src0->data == tensor->data) ||
        tensor->op == GGML_OP_VIEW ||
        force_inplace;
    const size_t size = ggml_nbytes(tensor);

    CUDA_CHECK(cudaSetDevice(g_main_device));
    if (inplace && (tensor->src0->backend == GGML_BACKEND_GPU || tensor->src0->backend == GGML_BACKEND_GPU_SPLIT)) {
        struct ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu * ) tensor->src0->extra;
        char * src0_ddc = (char *) src0_extra->data_device[g_main_device];
        size_t offset = 0;
        if (tensor->op == GGML_OP_VIEW) {
            memcpy(&offset, tensor->opt[0]->data, sizeof(size_t));
        }
        extra->data_device[g_main_device] = src0_ddc + offset;
    } else if (tensor->op == GGML_OP_CPY) {
        struct ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu * ) tensor->src1->extra;
        void * src1_ddv = src1_extra->data_device[g_main_device];
        extra->data_device[g_main_device] = src1_ddv;
    } else if (scratch) {
        GGML_ASSERT(size <= g_scratch_size);
        if (g_scratch_offset + size > g_scratch_size) {
            g_scratch_offset = 0;
        }

        char * data = (char *) g_scratch_buffer;
        if (data == nullptr) {
            CUDA_CHECK(cudaMalloc(&data, g_scratch_size));
            g_scratch_buffer = data;
        }
        extra->data_device[g_main_device] = data + g_scratch_offset;

        g_scratch_offset += size;

        GGML_ASSERT(g_scratch_offset <= g_scratch_size);
    } else { // allocate new buffers outside of scratch
        void * data;
        CUDA_CHECK(cudaMalloc(&data, size));
        CUDA_CHECK(cudaMemset(data, 0, size));
        extra->data_device[g_main_device] = data;
    }

    tensor->extra = extra;
}

void ggml_cuda_assign_buffers(struct ggml_tensor * tensor) {
    ggml_cuda_assign_buffers_impl(tensor, true, false);
}

void ggml_cuda_assign_buffers_no_scratch(struct ggml_tensor * tensor) {
    ggml_cuda_assign_buffers_impl(tensor, false, false);
}

void ggml_cuda_assign_buffers_force_inplace(struct ggml_tensor * tensor) {
    ggml_cuda_assign_buffers_impl(tensor, false, true);
}

void ggml_cuda_set_main_device(int main_device) {
    if (main_device >= g_device_count) {
        fprintf(stderr, "warning: cannot set main_device=%d because there are only %d devices. Using device %d instead.\n",
                main_device, g_device_count, g_main_device);
        return;
    }
    g_main_device = main_device;
    if (g_device_count > 1) {
        cudaDeviceProp prop;
        CUDA_CHECK(cudaGetDeviceProperties(&prop, g_main_device));
        fprintf(stderr, "%s: using device %d (%s) as main device\n", __func__, g_main_device, prop.name);
    }
}

void ggml_cuda_set_scratch_size(size_t scratch_size) {
    g_scratch_size = scratch_size;
}

void ggml_cuda_free_scratch() {
    if (g_scratch_buffer == nullptr) {
        return;
    }

    CUDA_CHECK(cudaFree(g_scratch_buffer));
    g_scratch_buffer = nullptr;
}

bool ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor){
    ggml_cuda_func_t func;
    const bool any_on_device = tensor->backend == GGML_BACKEND_GPU
        || (tensor->src0 != nullptr && (tensor->src0->backend == GGML_BACKEND_GPU || tensor->src0->backend == GGML_BACKEND_GPU_SPLIT))
        || (tensor->src1 != nullptr && tensor->src1->backend == GGML_BACKEND_GPU);

    switch (tensor->op) {
        case GGML_OP_ADD:
            if (!any_on_device) {
                return false;
            }
            func = ggml_cuda_add;
            break;
        case GGML_OP_MUL:
            if (!any_on_device) {
                return false;
            }
            func = ggml_cuda_mul;
            break;
        case GGML_OP_SILU:
            if (!any_on_device) {
                return false;
            }
            func = ggml_cuda_silu;
            break;
        case GGML_OP_RMS_NORM:
            if (!any_on_device) {
                return false;
            }
            func = ggml_cuda_rms_norm;
            break;
        case GGML_OP_MUL_MAT:
            if (!any_on_device && !ggml_cuda_can_mul_mat(tensor->src0, tensor->src1, tensor)) {
                return false;
            }
            func = ggml_cuda_mul_mat;
            break;
        case GGML_OP_SCALE:
            if (!any_on_device) {
                return false;
            }
            func = ggml_cuda_scale;
            break;
        case GGML_OP_CPY:
            if (!any_on_device) {
                return false;
            }
            func = ggml_cuda_cpy;
            break;
        case GGML_OP_RESHAPE:
        case GGML_OP_VIEW:
        case GGML_OP_PERMUTE:
        case GGML_OP_TRANSPOSE:
            if (!any_on_device) {
                return false;
            }
            func = ggml_cuda_nop;
            break;
        case GGML_OP_DIAG_MASK_INF:
            if (!any_on_device) {
                return false;
            }
            func = ggml_cuda_diag_mask_inf;
            break;
        case GGML_OP_SOFT_MAX:
            if (!any_on_device) {
                return false;
            }
            func = ggml_cuda_soft_max;
            break;
        case GGML_OP_ROPE:
            if (!any_on_device) {
                return false;
            }
            func = ggml_cuda_rope;
            break;
        default:
            return false;
    }

    if (params->ith != 0) {
        return true;
    }
    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
        return true;
    }
    func(tensor->src0, tensor->src1, tensor);
    return true;
}