llama_cpp 0.2.1 → 0.2.2

data/ext/llama_cpp/src/ggml-cuda.cu

@@ -13,6 +13,10 @@
 #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)                                                                 \
@@ -46,7 +50,15 @@ static_assert(sizeof(half) == sizeof(ggml_fp16_t), "wrong fp16 size");
     } while (0)
 #endif // CUDART_VERSION >= 11
 
-typedef void (*dequantize_kernel_t)(const void * vx, const int ib, const int iqs, float & v0, float & v1);
+#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);
@@ -167,6 +179,12 @@ static_assert(sizeof(block_q6_K) == sizeof(ggml_fp16_t) + 13*QK_K/16, "wrong q6_
 #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;
 
@@ -224,82 +242,106 @@ static __global__ void rms_norm_f32(const float * x, float * dst, const int ncol
     }
 }
 
-static __device__ void dequantize_q4_0(const void * vx, const int ib, const int iqs, float & v0, float & v1){
+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 float d = x[ib].d;
+    const dfloat d = x[ib].d;
 
-    const uint8_t vui = x[ib].qs[iqs];
+    const int vui = x[ib].qs[iqs];
 
-    const int8_t vi0 = vui & 0xF;
-    const int8_t vi1 = vui >> 4;
+    v.x = vui & 0xF;
+    v.y = vui >> 4;
 
-    v0 = (vi0 - 8)*d;
-    v1 = (vi1 - 8)*d;
+#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__ void dequantize_q4_1(const void * vx, const int ib, const int iqs, float & v0, float & v1){
+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 float d = x[ib].d;
-    const float m = x[ib].m;
+    const dfloat d = x[ib].d;
+    const dfloat m = x[ib].m;
 
-    const uint8_t vui = x[ib].qs[iqs];
+    const int vui = x[ib].qs[iqs];
 
-    const int8_t vi0 = vui & 0xF;
-    const int8_t vi1 = vui >> 4;
+    v.x = vui & 0xF;
+    v.y = vui >> 4;
 
-    v0 = vi0*d + m;
-    v1 = vi1*d + m;
+#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__ void dequantize_q5_0(const void * vx, const int ib, const int iqs, float & v0, float & v1){
+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 float d = x[ib].d;
+    const dfloat d = x[ib].d;
 
     uint32_t qh;
     memcpy(&qh, x[ib].qh, sizeof(qh));
 
-    const uint8_t xh_0 = ((qh >> (iqs +  0)) << 4) & 0x10;
-    const uint8_t xh_1 = ((qh >> (iqs + 12))     ) & 0x10;
+    const int xh_0 = ((qh >> (iqs +  0)) << 4) & 0x10;
+    const int xh_1 = ((qh >> (iqs + 12))     ) & 0x10;
 
-    const int32_t x0 = ((x[ib].qs[iqs] & 0xf) | xh_0) - 16;
-    const int32_t x1 = ((x[ib].qs[iqs] >>  4) | xh_1) - 16;
+    v.x = ((x[ib].qs[iqs] & 0xf) | xh_0);
+    v.y = ((x[ib].qs[iqs] >>  4) | xh_1);
 
-    v0 = x0*d;
-    v1 = x1*d;
+#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__ void dequantize_q5_1(const void * vx, const int ib, const int iqs, float & v0, float & v1){
+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 float d = x[ib].d;
-    const float m = x[ib].m;
+    const dfloat d = x[ib].d;
+    const dfloat m = x[ib].m;
 
     uint32_t qh;
     memcpy(&qh, x[ib].qh, sizeof(qh));
 
-    const uint8_t xh_0 = ((qh >> (iqs +  0)) << 4) & 0x10;
-    const uint8_t xh_1 = ((qh >> (iqs + 12))     ) & 0x10;
+    const int xh_0 = ((qh >> (iqs +  0)) << 4) & 0x10;
+    const int xh_1 = ((qh >> (iqs + 12))     ) & 0x10;
 
-    const int32_t x0 = ((x[ib].qs[iqs] & 0xf) | xh_0);
-    const int32_t x1 = ((x[ib].qs[iqs] >>  4) | xh_1);
+    v.x = ((x[ib].qs[iqs] & 0xf) | xh_0);
+    v.y = ((x[ib].qs[iqs] >>  4) | xh_1);
 
-    v0 = x0*d + m;
-    v1 = x1*d + m;
+#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__ void dequantize_q8_0(const void * vx, const int ib, const int iqs, float & v0, float & v1){
+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 float d = x[ib].d;
+    const dfloat d = x[ib].d;
 
-    const int8_t vi0 = x[ib].qs[iqs + 0];
-    const int8_t vi1 = x[ib].qs[iqs + 1];
+    v.x = x[ib].qs[iqs + 0];
+    v.y = x[ib].qs[iqs + 1];
 
-    v0 = vi0*d;
-    v1 = vi1*d;
+#ifdef GGML_CUDA_DMMV_F16
+    v = __hmul2(v, {d, d});
+#else
+    v.x *= d;
+    v.y *= d;
+#endif // GGML_CUDA_DMMV_F16
 }
 
 //================================== k-quants
@@ -326,37 +368,6 @@ static __global__ void dequantize_block_q2_K(const void * vx, float * yy) {
 
 }
 
-static __device__ void vec_dot_q2_K(const void * vx, const int ib, const int iqs, const float * yy, float & result) {
-
-    const block_q2_K * x = (const block_q2_K *) vx;
-
-    // if n is 0, we want to do the lower 128, else the upper 128,
-    // covering y[l+0],  y[l+32], y[l+64], y[l+96] and
-    //          y[l+16], y[l+48], y[l+80], y[l+112]
-    int n = iqs/128;                // 0 or 1
-    int r = iqs - 128*n;            // 0...120 in steps of 8
-    int l = r/8;                    // 0...15 in steps of 1
-
-    const float   * y = yy + 128*n + l;
-    const uint8_t * q = x[ib].qs + 32*n + l;
-    const uint8_t * s = x[ib].scales + 8*n;
-
-    const float dall = x[ib].d;
-    const float dmin = x[ib].dmin;
-
-    float sum = y[  0] * (dall * ((s[0] & 0xF) * ((q[ 0] >> 0) & 3)) - dmin * (s[0] >> 4))
-              + y[ 32] * (dall * ((s[2] & 0xF) * ((q[ 0] >> 2) & 3)) - dmin * (s[2] >> 4))
-              + y[ 64] * (dall * ((s[4] & 0xF) * ((q[ 0] >> 4) & 3)) - dmin * (s[4] >> 4))
-              + y[ 96] * (dall * ((s[6] & 0xF) * ((q[ 0] >> 6) & 3)) - dmin * (s[6] >> 4))
-              + y[ 16] * (dall * ((s[1] & 0xF) * ((q[16] >> 0) & 3)) - dmin * (s[1] >> 4))
-              + y[ 48] * (dall * ((s[3] & 0xF) * ((q[16] >> 2) & 3)) - dmin * (s[3] >> 4))
-              + y[ 80] * (dall * ((s[5] & 0xF) * ((q[16] >> 4) & 3)) - dmin * (s[5] >> 4))
-              + y[112] * (dall * ((s[7] & 0xF) * ((q[16] >> 6) & 3)) - dmin * (s[7] >> 4));
-
-    result = sum;
-
-}
-
 static __global__ void dequantize_block_q3_K(const void * vx, float * yy) {
 
     int r = threadIdx.x/4;
@@ -388,51 +399,6 @@ static __global__ void dequantize_block_q3_K(const void * vx, float * yy) {
 
 }
 
-static __device__ void vec_dot_q3_K(const void * vx, const int ib, const int iqs, const float * yy, float & result) {
-
-    const block_q3_K * x = (const block_q3_K *) vx;
-
-    const uint32_t kmask1 = 0x03030303;
-    const uint32_t kmask2 = 0x0f0f0f0f;
-
-    uint32_t aux[3];
-    uint32_t utmp[4];
-
-    // if n is 0, we want to do the lower 128, else the upper 128,
-    // covering y[l+0],  y[l+32], y[l+64], y[l+96] and
-    //          y[l+16], y[l+48], y[l+80], y[l+112]
-    int n = iqs/128;                // 0 or 1
-    int r = iqs - 128*n;            // 0...120 in steps of 8
-    int l = r/8;                    // 0...15 in steps of 1
-
-    const float   * y = yy + 128*n + l;
-    const uint8_t * q = x[ib].qs + 32*n + l;
-    const uint8_t * hm = x[ib].hmask + l;
-    const int8_t  * s = (const int8_t *)utmp + 8*n;
-
-    memcpy(aux, x[ib].scales, 12);
-    utmp[3] = ((aux[1] >> 4) & kmask2) | (((aux[2] >> 6) & kmask1) << 4);
-    utmp[2] = ((aux[0] >> 4) & kmask2) | (((aux[2] >> 4) & kmask1) << 4);
-    utmp[1] = (aux[1] & kmask2) | (((aux[2] >> 2) & kmask1) << 4);
-    utmp[0] = (aux[0] & kmask2) | (((aux[2] >> 0) & kmask1) << 4);
-
-    const float dall = x[ib].d;
-
-    const uint8_t m = 1 << (4*n);
-
-    float sum = y[  0] * (s[0] - 32) * (((q[ 0] >> 0) & 3) - (hm[ 0] & (m << 0) ? 0 : 4))
-              + y[ 32] * (s[2] - 32) * (((q[ 0] >> 2) & 3) - (hm[ 0] & (m << 1) ? 0 : 4))
-              + y[ 64] * (s[4] - 32) * (((q[ 0] >> 4) & 3) - (hm[ 0] & (m << 2) ? 0 : 4))
-              + y[ 96] * (s[6] - 32) * (((q[ 0] >> 6) & 3) - (hm[ 0] & (m << 3) ? 0 : 4))
-              + y[ 16] * (s[1] - 32) * (((q[16] >> 0) & 3) - (hm[16] & (m << 0) ? 0 : 4))
-              + y[ 48] * (s[3] - 32) * (((q[16] >> 2) & 3) - (hm[16] & (m << 1) ? 0 : 4))
-              + y[ 80] * (s[5] - 32) * (((q[16] >> 4) & 3) - (hm[16] & (m << 2) ? 0 : 4))
-              + y[112] * (s[7] - 32) * (((q[16] >> 6) & 3) - (hm[16] & (m << 3) ? 0 : 4));
-
-    result = sum * dall;
-
-}
-
 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;
@@ -479,38 +445,6 @@ static __global__ void dequantize_block_q4_K(const void * vx, float * yy) {
     }
 }
 
-static __device__ void vec_dot_q4_K(const void * vx, const int ib, const int iqs, const float * yy, float & result) {
-
-    const block_q4_K * x = (const block_q4_K *) vx;
-
-                                    // iqs is in 0...248 in steps of 8 =>
-    const int j  = iqs / 64;        // j  is in 0...3
-    const int ir = (iqs - 64*j)/2;  // ir is in 0...28 in steps of 4
-    const int is = 2*j;             // is is in 0...6 in steps of 2
-
-    const float   * y = yy + 64*j + ir;
-    const uint8_t * q = x[ib].qs + 32*j + ir;
-
-    const float dall = x[ib].d;
-    const float dmin = x[ib].dmin;
-
-    uint8_t sc, m;
-    get_scale_min_k4(is + 0, x[ib].scales, sc, m);
-    const float d1 = dall * sc;
-    const float m1 = dmin * m;
-    get_scale_min_k4(is + 1, x[ib].scales, sc, m);
-    const float d2 = dall * sc;
-    const float m2 = dmin * m;
-
-    float sum = 0;
-    for (int k = 0; k < 4; ++k) {
-        sum += y[k +  0] * (d1 * (q[k] & 0xF) - m1);
-        sum += y[k + 32] * (d2 * (q[k] >>  4) - m2);
-    }
-    result = sum;
-
-}
-
 static __global__ void dequantize_block_q5_K(const void * vx, float * yy) {
     const block_q5_K * x = (const block_q5_K *) vx;
 
@@ -544,43 +478,6 @@ static __global__ void dequantize_block_q5_K(const void * vx, float * yy) {
     y[33] = d2 * ((ql[ 1] >>  4) + (qh[ 1] & hm ? 16 : 0)) - m2;
 }
 
-static __device__ void vec_dot_q5_K(const void * vx, const int ib, const int iqs, const float * yy, float & result) {
-
-    const block_q5_K * x = (const block_q5_K *) vx;
-
-                                    // iqs is in 0...248 in steps of 8 =>
-    const int j  = iqs / 64;        // j  is in 0...3
-    const int ir = (iqs - 64*j)/2;  // ir is in 0...28 in steps of 4
-    const int is = 2*j;             // is is in 0...6 in steps of 2
-
-    const float   * y  = yy + 64*j + ir;
-    const uint8_t * ql = x[ib].qs + 32*j + ir;
-    const uint8_t * qh = x[ib].qh + ir;
-
-    const float dall = x[ib].d;
-    const float dmin = x[ib].dmin;
-
-    uint8_t sc, m;
-    get_scale_min_k4(is + 0, x[ib].scales, sc, m);
-    const float d1 = dall * sc;
-    const float m1 = dmin * m;
-    get_scale_min_k4(is + 1, x[ib].scales, sc, m);
-    const float d2 = dall * sc;
-    const float m2 = dmin * m;
-
-    uint8_t   hm  = 1 << is;
-    float sum = 0;
-    for (int k = 0; k < 4; ++k) {
-        sum += y[k +  0] * (d1 * ((ql[k] & 0xF) + (qh[k] & hm ? 16 : 0)) - m1);
-    }
-    hm <<= 1;
-    for (int k = 0; k < 4; ++k) {
-        sum += y[k + 32] * (d2 * ((ql[k] >>  4) + (qh[k] & hm ? 16 : 0)) - m2);
-    }
-    result = sum;
-
-}
-
 static __global__ void dequantize_block_q6_K(const void * vx, float * yy) {
     const block_q6_K * x = (const block_q6_K *) vx;
 
@@ -606,38 +503,395 @@ static __global__ void dequantize_block_q6_K(const void * vx, float * yy) {
     y[96] = d * sc[6] * ((int8_t)((ql[32]  >> 4) | (((qh >> 6) & 3) << 4)) - 32);
 }
 
-static __device__ void vec_dot_q6_K(const void * vx, const int ib, const int iqs, const float * yy, float & result) {
+static __global__ void dequantize_mul_mat_vec_q2_k(const void * vx, const float * yy, float * dst, const int ncols, int nrows) {
 
-    const block_q6_K * x = (const block_q6_K *) vx;
+    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;
+
+    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;
+
+    float tmp = 0; // partial sum for thread in warp
+
+    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;
+
+    }
+
+    // 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_q3_k(const void * vx, const float * yy, float * dst, const int ncols, int nrows) {
+
+    const uint16_t kmask1 = 0x0303;
+    const uint16_t kmask2 = 0x0f0f;
+
+    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;
+
+    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;
+
+    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 * 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;
+
+    }
+
+    // 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_q4_k(const void * vx, const float * yy, float * dst, const int ncols, int nrows) {
+
+    const uint16_t kmask1 = 0x3f3f;
+    const uint16_t kmask2 = 0x0f0f;
+    const uint16_t kmask3 = 0xc0c0;
+
+    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 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;
+
+    const block_q4_K * x = (const block_q4_K *)vx + ib0;
+
+    float tmp = 0; // partial sum for thread in warp
 
-    const int ip = iqs / 128;        // 0 or 1
-    const int il = (iqs - 128*ip)/8; // 0...15
-    const int is = 8*ip;
+    for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
 
-    const float * y = yy + 128*ip + il;
+        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 d = x[ib].d;
+        const float dall = x[i].d;
+        const float dmin = x[i].dmin;
 
-    const uint8_t * ql = x[ib].ql + 64*ip + il;
-    const uint8_t * qh = x[ib].qh + 32*ip + il;
-    const int8_t  * sc = x[ib].scales + is;
+        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);
 
-    result = y[  0] * d * sc[0] * ((int8_t)((ql[ 0] & 0xF) | (((qh[ 0] >> 0) & 3) << 4)) - 32)
-           + y[ 32] * d * sc[2] * ((int8_t)((ql[32] & 0xF) | (((qh[ 0] >> 2) & 3) << 4)) - 32)
-           + y[ 64] * d * sc[4] * ((int8_t)((ql[ 0]  >> 4) | (((qh[ 0] >> 4) & 3) << 4)) - 32)
-           + y[ 96] * d * sc[6] * ((int8_t)((ql[32]  >> 4) | (((qh[ 0] >> 6) & 3) << 4)) - 32)
-           + y[ 16] * d * sc[1] * ((int8_t)((ql[16] & 0xF) | (((qh[16] >> 0) & 3) << 4)) - 32)
-           + y[ 48] * d * sc[3] * ((int8_t)((ql[48] & 0xF) | (((qh[16] >> 2) & 3) << 4)) - 32)
-           + y[ 80] * d * sc[5] * ((int8_t)((ql[16]  >> 4) | (((qh[16] >> 4) & 3) << 4)) - 32)
-           + y[112] * d * sc[7] * ((int8_t)((ql[48]  >> 4) | (((qh[16] >> 6) & 3) << 4)) - 32);
+        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;
+
+    }
+
+    // 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 uint16_t kmask1 = 0x3f3f;
+    const uint16_t kmask2 = 0x0f0f;
+    const uint16_t kmask3 = 0xc0c0;
+
+    //const int row = blockIdx.x*blockDim.y + threadIdx.y;
+    const int row = blockIdx.x;
+    const int num_blocks_per_row = ncols / QK_K;
+    const int ib0 = row*num_blocks_per_row;
+
+    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;
+
+    const block_q5_K * x = (const block_q5_K *)vx + ib0;
+
+    float tmp = 0; // partial sum for thread in warp
+
+    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;
+
+    }
+
+    // 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_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;
+
+    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
+
+    }
+
+    // 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, float & v0, float & v1){
+static __device__ void convert_f16(const void * vx, const int ib, const int iqs, dfloat2 & v){
     const half * x = (const half *) vx;
 
-    v0 = __half2float(x[ib + iqs + 0]);
-    v1 = __half2float(x[ib + iqs + 1]);
+    // 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>
@@ -654,13 +908,15 @@ static __global__ void dequantize_block(const void * vx, float * y, const int k)
     const int y_offset = qr == 1 ? 1 : qk/2;
 
     // dequantize
-    float & v0 = y[iybs + iqs + 0];
-    float & v1 = y[iybs + iqs + y_offset];
-    dequantize_kernel(vx, ib, iqs, v0, v1);
+    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 float * y, float * dst, const int ncols, const int nrows) {
+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;
@@ -675,7 +931,12 @@ static __global__ void dequantize_mul_mat_vec(const void * vx, const float * y,
     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;
 
-    float tmp = 0.0f; // partial sum for thread in warp
+// 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;
@@ -689,14 +950,21 @@ static __global__ void dequantize_mul_mat_vec(const void * vx, const float * y,
             // process 2 vals per j iter
 
             // dequantize
-            float v0, v1;
-            dequantize_kernel(vx, ib, iqs + j/qr, v0, v1);
             // 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
-            tmp += v0 * y[iybs + iqs + j/qr + 0];
-            tmp += v1 * y[iybs + iqs + j/qr + y_offset];
             // 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
         }
     }
 
@@ -708,47 +976,11 @@ static __global__ void dequantize_mul_mat_vec(const void * vx, const float * y,
     }
 
     if (tid == 0) {
+#ifdef GGML_CUDA_DMMV_F16
+        dst[row] = tmp.x + tmp.y;
+#else
         dst[row] = tmp;
-    }
-}
-
-template <int n_thread, dot_kernel_k_t dot_kernel>
-static __global__ void dequantize_mul_mat_vec_k(const void * vx, const float * y, float * dst, const int ncols, const int nrows) {
-    const int row = blockIdx.y*blockDim.y + threadIdx.y;
-
-    if (row >= nrows) {
-        return;
-    }
-
-    const int tid = threadIdx.x;
-
-    const int iter_stride = QK_K;
-    const int vals_per_iter = iter_stride / n_thread;
-    const int num_blocks_per_row = ncols / QK_K;
-    const int ib0 = row*num_blocks_per_row;
-
-    float tmp = 0; // partial sum for thread in warp
-
-    for (int i = 0; i < ncols; i += iter_stride) {
-        const int col = i + vals_per_iter*tid;
-        const int ib = ib0 + col/QK_K; // x block index
-        const int iqs = col%QK_K; // x quant index
-        const int iybs = col - col%QK_K; // y block start index
-
-        float v;
-        dot_kernel(vx, ib, iqs, y + iybs, v);
-        tmp += v;
-    }
-
-    // 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;
+#endif // GGML_CUDA_DMMV_F16
     }
 }
 
@@ -1043,7 +1275,7 @@ static void dequantize_row_q6_K_cuda(const void * vx, float * y, const int k, cu
     dequantize_block_q6_K<<<nb, 64, 0, stream>>>(vx, y);
 }
 
-static void dequantize_mul_mat_vec_q4_0_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
+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);
@@ -1052,7 +1284,7 @@ static void dequantize_mul_mat_vec_q4_0_cuda(const void * vx, const float * y, f
         <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
 }
 
-static void dequantize_mul_mat_vec_q4_1_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
+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);
@@ -1061,7 +1293,7 @@ static void dequantize_mul_mat_vec_q4_1_cuda(const void * vx, const float * y, f
         <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
 }
 
-static void dequantize_mul_mat_vec_q5_0_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
+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);
@@ -1070,7 +1302,7 @@ static void dequantize_mul_mat_vec_q5_0_cuda(const void * vx, const float * y, f
         <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
 }
 
-static void dequantize_mul_mat_vec_q5_1_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
+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);
@@ -1079,7 +1311,7 @@ static void dequantize_mul_mat_vec_q5_1_cuda(const void * vx, const float * y, f
         <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
 }
 
-static void dequantize_mul_mat_vec_q8_0_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
+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);
@@ -1090,47 +1322,44 @@ static void dequantize_mul_mat_vec_q8_0_cuda(const void * vx, const float * y, f
 
 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;
+    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_k<32, vec_dot_q2_K><<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
+    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;
+    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_k<32, vec_dot_q3_K><<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
+    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;
+    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_k<32, vec_dot_q4_K><<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
+    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 int ny = 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_k<32, vec_dot_q5_K><<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
+    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;
+    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_k<32, vec_dot_q6_K><<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
+    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) {
@@ -1138,7 +1367,7 @@ static void convert_fp16_to_fp32_cuda(const void * vx, float * y, const int k, c
     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 float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
+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);
@@ -1306,19 +1535,13 @@ 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;
 
-#define GGML_CUDA_MAX_STREAMS 8 // Set this to 1 for reproducible matrix multiplication.
-#define GGML_CUDA_MAX_EVENTS 64
-
 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][GGML_CUDA_MAX_STREAMS] = { nullptr };
-
-static cudaStream_t g_cudaStreams_memcpy_src1[GGML_CUDA_MAX_DEVICES][GGML_CUDA_MAX_STREAMS] = { nullptr };
-static cudaEvent_t g_cudaEvents_memcpy_src1[GGML_CUDA_MAX_DEVICES][GGML_CUDA_MAX_EVENTS] = { nullptr };
+static cudaStream_t g_cudaStreams_main[GGML_CUDA_MAX_DEVICES] = { nullptr };
 
 void ggml_init_cublas() {
     static bool initialized = false;
@@ -1342,15 +1565,8 @@ void ggml_init_cublas() {
         for (int id = 0; id < g_device_count; ++id) {
             CUDA_CHECK(cudaSetDevice(id));
 
-            // create streams
-            for (int i = 0; i < GGML_CUDA_MAX_STREAMS; ++i) {
-                CUDA_CHECK(cudaStreamCreateWithFlags(&g_cudaStreams_main[id][i], cudaStreamNonBlocking));
-                CUDA_CHECK(cudaStreamCreateWithFlags(&g_cudaStreams_memcpy_src1[id][i], cudaStreamNonBlocking));
-            }
-            // create events
-            for (int i = 0; i < GGML_CUDA_MAX_EVENTS; ++i) {
-                CUDA_CHECK(cudaEventCreateWithFlags(&g_cudaEvents_memcpy_src1[id][i], cudaEventDisableTiming));
-            }
+            // create main stream
+            CUDA_CHECK(cudaStreamCreateWithFlags(&g_cudaStreams_main[id], cudaStreamNonBlocking));
 
             // create cublas handle
             CUBLAS_CHECK(cublasCreate(&g_cublas_handles[id]));
@@ -1566,21 +1782,40 @@ inline void ggml_cuda_op_dequantize_mul_mat_vec(
     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_ddf_i, dst_ddf_i, ne00, nrows, cudaStream_main);
+            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_ddf_i, dst_ddf_i, ne00, nrows, cudaStream_main);
+            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_ddf_i, dst_ddf_i, ne00, nrows, cudaStream_main);
+            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_ddf_i, dst_ddf_i, ne00, nrows, cudaStream_main);
+            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_ddf_i, dst_ddf_i, ne00, nrows, cudaStream_main);
+            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);
@@ -1598,7 +1833,7 @@ inline void ggml_cuda_op_dequantize_mul_mat_vec(
             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_ddf_i, dst_ddf_i, ne00, nrows, cudaStream_main);
+            convert_mul_mat_vec_f16_cuda(src0_ddq_i, src1_dfloat, dst_ddf_i, ne00, nrows, cudaStream_main);
             break;
         default:
             GGML_ASSERT(false);
@@ -1606,6 +1841,12 @@ inline void ggml_cuda_op_dequantize_mul_mat_vec(
     }
     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;
@@ -1817,6 +2058,12 @@ static void ggml_cuda_op(const ggml_tensor * src0, const ggml_tensor * src1, ggm
     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;
@@ -1915,9 +2162,7 @@ static void ggml_cuda_op(const ggml_tensor * src0, const ggml_tensor * src1, ggm
                 }
                 const int64_t i11 = i13*ne12 + i12;
 
-                cudaStream_t cudaStream_main        =        g_cudaStreams_main[id][i0 % GGML_CUDA_MAX_STREAMS];
-                cudaStream_t cudaStream_memcpy_src1 = g_cudaStreams_memcpy_src1[id][i0 % GGML_CUDA_MAX_STREAMS];
-                cudaEvent_t  cudaEvent_memcpy_src1  =  g_cudaEvents_memcpy_src1[id][i0 % GGML_CUDA_MAX_EVENTS];
+                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;
@@ -1945,14 +2190,14 @@ static void ggml_cuda_op(const ggml_tensor * src0, const ggml_tensor * src1, ggm
                     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_memcpy_src1));
+                        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_memcpy_src1));
+                                                    cudaMemcpyDeviceToDevice, cudaStream_main));
                         }
                     } else if (src1_on_device && !src1_is_contiguous) {
                         GGML_ASSERT(!split);
@@ -1961,7 +2206,6 @@ static void ggml_cuda_op(const ggml_tensor * src0, const ggml_tensor * src1, ggm
                         GGML_ASSERT(false);
                     }
                 }
-                CUDA_CHECK(cudaEventRecord(cudaEvent_memcpy_src1, cudaStream_memcpy_src1));
 
                 if (!src0_on_device || !src0_is_contiguous) {
                     if (src0_is_f32) {
@@ -1977,9 +2221,6 @@ static void ggml_cuda_op(const ggml_tensor * src0, const ggml_tensor * src1, ggm
                     CUDA_CHECK(cudaGetLastError());
                 }
 
-                // wait with main stream until src1 memcpy is done
-                CUDA_CHECK(cudaStreamWaitEvent(cudaStream_main, cudaEvent_memcpy_src1, 0));
-
                 // 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);
 
@@ -2017,8 +2258,13 @@ static void ggml_cuda_op(const ggml_tensor * src0, const ggml_tensor * src1, ggm
 
     // 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]);
         }
@@ -2084,7 +2330,7 @@ void ggml_cuda_mul_mat_vec_p021(const ggml_tensor * src0, const ggml_tensor * sr
     const int64_t ne02 = src0->ne[2];
 
     CUDA_CHECK(cudaSetDevice(g_main_device));
-    cudaStream_t cudaStream_main = g_cudaStreams_main[g_main_device][0];
+    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];
@@ -2096,8 +2342,6 @@ void ggml_cuda_mul_mat_vec_p021(const ggml_tensor * src0, const ggml_tensor * sr
     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);
-
-    CUDA_CHECK(cudaDeviceSynchronize());
 }
 
 void ggml_cuda_mul_mat_vec_nc(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst){
@@ -2115,7 +2359,7 @@ void ggml_cuda_mul_mat_vec_nc(const ggml_tensor * src0, const ggml_tensor * src1
     const int64_t nb02 = src0->nb[2];
 
     CUDA_CHECK(cudaSetDevice(g_main_device));
-    cudaStream_t cudaStream_main = g_cudaStreams_main[g_main_device][0];
+    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];
@@ -2130,8 +2374,6 @@ void ggml_cuda_mul_mat_vec_nc(const ggml_tensor * src0, const ggml_tensor * src1
     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);
-
-    CUDA_CHECK(cudaDeviceSynchronize());
 }
 
 void ggml_cuda_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
@@ -2187,7 +2429,7 @@ void ggml_cuda_cpy(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tens
     const int64_t nb12 = src1->nb[2];
 
     CUDA_CHECK(cudaSetDevice(g_main_device));
-    cudaStream_t cudaStream_main = g_cudaStreams_main[g_main_device][0];
+    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;
@@ -2205,8 +2447,6 @@ void ggml_cuda_cpy(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tens
         GGML_ASSERT(false);
     }
 
-    CUDA_CHECK(cudaDeviceSynchronize());
-
     (void) dst;
 }