llama_cpp 0.2.2 → 0.3.0

data/ext/llama_cpp/src/k_quants.c

@@ -261,6 +261,7 @@ static float make_qkx1_quants(int n, int nmax, const float * restrict x, uint8_t
     return scale;
 }
 
+#if QK_K == 256
 static inline void get_scale_min_k4(int j, const uint8_t * restrict q, uint8_t * restrict d, uint8_t * restrict m) {
     if (j < 4) {
         *d = q[j] & 63; *m = q[j + 4] & 63;
@@ -269,6 +270,7 @@ static inline void get_scale_min_k4(int j, const uint8_t * restrict q, uint8_t *
         *m = (q[j+4] >>  4) | ((q[j-0] >> 6) << 4);
     }
 }
+#endif
 
 //========================- 2-bit (de)-quantization
 
@@ -330,11 +332,17 @@ void quantize_row_q2_K_reference(const float * restrict x, block_q2_K * restrict
             }
         }
 
+#if QK_K == 256
         for (int j = 0; j < QK_K; j += 128) {
             for (int l = 0; l < 32; ++l) {
                 y[i].qs[j/4 + l] = L[j + l] | (L[j + l + 32] << 2) | (L[j + l + 64] << 4) | (L[j + l + 96] << 6);
             }
         }
+#else
+        for (int l = 0; l < 16; ++l) {
+            y[i].qs[l] = L[l] | (L[l + 16] << 2) | (L[l + 32] << 4) | (L[l + 48] << 6);
+        }
+#endif
 
         x += QK_K;
 
@@ -352,6 +360,7 @@ void dequantize_row_q2_K(const block_q2_K * restrict x, float * restrict y, int
 
         const uint8_t * q = x[i].qs;
 
+#if QK_K == 256
         int is = 0;
         float dl, ml;
         for (int n = 0; n < QK_K; n += 128) {
@@ -370,7 +379,19 @@ void dequantize_row_q2_K(const block_q2_K * restrict x, float * restrict y, int
             }
             q += 32;
         }
-
+#else
+        float dl1 = d * (x[i].scales[0] & 0xF), ml1 = min * (x[i].scales[0] >> 4);
+        float dl2 = d * (x[i].scales[1] & 0xF), ml2 = min * (x[i].scales[1] >> 4);
+        float dl3 = d * (x[i].scales[2] & 0xF), ml3 = min * (x[i].scales[2] >> 4);
+        float dl4 = d * (x[i].scales[3] & 0xF), ml4 = min * (x[i].scales[3] >> 4);
+        for (int l = 0; l < 16; ++l) {
+            y[l+ 0] = dl1 * ((int8_t)((q[l] >> 0) & 3)) - ml1;
+            y[l+16] = dl2 * ((int8_t)((q[l] >> 2) & 3)) - ml2;
+            y[l+32] = dl3 * ((int8_t)((q[l] >> 4) & 3)) - ml3;
+            y[l+48] = dl4 * ((int8_t)((q[l] >> 6) & 3)) - ml4;
+        }
+        y += QK_K;
+#endif
     }
 }
 
@@ -412,6 +433,7 @@ void quantize_row_q3_K_reference(const float * restrict x, block_q3_K * restrict
             }
         }
 
+#if QK_K == 256
         memset(y[i].scales, 0, 12);
         if (max_scale) {
             float iscale = -32.f/max_scale;
@@ -445,9 +467,39 @@ void quantize_row_q3_K_reference(const float * restrict x, block_q3_K * restrict
                 L[16*j + ii] = l + 4;
             }
         }
+#else
+        if (max_scale) {
+            float iscale = -8.f/max_scale;
+            for (int j = 0; j < QK_K/16; j+=2) {
+                int l1 = nearest_int(iscale*scales[j]);
+                l1 = 8 + MAX(-8, MIN(7, l1));
+                int l2 = nearest_int(iscale*scales[j+1]);
+                l2 = 8 + MAX(-8, MIN(7, l2));
+                y[i].scales[j/2] = l1 | (l2 << 4);
+            }
+            y[i].d = ggml_fp32_to_fp16(1/iscale);
+        } else {
+            for (int j = 0; j < QK_K/16; j+=2) {
+                y[i].scales[j/2] = 0;
+            }
+            y[i].d = ggml_fp32_to_fp16(0.f);
+        }
+        for (int j = 0; j < QK_K/16; ++j) {
+            int s = j%2 == 0 ? y[i].scales[j/2] & 0xF : y[i].scales[j/2] >> 4;
+            float d = ggml_fp16_to_fp32(y[i].d) * (s - 8);
+            if (!d) {
+                continue;
+            }
+            for (int ii = 0; ii < 16; ++ii) {
+                int l = nearest_int(x[16*j + ii]/d);
+                l = MAX(-4, MIN(3, l));
+                L[16*j + ii] = l + 4;
+            }
+        }
+#endif
 
         memset(y[i].hmask, 0, QK_K/8);
-        // We put the high-bit for the 1st 32 quants into bit 0, the next 32 into bit 1, etc.
+        // We put the high-bit for the 1st 8 quants into bit 0, the next 8 into bit 1, etc.
         int m = 0;
         uint8_t hm = 1;
         for (int j = 0; j < QK_K; ++j) {
@@ -459,19 +511,25 @@ void quantize_row_q3_K_reference(const float * restrict x, block_q3_K * restrict
                 m = 0; hm <<= 1;
             }
         }
+#if QK_K == 256
         for (int j = 0; j < QK_K; j += 128) {
             for (int l = 0; l < 32; ++l) {
                 y[i].qs[j/4 + l] = L[j + l] | (L[j + l + 32] << 2) | (L[j + l + 64] << 4) | (L[j + l + 96] << 6);
             }
         }
+#else
+        for (int l = 0; l < 16; ++l) {
+            y[i].qs[l] = L[l] | (L[l + 16] << 2) | (L[l + 32] << 4) | (L[l + 48] << 6);
+        }
+#endif
 
         x += QK_K;
     }
 }
 
+#if QK_K == 256
 void dequantize_row_q3_K(const block_q3_K * restrict x, float * restrict y, int k) {
     assert(k % QK_K == 0);
-    assert(QK_K == 256);
     const int nb = k / QK_K;
 
     const uint32_t kmask1 = 0x03030303;
@@ -519,6 +577,39 @@ void dequantize_row_q3_K(const block_q3_K * restrict x, float * restrict y, int
 
     }
 }
+#else
+void dequantize_row_q3_K(const block_q3_K * restrict x, float * restrict y, int k) {
+    assert(k % QK_K == 0);
+    assert(QK_K == 64);
+    const int nb = k / QK_K;
+
+    for (int i = 0; i < nb; i++) {
+
+        const float d_all = ggml_fp16_to_fp32(x[i].d);
+
+        const uint8_t * restrict q = x[i].qs;
+        const uint8_t * restrict hm = x[i].hmask;
+
+        const float d1 = d_all * ((x[i].scales[0] & 0xF) - 8);
+        const float d2 = d_all * ((x[i].scales[0] >>  4) - 8);
+        const float d3 = d_all * ((x[i].scales[1] & 0xF) - 8);
+        const float d4 = d_all * ((x[i].scales[1] >>  4) - 8);
+
+        for (int l=0; l<8; ++l) {
+            uint8_t h = hm[l];
+            y[l+ 0] = d1 * ((int8_t)((q[l+0] >> 0) & 3) - ((h & 0x01) ? 0 : 4));
+            y[l+ 8] = d1 * ((int8_t)((q[l+8] >> 0) & 3) - ((h & 0x02) ? 0 : 4));
+            y[l+16] = d2 * ((int8_t)((q[l+0] >> 2) & 3) - ((h & 0x04) ? 0 : 4));
+            y[l+24] = d2 * ((int8_t)((q[l+8] >> 2) & 3) - ((h & 0x08) ? 0 : 4));
+            y[l+32] = d3 * ((int8_t)((q[l+0] >> 4) & 3) - ((h & 0x10) ? 0 : 4));
+            y[l+40] = d3 * ((int8_t)((q[l+8] >> 4) & 3) - ((h & 0x20) ? 0 : 4));
+            y[l+48] = d4 * ((int8_t)((q[l+0] >> 6) & 3) - ((h & 0x40) ? 0 : 4));
+            y[l+56] = d4 * ((int8_t)((q[l+8] >> 6) & 3) - ((h & 0x80) ? 0 : 4));
+        }
+        y += QK_K;
+    }
+}
+#endif
 
 void quantize_row_q3_K(const float * restrict x, void * restrict vy, int k) {
     quantize_row_q3_K_reference(x, vy, k);
@@ -563,6 +654,7 @@ void quantize_row_q4_K_reference(const float * restrict x, block_q4_K * restrict
             }
         }
 
+#if QK_K == 256
         float inv_scale = max_scale > 0 ? 63.f/max_scale : 0.f;
         float inv_min   = max_min   > 0 ? 63.f/max_min   : 0.f;
         for (int j = 0; j < QK_K/32; ++j) {
@@ -594,9 +686,43 @@ void quantize_row_q4_K_reference(const float * restrict x, block_q4_K * restrict
                 L[32*j + ii] = l;
             }
         }
+#else
+        const float s_factor = 15.f;
+        float inv_scale = max_scale > 0 ? s_factor/max_scale : 0.f;
+        float inv_min   = max_min   > 0 ? s_factor/max_min   : 0.f;
+        int d1 = nearest_int(inv_scale*scales[0]);
+        int m1 = nearest_int(inv_min*mins[0]);
+        int d2 = nearest_int(inv_scale*scales[1]);
+        int m2 = nearest_int(inv_min*mins[1]);
+        y[i].scales[0] = d1 | (m1 << 4);
+        y[i].scales[1] = d2 | (m2 << 4);
+        y[i].d[0] = ggml_fp32_to_fp16(max_scale/s_factor);
+        y[i].d[1] = ggml_fp32_to_fp16(max_min/s_factor);
+
+        float sumlx = 0;
+        int   suml2 = 0;
+        for (int j = 0; j < QK_K/32; ++j) {
+            const uint8_t sd = y[i].scales[j] & 0xF;
+            const uint8_t sm = y[i].scales[j] >>  4;
+            const float d = ggml_fp16_to_fp32(y[i].d[0]) * sd;
+            if (!d) continue;
+            const float m = ggml_fp16_to_fp32(y[i].d[1]) * sm;
+            for (int ii = 0; ii < 32; ++ii) {
+                int l = nearest_int((x[32*j + ii] + m)/d);
+                l = MAX(0, MIN(15, l));
+                L[32*j + ii] = l;
+                sumlx += (x[32*j + ii] + m)*l*sd;
+                suml2 += l*l*sd*sd;
+            }
+        }
+        if (suml2) {
+            y[i].d[0] = ggml_fp32_to_fp16(sumlx/suml2);
+        }
+#endif
         uint8_t * q = y[i].qs;
         for (int j = 0; j < QK_K; j += 64) {
-            for (int l = 0; l < 32; ++l) *q++ = L[j + l] | (L[j + l + 32] << 4);
+            for (int l = 0; l < 32; ++l) q[l] = L[j + l] | (L[j + l + 32] << 4);
+            q += 32;
         }
 
         x += QK_K;
@@ -610,11 +736,13 @@ void dequantize_row_q4_K(const block_q4_K * restrict x, float * restrict y, int
 
     for (int i = 0; i < nb; i++) {
 
-        const float d = ggml_fp16_to_fp32(x[i].d);
-        const float min = ggml_fp16_to_fp32(x[i].dmin);
-
         const uint8_t * q = x[i].qs;
 
+#if QK_K == 256
+
+        const float d   = ggml_fp16_to_fp32(x[i].d);
+        const float min = ggml_fp16_to_fp32(x[i].dmin);
+
         int is = 0;
         uint8_t sc, m;
         for (int j = 0; j < QK_K; j += 64) {
@@ -626,6 +754,17 @@ void dequantize_row_q4_K(const block_q4_K * restrict x, float * restrict y, int
             for (int l = 0; l < 32; ++l) *y++ = d2 * (q[l]  >> 4) - m2;
             q += 32; is += 2;
         }
+#else
+        const float dall = ggml_fp16_to_fp32(x[i].d[0]);
+        const float mall = ggml_fp16_to_fp32(x[i].d[1]);
+        const float d1 = dall * (x[i].scales[0] & 0xF), m1 = mall * (x[i].scales[0] >> 4);
+        const float d2 = dall * (x[i].scales[1] & 0xF), m2 = mall * (x[i].scales[1] >> 4);
+        for (int l = 0; l < 32; ++l) {
+            y[l+ 0] = d1 * (q[l] & 0xF) - m1;
+            y[l+32] = d2 * (q[l] >>  4) - m2;
+        }
+        y += QK_K;
+#endif
 
     }
 }
@@ -653,12 +792,19 @@ void quantize_row_q5_K_reference(const float * restrict x, block_q5_K * restrict
     assert(k % QK_K == 0);
     const int nb = k / QK_K;
 
+#if QK_K == 256
     uint8_t L[QK_K];
     float mins[QK_K/32];
     float scales[QK_K/32];
+#else
+    int8_t L[QK_K];
+    float scales[QK_K/16];
+#endif
 
     for (int i = 0; i < nb; i++) {
 
+#if QK_K == 256
+
         float max_scale = 0; // as we are deducting the min, scales are always positive
         float max_min = 0;
         for (int j = 0; j < QK_K/32; ++j) {
@@ -725,6 +871,52 @@ void quantize_row_q5_K_reference(const float * restrict x, block_q5_K * restrict
             m1 <<= 2; m2 <<= 2;
             ql += 32;
         }
+#else
+        float max_scale = 0, amax = 0;
+        for (int j = 0; j < QK_K/16; ++j) {
+            scales[j] = make_qx_quants(16, 16, x + 16*j, L + 16*j, 1);
+            float abs_scale = fabsf(scales[j]);
+            if (abs_scale > amax) {
+                amax = abs_scale;
+                max_scale = scales[j];
+            }
+        }
+
+        float iscale = -128.f/max_scale;
+        for (int j = 0; j < QK_K/16; ++j) {
+            int l = nearest_int(iscale*scales[j]);
+            y[i].scales[j] = MAX(-128, MIN(127, l));
+        }
+        y[i].d = ggml_fp32_to_fp16(1/iscale);
+
+        for (int j = 0; j < QK_K/16; ++j) {
+            const float d = ggml_fp16_to_fp32(y[i].d) * y[i].scales[j];
+            if (!d) continue;
+            for (int ii = 0; ii < 16; ++ii) {
+                int l = nearest_int(x[16*j + ii]/d);
+                l = MAX(-16, MIN(15, l));
+                L[16*j + ii] = l + 16;
+            }
+        }
+
+        uint8_t * restrict qh = y[i].qh;
+        uint8_t * restrict ql = y[i].qs;
+        memset(qh, 0, QK_K/8);
+
+        for (int j = 0; j < 32; ++j) {
+            int jm = j%8;
+            int is = j/8;
+            int l1 = L[j];
+            if (l1 > 15) {
+                l1 -= 16; qh[jm] |= (1 << is);
+            }
+            int l2 = L[j + 32];
+            if (l2 > 15) {
+                l2 -= 16; qh[jm] |= (1 << (4 + is));
+            }
+            ql[j] = l1 | (l2 << 4);
+        }
+#endif
 
         x += QK_K;
 
@@ -737,12 +929,14 @@ void dequantize_row_q5_K(const block_q5_K * restrict x, float * restrict y, int
 
     for (int i = 0; i < nb; i++) {
 
-        const float d = ggml_fp16_to_fp32(x[i].d);
-        const float min = ggml_fp16_to_fp32(x[i].dmin);
-
         const uint8_t * ql = x[i].qs;
         const uint8_t * qh = x[i].qh;
 
+#if QK_K == 256
+
+        const float d = ggml_fp16_to_fp32(x[i].d);
+        const float min = ggml_fp16_to_fp32(x[i].dmin);
+
         int is = 0;
         uint8_t sc, m;
         uint8_t u1 = 1, u2 = 2;
@@ -756,6 +950,21 @@ void dequantize_row_q5_K(const block_q5_K * restrict x, float * restrict y, int
             ql += 32; is += 2;
             u1 <<= 2; u2 <<= 2;
         }
+#else
+        float d = ggml_fp16_to_fp32(x[i].d);
+        const int8_t * restrict s = x[i].scales;
+        for (int l = 0; l < 8; ++l) {
+            y[l+ 0] = d * s[0] * ((ql[l+ 0] & 0xF) - (qh[l] & 0x01 ? 0 : 16));
+            y[l+ 8] = d * s[0] * ((ql[l+ 8] & 0xF) - (qh[l] & 0x02 ? 0 : 16));
+            y[l+16] = d * s[1] * ((ql[l+16] & 0xF) - (qh[l] & 0x04 ? 0 : 16));
+            y[l+24] = d * s[1] * ((ql[l+24] & 0xF) - (qh[l] & 0x08 ? 0 : 16));
+            y[l+32] = d * s[2] * ((ql[l+ 0] >>  4) - (qh[l] & 0x10 ? 0 : 16));
+            y[l+40] = d * s[2] * ((ql[l+ 8] >>  4) - (qh[l] & 0x20 ? 0 : 16));
+            y[l+48] = d * s[3] * ((ql[l+16] >>  4) - (qh[l] & 0x40 ? 0 : 16));
+            y[l+56] = d * s[3] * ((ql[l+24] >>  4) - (qh[l] & 0x80 ? 0 : 16));
+        }
+        y += QK_K;
+#endif
     }
 }
 
@@ -823,6 +1032,7 @@ void quantize_row_q6_K_reference(const float * restrict x, block_q6_K * restrict
 
         uint8_t * restrict ql = y[i].ql;
         uint8_t * restrict qh = y[i].qh;
+#if QK_K == 256
         for (int j = 0; j < QK_K; j += 128) {
             for (int l = 0; l < 32; ++l) {
                 const uint8_t q1 = L[j + l +  0] & 0xF;
@@ -836,6 +1046,16 @@ void quantize_row_q6_K_reference(const float * restrict x, block_q6_K * restrict
             ql += 64;
             qh += 32;
         }
+#else
+        for (int l = 0; l < 32; ++l) {
+            const uint8_t q1 = L[l +  0] & 0xF;
+            const uint8_t q2 = L[l + 32] & 0xF;
+            ql[l] = q1 | (q2 << 4);
+        }
+        for (int l = 0; l < 16; ++l) {
+            qh[l] = (L[l] >> 4) | ((L[l + 16] >> 4) << 2) | ((L[l + 32] >> 4) << 4) | ((L[l + 48] >> 4) << 6);
+        }
+#endif
 
         x += QK_K;
 
@@ -854,6 +1074,7 @@ void dequantize_row_q6_K(const block_q6_K * restrict x, float * restrict y, int
         const uint8_t * restrict qh = x[i].qh;
         const int8_t  * restrict sc = x[i].scales;
 
+#if QK_K == 256
         for (int n = 0; n < QK_K; n += 128) {
             for (int l = 0; l < 32; ++l) {
                 int is = l/16;
@@ -871,6 +1092,19 @@ void dequantize_row_q6_K(const block_q6_K * restrict x, float * restrict y, int
             qh += 32;
             sc += 8;
         }
+#else
+        for (int l = 0; l < 16; ++l) {
+            const int8_t q1 = (int8_t)((ql[l+ 0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32;
+            const int8_t q2 = (int8_t)((ql[l+16] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32;
+            const int8_t q3 = (int8_t)((ql[l+ 0]  >> 4) | (((qh[l] >> 4) & 3) << 4)) - 32;
+            const int8_t q4 = (int8_t)((ql[l+16]  >> 4) | (((qh[l] >> 6) & 3) << 4)) - 32;
+            y[l+ 0] = d * sc[0] * q1;
+            y[l+16] = d * sc[1] * q2;
+            y[l+32] = d * sc[2] * q3;
+            y[l+48] = d * sc[3] * q4;
+        }
+        y  += 64;
+#endif
 
     }
 }
@@ -1002,6 +1236,7 @@ static inline __m128i get_scale_shuffle(int i) {
 }
 #endif
 
+#if QK_K == 256
 void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
 
     const block_q2_K * restrict x = vx;
@@ -1158,6 +1393,112 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri
 
     *s = hsum_float_8(acc);
 
+#elif defined __AVX__
+
+    const __m128i m3 = _mm_set1_epi8(0x3);
+    const __m128i m4 = _mm_set1_epi8(0xF);
+    const __m128i m2 = _mm_set1_epi8(0x2);
+
+    __m256 acc = _mm256_setzero_ps();
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float dall = y[i].d * ggml_fp16_to_fp32(x[i].d);
+        const float dmin = -y[i].d * ggml_fp16_to_fp32(x[i].dmin);
+
+        const uint8_t * restrict q2 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        // load mins and scales from block_q2_K.scales[QK_K/16]
+        const __m128i mins_and_scales = _mm_loadu_si128((const __m128i*)x[i].scales);
+        const __m128i scales16 = _mm_and_si128(mins_and_scales, m4);
+        const __m128i mins16 = _mm_and_si128(_mm_srli_epi16(mins_and_scales, 4), m4);
+        const __m128i mins_0 = _mm_cvtepi8_epi16(mins16);
+        const __m128i mins_1 = _mm_cvtepi8_epi16(_mm_unpackhi_epi64(mins16, mins16));
+
+        // summs = y[i].bsums * (x[i].scales >> 4) in 16bits*8*2 to 32bits*4*2
+        const __m128i summs_0 = _mm_madd_epi16(mins_0, _mm_loadu_si128((const __m128i*)&y[i].bsums[0]));
+        const __m128i summs_1 = _mm_madd_epi16(mins_1, _mm_loadu_si128((const __m128i*)&y[i].bsums[8]));
+
+        // sumf += -dmin * summs in 32bits*8
+        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&dmin), _mm256_cvtepi32_ps(_mm256_set_m128i(summs_1, summs_0))), acc);
+
+        const __m128i scales_0 = _mm_cvtepi8_epi16(scales16);
+        const __m128i scales_1 = _mm_cvtepi8_epi16(_mm_unpackhi_epi64(scales16, scales16));
+        const __m128i scales[2] = { scales_0, scales_1 };
+
+        __m128i sumi_0 = _mm_setzero_si128();
+        __m128i sumi_1 = _mm_setzero_si128();
+
+        for (int j = 0; j < QK_K/128; ++j) {
+
+            // load Q8 quants int8*16*8 from block_q8_K.qs[QK_K]
+            const __m128i q8_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_2 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_3 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_4 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_5 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_6 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_7 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+
+            // load 2bits*16*8 from block_q2_K.qs[QK_K/4]
+            __m128i q2bits = _mm_loadu_si128((const __m128i*)q2); q2 += 16;
+            const __m128i q2_0 = _mm_and_si128(q2bits, m3);
+            const __m128i q2_2 = _mm_and_si128(_mm_srli_epi16(q2bits, 2), m3);
+            const __m128i q2_4 = _mm_and_si128(_mm_srli_epi16(q2bits, 4), m3);
+            const __m128i q2_6 = _mm_and_si128(_mm_srli_epi16(q2bits, 6), m3);
+            q2bits = _mm_loadu_si128((const __m128i*)q2); q2 += 16;
+            const __m128i q2_1 = _mm_and_si128(q2bits, m3);
+            const __m128i q2_3 = _mm_and_si128(_mm_srli_epi16(q2bits, 2), m3);
+            const __m128i q2_5 = _mm_and_si128(_mm_srli_epi16(q2bits, 4), m3);
+            const __m128i q2_7 = _mm_and_si128(_mm_srli_epi16(q2bits, 6), m3);
+
+            // isuml = q8[l] * ((q2[l] >> shift) & 3) in 8bits*16*8 to 16bits*8*8
+            __m128i p0 = _mm_maddubs_epi16(q2_0, q8_0);
+            __m128i p1 = _mm_maddubs_epi16(q2_1, q8_1);
+            __m128i p2 = _mm_maddubs_epi16(q2_2, q8_2);
+            __m128i p3 = _mm_maddubs_epi16(q2_3, q8_3);
+            __m128i p4 = _mm_maddubs_epi16(q2_4, q8_4);
+            __m128i p5 = _mm_maddubs_epi16(q2_5, q8_5);
+            __m128i p6 = _mm_maddubs_epi16(q2_6, q8_6);
+            __m128i p7 = _mm_maddubs_epi16(q2_7, q8_7);
+
+            // isum += (x[i].scales[is++] & 0xF) * isuml in 16bits*8*8 to 32bits*4*8
+            __m128i shuffle = _mm_set1_epi16(0x0100);
+            p0 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p0);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            p1 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p1);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            p2 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p2);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            p3 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p3);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            p4 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p4);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            p5 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p5);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            p6 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p6);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            p7 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p7);
+
+            p0 = _mm_add_epi32(p0, p1);
+            p2 = _mm_add_epi32(p2, p3);
+            p4 = _mm_add_epi32(p4, p5);
+            p6 = _mm_add_epi32(p6, p7);
+
+            // isum in 32bits*4*2
+            sumi_0 = _mm_add_epi32(sumi_0, _mm_add_epi32(p0, p2));
+            sumi_1 = _mm_add_epi32(sumi_1, _mm_add_epi32(p4, p6));
+        }
+
+        // sumf += dall * isum - dmin * summs in 32bits
+        __m256i sumi = _mm256_set_m128i(sumi_1, sumi_0);
+        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&dall), _mm256_cvtepi32_ps(sumi)), acc);
+    }
+
+    *s = hsum_float_8(acc);
+
 #else
 
     float sumf = 0;
@@ -1201,6 +1542,168 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri
 #endif
 }
 
+#else
+
+void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
+
+    const block_q2_K * restrict x = vx;
+    const block_q8_K * restrict y = vy;
+
+    const int nb = n / QK_K;
+
+#ifdef __ARM_NEON
+
+    const uint8x16_t m3 = vdupq_n_u8(0x3);
+    const int32x4_t  vzero = vdupq_n_s32(0);
+
+    int8x16x4_t q2bytes;
+
+    uint32_t aux32[2];
+    const uint8_t * scales = (const uint8_t *)aux32;
+
+    float sum = 0;
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * (float)x[i].d;
+        const float dmin = -y[i].d * (float)x[i].dmin;
+
+        const uint8_t * restrict q2 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+        const uint32_t * restrict sc = (const uint32_t *)x[i].scales;
+
+        aux32[0] = sc[0] & 0x0f0f0f0f;
+        aux32[1] = (sc[0] >> 4) & 0x0f0f0f0f;
+
+        sum += dmin * (scales[4] * y[i].bsums[0] + scales[5] * y[i].bsums[1] + scales[6] * y[i].bsums[2] + scales[7] * y[i].bsums[3]);
+
+        int isum1 = 0, isum2 = 0;
+
+        const uint8x16_t q2bits = vld1q_u8(q2);
+
+        const int8x16x4_t q8bytes = vld1q_s8_x4(q8);
+
+        q2bytes.val[0] = vreinterpretq_s8_u8(vandq_u8(q2bits, m3));
+        q2bytes.val[1] = vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q2bits, 2), m3));
+        q2bytes.val[2] = vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q2bits, 4), m3));
+        q2bytes.val[3] = vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q2bits, 6), m3));
+
+#if defined(__ARM_FEATURE_DOTPROD)
+        isum1 += vaddvq_s32(vdotq_s32(vzero, q2bytes.val[0], q8bytes.val[0])) * scales[0];
+        isum2 += vaddvq_s32(vdotq_s32(vzero, q2bytes.val[1], q8bytes.val[1])) * scales[1];
+        isum1 += vaddvq_s32(vdotq_s32(vzero, q2bytes.val[2], q8bytes.val[2])) * scales[2];
+        isum2 += vaddvq_s32(vdotq_s32(vzero, q2bytes.val[3], q8bytes.val[3])) * scales[3];
+#else
+        const int16x8_t p1 = vaddq_s16(vmull_s8(vget_low_s8 (q2bytes.val[0]), vget_low_s8 (q8bytes.val[0])),
+                                       vmull_s8(vget_high_s8(q2bytes.val[0]), vget_high_s8(q8bytes.val[0])));
+        const int16x8_t p2 = vaddq_s16(vmull_s8(vget_low_s8 (q2bytes.val[1]), vget_low_s8 (q8bytes.val[1])),
+                                       vmull_s8(vget_high_s8(q2bytes.val[1]), vget_high_s8(q8bytes.val[1])));
+        isum1 += vaddvq_s16(p1) * scales[0];
+        isum2 += vaddvq_s16(p2) * scales[1];
+
+        const int16x8_t p3 = vaddq_s16(vmull_s8(vget_low_s8 (q2bytes.val[2]), vget_low_s8 (q8bytes.val[2])),
+                                       vmull_s8(vget_high_s8(q2bytes.val[2]), vget_high_s8(q8bytes.val[2])));
+        const int16x8_t p4 = vaddq_s16(vmull_s8(vget_low_s8 (q2bytes.val[3]), vget_low_s8 (q8bytes.val[3])),
+                                       vmull_s8(vget_high_s8(q2bytes.val[3]), vget_high_s8(q8bytes.val[3])));
+        isum1 += vaddvq_s16(p3) * scales[2];
+        isum2 += vaddvq_s16(p4) * scales[3];
+#endif
+        sum += d * (isum1 + isum2);
+
+    }
+
+    *s = sum;
+
+#elif defined __AVX2__
+
+    const __m256i m3 = _mm256_set1_epi8(3);
+
+    __m256 acc = _mm256_setzero_ps();
+
+    uint32_t ud, um;
+    const uint8_t * restrict db = (const uint8_t *)&ud;
+    const uint8_t * restrict mb = (const uint8_t *)&um;
+
+    float summs = 0;
+
+    // TODO: optimize this
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * ggml_fp16_to_fp32(x[i].d);
+        const float dmin = -y[i].d * ggml_fp16_to_fp32(x[i].dmin);
+
+        const uint8_t * restrict q2 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        const uint32_t * restrict sc = (const uint32_t *)x[i].scales;
+        ud = (sc[0] >> 0) & 0x0f0f0f0f;
+        um = (sc[0] >> 4) & 0x0f0f0f0f;
+
+        int32_t smin = mb[0] * y[i].bsums[0] + mb[1] * y[i].bsums[1] + mb[2] * y[i].bsums[2] + mb[3] * y[i].bsums[3];
+        summs += dmin * smin;
+
+        const __m128i q2bits = _mm_loadu_si128((const __m128i*)q2);
+        const __m256i q2_0 = _mm256_and_si256(_mm256_set_m128i(_mm_srli_epi16(q2bits, 2), q2bits), m3);
+        const __m256i q2_1 = _mm256_and_si256(_mm256_set_m128i(_mm_srli_epi16(q2bits, 6), _mm_srli_epi16(q2bits, 4)), m3);
+
+        const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)(q8+ 0));
+        const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)(q8+32));
+
+        const __m256i p0 = _mm256_maddubs_epi16(q2_0, q8_0);
+        const __m256i p1 = _mm256_maddubs_epi16(q2_1, q8_1);
+
+        const __m256i p_0 = _mm256_cvtepi16_epi32(_mm256_extracti128_si256(p0, 0));
+        const __m256i p_1 = _mm256_cvtepi16_epi32(_mm256_extracti128_si256(p0, 1));
+        const __m256i p_2 = _mm256_cvtepi16_epi32(_mm256_extracti128_si256(p1, 0));
+        const __m256i p_3 = _mm256_cvtepi16_epi32(_mm256_extracti128_si256(p1, 1));
+
+        acc = _mm256_fmadd_ps(_mm256_set1_ps(d * db[0]), _mm256_cvtepi32_ps(p_0), acc);
+        acc = _mm256_fmadd_ps(_mm256_set1_ps(d * db[1]), _mm256_cvtepi32_ps(p_1), acc);
+        acc = _mm256_fmadd_ps(_mm256_set1_ps(d * db[2]), _mm256_cvtepi32_ps(p_2), acc);
+        acc = _mm256_fmadd_ps(_mm256_set1_ps(d * db[3]), _mm256_cvtepi32_ps(p_3), acc);
+    }
+
+    *s = hsum_float_8(acc) + summs;
+
+#else
+
+    float sumf = 0;
+
+    int isum[4];
+
+    for (int i = 0; i < nb; ++i) {
+
+        const uint8_t * q2 = x[i].qs;
+        const  int8_t * q8 = y[i].qs;
+        const uint8_t * sc = x[i].scales;
+
+        int summs = 0;
+        for (int j = 0; j < QK_K/16; ++j) {
+            summs += y[i].bsums[j] * (sc[j] >> 4);
+        }
+
+        const float dall = y[i].d * ggml_fp16_to_fp32(x[i].d);
+        const float dmin = y[i].d * ggml_fp16_to_fp32(x[i].dmin);
+
+        isum[0] = isum[1] = isum[2] = isum[3] = 0;
+        for (int l =  0; l < 16; ++l) {
+            isum[0] += q8[l+ 0] * ((q2[l] >> 0) & 3);
+            isum[1] += q8[l+16] * ((q2[l] >> 2) & 3);
+            isum[2] += q8[l+32] * ((q2[l] >> 4) & 3);
+            isum[3] += q8[l+48] * ((q2[l] >> 6) & 3);
+        }
+        for (int l = 0; l < 4; ++l) {
+            isum[l] *= (sc[l] & 0xF);
+        }
+        sumf += dall * (isum[0] + isum[1] + isum[2] + isum[3]) - dmin * summs;
+    }
+    *s = sumf;
+#endif
+}
+#endif
+
+#if QK_K == 256
 void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
     assert(n % QK_K == 0);
 
@@ -1434,34 +1937,176 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri
 
     *s = hsum_float_8(acc);
 
-#else
-    // scalar version
-    // This function is written like this so the compiler can manage to vectorize most of it
-    // Using -Ofast, GCC and clang manage to produce code that is within a factor of 2 or so from the
-    // manually vectorized version above. Every other version I tried would run at least 4 times slower.
-    // The ideal situation would be if we could just write the code once, and the compiler would
-    // automatically produce the best possible set of machine instructions, instead of us having to manually
-    // write vectorized versions for AVX, ARM_NEON, etc.
+#elif defined __AVX__
 
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
+    const __m128i m3 = _mm_set1_epi8(3);
+    const __m128i mone = _mm_set1_epi8(1);
+    const __m128i m32 = _mm_set1_epi8(32);
+    const __m128i m2 = _mm_set1_epi8(2);
 
-    uint32_t auxs[4];
-    const int8_t * scales = (const int8_t*)auxs;
+    __m256 acc = _mm256_setzero_ps();
+
+    uint32_t *aux;
 
-    float sumf = 0;
     for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * ggml_fp16_to_fp32(x[i].d);
+
         const uint8_t * restrict q3 = x[i].qs;
-        const uint8_t * restrict hm = x[i].hmask;
-        const  int8_t * restrict q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * restrict a = aux8;
-        uint8_t m = 1;
-        for (int j = 0; j < QK_K; j += 128) {
-            for (int l = 0; l < 32; ++l) a[l] = q3[l] & 3;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        // Set up scales
+        aux = (uint32_t *)x[i].scales;
+        __m128i scales128 = _mm_set_epi32(
+                ((aux[1] >> 4) & kmask2) | (((aux[2] >> 6) & kmask1) << 4),
+                ((aux[0] >> 4) & kmask2) | (((aux[2] >> 4) & kmask1) << 4),
+                (aux[1] & kmask2) | (((aux[2] >> 2) & kmask1) << 4),
+                (aux[0] & kmask2) | (((aux[2] >> 0) & kmask1) << 4));
+        scales128 = _mm_sub_epi8(scales128, m32);
+        const __m128i scales_0 = _mm_cvtepi8_epi16(scales128);
+        const __m128i scales_1 = _mm_cvtepi8_epi16(_mm_unpackhi_epi64(scales128, scales128));
+        const __m128i scales[2] = { scales_0, scales_1 };
+
+        // high bit *128*2 from block_q3_K.hmask[QK_K/8]
+        const __m128i hbits_0 = _mm_loadu_si128((const __m128i*)&x[i].hmask[0]);
+        const __m128i hbits_1 = _mm_loadu_si128((const __m128i*)&x[i].hmask[16]);
+
+        // integer accumulator
+        __m128i sumi_0 = _mm_setzero_si128();
+        __m128i sumi_1 = _mm_setzero_si128();
+
+        for (int j = 0; j < QK_K/128; ++j) {
+            // load low 2 bits *64*2 from block_q3_K.qs[QK_K/4]
+            const __m128i q3bits_0 = _mm_loadu_si128((const __m128i*)q3); q3 += 16;
+            const __m128i q3bits_1 = _mm_loadu_si128((const __m128i*)q3); q3 += 16;
+
+            // prepare low and high bits
+            const int bit = j << 2;
+
+            const __m128i q3l_0 = _mm_and_si128(q3bits_0, m3);
+            const __m128i q3l_1 = _mm_and_si128(q3bits_1, m3);
+            const __m128i q3h_0 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_0, _mm_slli_epi16(mone, bit)), bit), 2);
+            const __m128i q3h_1 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_1, _mm_slli_epi16(mone, bit)), bit), 2);
+
+            const __m128i q3l_2 = _mm_and_si128(_mm_srli_epi16(q3bits_0, 2), m3);
+            const __m128i q3l_3 = _mm_and_si128(_mm_srli_epi16(q3bits_1, 2), m3);
+            const __m128i q3h_2 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_0, _mm_slli_epi16(mone, bit+1)), bit+1), 2);
+            const __m128i q3h_3 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_1, _mm_slli_epi16(mone, bit+1)), bit+1), 2);
+
+            const __m128i q3l_4 = _mm_and_si128(_mm_srli_epi16(q3bits_0, 4), m3);
+            const __m128i q3l_5 = _mm_and_si128(_mm_srli_epi16(q3bits_1, 4), m3);
+            const __m128i q3h_4 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_0, _mm_slli_epi16(mone, bit+2)), bit+2), 2);
+            const __m128i q3h_5 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_1, _mm_slli_epi16(mone, bit+2)), bit+2), 2);
+
+            const __m128i q3l_6 = _mm_and_si128(_mm_srli_epi16(q3bits_0, 6), m3);
+            const __m128i q3l_7 = _mm_and_si128(_mm_srli_epi16(q3bits_1, 6), m3);
+            const __m128i q3h_6 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_0, _mm_slli_epi16(mone, bit+3)), bit+3), 2);
+            const __m128i q3h_7 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_1, _mm_slli_epi16(mone, bit+3)), bit+3), 2);
+
+            // load Q8 quants from block_q8_K.qs[QK_K]
+            const __m128i q8_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_2 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_3 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_4 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_5 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_6 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_7 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+
+            // Dot product: we multiply the 2 low bits and 1 high bit part separately, so we can use _mm256_maddubs_epi16,
+            // and then subtract. The high bit part has the 2 already subtracted (and so, it is zero if the high bit was not set,
+            // and 2 if the high bit was set)
+            __m128i q8s_0 = _mm_maddubs_epi16(q3h_0, q8_0);
+            __m128i q8s_1 = _mm_maddubs_epi16(q3h_1, q8_1);
+            __m128i q8s_2 = _mm_maddubs_epi16(q3h_2, q8_2);
+            __m128i q8s_3 = _mm_maddubs_epi16(q3h_3, q8_3);
+            __m128i q8s_4 = _mm_maddubs_epi16(q3h_4, q8_4);
+            __m128i q8s_5 = _mm_maddubs_epi16(q3h_5, q8_5);
+            __m128i q8s_6 = _mm_maddubs_epi16(q3h_6, q8_6);
+            __m128i q8s_7 = _mm_maddubs_epi16(q3h_7, q8_7);
+
+            __m128i p16_0 = _mm_maddubs_epi16(q3l_0, q8_0);
+            __m128i p16_1 = _mm_maddubs_epi16(q3l_1, q8_1);
+            __m128i p16_2 = _mm_maddubs_epi16(q3l_2, q8_2);
+            __m128i p16_3 = _mm_maddubs_epi16(q3l_3, q8_3);
+            __m128i p16_4 = _mm_maddubs_epi16(q3l_4, q8_4);
+            __m128i p16_5 = _mm_maddubs_epi16(q3l_5, q8_5);
+            __m128i p16_6 = _mm_maddubs_epi16(q3l_6, q8_6);
+            __m128i p16_7 = _mm_maddubs_epi16(q3l_7, q8_7);
+
+            p16_0 = _mm_sub_epi16(p16_0, q8s_0);
+            p16_1 = _mm_sub_epi16(p16_1, q8s_1);
+            p16_2 = _mm_sub_epi16(p16_2, q8s_2);
+            p16_3 = _mm_sub_epi16(p16_3, q8s_3);
+            p16_4 = _mm_sub_epi16(p16_4, q8s_4);
+            p16_5 = _mm_sub_epi16(p16_5, q8s_5);
+            p16_6 = _mm_sub_epi16(p16_6, q8s_6);
+            p16_7 = _mm_sub_epi16(p16_7, q8s_7);
+
+            // multiply with scales
+            __m128i shuffle = _mm_set1_epi16(0x0100);
+            p16_0 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_0);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            p16_1 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_1);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            p16_2 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_2);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            p16_3 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_3);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            p16_4 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_4);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            p16_5 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_5);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            p16_6 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_6);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            p16_7 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_7);
+
+            // accumulate
+            p16_0 = _mm_add_epi32(p16_0, p16_1);
+            p16_2 = _mm_add_epi32(p16_2, p16_3);
+            p16_4 = _mm_add_epi32(p16_4, p16_5);
+            p16_6 = _mm_add_epi32(p16_6, p16_7);
+            sumi_0 = _mm_add_epi32(sumi_0, _mm_add_epi32(p16_0, p16_2));
+            sumi_1 = _mm_add_epi32(sumi_1, _mm_add_epi32(p16_4, p16_6));
+
+        }
+
+        // multiply with block scale and accumulate
+        __m256i sumi = _mm256_set_m128i(sumi_1, sumi_0);
+        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(sumi)), acc);
+
+    }
+
+    *s = hsum_float_8(acc);
+
+#else
+    // scalar version
+    // This function is written like this so the compiler can manage to vectorize most of it
+    // Using -Ofast, GCC and clang manage to produce code that is within a factor of 2 or so from the
+    // manually vectorized version above. Every other version I tried would run at least 4 times slower.
+    // The ideal situation would be if we could just write the code once, and the compiler would
+    // automatically produce the best possible set of machine instructions, instead of us having to manually
+    // write vectorized versions for AVX, ARM_NEON, etc.
+
+    int8_t  aux8[QK_K];
+    int16_t aux16[8];
+    float   sums [8];
+    int32_t aux32[8];
+    memset(sums, 0, 8*sizeof(float));
+
+    uint32_t auxs[4];
+    const int8_t * scales = (const int8_t*)auxs;
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        const uint8_t * restrict q3 = x[i].qs;
+        const uint8_t * restrict hm = x[i].hmask;
+        const  int8_t * restrict q8 = y[i].qs;
+        memset(aux32, 0, 8*sizeof(int32_t));
+        int8_t * restrict a = aux8;
+        uint8_t m = 1;
+        for (int j = 0; j < QK_K; j += 128) {
+            for (int l = 0; l < 32; ++l) a[l] = q3[l] & 3;
             for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
             a += 32; m <<= 1;
             for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 2) & 3;
@@ -1501,6 +2146,206 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri
 
 }
 
+#else
+
+void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
+    assert(n % QK_K == 0);
+
+    const block_q3_K * restrict x = vx;
+    const block_q8_K * restrict y = vy;
+
+    const int nb = n / QK_K;
+
+#ifdef __ARM_NEON
+
+#ifdef __ARM_FEATURE_DOTPROD
+    const int32x4_t  vzero = vdupq_n_s32(0);
+#endif
+
+    const uint8x16_t m3b = vdupq_n_u8(0x3);
+    const uint8x16_t mh  = vdupq_n_u8(4);
+
+    int8x16x4_t q3bytes;
+
+    uint16_t aux16[2];
+    int8_t * scales = (int8_t *)aux16;
+
+    float sum = 0;
+
+    for (int i = 0; i < nb; ++i) {
+
+        uint8x16x4_t q3h;
+
+        const uint8x8_t  hbits    = vld1_u8(x[i].hmask);
+        const uint8x16_t q3bits   = vld1q_u8(x[i].qs);
+        const int8x16x4_t q8bytes = vld1q_s8_x4(y[i].qs);
+
+        const uint16_t a = *(const uint16_t *)x[i].scales;
+        aux16[0] = a & 0x0f0f;
+        aux16[1] = (a >> 4) & 0x0f0f;
+
+        for (int j = 0; j < 4; ++j) scales[j] -= 8;
+
+        int32_t isum = -4*(scales[0] * y[i].bsums[0] + scales[2] * y[i].bsums[1] + scales[1] * y[i].bsums[2] + scales[3] * y[i].bsums[3]);
+
+        const float d = y[i].d * (float)x[i].d;
+
+        const uint8x16_t htmp = vcombine_u8(hbits, vshr_n_u8(hbits, 1));
+        q3h.val[0] = vandq_u8(mh, vshlq_n_u8(htmp, 2));
+        q3h.val[1] = vandq_u8(mh, htmp);
+        q3h.val[2] = vandq_u8(mh, vshrq_n_u8(htmp, 2));
+        q3h.val[3] = vandq_u8(mh, vshrq_n_u8(htmp, 4));
+
+        q3bytes.val[0] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q3bits, m3b),                q3h.val[0]));
+        q3bytes.val[1] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(vshrq_n_u8(q3bits, 2), m3b), q3h.val[1]));
+        q3bytes.val[2] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(vshrq_n_u8(q3bits, 4), m3b), q3h.val[2]));
+        q3bytes.val[3] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q3bits, 6),                q3h.val[3]));
+
+#if defined(__ARM_FEATURE_DOTPROD)
+        isum += vaddvq_s32(vdotq_s32(vzero, q3bytes.val[0], q8bytes.val[0])) * scales[0];
+        isum += vaddvq_s32(vdotq_s32(vzero, q3bytes.val[1], q8bytes.val[1])) * scales[2];
+        isum += vaddvq_s32(vdotq_s32(vzero, q3bytes.val[2], q8bytes.val[2])) * scales[1];
+        isum += vaddvq_s32(vdotq_s32(vzero, q3bytes.val[3], q8bytes.val[3])) * scales[3];
+#else
+        const int16x8_t p0 = vaddq_s16(vmull_s8(vget_low_s8 (q3bytes.val[0]), vget_low_s8 (q8bytes.val[0])),
+                                       vmull_s8(vget_high_s8(q3bytes.val[0]), vget_high_s8(q8bytes.val[0])));
+        const int16x8_t p1 = vaddq_s16(vmull_s8(vget_low_s8 (q3bytes.val[1]), vget_low_s8 (q8bytes.val[1])),
+                                       vmull_s8(vget_high_s8(q3bytes.val[1]), vget_high_s8(q8bytes.val[1])));
+        const int16x8_t p2 = vaddq_s16(vmull_s8(vget_low_s8 (q3bytes.val[2]), vget_low_s8 (q8bytes.val[2])),
+                                       vmull_s8(vget_high_s8(q3bytes.val[2]), vget_high_s8(q8bytes.val[2])));
+        const int16x8_t p3 = vaddq_s16(vmull_s8(vget_low_s8 (q3bytes.val[3]), vget_low_s8 (q8bytes.val[3])),
+                                       vmull_s8(vget_high_s8(q3bytes.val[3]), vget_high_s8(q8bytes.val[3])));
+        isum += vaddvq_s16(p0) * scales[0] + vaddvq_s16(p1) * scales[2] + vaddvq_s16(p2) * scales[1] + vaddvq_s16(p3) * scales[3];
+#endif
+
+        sum += d * isum;
+
+    }
+
+    *s = sum;
+
+#elif defined __AVX2__
+
+    const __m256i m3 = _mm256_set1_epi8(3);
+    const __m256i m1 = _mm256_set1_epi8(1);
+
+    __m256 acc = _mm256_setzero_ps();
+
+    uint64_t aux64;
+
+    uint16_t aux16[2];
+    const int8_t * aux8 = (const int8_t *)aux16;
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * ggml_fp16_to_fp32(x[i].d);
+
+        const uint8_t * restrict q3 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        const uint16_t a = *(const uint16_t *)x[i].scales;
+        aux16[0] = a & 0x0f0f;
+        aux16[1] = (a >> 4) & 0x0f0f;
+
+        const __m256i scale_0 = _mm256_set_m128i(_mm_set1_epi16(aux8[2] - 8), _mm_set1_epi16(aux8[0] - 8));
+        const __m256i scale_1 = _mm256_set_m128i(_mm_set1_epi16(aux8[3] - 8), _mm_set1_epi16(aux8[1] - 8));
+
+        memcpy(&aux64, x[i].hmask, 8);
+
+        const __m128i haux = _mm_set_epi64x(aux64 >> 1, aux64 >> 0);
+        __m256i q3h_0 = _mm256_set_m128i(_mm_srli_epi16(haux, 2), haux);
+        __m256i q3h_1 = _mm256_srli_epi16(q3h_0, 4);
+        q3h_0 = _mm256_slli_epi16(_mm256_andnot_si256(q3h_0, m1), 2);
+        q3h_1 = _mm256_slli_epi16(_mm256_andnot_si256(q3h_1, m1), 2);
+
+        // load low 2 bits
+        const __m128i q3bits = _mm_loadu_si128((const __m128i*)q3);
+
+        // prepare low and high bits
+        const __m256i q3aux  = _mm256_set_m128i(_mm_srli_epi16(q3bits, 2), q3bits);
+        const __m256i q3l_0 = _mm256_and_si256(q3aux, m3);
+        const __m256i q3l_1 = _mm256_and_si256(_mm256_srli_epi16(q3aux, 4), m3);
+
+        // load Q8 quants
+        const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)(q8+ 0));
+        const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)(q8+32));
+
+        // Dot product: we multiply the 2 low bits and 1 high bit part separately, so we can use _mm256_maddubs_epi16,
+        // and then subtract. The high bit part has the 2 already subtracted (and so, it is zero if the high bit was not set,
+        // and 2 if the high bit was set)
+        const __m256i q8s_0 = _mm256_maddubs_epi16(q3h_0, q8_0);
+        const __m256i q8s_1 = _mm256_maddubs_epi16(q3h_1, q8_1);
+
+        __m256i p16_0 = _mm256_maddubs_epi16(q3l_0, q8_0);
+        __m256i p16_1 = _mm256_maddubs_epi16(q3l_1, q8_1);
+
+        p16_0 = _mm256_sub_epi16(p16_0, q8s_0);
+        p16_1 = _mm256_sub_epi16(p16_1, q8s_1);
+
+        // multiply with scales
+        p16_0 = _mm256_madd_epi16(scale_0, p16_0);
+        p16_1 = _mm256_madd_epi16(scale_1, p16_1);
+
+        p16_0 = _mm256_add_epi32(p16_0, p16_1);
+
+        // multiply with block scale and accumulate
+        acc = _mm256_fmadd_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(p16_0), acc);
+
+    }
+
+    *s = hsum_float_8(acc);
+
+#else
+
+    int8_t  aux8[QK_K];
+    int16_t aux16[8];
+    float   sums [8];
+    int32_t aux32[8];
+    int32_t scales[4];
+    memset(sums, 0, 8*sizeof(float));
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        const uint8_t * restrict q3 = x[i].qs;
+        const uint8_t * restrict hm = x[i].hmask;
+        const  int8_t * restrict q8 = y[i].qs;
+        int8_t * restrict a = aux8;
+        for (int l = 0; l < 8; ++l) {
+            a[l+ 0] = (int8_t)((q3[l+0] >> 0) & 3) - (hm[l] & 0x01 ? 0 : 4);
+            a[l+ 8] = (int8_t)((q3[l+8] >> 0) & 3) - (hm[l] & 0x02 ? 0 : 4);
+            a[l+16] = (int8_t)((q3[l+0] >> 2) & 3) - (hm[l] & 0x04 ? 0 : 4);
+            a[l+24] = (int8_t)((q3[l+8] >> 2) & 3) - (hm[l] & 0x08 ? 0 : 4);
+            a[l+32] = (int8_t)((q3[l+0] >> 4) & 3) - (hm[l] & 0x10 ? 0 : 4);
+            a[l+40] = (int8_t)((q3[l+8] >> 4) & 3) - (hm[l] & 0x20 ? 0 : 4);
+            a[l+48] = (int8_t)((q3[l+0] >> 6) & 3) - (hm[l] & 0x40 ? 0 : 4);
+            a[l+56] = (int8_t)((q3[l+8] >> 6) & 3) - (hm[l] & 0x80 ? 0 : 4);
+        }
+
+        scales[0] = (x[i].scales[0] & 0xF) - 8;
+        scales[1] = (x[i].scales[0] >>  4) - 8;
+        scales[2] = (x[i].scales[1] & 0xF) - 8;
+        scales[3] = (x[i].scales[1] >>  4) - 8;
+
+        memset(aux32, 0, 8*sizeof(int32_t));
+        for (int j = 0; j < QK_K/16; ++j) {
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] += q8[l] * a[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux32[l] += scales[j] * aux16[l];
+        }
+        const float d = ggml_fp16_to_fp32(x[i].d) * y[i].d;
+        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
+    }
+    for (int l = 0; l < 8; ++l) sumf += sums[l];
+    *s = sumf;
+
+#endif
+
+}
+#endif
+
+#if QK_K == 256
 void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
     assert(n % QK_K == 0);
 
@@ -1614,9 +2459,6 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri
         const float d = y[i].d * ggml_fp16_to_fp32(x[i].d);
         const float dmin = -y[i].d * ggml_fp16_to_fp32(x[i].dmin);
 
-        const uint8_t * restrict q4 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
-
         memcpy(utmp, x[i].scales, 12);
         utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
         const uint32_t uaux = utmp[1] & kmask1;
@@ -1624,6 +2466,9 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri
         utmp[2] = uaux;
         utmp[0] &= kmask1;
 
+        const uint8_t * restrict q4 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+
         const __m256i mins_and_scales = _mm256_cvtepu8_epi16(_mm_set_epi32(utmp[3], utmp[2], utmp[1], utmp[0]));
 
         const __m256i q8sums = _mm256_loadu_si256((const __m256i*)y[i].bsums);
@@ -1667,6 +2512,88 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri
 
     *s = hsum_float_8(acc) + _mm_cvtss_f32(acc_m);
 
+#elif defined __AVX__
+
+    const __m128i m4 = _mm_set1_epi8(0xF);
+    const __m128i m2 = _mm_set1_epi8(0x2);
+
+    __m256 acc = _mm256_setzero_ps();
+    __m128 acc_m = _mm_setzero_ps();
+
+   for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * ggml_fp16_to_fp32(x[i].d);
+        const float dmin = -y[i].d * ggml_fp16_to_fp32(x[i].dmin);
+
+        const uint8_t * restrict q4 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        memcpy(utmp, x[i].scales, 12);
+        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
+        const uint32_t uaux = utmp[1] & kmask1;
+        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
+        utmp[2] = uaux;
+        utmp[0] &= kmask1;
+
+        const __m128i utmps = _mm_set_epi32(utmp[3], utmp[2], utmp[1], utmp[0]);
+        const __m128i scales = _mm_cvtepu8_epi16(utmps);
+        const __m128i mins = _mm_cvtepu8_epi16(_mm_unpackhi_epi64(utmps, utmps));
+
+        const __m128i q8sums_0 = _mm_loadu_si128((const __m128i*)&y[i].bsums[0]);
+        const __m128i q8sums_1 = _mm_loadu_si128((const __m128i*)&y[i].bsums[8]);
+        const __m128i q8s = _mm_hadd_epi16(q8sums_0, q8sums_1);
+        const __m128i prod = _mm_madd_epi16(mins, q8s);
+        acc_m = _mm_add_ps(_mm_mul_ps(_mm_set1_ps(dmin), _mm_cvtepi32_ps(prod)), acc_m);
+
+        __m128i sumi_0 = _mm_setzero_si128();
+        __m128i sumi_1 = _mm_setzero_si128();
+
+        __m128i shuffle = _mm_set1_epi16(0x0100);
+        for (int j = 0; j < QK_K/64; ++j) {
+
+            const __m128i scale_l = _mm_shuffle_epi8(scales, shuffle);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            const __m128i scale_h = _mm_shuffle_epi8(scales, shuffle);
+            shuffle = _mm_add_epi16(shuffle, m2);
+
+            __m128i q4bits = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
+            const __m128i q4l_0 = _mm_and_si128(q4bits, m4);
+            const __m128i q4h_0 = _mm_and_si128(_mm_srli_epi16(q4bits, 4), m4);
+            q4bits = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
+            const __m128i q4l_1 = _mm_and_si128(q4bits, m4);
+            const __m128i q4h_1 = _mm_and_si128(_mm_srli_epi16(q4bits, 4), m4);
+
+            const __m128i q8l_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            __m128i p16l = _mm_maddubs_epi16(q4l_0, q8l_0);
+            p16l = _mm_madd_epi16(scale_l, p16l);
+            sumi_0 = _mm_add_epi32(sumi_0, p16l);
+            const __m128i q8l_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            p16l = _mm_maddubs_epi16(q4l_1, q8l_1);
+            p16l = _mm_madd_epi16(scale_l, p16l);
+            sumi_1 = _mm_add_epi32(sumi_1, p16l);
+
+            const __m128i q8h_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            __m128i p16h = _mm_maddubs_epi16(q4h_0, q8h_0);
+            p16h = _mm_madd_epi16(scale_h, p16h);
+            sumi_0 = _mm_add_epi32(sumi_0, p16h);
+            const __m128i q8h_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            p16h = _mm_maddubs_epi16(q4h_1, q8h_1);
+            p16h = _mm_madd_epi16(scale_h, p16h);
+            sumi_1 = _mm_add_epi32(sumi_1, p16h);
+
+        }
+
+        __m256 vd = _mm256_set1_ps(d);
+        __m256i sumi = _mm256_set_m128i(sumi_1, sumi_0);
+        acc = _mm256_add_ps(_mm256_mul_ps(vd, _mm256_cvtepi32_ps(sumi)), acc);
+
+    }
+
+    acc_m = _mm_add_ps(acc_m, _mm_movehl_ps(acc_m, acc_m));
+    acc_m = _mm_add_ss(acc_m, _mm_movehdup_ps(acc_m));
+
+    *s = hsum_float_8(acc) + _mm_cvtss_f32(acc_m);
+
 #else
 
 
@@ -1726,7 +2653,176 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri
     *s = sumf;
 #endif
 }
+#else
+void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
+    assert(n % QK_K == 0);
+
+    const block_q4_K * restrict x = vx;
+    const block_q8_K * restrict y = vy;
+
+    const int nb = n / QK_K;
+
+#ifdef __ARM_NEON
+
+    const uint8x16_t m4b = vdupq_n_u8(0xf);
+
+#ifdef __ARM_FEATURE_DOTPROD
+    const int32x4_t mzero = vdupq_n_s32(0);
+#endif
+
+    float sumf = 0;
+
+    int8x16x2_t q4bytes;
+    int8x16x4_t q8bytes;
+
+    float sum_mins = 0.f;
+
+    uint16_t aux16[2];
+    const uint8_t * restrict scales = (const uint8_t *)aux16;
+
+    for (int i = 0; i < nb; ++i) {
+
+        const uint8_t * restrict q4 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        const uint16_t * restrict a = (const uint16_t *)x[i].scales;
+        aux16[0] = a[0] & 0x0f0f;
+        aux16[1] = (a[0] >> 4) & 0x0f0f;
+
+        const int32_t summi = scales[2] * (y[i].bsums[0] + y[i].bsums[1]) + scales[3] * (y[i].bsums[2] + y[i].bsums[3]);
+        sum_mins += y[i].d * (float)x[i].d[1] * summi;
+
+        const float d = y[i].d * (float)x[i].d[0];
+
+        const uint8x16x2_t q4bits = vld1q_u8_x2(q4);
+
+#ifdef __ARM_FEATURE_DOTPROD
+        q8bytes = vld1q_s8_x4(q8);
+        q4bytes.val[0] = vreinterpretq_s8_u8(vandq_u8  (q4bits.val[0], m4b));
+        q4bytes.val[1] = vreinterpretq_s8_u8(vandq_u8  (q4bits.val[1], m4b));
+
+        const int32x4_t p1 = vdotq_s32(vdotq_s32(mzero, q4bytes.val[0], q8bytes.val[0]), q4bytes.val[1], q8bytes.val[1]);
+        const int32_t sumi1 = vaddvq_s32(p1) * scales[0];
+
+        q4bytes.val[0] = vreinterpretq_s8_u8(vshrq_n_u8(q4bits.val[0], 4));
+        q4bytes.val[1] = vreinterpretq_s8_u8(vshrq_n_u8(q4bits.val[1], 4));
+
+        const int32x4_t p2 = vdotq_s32(vdotq_s32(mzero, q4bytes.val[0], q8bytes.val[2]), q4bytes.val[1], q8bytes.val[3]);
+        const int32_t sumi2 = vaddvq_s32(p2) * scales[1];
+
+#else
+        q8bytes = vld1q_s8_x4(q8);
+        q4bytes.val[0] = vreinterpretq_s8_u8(vandq_u8  (q4bits.val[0], m4b));
+        q4bytes.val[1] = vreinterpretq_s8_u8(vandq_u8  (q4bits.val[1], m4b));
+        const int16x8_t p0 = vaddq_s16(vmull_s8(vget_low_s8 (q4bytes.val[0]), vget_low_s8 (q8bytes.val[0])),
+                                       vmull_s8(vget_high_s8(q4bytes.val[0]), vget_high_s8(q8bytes.val[0])));
+        const int16x8_t p1 = vaddq_s16(vmull_s8(vget_low_s8 (q4bytes.val[1]), vget_low_s8 (q8bytes.val[1])),
+                                       vmull_s8(vget_high_s8(q4bytes.val[1]), vget_high_s8(q8bytes.val[1])));
+        int32_t sumi1 = vaddvq_s16(vaddq_s16(p0, p1)) * scales[0];
+
+        q4bytes.val[0] = vreinterpretq_s8_u8(vshrq_n_u8(q4bits.val[0], 4));
+        q4bytes.val[1] = vreinterpretq_s8_u8(vshrq_n_u8(q4bits.val[1], 4));
+        const int16x8_t p2 = vaddq_s16(vmull_s8(vget_low_s8 (q4bytes.val[0]), vget_low_s8 (q8bytes.val[2])),
+                                       vmull_s8(vget_high_s8(q4bytes.val[0]), vget_high_s8(q8bytes.val[2])));
+        const int16x8_t p3 = vaddq_s16(vmull_s8(vget_low_s8 (q4bytes.val[1]), vget_low_s8 (q8bytes.val[3])),
+                                       vmull_s8(vget_high_s8(q4bytes.val[1]), vget_high_s8(q8bytes.val[3])));
+        int32_t sumi2 = vaddvq_s16(vaddq_s16(p2, p3)) * scales[1];
+
+#endif
+        sumf += d * (sumi1 + sumi2);
+
+    }
+
+    *s = sumf - sum_mins;
+
+#elif defined __AVX2__
+
+    const __m256i m4 = _mm256_set1_epi8(0xF);
+
+    __m256 acc = _mm256_setzero_ps();
+
+    float summs = 0;
+
+    uint16_t aux16[2];
+    const uint8_t * scales = (const uint8_t *)aux16;
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = ggml_fp16_to_fp32(x[i].d[0]) * y[i].d;
+        const float m = ggml_fp16_to_fp32(x[i].d[1]) * y[i].d;
+        const __m256 vd = _mm256_set1_ps(d);
+
+        const uint16_t * a = (const uint16_t *)x[i].scales;
+        aux16[0] = a[0] & 0x0f0f;
+        aux16[1] = (a[0] >> 4) & 0x0f0f;
+
+        summs += m * (scales[2] * (y[i].bsums[0] + y[i].bsums[1]) + scales[3] * (y[i].bsums[2] + y[i].bsums[3]));
+
+        const uint8_t * restrict q4 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        const __m256i q4bits = _mm256_loadu_si256((const __m256i*)q4);
+        const __m256i q4l = _mm256_and_si256(q4bits, m4);
+        const __m256i q4h = _mm256_and_si256(_mm256_srli_epi16(q4bits, 4), m4);
 
+        const __m256i q8l = _mm256_loadu_si256((const __m256i*)(q8+ 0));
+        const __m256i q8h = _mm256_loadu_si256((const __m256i*)(q8+32));
+
+        const __m256i p16l = _mm256_maddubs_epi16(q4l, q8l);
+        const __m256i p16h = _mm256_maddubs_epi16(q4h, q8h);
+
+        const __m256i p32l = _mm256_madd_epi16(_mm256_set1_epi16(scales[0]), p16l);
+        acc = _mm256_fmadd_ps(vd, _mm256_cvtepi32_ps(p32l), acc);
+
+        const __m256i p32h = _mm256_madd_epi16(_mm256_set1_epi16(scales[1]), p16h);
+        acc = _mm256_fmadd_ps(vd, _mm256_cvtepi32_ps(p32h), acc);
+
+    }
+
+    *s = hsum_float_8(acc) - summs;
+
+#else
+
+    uint8_t aux8[QK_K];
+    int16_t aux16[16];
+    float   sums [8];
+    memset(sums, 0, 8*sizeof(float));
+
+    uint16_t s16[2];
+    const uint8_t * restrict scales = (const uint8_t *)s16;
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        const uint8_t * restrict q4 = x[i].qs;
+        const  int8_t * restrict q8 = y[i].qs;
+        uint8_t * restrict a = aux8;
+        for (int l = 0; l < 32; ++l) a[l+ 0] = q4[l] & 0xF;
+        for (int l = 0; l < 32; ++l) a[l+32] = q4[l]  >> 4;
+
+        const uint16_t * restrict b = (const uint16_t *)x[i].scales;
+        s16[0] = b[0] & 0x0f0f;
+        s16[1] = (b[0] >> 4) & 0x0f0f;
+
+        sumf -= y[i].d * ggml_fp16_to_fp32(x[i].d[1]) * (scales[2] * (y[i].bsums[0] + y[i].bsums[1]) + scales[3] * (y[i].bsums[2] + y[i].bsums[3]));
+
+        const float d = y[i].d * ggml_fp16_to_fp32(x[i].d[0]);
+
+        for (int j = 0; j < QK_K/32; ++j) {
+            for (int l = 0; l < 16; ++l) aux16[l] = q8[l] * a[l];
+            q8 += 16; a += 16;
+            for (int l = 0; l < 16; ++l) aux16[l] += q8[l] * a[l];
+            q8 += 16; a += 16;
+            const float dl = d * scales[j];
+            for (int l = 0; l < 8; ++l) sums[l] += dl * (aux16[l] + aux16[l+8]);
+        }
+    }
+    for (int l = 0; l < 8; ++l) sumf += sums[l];
+    *s = sumf;
+#endif
+}
+#endif
+
+#if QK_K == 256
 void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
     assert(n % QK_K == 0);
 
@@ -1840,18 +2936,23 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri
 
    for (int i = 0; i < nb; ++i) {
 
-        const float d = y[i].d * ggml_fp16_to_fp32(x[i].d);
-        const float dmin = -y[i].d * ggml_fp16_to_fp32(x[i].dmin);
-
         const uint8_t * restrict q5 = x[i].qs;
         const int8_t  * restrict q8 = y[i].qs;
 
+#if QK_K == 256
+        const float d = y[i].d * ggml_fp16_to_fp32(x[i].d);
+        const float dmin = -y[i].d * ggml_fp16_to_fp32(x[i].dmin);
+
         memcpy(utmp, x[i].scales, 12);
         utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
         const uint32_t uaux = utmp[1] & kmask1;
         utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
         utmp[2] = uaux;
         utmp[0] &= kmask1;
+#else
+        // TODO
+        const float d = 0, dmin = 0;
+#endif
 
         const __m256i mins_and_scales = _mm256_cvtepu8_epi16(_mm_set_epi32(utmp[3], utmp[2], utmp[1], utmp[0]));
 
@@ -1876,33 +2977,133 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri
             const __m256i scale_0 = _mm256_shuffle_epi8(scales, get_scale_shuffle_k4(2*j+0));
             const __m256i scale_1 = _mm256_shuffle_epi8(scales, get_scale_shuffle_k4(2*j+1));
 
-            const __m256i q5bits = _mm256_loadu_si256((const __m256i*)q5); q5 += 32;
+            const __m256i q5bits = _mm256_loadu_si256((const __m256i*)q5); q5 += 32;
+
+            const __m256i q5l_0 = _mm256_and_si256(q5bits, m4);
+            const __m256i q5h_0 = _mm256_slli_epi16(_mm256_srli_epi16(_mm256_and_si256(hbits, hmask), bit++), 4);
+            const __m256i q5_0  = _mm256_add_epi8(q5l_0, q5h_0);
+            hmask = _mm256_slli_epi16(hmask, 1);
+
+            const __m256i q5l_1 = _mm256_and_si256(_mm256_srli_epi16(q5bits, 4), m4);
+            const __m256i q5h_1 = _mm256_slli_epi16(_mm256_srli_epi16(_mm256_and_si256(hbits, hmask), bit++), 4);
+            const __m256i q5_1  = _mm256_add_epi8(q5l_1, q5h_1);
+            hmask = _mm256_slli_epi16(hmask, 1);
+
+            const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+            const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+
+            __m256i p16_0 = _mm256_maddubs_epi16(q5_0, q8_0);
+            __m256i p16_1 = _mm256_maddubs_epi16(q5_1, q8_1);
+
+            p16_0 = _mm256_madd_epi16(scale_0, p16_0);
+            p16_1 = _mm256_madd_epi16(scale_1, p16_1);
+
+            sumi = _mm256_add_epi32(sumi, _mm256_add_epi32(p16_0, p16_1));
+
+        }
+
+        __m256 vd = _mm256_set1_ps(d);
+        acc = _mm256_fmadd_ps(vd, _mm256_cvtepi32_ps(sumi), acc);
+
+    }
+
+    *s = hsum_float_8(acc) + summs;
+
+#elif defined __AVX__
+
+    const __m128i m4 = _mm_set1_epi8(0xF);
+    const __m128i mzero = _mm_setzero_si128();
+    const __m128i mone  = _mm_set1_epi8(1);
+    const __m128i m2 = _mm_set1_epi8(2);
+
+    __m256 acc = _mm256_setzero_ps();
+
+    float summs = 0.f;
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * ggml_fp16_to_fp32(x[i].d);
+        const float dmin = -y[i].d * ggml_fp16_to_fp32(x[i].dmin);
+
+        const uint8_t * restrict q5 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        memcpy(utmp, x[i].scales, 12);
+        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
+        const uint32_t uaux = utmp[1] & kmask1;
+        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
+        utmp[2] = uaux;
+        utmp[0] &= kmask1;
+
+        const __m128i utmps = _mm_set_epi32(utmp[3], utmp[2], utmp[1], utmp[0]);
+        const __m128i scales = _mm_cvtepu8_epi16(utmps);
+        const __m128i mins = _mm_cvtepu8_epi16(_mm_unpackhi_epi64(utmps, utmps));
 
-            const __m256i q5l_0 = _mm256_and_si256(q5bits, m4);
-            const __m256i q5h_0 = _mm256_slli_epi16(_mm256_srli_epi16(_mm256_and_si256(hbits, hmask), bit++), 4);
-            const __m256i q5_0  = _mm256_add_epi8(q5l_0, q5h_0);
-            hmask = _mm256_slli_epi16(hmask, 1);
+        const __m128i q8sums_0 = _mm_loadu_si128((const __m128i*)&y[i].bsums[0]);
+        const __m128i q8sums_1 = _mm_loadu_si128((const __m128i*)&y[i].bsums[8]);
+        const __m128i q8s = _mm_hadd_epi16(q8sums_0, q8sums_1);
+        const __m128i prod = _mm_madd_epi16(mins, q8s);
+        const __m128i hsum = _mm_hadd_epi32(_mm_hadd_epi32(prod, mzero), mzero);
+        summs += dmin * _mm_extract_epi32(hsum, 0);
 
-            const __m256i q5l_1 = _mm256_and_si256(_mm256_srli_epi16(q5bits, 4), m4);
-            const __m256i q5h_1 = _mm256_slli_epi16(_mm256_srli_epi16(_mm256_and_si256(hbits, hmask), bit++), 4);
-            const __m256i q5_1  = _mm256_add_epi8(q5l_1, q5h_1);
-            hmask = _mm256_slli_epi16(hmask, 1);
+        const __m128i hbits_0 = _mm_loadu_si128((const __m128i*)&x[i].qh[0]);
+        const __m128i hbits_1 = _mm_loadu_si128((const __m128i*)&x[i].qh[16]);
+        __m128i hmask = mone;
 
-            const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
-            const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+        __m128i sumi_0 = _mm_setzero_si128();
+        __m128i sumi_1 = _mm_setzero_si128();
 
-            __m256i p16_0 = _mm256_maddubs_epi16(q5_0, q8_0);
-            __m256i p16_1 = _mm256_maddubs_epi16(q5_1, q8_1);
+        int bit = 0;
 
-            p16_0 = _mm256_madd_epi16(scale_0, p16_0);
-            p16_1 = _mm256_madd_epi16(scale_1, p16_1);
+        __m128i shuffle = _mm_set1_epi16(0x0100);
+        for (int j = 0; j < QK_K/64; ++j) {
 
-            sumi = _mm256_add_epi32(sumi, _mm256_add_epi32(p16_0, p16_1));
+            const __m128i scale_0 = _mm_shuffle_epi8(scales, shuffle);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            const __m128i scale_1 = _mm_shuffle_epi8(scales, shuffle);
+            shuffle = _mm_add_epi16(shuffle, m2);
+
+            const __m128i q5bits_0 = _mm_loadu_si128((const __m128i*)q5); q5 += 16;
+            const __m128i q5bits_1 = _mm_loadu_si128((const __m128i*)q5); q5 += 16;
+
+            __m128i q5l_0 = _mm_and_si128(q5bits_0, m4);
+            __m128i q5l_1 = _mm_and_si128(q5bits_1, m4);
+            __m128i q5h_0 = _mm_slli_epi16(_mm_srli_epi16(_mm_and_si128(hbits_0, hmask), bit), 4);
+            __m128i q5h_1 = _mm_slli_epi16(_mm_srli_epi16(_mm_and_si128(hbits_1, hmask), bit++), 4);
+            __m128i q5_0  = _mm_add_epi8(q5l_0, q5h_0);
+            __m128i q5_1  = _mm_add_epi8(q5l_1, q5h_1);
+            hmask = _mm_slli_epi16(hmask, 1);
+
+            __m128i q8_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            __m128i q8_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            __m128i p16_0 = _mm_maddubs_epi16(q5_0, q8_0);
+            __m128i p16_1 = _mm_maddubs_epi16(q5_1, q8_1);
+            p16_0 = _mm_madd_epi16(scale_0, p16_0);
+            p16_1 = _mm_madd_epi16(scale_0, p16_1);
+
+            q5l_0 = _mm_and_si128(_mm_srli_epi16(q5bits_0, 4), m4);
+            q5l_1 = _mm_and_si128(_mm_srli_epi16(q5bits_1, 4), m4);
+            q5h_0 = _mm_slli_epi16(_mm_srli_epi16(_mm_and_si128(hbits_0, hmask), bit), 4);
+            q5h_1 = _mm_slli_epi16(_mm_srli_epi16(_mm_and_si128(hbits_1, hmask), bit++), 4);
+            q5_0  = _mm_add_epi8(q5l_0, q5h_0);
+            q5_1  = _mm_add_epi8(q5l_1, q5h_1);
+            hmask = _mm_slli_epi16(hmask, 1);
+
+            q8_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            q8_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            __m128i p16_2 = _mm_maddubs_epi16(q5_0, q8_0);
+            __m128i p16_3 = _mm_maddubs_epi16(q5_1, q8_1);
+            p16_2 = _mm_madd_epi16(scale_1, p16_2);
+            p16_3 = _mm_madd_epi16(scale_1, p16_3);
+
+            sumi_0 = _mm_add_epi32(sumi_0, _mm_add_epi32(p16_0, p16_2));
+            sumi_1 = _mm_add_epi32(sumi_1, _mm_add_epi32(p16_1, p16_3));
 
         }
 
         __m256 vd = _mm256_set1_ps(d);
-        acc = _mm256_fmadd_ps(vd, _mm256_cvtepi32_ps(sumi), acc);
+        __m256i sumi = _mm256_set_m128i(sumi_1, sumi_0);
+        acc = _mm256_add_ps(_mm256_mul_ps(vd, _mm256_cvtepi32_ps(sumi)), acc);
 
     }
 
@@ -1972,8 +3173,169 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri
 #endif
 }
 
+#else
+
+void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
+    assert(n % QK_K == 0);
+
+    const block_q5_K * restrict x = vx;
+    const block_q8_K * restrict y = vy;
+
+    const int nb = n / QK_K;
+
+#ifdef __ARM_NEON
+
+    const uint8x16_t m4b = vdupq_n_u8(0xf);
+    const int32x4_t mzero = vdupq_n_s32(0);
+    const uint8x16_t mh = vdupq_n_u8(16);
+
+    int8x16x4_t q5bytes;
+    uint8x16x4_t q5h;
+
+    float sumf = 0;
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * (float)x[i].d;
+        const int8_t * sc = x[i].scales;
+
+        const uint8_t * restrict q5 = x[i].qs;
+        const uint8_t * restrict qh = x[i].qh;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        const uint8x8_t qhbits = vld1_u8(qh);
+
+        const uint8x16x2_t q5bits = vld1q_u8_x2(q5);
+        const int8x16x4_t q8bytes = vld1q_s8_x4(q8);
+
+        const uint8x16_t htmp = vcombine_u8(qhbits, vshr_n_u8(qhbits, 1));
+        q5h.val[0] = vbicq_u8(mh, vshlq_n_u8(htmp, 4));
+        q5h.val[1] = vbicq_u8(mh, vshlq_n_u8(htmp, 2));
+        q5h.val[2] = vbicq_u8(mh, htmp);
+        q5h.val[3] = vbicq_u8(mh, vshrq_n_u8(htmp, 2));
+
+        q5bytes.val[0] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(q5bits.val[0], m4b)), vreinterpretq_s8_u8(q5h.val[0]));
+        q5bytes.val[1] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(q5bits.val[1], m4b)), vreinterpretq_s8_u8(q5h.val[1]));
+        q5bytes.val[2] = vsubq_s8(vreinterpretq_s8_u8(vshrq_n_u8(q5bits.val[0], 4)), vreinterpretq_s8_u8(q5h.val[2]));
+        q5bytes.val[3] = vsubq_s8(vreinterpretq_s8_u8(vshrq_n_u8(q5bits.val[1], 4)), vreinterpretq_s8_u8(q5h.val[3]));
+
+#if defined(__ARM_FEATURE_DOTPROD)
+
+        int32_t sumi1 = sc[0] * vaddvq_s32(vdotq_s32(mzero, q5bytes.val[0], q8bytes.val[0]));
+        int32_t sumi2 = sc[1] * vaddvq_s32(vdotq_s32(mzero, q5bytes.val[1], q8bytes.val[1]));
+        int32_t sumi3 = sc[2] * vaddvq_s32(vdotq_s32(mzero, q5bytes.val[2], q8bytes.val[2]));
+        int32_t sumi4 = sc[3] * vaddvq_s32(vdotq_s32(mzero, q5bytes.val[3], q8bytes.val[3]));
+
+        sumf += d * (sumi1 + sumi2 + sumi3 + sumi4);
+
+#else
+
+        const int16x8_t p0 = vaddq_s16(vmull_s8(vget_low_s8 (q5bytes.val[0]), vget_low_s8 (q8bytes.val[0])),
+                                       vmull_s8(vget_high_s8(q5bytes.val[0]), vget_high_s8(q8bytes.val[0])));
+        const int16x8_t p1 = vaddq_s16(vmull_s8(vget_low_s8 (q5bytes.val[1]), vget_low_s8 (q8bytes.val[1])),
+                                       vmull_s8(vget_high_s8(q5bytes.val[1]), vget_high_s8(q8bytes.val[1])));
+        int32_t sumi = sc[0] * vaddvq_s16(p0) + sc[1] * vaddvq_s16(p1);
+
+        const int16x8_t p2 = vaddq_s16(vmull_s8(vget_low_s8 (q5bytes.val[2]), vget_low_s8 (q8bytes.val[2])),
+                                       vmull_s8(vget_high_s8(q5bytes.val[2]), vget_high_s8(q8bytes.val[2])));
+        const int16x8_t p3 = vaddq_s16(vmull_s8(vget_low_s8 (q5bytes.val[3]), vget_low_s8 (q8bytes.val[3])),
+                                       vmull_s8(vget_high_s8(q5bytes.val[3]), vget_high_s8(q8bytes.val[3])));
+        sumi += sc[2] * vaddvq_s16(p2) + sc[3] * vaddvq_s16(p3);
+
+        sumf += d*sumi;
+#endif
+
+    }
+
+    *s = sumf;
+
+#elif defined __AVX2__
+
+    const __m256i m4 = _mm256_set1_epi8(0xF);
+    const __m256i mone  = _mm256_set1_epi8(1);
+
+    __m256 acc = _mm256_setzero_ps();
+
+    for (int i = 0; i < nb; ++i) {
+
+        const uint8_t * restrict q5 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        const float d = y[i].d * ggml_fp16_to_fp32(x[i].d);
+
+        const __m256i q5bits = _mm256_loadu_si256((const __m256i*)q5);
+
+        const __m256i scale_l = _mm256_set_m128i(_mm_set1_epi16(x[i].scales[1]), _mm_set1_epi16(x[i].scales[0]));
+        const __m256i scale_h = _mm256_set_m128i(_mm_set1_epi16(x[i].scales[3]), _mm_set1_epi16(x[i].scales[2]));
+
+        int64_t aux64;
+        memcpy(&aux64, x[i].qh, 8);
+        const __m128i haux128 = _mm_set_epi64x(aux64 >> 1, aux64);
+        const __m256i haux256 = _mm256_set_m128i(_mm_srli_epi16(haux128, 2), haux128);
+
+        const __m256i q5h_0 = _mm256_slli_epi16(_mm256_andnot_si256(haux256, mone), 4);
+        const __m256i q5h_1 = _mm256_slli_epi16(_mm256_andnot_si256(_mm256_srli_epi16(haux256, 4), mone), 4);
+
+        const __m256i q5l_0 = _mm256_and_si256(q5bits, m4);
+        const __m256i q5l_1 = _mm256_and_si256(_mm256_srli_epi16(q5bits, 4), m4);
+
+        const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)(q8+ 0));
+        const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)(q8+32));
+
+        const __m256i p16_0 = _mm256_madd_epi16(scale_l, _mm256_maddubs_epi16(q5l_0, q8_0));
+        const __m256i p16_1 = _mm256_madd_epi16(scale_h, _mm256_maddubs_epi16(q5l_1, q8_1));
+        const __m256i s16_0 = _mm256_madd_epi16(scale_l, _mm256_maddubs_epi16(q5h_0, q8_0));
+        const __m256i s16_1 = _mm256_madd_epi16(scale_h, _mm256_maddubs_epi16(q5h_1, q8_1));
+
+        const __m256i dot = _mm256_sub_epi32(_mm256_add_epi32(p16_0, p16_1), _mm256_add_epi32(s16_0, s16_1));
+
+        acc = _mm256_fmadd_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(dot), acc);
+
+    }
+
+    *s = hsum_float_8(acc);
+
+#else
+
+
+    uint8_t aux8[QK_K];
+    int16_t aux16[16];
+    float   sums [8];
+    memset(sums, 0, 8*sizeof(float));
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        const uint8_t * restrict q4 = x[i].qs;
+        const uint8_t * restrict hm = x[i].qh;
+        const  int8_t * restrict q8 = y[i].qs;
+        uint8_t * restrict a = aux8;
+        for (int l = 0; l < 32; ++l) {
+            a[l+ 0] = q4[l] & 0xF;
+            a[l+32] = q4[l]  >> 4;
+        }
+        for (int is = 0; is < 8; ++is) {
+            uint8_t m = 1 << is;
+            for (int l = 0; l < 8; ++l) a[8*is + l] -= (hm[l] & m ? 0 : 16);
+        }
+
+        const float d = y[i].d * ggml_fp16_to_fp32(x[i].d);
+        const int8_t * restrict sc = x[i].scales;
+
+        for (int j = 0; j < QK_K/16; ++j) {
+            const float dl = d * sc[j];
+            for (int l = 0; l < 16; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l <  8; ++l) sums[l] += dl * (aux16[l] + aux16[8+l]);
+            q8 += 16; a += 16;
+        }
+    }
+    for (int l = 0; l < 8; ++l) sumf += sums[l];
+    *s = sumf;
+#endif
+}
+#endif
 
 
+#if QK_K == 256
 void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
     assert(n % QK_K == 0);
 
@@ -2198,6 +3560,124 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri
 
     *s = hsum_float_8(acc);
 
+#elif defined __AVX__
+
+    const __m128i m4 = _mm_set1_epi8(0xF);
+    const __m128i m3 = _mm_set1_epi8(3);
+    const __m128i m32s = _mm_set1_epi8(32);
+    const __m128i m2 = _mm_set1_epi8(2);
+
+    __m256 acc = _mm256_setzero_ps();
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * ggml_fp16_to_fp32(x[i].d);
+
+        const uint8_t * restrict q4 = x[i].ql;
+        const uint8_t * restrict qh = x[i].qh;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        const __m128i scales = _mm_loadu_si128((const __m128i*)x[i].scales);
+
+        __m128i sumi_0 = _mm_setzero_si128();
+        __m128i sumi_1 = _mm_setzero_si128();
+
+        __m128i shuffle = _mm_set_epi64x(0x0101010101010101, 0x0000000000000000);
+        for (int j = 0; j < QK_K/128; ++j) {
+
+            const __m128i q4bitsH_0 = _mm_loadu_si128((const __m128i*)qh); qh += 16;
+            const __m128i q4bitsH_1 = _mm_loadu_si128((const __m128i*)qh); qh += 16;
+
+            const __m128i q4h_0 = _mm_slli_epi16(_mm_and_si128(q4bitsH_0, m3), 4);
+            const __m128i q4h_1 = _mm_slli_epi16(_mm_and_si128(q4bitsH_1, m3), 4);
+            const __m128i q4h_2 = _mm_slli_epi16(_mm_and_si128(_mm_srli_epi16(q4bitsH_0, 2), m3), 4);
+            const __m128i q4h_3 = _mm_slli_epi16(_mm_and_si128(_mm_srli_epi16(q4bitsH_1, 2), m3), 4);
+            const __m128i q4h_4 = _mm_slli_epi16(_mm_and_si128(_mm_srli_epi16(q4bitsH_0, 4), m3), 4);
+            const __m128i q4h_5 = _mm_slli_epi16(_mm_and_si128(_mm_srli_epi16(q4bitsH_1, 4), m3), 4);
+            const __m128i q4h_6 = _mm_slli_epi16(_mm_and_si128(_mm_srli_epi16(q4bitsH_0, 6), m3), 4);
+            const __m128i q4h_7 = _mm_slli_epi16(_mm_and_si128(_mm_srli_epi16(q4bitsH_1, 6), m3), 4);
+
+            const __m128i q4bits1_0 = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
+            const __m128i q4bits1_1 = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
+            const __m128i q4bits2_0 = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
+            const __m128i q4bits2_1 = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
+
+            const __m128i q4_0 = _mm_or_si128(_mm_and_si128(q4bits1_0, m4), q4h_0);
+            const __m128i q4_1 = _mm_or_si128(_mm_and_si128(q4bits1_1, m4), q4h_1);
+            const __m128i q4_2 = _mm_or_si128(_mm_and_si128(q4bits2_0, m4), q4h_2);
+            const __m128i q4_3 = _mm_or_si128(_mm_and_si128(q4bits2_1, m4), q4h_3);
+            const __m128i q4_4 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(q4bits1_0, 4), m4), q4h_4);
+            const __m128i q4_5 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(q4bits1_1, 4), m4), q4h_5);
+            const __m128i q4_6 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(q4bits2_0, 4), m4), q4h_6);
+            const __m128i q4_7 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(q4bits2_1, 4), m4), q4h_7);
+
+            const __m128i q8_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_2 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_3 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_4 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_5 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_6 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_7 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+
+            __m128i q8s_0 = _mm_maddubs_epi16(m32s, q8_0);
+            __m128i q8s_1 = _mm_maddubs_epi16(m32s, q8_1);
+            __m128i q8s_2 = _mm_maddubs_epi16(m32s, q8_2);
+            __m128i q8s_3 = _mm_maddubs_epi16(m32s, q8_3);
+            __m128i q8s_4 = _mm_maddubs_epi16(m32s, q8_4);
+            __m128i q8s_5 = _mm_maddubs_epi16(m32s, q8_5);
+            __m128i q8s_6 = _mm_maddubs_epi16(m32s, q8_6);
+            __m128i q8s_7 = _mm_maddubs_epi16(m32s, q8_7);
+
+            __m128i p16_0 = _mm_maddubs_epi16(q4_0, q8_0);
+            __m128i p16_1 = _mm_maddubs_epi16(q4_1, q8_1);
+            __m128i p16_2 = _mm_maddubs_epi16(q4_2, q8_2);
+            __m128i p16_3 = _mm_maddubs_epi16(q4_3, q8_3);
+            __m128i p16_4 = _mm_maddubs_epi16(q4_4, q8_4);
+            __m128i p16_5 = _mm_maddubs_epi16(q4_5, q8_5);
+            __m128i p16_6 = _mm_maddubs_epi16(q4_6, q8_6);
+            __m128i p16_7 = _mm_maddubs_epi16(q4_7, q8_7);
+
+            p16_0 = _mm_sub_epi16(p16_0, q8s_0);
+            p16_1 = _mm_sub_epi16(p16_1, q8s_1);
+            p16_2 = _mm_sub_epi16(p16_2, q8s_2);
+            p16_3 = _mm_sub_epi16(p16_3, q8s_3);
+            p16_4 = _mm_sub_epi16(p16_4, q8s_4);
+            p16_5 = _mm_sub_epi16(p16_5, q8s_5);
+            p16_6 = _mm_sub_epi16(p16_6, q8s_6);
+            p16_7 = _mm_sub_epi16(p16_7, q8s_7);
+
+            const __m128i scale_0 = _mm_shuffle_epi8(scales, shuffle);
+            shuffle = _mm_add_epi8(shuffle, m2);
+            const __m128i scale_1 = _mm_shuffle_epi8(scales, shuffle);
+            shuffle = _mm_add_epi8(shuffle, m2);
+            const __m128i scale_2 = _mm_shuffle_epi8(scales, shuffle);
+            shuffle = _mm_add_epi8(shuffle, m2);
+            const __m128i scale_3 = _mm_shuffle_epi8(scales, shuffle);
+            shuffle = _mm_add_epi8(shuffle, m2);
+
+            p16_0 = _mm_madd_epi16(_mm_cvtepi8_epi16(scale_0), p16_0);
+            p16_1 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_unpackhi_epi64(scale_0, scale_0)), p16_1);
+            p16_2 = _mm_madd_epi16(_mm_cvtepi8_epi16(scale_1), p16_2);
+            p16_3 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_unpackhi_epi64(scale_1, scale_1)), p16_3);
+            p16_4 = _mm_madd_epi16(_mm_cvtepi8_epi16(scale_2), p16_4);
+            p16_5 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_unpackhi_epi64(scale_2, scale_2)), p16_5);
+            p16_6 = _mm_madd_epi16(_mm_cvtepi8_epi16(scale_3), p16_6);
+            p16_7 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_unpackhi_epi64(scale_3, scale_3)), p16_7);
+
+            sumi_0 = _mm_add_epi32(sumi_0, _mm_add_epi32(p16_0, p16_2));
+            sumi_1 = _mm_add_epi32(sumi_1, _mm_add_epi32(p16_1, p16_3));
+            sumi_0 = _mm_add_epi32(sumi_0, _mm_add_epi32(p16_4, p16_6));
+            sumi_1 = _mm_add_epi32(sumi_1, _mm_add_epi32(p16_5, p16_7));
+
+        }
+
+        __m256i sumi = _mm256_set_m128i(sumi_1, sumi_0);
+        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(sumi)), acc);
+    }
+
+    *s = hsum_float_8(acc);
+
 #else
 
     int8_t  aux8[QK_K];
@@ -2242,3 +3722,179 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri
     *s = sumf;
 #endif
 }
+
+#else
+
+void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
+    assert(n % QK_K == 0);
+
+    const block_q6_K * restrict x = vx;
+    const block_q8_K * restrict y = vy;
+
+    const int nb = n / QK_K;
+
+#ifdef __ARM_NEON
+
+    float sum = 0;
+
+    const uint8x16_t m4b = vdupq_n_u8(0xF);
+    const int32x4_t  vzero = vdupq_n_s32(0);
+    const int8x16_t  m32s = vdupq_n_s8(32);
+
+    const uint8x16_t mone = vdupq_n_u8(3);
+
+    int8x16x4_t q6bytes;
+    uint8x16x4_t q6h;
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d_all = (float)x[i].d;
+
+        const uint8_t * restrict q6 = x[i].ql;
+        const uint8_t * restrict qh = x[i].qh;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        const int8_t * restrict scale = x[i].scales;
+
+        int32_t isum = 0;
+
+        uint8x16_t   qhbits = vld1q_u8(qh);
+        uint8x16x2_t q6bits = vld1q_u8_x2(q6);
+        int8x16x4_t q8bytes = vld1q_s8_x4(q8);
+
+        q6h.val[0] = vshlq_n_u8(vandq_u8(mone, qhbits), 4);
+        uint8x16_t shifted = vshrq_n_u8(qhbits, 2);
+        q6h.val[1] = vshlq_n_u8(vandq_u8(mone, shifted), 4);
+        shifted = vshrq_n_u8(qhbits, 4);
+        q6h.val[2] = vshlq_n_u8(vandq_u8(mone, shifted), 4);
+        shifted = vshrq_n_u8(qhbits, 6);
+        q6h.val[3] = vshlq_n_u8(vandq_u8(mone, shifted), 4);
+
+        q6bytes.val[0] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[0], m4b), q6h.val[0])), m32s);
+        q6bytes.val[1] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[1], m4b), q6h.val[1])), m32s);
+        q6bytes.val[2] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[0], 4), q6h.val[2])), m32s);
+        q6bytes.val[3] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[1], 4), q6h.val[3])), m32s);
+
+#if defined(__ARM_FEATURE_DOTPROD)
+
+        isum += vaddvq_s32(vdotq_s32(vzero, q6bytes.val[0], q8bytes.val[0])) * scale[0] +
+                vaddvq_s32(vdotq_s32(vzero, q6bytes.val[1], q8bytes.val[1])) * scale[1] +
+                vaddvq_s32(vdotq_s32(vzero, q6bytes.val[2], q8bytes.val[2])) * scale[2] +
+                vaddvq_s32(vdotq_s32(vzero, q6bytes.val[3], q8bytes.val[3])) * scale[3];
+#else
+
+        int16x8_t p0 = vaddq_s16(vmull_s8(vget_low_s8 (q6bytes.val[0]), vget_low_s8 (q8bytes.val[0])),
+                                 vmull_s8(vget_high_s8(q6bytes.val[0]), vget_high_s8(q8bytes.val[0])));
+        int16x8_t p1 = vaddq_s16(vmull_s8(vget_low_s8 (q6bytes.val[1]), vget_low_s8 (q8bytes.val[1])),
+                                 vmull_s8(vget_high_s8(q6bytes.val[1]), vget_high_s8(q8bytes.val[1])));
+        isum += vaddvq_s16(p0) * scale[0] + vaddvq_s16(p1) * scale[1];
+
+        int16x8_t p2 = vaddq_s16(vmull_s8(vget_low_s8 (q6bytes.val[2]), vget_low_s8 (q8bytes.val[2])),
+                                 vmull_s8(vget_high_s8(q6bytes.val[2]), vget_high_s8(q8bytes.val[2])));
+        int16x8_t p3 = vaddq_s16(vmull_s8(vget_low_s8 (q6bytes.val[3]), vget_low_s8 (q8bytes.val[3])),
+                                 vmull_s8(vget_high_s8(q6bytes.val[3]), vget_high_s8(q8bytes.val[3])));
+        isum += vaddvq_s16(p2) * scale[2] + vaddvq_s16(p3) * scale[3];
+#endif
+
+        sum += isum * d_all * y[i].d;
+
+    }
+    *s = sum;
+
+#elif defined __AVX2__
+
+    const __m256i m4 = _mm256_set1_epi8(0xF);
+    const __m256i m2 = _mm256_set1_epi8(3);
+    const __m256i m32s = _mm256_set1_epi8(32);
+
+    __m256 acc = _mm256_setzero_ps();
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * ggml_fp16_to_fp32(x[i].d);
+
+        const uint8_t * restrict q4 = x[i].ql;
+        const uint8_t * restrict qh = x[i].qh;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        const __m64 scales_1 = _mm_set1_pi8(x[i].scales[0]);
+        const __m64 scales_2 = _mm_set1_pi8(x[i].scales[1]);
+        const __m64 scales_3 = _mm_set1_pi8(x[i].scales[2]);
+        const __m64 scales_4 = _mm_set1_pi8(x[i].scales[3]);
+
+        __m256i sumi = _mm256_setzero_si256();
+
+        const __m128i scale_0 = _mm_set_epi64(scales_2, scales_1);
+        const __m128i scale_1 = _mm_set_epi64(scales_4, scales_3);
+
+        const __m256i q4bits1 = _mm256_loadu_si256((const __m256i*)q4);
+        const __m128i q4bitsH = _mm_loadu_si128((const __m128i*)qh);
+
+        const __m256i q4h_0 = _mm256_slli_epi16(_mm256_and_si256(_mm256_set_m128i(_mm_srli_epi16(q4bitsH, 2), q4bitsH), m2), 4);
+        const __m256i q4h_1 = _mm256_slli_epi16(_mm256_and_si256(_mm256_set_m128i(_mm_srli_epi16(q4bitsH, 6), _mm_srli_epi16(q4bitsH, 4)), m2), 4);
+
+        const __m256i q4_0 = _mm256_or_si256(_mm256_and_si256(q4bits1, m4), q4h_0);
+        const __m256i q4_1 = _mm256_or_si256(_mm256_and_si256(_mm256_srli_epi16(q4bits1, 4), m4), q4h_1);
+
+        const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)(q8+ 0));
+        const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)(q8+32));
+
+        __m256i q8s_0 = _mm256_maddubs_epi16(m32s, q8_0);
+        __m256i q8s_1 = _mm256_maddubs_epi16(m32s, q8_1);
+
+        __m256i p16_0 = _mm256_maddubs_epi16(q4_0, q8_0);
+        __m256i p16_1 = _mm256_maddubs_epi16(q4_1, q8_1);
+
+        p16_0 = _mm256_sub_epi16(p16_0, q8s_0);
+        p16_1 = _mm256_sub_epi16(p16_1, q8s_1);
+
+        p16_0 = _mm256_madd_epi16(_mm256_cvtepi8_epi16(scale_0), p16_0);
+        p16_1 = _mm256_madd_epi16(_mm256_cvtepi8_epi16(scale_1), p16_1);
+
+        sumi = _mm256_add_epi32(sumi, _mm256_add_epi32(p16_0, p16_1));
+
+        acc = _mm256_fmadd_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(sumi), acc);
+    }
+
+    *s = hsum_float_8(acc);
+
+#else
+
+    int8_t  aux8[QK_K];
+    int16_t aux16[8];
+    float   sums [8];
+    int32_t aux32[8];
+    memset(sums, 0, 8*sizeof(float));
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        const uint8_t * restrict q4 = x[i].ql;
+        const uint8_t * restrict qh = x[i].qh;
+        const  int8_t * restrict q8 = y[i].qs;
+        memset(aux32, 0, 8*sizeof(int32_t));
+        int8_t * restrict a = aux8;
+        for (int l = 0; l < 16; ++l) {
+            a[l+ 0] = (int8_t)((q4[l+ 0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32;
+            a[l+16] = (int8_t)((q4[l+16] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32;
+            a[l+32] = (int8_t)((q4[l+ 0] >>  4) | (((qh[l] >> 4) & 3) << 4)) - 32;
+            a[l+48] = (int8_t)((q4[l+16] >>  4) | (((qh[l] >> 6) & 3) << 4)) - 32;
+        }
+        int is = 0;
+        for (int j = 0; j < QK_K/16; ++j) {
+            int scale = x[i].scales[is++];
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+        }
+        const float d = ggml_fp16_to_fp32(x[i].d) * y[i].d;
+        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
+    }
+    for (int l = 0; l < 8; ++l) sumf += sums[l];
+    *s = sumf;
+#endif
+}
+
+#endif