@fugood/llama.node 0.2.2 → 0.3.0

package/src/llama.cpp/{ggml-quants.c → ggml/src/ggml-quants.c}

@@ -4,8 +4,6 @@
 #include "ggml-quants.h"
 #include "ggml-impl.h"
 
-#define GGML_COMMON_IMPL_C
-#include "ggml-common.h"
 
 #include <math.h>
 #include <string.h>
@@ -660,7 +658,7 @@ static inline __m128i packNibbles( __m256i bytes ) {
 #endif  //__loongarch_asx
 
 // reference implementation for deterministic creation of model files
-void quantize_row_q4_0_reference(const float * restrict x, block_q4_0 * restrict y, int64_t k) {
+void quantize_row_q4_0_ref(const float * restrict x, block_q4_0 * restrict y, int64_t k) {
     static const int qk = QK4_0;
 
     assert(k % qk == 0);
@@ -698,11 +696,11 @@ void quantize_row_q4_0_reference(const float * restrict x, block_q4_0 * restrict
 }
 
 void quantize_row_q4_0(const float * restrict x, void * restrict y, int64_t k) {
-    quantize_row_q4_0_reference(x, y, k);
+    quantize_row_q4_0_ref(x, y, k);
 }
 
 
-void quantize_row_q4_1_reference(const float * restrict x, block_q4_1 * restrict y, int64_t k) {
+void quantize_row_q4_1_ref(const float * restrict x, block_q4_1 * restrict y, int64_t k) {
     const int qk = QK4_1;
 
     assert(k % qk == 0);
@@ -740,10 +738,10 @@ void quantize_row_q4_1_reference(const float * restrict x, block_q4_1 * restrict
 }
 
 void quantize_row_q4_1(const float * restrict x, void * restrict y, int64_t k) {
-    quantize_row_q4_1_reference(x, y, k);
+    quantize_row_q4_1_ref(x, y, k);
 }
 
-void quantize_row_q5_0_reference(const float * restrict x, block_q5_0 * restrict y, int64_t k) {
+void quantize_row_q5_0_ref(const float * restrict x, block_q5_0 * restrict y, int64_t k) {
     static const int qk = QK5_0;
 
     assert(k % qk == 0);
@@ -788,10 +786,10 @@ void quantize_row_q5_0_reference(const float * restrict x, block_q5_0 * restrict
 }
 
 void quantize_row_q5_0(const float * restrict x, void * restrict y, int64_t k) {
-    quantize_row_q5_0_reference(x, y, k);
+    quantize_row_q5_0_ref(x, y, k);
 }
 
-void quantize_row_q5_1_reference(const float * restrict x, block_q5_1 * restrict y, int64_t k) {
+void quantize_row_q5_1_ref(const float * restrict x, block_q5_1 * restrict y, int64_t k) {
     const int qk = QK5_1;
 
     assert(k % qk == 0);
@@ -836,11 +834,11 @@ void quantize_row_q5_1_reference(const float * restrict x, block_q5_1 * restrict
 }
 
 void quantize_row_q5_1(const float * restrict x, void * restrict y, int64_t k) {
-    quantize_row_q5_1_reference(x, y, k);
+    quantize_row_q5_1_ref(x, y, k);
 }
 
 // reference implementation for deterministic creation of model files
-void quantize_row_q8_0_reference(const float * restrict x, block_q8_0 * restrict y, int64_t k) {
+void quantize_row_q8_0_ref(const float * restrict x, block_q8_0 * restrict y, int64_t k) {
     assert(k % QK8_0 == 0);
     const int nb = k / QK8_0;
 
@@ -1078,6 +1076,7 @@ void quantize_row_q8_0(const float * restrict x, void * restrict vy, int64_t k)
         }
         vec_xst(vec_pack(vec_pack(vi[0], vi[1]), vec_pack(vi[2], vi[3])),  0, &y[i].qs[0]);
         vec_xst(vec_pack(vec_pack(vi[4], vi[5]), vec_pack(vi[6], vi[7])), 16, &y[i].qs[0]);
+    }
 
 #elif defined(__loongarch_asx)
     for (int i = 0; i < nb; i++) {
@@ -1145,12 +1144,12 @@ void quantize_row_q8_0(const float * restrict x, void * restrict vy, int64_t k)
 #else
     GGML_UNUSED(nb);
     // scalar
-    quantize_row_q8_0_reference(x, y, k);
+    quantize_row_q8_0_ref(x, y, k);
 #endif
 }
 
 // reference implementation for deterministic creation of model files
-void quantize_row_q8_1_reference(const float * restrict x, block_q8_1 * restrict y, int64_t k) {
+void quantize_row_q8_1_ref(const float * restrict x, block_q8_1 * restrict y, int64_t k) {
     assert(QK8_1 == 32);
     assert(k % QK8_1 == 0);
     const int nb = k / QK8_1;
@@ -1437,6 +1436,7 @@ void quantize_row_q8_1(const float * restrict x, void * restrict vy, int64_t k)
         accv = vec_add(accv, vec_sld(accv, accv, 4));
         accv = vec_add(accv, vec_sld(accv, accv, 8));
         y[i].s = GGML_FP32_TO_FP16(d * vec_extract(accv, 0));
+    }
 
 #elif defined(__loongarch_asx)
     for (int i = 0; i < nb; i++) {
@@ -1508,7 +1508,7 @@ void quantize_row_q8_1(const float * restrict x, void * restrict vy, int64_t k)
 #else
     GGML_UNUSED(nb);
     // scalar
-    quantize_row_q8_1_reference(x, y, k);
+    quantize_row_q8_1_ref(x, y, k);
 #endif
 }
 
@@ -1899,7 +1899,7 @@ static inline void get_scale_min_k4(int j, const uint8_t * restrict q, uint8_t *
 
 //========================- 2-bit (de)-quantization
 
-void quantize_row_q2_K_reference(const float * restrict x, block_q2_K * restrict y, int64_t k) {
+void quantize_row_q2_K_ref(const float * restrict x, block_q2_K * restrict y, int64_t k) {
     assert(k % QK_K == 0);
     const int nb = k / QK_K;
 
@@ -2002,7 +2002,7 @@ void dequantize_row_q2_K(const block_q2_K * restrict x, float * restrict y, int6
 }
 
 void quantize_row_q2_K(const float * restrict x, void * restrict vy, int64_t k) {
-    quantize_row_q2_K_reference(x, vy, k);
+    quantize_row_q2_K_ref(x, vy, k);
 }
 
 static float make_qkx3_quants(int n, int nmax, const float * restrict x, const float * restrict weights,
@@ -2226,7 +2226,7 @@ static void quantize_row_q2_K_impl(const float * restrict x, block_q2_K * restri
 size_t quantize_q2_K(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
     size_t row_size = ggml_row_size(GGML_TYPE_Q2_K, n_per_row);
     if (!quant_weights) {
-        quantize_row_q2_K_reference(src, dst, (int64_t)nrow*n_per_row);
+        quantize_row_q2_K_ref(src, dst, (int64_t)nrow*n_per_row);
     }
     else {
         char * qrow = (char *)dst;
@@ -2241,7 +2241,7 @@ size_t quantize_q2_K(const float * restrict src, void * restrict dst, int64_t nr
 
 //========================= 3-bit (de)-quantization
 
-void quantize_row_q3_K_reference(const float * restrict x, block_q3_K * restrict y, int64_t k) {
+void quantize_row_q3_K_ref(const float * restrict x, block_q3_K * restrict y, int64_t k) {
     assert(k % QK_K == 0);
     const int nb = k / QK_K;
 
@@ -2368,7 +2368,7 @@ void dequantize_row_q3_K(const block_q3_K * restrict x, float * restrict y, int6
 }
 
 void quantize_row_q3_K(const float * restrict x, void * restrict vy, int64_t k) {
-    quantize_row_q3_K_reference(x, vy, k);
+    quantize_row_q3_K_ref(x, vy, k);
 }
 
 static void quantize_row_q3_K_impl(const float * restrict x, block_q3_K * restrict y, int64_t n_per_row, const float * restrict quant_weights) {
@@ -2458,7 +2458,7 @@ static void quantize_row_q3_K_impl(const float * restrict x, block_q3_K * restri
 size_t quantize_q3_K(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
     size_t row_size = ggml_row_size(GGML_TYPE_Q3_K, n_per_row);
     if (!quant_weights) {
-        quantize_row_q3_K_reference(src, dst, (int64_t)nrow*n_per_row);
+        quantize_row_q3_K_ref(src, dst, (int64_t)nrow*n_per_row);
     }
     else {
         char * qrow = (char *)dst;
@@ -2473,7 +2473,7 @@ size_t quantize_q3_K(const float * restrict src, void * restrict dst, int64_t nr
 
 // ====================== 4-bit (de)-quantization
 
-void quantize_row_q4_K_reference(const float * restrict x, block_q4_K * restrict y, int64_t k) {
+void quantize_row_q4_K_ref(const float * restrict x, block_q4_K * restrict y, int64_t k) {
     assert(k % QK_K == 0);
     const int nb = k / QK_K;
 
@@ -2572,7 +2572,7 @@ void dequantize_row_q4_K(const block_q4_K * restrict x, float * restrict y, int6
 void quantize_row_q4_K(const float * restrict x, void * restrict vy, int64_t k) {
     assert(k % QK_K == 0);
     block_q4_K * restrict y = vy;
-    quantize_row_q4_K_reference(x, y, k);
+    quantize_row_q4_K_ref(x, y, k);
 }
 
 static void quantize_row_q4_K_impl(const float * restrict x, block_q4_K * restrict y, int64_t n_per_row, const float * quant_weights) {
@@ -2651,7 +2651,7 @@ static void quantize_row_q4_K_impl(const float * restrict x, block_q4_K * restri
 size_t quantize_q4_K(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
     size_t row_size = ggml_row_size(GGML_TYPE_Q4_K, n_per_row);
     if (!quant_weights) {
-        quantize_row_q4_K_reference(src, dst, (int64_t)nrow*n_per_row);
+        quantize_row_q4_K_ref(src, dst, (int64_t)nrow*n_per_row);
     }
     else {
         char * qrow = (char *)dst;
@@ -2666,7 +2666,7 @@ size_t quantize_q4_K(const float * restrict src, void * restrict dst, int64_t nr
 
 // ====================== 5-bit (de)-quantization
 
-void quantize_row_q5_K_reference(const float * restrict x, block_q5_K * restrict y, int64_t k) {
+void quantize_row_q5_K_ref(const float * restrict x, block_q5_K * restrict y, int64_t k) {
     assert(k % QK_K == 0);
     const int64_t nb = k / QK_K;
 
@@ -2783,7 +2783,7 @@ void dequantize_row_q5_K(const block_q5_K * restrict x, float * restrict y, int6
 void quantize_row_q5_K(const float * restrict x, void * restrict vy, int64_t k) {
     assert(k % QK_K == 0);
     block_q5_K * restrict y = vy;
-    quantize_row_q5_K_reference(x, y, k);
+    quantize_row_q5_K_ref(x, y, k);
 }
 
 static void quantize_row_q5_K_impl(const float * restrict x, block_q5_K * restrict y, int64_t n_per_row, const float * quant_weights) {
@@ -2882,7 +2882,7 @@ static void quantize_row_q5_K_impl(const float * restrict x, block_q5_K * restri
 size_t quantize_q5_K(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
     size_t row_size = ggml_row_size(GGML_TYPE_Q5_K, n_per_row);
     if (!quant_weights) {
-        quantize_row_q5_K_reference(src, dst, (int64_t)nrow*n_per_row);
+        quantize_row_q5_K_ref(src, dst, (int64_t)nrow*n_per_row);
     }
     else {
         char * qrow = (char *)dst;
@@ -2897,7 +2897,7 @@ size_t quantize_q5_K(const float * restrict src, void * restrict dst, int64_t nr
 
 // ====================== 6-bit (de)-quantization
 
-void quantize_row_q6_K_reference(const float * restrict x, block_q6_K * restrict y, int64_t k) {
+void quantize_row_q6_K_ref(const float * restrict x, block_q6_K * restrict y, int64_t k) {
     assert(k % QK_K == 0);
     const int64_t nb = k / QK_K;
 
@@ -3001,7 +3001,7 @@ void dequantize_row_q6_K(const block_q6_K * restrict x, float * restrict y, int6
 void quantize_row_q6_K(const float * restrict x, void * restrict vy, int64_t k) {
     assert(k % QK_K == 0);
     block_q6_K * restrict y = vy;
-    quantize_row_q6_K_reference(x, y, k);
+    quantize_row_q6_K_ref(x, y, k);
 }
 
 static void quantize_row_q6_K_impl(const float * restrict x, block_q6_K * restrict y, int64_t n_per_row, const float * quant_weights) {
@@ -3091,7 +3091,7 @@ static void quantize_row_q6_K_impl(const float * restrict x, block_q6_K * restri
 size_t quantize_q6_K(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
     size_t row_size = ggml_row_size(GGML_TYPE_Q6_K, n_per_row);
     if (!quant_weights) {
-        quantize_row_q6_K_reference(src, dst, (int64_t)nrow*n_per_row);
+        quantize_row_q6_K_ref(src, dst, (int64_t)nrow*n_per_row);
     }
     else {
         char * qrow = (char *)dst;
@@ -3108,7 +3108,7 @@ static void quantize_row_q4_0_impl(const float * restrict x, block_q4_0 * restri
     static_assert(QK4_0 == 32, "QK4_0 must be 32");
 
     if (!quant_weights) {
-        quantize_row_q4_0_reference(x, y, n_per_row);
+        quantize_row_q4_0_ref(x, y, n_per_row);
         return;
     }
 
@@ -3134,7 +3134,7 @@ static void quantize_row_q4_0_impl(const float * restrict x, block_q4_0 * restri
 
 size_t quantize_q4_0(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
     if (!quant_weights) {
-        quantize_row_q4_0_reference(src, dst, (int64_t)nrow*n_per_row);
+        quantize_row_q4_0_ref(src, dst, (int64_t)nrow*n_per_row);
         return nrow * ggml_row_size(GGML_TYPE_Q4_0, n_per_row);
     }
     size_t row_size = ggml_row_size(GGML_TYPE_Q4_0, n_per_row);
@@ -3151,7 +3151,7 @@ static void quantize_row_q4_1_impl(const float * restrict x, block_q4_1 * restri
     static_assert(QK4_1 == 32, "QK4_1 must be 32");
 
     if (!quant_weights) {
-        quantize_row_q4_1_reference(x, y, n_per_row);
+        quantize_row_q4_1_ref(x, y, n_per_row);
         return;
     }
 
@@ -3179,7 +3179,7 @@ static void quantize_row_q4_1_impl(const float * restrict x, block_q4_1 * restri
 
 size_t quantize_q4_1(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
     if (!quant_weights) {
-        quantize_row_q4_1_reference(src, dst, (int64_t)nrow*n_per_row);
+        quantize_row_q4_1_ref(src, dst, (int64_t)nrow*n_per_row);
         return nrow * ggml_row_size(GGML_TYPE_Q4_1, n_per_row);
     }
     size_t row_size = ggml_row_size(GGML_TYPE_Q4_1, n_per_row);
@@ -3196,7 +3196,7 @@ static void quantize_row_q5_0_impl(const float * restrict x, block_q5_0 * restri
     static_assert(QK5_0 == 32, "QK5_0 must be 32");
 
     if (!quant_weights) {
-        quantize_row_q5_0_reference(x, y, n_per_row);
+        quantize_row_q5_0_ref(x, y, n_per_row);
         return;
     }
 
@@ -3233,7 +3233,7 @@ static void quantize_row_q5_0_impl(const float * restrict x, block_q5_0 * restri
 
 size_t quantize_q5_0(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
     if (!quant_weights) {
-        quantize_row_q5_0_reference(src, dst, (int64_t)nrow*n_per_row);
+        quantize_row_q5_0_ref(src, dst, (int64_t)nrow*n_per_row);
         return nrow * ggml_row_size(GGML_TYPE_Q5_0, n_per_row);
     }
     size_t row_size = ggml_row_size(GGML_TYPE_Q5_0, n_per_row);
@@ -3250,7 +3250,7 @@ static void quantize_row_q5_1_impl(const float * restrict x, block_q5_1 * restri
     static_assert(QK5_1 == 32, "QK5_1 must be 32");
 
     if (!quant_weights) {
-        quantize_row_q5_1_reference(x, y, n_per_row);
+        quantize_row_q5_1_ref(x, y, n_per_row);
         return;
     }
 
@@ -3286,7 +3286,7 @@ static void quantize_row_q5_1_impl(const float * restrict x, block_q5_1 * restri
 
 size_t quantize_q5_1(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
     if (!quant_weights) {
-        quantize_row_q5_1_reference(src, dst, (int64_t)nrow*n_per_row);
+        quantize_row_q5_1_ref(src, dst, (int64_t)nrow*n_per_row);
         return nrow * ggml_row_size(GGML_TYPE_Q5_1, n_per_row);
     }
     size_t row_size = ggml_row_size(GGML_TYPE_Q5_1, n_per_row);
@@ -3302,7 +3302,7 @@ size_t quantize_q5_1(const float * restrict src, void * restrict dst, int64_t nr
 size_t quantize_q8_0(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
     (void)quant_weights; // not used
     const size_t row_size = ggml_row_size(GGML_TYPE_Q8_0, n_per_row);
-    quantize_row_q8_0_reference(src, dst, (int64_t)nrow*n_per_row);
+    quantize_row_q8_0_ref(src, dst, (int64_t)nrow*n_per_row);
     return nrow * row_size;
 }
 
@@ -3590,7 +3590,7 @@ void dequantize_row_iq4_xs(const block_iq4_xs * restrict x, float * restrict y,
 
 //===================================== Q8_K ==============================================
 
-void quantize_row_q8_K_reference(const float * restrict x, block_q8_K * restrict y, int64_t k) {
+void quantize_row_q8_K_ref(const float * restrict x, block_q8_K * restrict y, int64_t k) {
     assert(k % QK_K == 0);
     const int64_t nb = k / QK_K;
 
@@ -3641,7 +3641,7 @@ void dequantize_row_q8_K(const block_q8_K * restrict x, float * restrict y, int6
 }
 
 void quantize_row_q8_K(const float * restrict x, void * restrict y, int64_t k) {
-    quantize_row_q8_K_reference(x, y, k);
+    quantize_row_q8_K_ref(x, y, k);
 }
 
 //===================================== Dot ptoducts =================================
@@ -3808,59 +3808,61 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r
         float32x4_t sumv1 = vextq_f32(sumv0, sumv0, 2);
         float32x4_t sumv2 = vzip1q_f32(sumv0, sumv1);
 
-        vst1_f32(s, vget_low_f32(sumv2));
+        vst1_f32(s,      vget_low_f32(sumv2));
         vst1_f32(s + bs, vget_high_f32(sumv2));
         return;
     }
 #endif
+
+    int ib = 0;
+    float sumf = 0;
+
 #if defined(__ARM_FEATURE_SVE)
-    const svbool_t ptrueh = svptrue_pat_b8(SV_VL16);
-    const svbool_t ptruel = svnot_b_z(svptrue_b8(), ptrueh);
+    if (svcntb() == QK8_0) {
+        const svbool_t ptrueh = svptrue_pat_b8(SV_VL16);
+        const svbool_t ptruel = svnot_b_z(svptrue_b8(), ptrueh);
 
-    svfloat32_t sumv0 = svdup_n_f32(0.0f);
-    svfloat32_t sumv1 = svdup_n_f32(0.0f);
+        svfloat32_t sumv0 = svdup_n_f32(0.0f);
+        svfloat32_t sumv1 = svdup_n_f32(0.0f);
 
-    assert(nb % 2 == 0); // TODO: handle odd nb
+        for (; ib + 1 < nb; ib += 2) {
+            const block_q4_0 * restrict x0 = &x[ib + 0];
+            const block_q4_0 * restrict x1 = &x[ib + 1];
+            const block_q8_0 * restrict y0 = &y[ib + 0];
+            const block_q8_0 * restrict y1 = &y[ib + 1];
 
-    for (int i = 0; i < nb; i += 2) {
-        const block_q4_0 * restrict x0 = &x[i + 0];
-        const block_q4_0 * restrict x1 = &x[i + 1];
-        const block_q8_0 * restrict y0 = &y[i + 0];
-        const block_q8_0 * restrict y1 = &y[i + 1];
+            // load x
+            const svuint8_t qx0r = svld1rq_u8(svptrue_b8(), x0->qs);
+            const svuint8_t qx1r = svld1rq_u8(svptrue_b8(), x1->qs);
 
-        // load x
-        const svuint8_t qx0r = svld1rq_u8(svptrue_b8(), x0->qs);
-        const svuint8_t qx1r = svld1rq_u8(svptrue_b8(), x1->qs);
+            // 4-bit -> 8-bit
+            const svint8_t qx0 = svreinterpret_s8_u8(svlsr_n_u8_m(ptruel, svand_n_u8_m(ptrueh, qx0r, 0x0F), 0x04));
+            const svint8_t qx1 = svreinterpret_s8_u8(svlsr_n_u8_m(ptruel, svand_n_u8_m(ptrueh, qx1r, 0x0F), 0x04));
 
-        // 4-bit -> 8-bit
-        const svint8_t qx0 = svreinterpret_s8_u8(svlsr_n_u8_m(ptruel, svand_n_u8_m(ptrueh, qx0r, 0x0F), 0x04));
-        const svint8_t qx1 = svreinterpret_s8_u8(svlsr_n_u8_m(ptruel, svand_n_u8_m(ptrueh, qx1r, 0x0F), 0x04));
+            // sub 8
+            const svint8_t qx0s = svsub_n_s8_x(svptrue_b8(), qx0, 8);
+            const svint8_t qx1s = svsub_n_s8_x(svptrue_b8(), qx1, 8);
 
-        // sub 8
-        const svint8_t qx0s = svsub_n_s8_x(svptrue_b8(), qx0, 8);
-        const svint8_t qx1s = svsub_n_s8_x(svptrue_b8(), qx1, 8);
+            // load y
+            const svint8_t qy0 = svld1_s8(svptrue_b8(), y0->qs);
+            const svint8_t qy1 = svld1_s8(svptrue_b8(), y1->qs);
 
-        // load y
-        const svint8_t qy0 = svld1_s8(svptrue_b8(), y0->qs);
-        const svint8_t qy1 = svld1_s8(svptrue_b8(), y1->qs);
+            // dot product
+            sumv0 = svmla_n_f32_x(svptrue_b32(), sumv0, svcvt_f32_s32_x(svptrue_b32(), svdot_s32(svdup_n_s32(0), qx0s, qy0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
+            sumv1 = svmla_n_f32_x(svptrue_b32(), sumv1, svcvt_f32_s32_x(svptrue_b32(), svdot_s32(svdup_n_s32(0), qx1s, qy1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
+        }
 
-        // dot product
-        sumv0 = svmla_n_f32_x(svptrue_b32(), sumv0, svcvt_f32_s32_x(svptrue_b32(), svdot_s32(svdup_n_s32(0), qx0s, qy0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
-        sumv1 = svmla_n_f32_x(svptrue_b32(), sumv1, svcvt_f32_s32_x(svptrue_b32(), svdot_s32(svdup_n_s32(0), qx1s, qy1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
+        sumf = svaddv_f32(svptrue_b32(), svadd_f32_x(svptrue_b32(), sumv0, sumv1));
     }
-
-    *s = svaddv_f32(svptrue_b32(), svadd_f32_x(svptrue_b32(), sumv0, sumv1));
 #elif defined(__ARM_NEON)
     float32x4_t sumv0 = vdupq_n_f32(0.0f);
     float32x4_t sumv1 = vdupq_n_f32(0.0f);
 
-    assert(nb % 2 == 0); // TODO: handle odd nb
-
-    for (int i = 0; i < nb; i += 2) {
-        const block_q4_0 * restrict x0 = &x[i + 0];
-        const block_q4_0 * restrict x1 = &x[i + 1];
-        const block_q8_0 * restrict y0 = &y[i + 0];
-        const block_q8_0 * restrict y1 = &y[i + 1];
+    for (; ib + 1 < nb; ib += 2) {
+        const block_q4_0 * restrict x0 = &x[ib + 0];
+        const block_q4_0 * restrict x1 = &x[ib + 1];
+        const block_q8_0 * restrict y0 = &y[ib + 0];
+        const block_q8_0 * restrict y1 = &y[ib + 1];
 
         const uint8x16_t m4b = vdupq_n_u8(0x0F);
         const int8x16_t  s8b = vdupq_n_s8(0x8);
@@ -3894,23 +3896,23 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r
         sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
     }
 
-    *s = vaddvq_f32(sumv0) + vaddvq_f32(sumv1);
+    sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1);
 #elif defined(__AVX2__)
     // Initialize accumulator with zeros
     __m256 acc = _mm256_setzero_ps();
 
     // Main loop
-    for (int i = 0; i < nb; ++i) {
+    for (; ib < nb; ++ib) {
         /* Compute combined scale for the block */
-        const __m256 d = _mm256_set1_ps( GGML_FP16_TO_FP32(x[i].d) * GGML_FP16_TO_FP32(y[i].d) );
+        const __m256 d = _mm256_set1_ps( GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d) );
 
-        __m256i qx = bytes_from_nibbles_32(x[i].qs);
+        __m256i qx = bytes_from_nibbles_32(x[ib].qs);
 
         // Now we have a vector with bytes in [ 0 .. 15 ] interval. Offset them into [ -8 .. +7 ] interval.
         const __m256i off = _mm256_set1_epi8( 8 );
         qx = _mm256_sub_epi8( qx, off );
 
-        __m256i qy = _mm256_loadu_si256((const __m256i *)y[i].qs);
+        __m256i qy = _mm256_loadu_si256((const __m256i *)y[ib].qs);
 
         const __m256 q = mul_sum_i8_pairs_float(qx, qy);
 
@@ -3918,28 +3920,28 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r
         acc = _mm256_fmadd_ps( d, q, acc );
     }
 
-    *s = hsum_float_8(acc);
+    sumf = hsum_float_8(acc);
 #elif defined(__AVX__)
     // Initialize accumulator with zeros
     __m256 acc = _mm256_setzero_ps();
 
     // Main loop
-    for (int i = 0; i < nb; ++i) {
+    for (; ib < nb; ++ib) {
         // Compute combined scale for the block
-        const __m256 d = _mm256_set1_ps( GGML_FP16_TO_FP32(x[i].d) * GGML_FP16_TO_FP32(y[i].d) );
+        const __m256 d = _mm256_set1_ps( GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d) );
 
         const __m128i lowMask = _mm_set1_epi8(0xF);
         const __m128i off = _mm_set1_epi8(8);
 
-        const __m128i tmp = _mm_loadu_si128((const __m128i *)x[i].qs);
+        const __m128i tmp = _mm_loadu_si128((const __m128i *)x[ib].qs);
 
         __m128i bx_0 = _mm_and_si128(lowMask, tmp);
-        __m128i by_0 = _mm_loadu_si128((const __m128i *)y[i].qs);
+        __m128i by_0 = _mm_loadu_si128((const __m128i *)y[ib].qs);
         bx_0 = _mm_sub_epi8(bx_0, off);
         const __m128i i32_0 = mul_sum_i8_pairs(bx_0, by_0);
 
         bx_0 = _mm_and_si128(lowMask, _mm_srli_epi64(tmp, 4));
-        by_0 = _mm_loadu_si128((const __m128i *)(y[i].qs + 16));
+        by_0 = _mm_loadu_si128((const __m128i *)(y[ib].qs + 16));
         bx_0 = _mm_sub_epi8(bx_0, off);
         const __m128i i32_1 = mul_sum_i8_pairs(bx_0, by_0);
 
@@ -3950,7 +3952,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r
         acc = _mm256_add_ps(_mm256_mul_ps( d, p ), acc);
     }
 
-    *s = hsum_float_8(acc);
+    sumf = hsum_float_8(acc);
 #elif defined(__SSSE3__)
     // set constants
     const __m128i lowMask = _mm_set1_epi8(0xF);
@@ -3962,94 +3964,40 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r
     __m128 acc_2 = _mm_setzero_ps();
     __m128 acc_3 = _mm_setzero_ps();
 
-    // First round without accumulation
-    {
-        _mm_prefetch(&x[0] + sizeof(block_q4_0), _MM_HINT_T0);
-        _mm_prefetch(&y[0] + sizeof(block_q8_0), _MM_HINT_T0);
-
-        // Compute combined scale for the block 0 and 1
-        const __m128 d_0_1 = _mm_set1_ps( GGML_FP16_TO_FP32(x[0].d) * GGML_FP16_TO_FP32(y[0].d) );
-
-        const __m128i tmp_0_1 = _mm_loadu_si128((const __m128i *)x[0].qs);
-
-        __m128i bx_0 = _mm_and_si128(lowMask, tmp_0_1);
-        __m128i by_0 = _mm_loadu_si128((const __m128i *)y[0].qs);
-        bx_0 = _mm_sub_epi8(bx_0, off);
-        const __m128i i32_0 = mul_sum_i8_pairs(bx_0, by_0);
-
-        __m128i bx_1 = _mm_and_si128(lowMask, _mm_srli_epi64(tmp_0_1, 4));
-        __m128i by_1 = _mm_loadu_si128((const __m128i *)(y[0].qs + 16));
-        bx_1 = _mm_sub_epi8(bx_1, off);
-        const __m128i i32_1 = mul_sum_i8_pairs(bx_1, by_1);
-
-        _mm_prefetch(&x[1] + sizeof(block_q4_0), _MM_HINT_T0);
-        _mm_prefetch(&y[1] + sizeof(block_q8_0), _MM_HINT_T0);
-
-        // Compute combined scale for the block 2 and 3
-        const __m128 d_2_3 = _mm_set1_ps( GGML_FP16_TO_FP32(x[1].d) * GGML_FP16_TO_FP32(y[1].d) );
-
-        const __m128i tmp_2_3 = _mm_loadu_si128((const __m128i *)x[1].qs);
-
-        __m128i bx_2 = _mm_and_si128(lowMask, tmp_2_3);
-        __m128i by_2 = _mm_loadu_si128((const __m128i *)y[1].qs);
-        bx_2 = _mm_sub_epi8(bx_2, off);
-        const __m128i i32_2 = mul_sum_i8_pairs(bx_2, by_2);
-
-        __m128i bx_3 = _mm_and_si128(lowMask, _mm_srli_epi64(tmp_2_3, 4));
-        __m128i by_3 = _mm_loadu_si128((const __m128i *)(y[1].qs + 16));
-        bx_3 = _mm_sub_epi8(bx_3, off);
-        const __m128i i32_3 = mul_sum_i8_pairs(bx_3, by_3);
-
-        // Convert int32_t to float
-        __m128 p0 = _mm_cvtepi32_ps(i32_0);
-        __m128 p1 = _mm_cvtepi32_ps(i32_1);
-        __m128 p2 = _mm_cvtepi32_ps(i32_2);
-        __m128 p3 = _mm_cvtepi32_ps(i32_3);
-
-        // Apply the scale
-        acc_0 = _mm_mul_ps( d_0_1, p0 );
-        acc_1 = _mm_mul_ps( d_0_1, p1 );
-        acc_2 = _mm_mul_ps( d_2_3, p2 );
-        acc_3 = _mm_mul_ps( d_2_3, p3 );
-    }
-
-    assert(nb % 2 == 0); // TODO: handle odd nb
-
-    // Main loop
-    for (int i = 2; i < nb; i+=2) {
-        _mm_prefetch(&x[i] + sizeof(block_q4_0), _MM_HINT_T0);
-        _mm_prefetch(&y[i] + sizeof(block_q8_0), _MM_HINT_T0);
+    for (; ib + 1 < nb; ib += 2) {
+        _mm_prefetch(&x[ib] + sizeof(block_q4_0), _MM_HINT_T0);
+        _mm_prefetch(&y[ib] + sizeof(block_q8_0), _MM_HINT_T0);
 
         // Compute combined scale for the block 0 and 1
-        const __m128 d_0_1 = _mm_set1_ps( GGML_FP16_TO_FP32(x[i].d) * GGML_FP16_TO_FP32(y[i].d) );
+        const __m128 d_0_1 = _mm_set1_ps( GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d) );
 
-        const __m128i tmp_0_1 = _mm_loadu_si128((const __m128i *)x[i].qs);
+        const __m128i tmp_0_1 = _mm_loadu_si128((const __m128i *)x[ib].qs);
 
         __m128i bx_0 = _mm_and_si128(lowMask, tmp_0_1);
-        __m128i by_0 = _mm_loadu_si128((const __m128i *)y[i].qs);
+        __m128i by_0 = _mm_loadu_si128((const __m128i *)y[ib].qs);
         bx_0 = _mm_sub_epi8(bx_0, off);
         const __m128i i32_0 = mul_sum_i8_pairs(bx_0, by_0);
 
         __m128i bx_1 = _mm_and_si128(lowMask, _mm_srli_epi64(tmp_0_1, 4));
-        __m128i by_1 = _mm_loadu_si128((const __m128i *)(y[i].qs + 16));
+        __m128i by_1 = _mm_loadu_si128((const __m128i *)(y[ib].qs + 16));
         bx_1 = _mm_sub_epi8(bx_1, off);
         const __m128i i32_1 = mul_sum_i8_pairs(bx_1, by_1);
 
-        _mm_prefetch(&x[i] + 2 * sizeof(block_q4_0), _MM_HINT_T0);
-        _mm_prefetch(&y[i] + 2 * sizeof(block_q8_0), _MM_HINT_T0);
+        _mm_prefetch(&x[ib] + 2 * sizeof(block_q4_0), _MM_HINT_T0);
+        _mm_prefetch(&y[ib] + 2 * sizeof(block_q8_0), _MM_HINT_T0);
 
         // Compute combined scale for the block 2 and 3
-        const __m128 d_2_3 = _mm_set1_ps( GGML_FP16_TO_FP32(x[i + 1].d) * GGML_FP16_TO_FP32(y[i + 1].d) );
+        const __m128 d_2_3 = _mm_set1_ps( GGML_FP16_TO_FP32(x[ib + 1].d) * GGML_FP16_TO_FP32(y[ib + 1].d) );
 
-        const __m128i tmp_2_3 = _mm_loadu_si128((const __m128i *)x[i + 1].qs);
+        const __m128i tmp_2_3 = _mm_loadu_si128((const __m128i *)x[ib + 1].qs);
 
         __m128i bx_2 = _mm_and_si128(lowMask, tmp_2_3);
-        __m128i by_2 = _mm_loadu_si128((const __m128i *)y[i + 1].qs);
+        __m128i by_2 = _mm_loadu_si128((const __m128i *)y[ib + 1].qs);
         bx_2 = _mm_sub_epi8(bx_2, off);
         const __m128i i32_2 = mul_sum_i8_pairs(bx_2, by_2);
 
         __m128i bx_3 = _mm_and_si128(lowMask, _mm_srli_epi64(tmp_2_3, 4));
-        __m128i by_3 = _mm_loadu_si128((const __m128i *)(y[i + 1].qs + 16));
+        __m128i by_3 = _mm_loadu_si128((const __m128i *)(y[ib + 1].qs + 16));
         bx_3 = _mm_sub_epi8(bx_3, off);
         const __m128i i32_3 = mul_sum_i8_pairs(bx_3, by_3);
 
@@ -4072,18 +4020,16 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r
         acc_3 = _mm_add_ps(p3_d, acc_3);
     }
 
-    *s = hsum_float_4x4(acc_0, acc_1, acc_2, acc_3);
+    sumf = hsum_float_4x4(acc_0, acc_1, acc_2, acc_3);
 #elif defined(__riscv_v_intrinsic)
-    float sumf = 0.0;
-
     size_t vl = __riscv_vsetvl_e8m1(qk/2);
 
-    for (int i = 0; i < nb; i++) {
+    for (; ib < nb; ++ib) {
         // load elements
-        vuint8mf2_t tx = __riscv_vle8_v_u8mf2(x[i].qs, vl);
+        vuint8mf2_t tx = __riscv_vle8_v_u8mf2(x[ib].qs, vl);
 
-        vint8mf2_t y0 = __riscv_vle8_v_i8mf2(y[i].qs, vl);
-        vint8mf2_t y1 = __riscv_vle8_v_i8mf2(y[i].qs+16, vl);
+        vint8mf2_t y0 = __riscv_vle8_v_i8mf2(y[ib].qs, vl);
+        vint8mf2_t y1 = __riscv_vle8_v_i8mf2(y[ib].qs+16, vl);
 
         // mask and store lower part of x, and then upper part
         vuint8mf2_t x_a = __riscv_vand_vx_u8mf2(tx, 0x0F, vl);
@@ -4106,30 +4052,29 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r
 
         int sumi = __riscv_vmv_x_s_i32m1_i32(vs2);
 
-        sumf += sumi*GGML_FP16_TO_FP32(x[i].d)*GGML_FP16_TO_FP32(y[i].d);
+        sumf += sumi*GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d);
     }
 
-    *s = sumf;
-
 #elif defined(__POWER9_VECTOR__)
     const vector signed char lowMask = vec_splats((signed char)0xF);
+    const vector signed int v0 = vec_splats((int32_t)0);
     const vector unsigned char v4 = vec_splats((unsigned char)0x4);
     const vector signed char v8 = vec_splats((signed char)0x8);
 
     vector float vsumf0 = vec_splats(0.0f);
 
-#pragma GCC unroll 4
-    for (int i = 0; i < nb; i++) {
-        __builtin_prefetch(x[i].qs, 0, 1);
-        __builtin_prefetch(y[i].qs, 0, 1);
+#pragma GCC unroll 8
+    for (; ib < nb; ++ib) {
+        __builtin_prefetch(x[ib].qs, 0, 1);
+        __builtin_prefetch(y[ib].qs, 0, 1);
 
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
-        vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[i].d));
+        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ib].d));
+        vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[ib].d));
         vector float vd = vec_mul(vxd, vyd);
 
-        vector signed char qxs = (vector signed char)vec_xl( 0, x[i].qs);
-        vector signed char q8y0 = vec_xl( 0, y[i].qs);
-        vector signed char q8y1 = vec_xl(16, y[i].qs);
+        vector signed char qxs = (vector signed char)vec_xl( 0, x[ib].qs);
+        vector signed char q8y0 = vec_xl( 0, y[ib].qs);
+        vector signed char q8y1 = vec_xl(16, y[ib].qs);
 
         vector signed char q4x0 = vec_and(qxs, lowMask);
         vector signed char q4x1 = vec_sr(qxs, v4);
@@ -4140,9 +4085,10 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r
         vector signed short qv0 = vec_add(vec_mule(q4x0, q8y0), vec_mulo(q4x0, q8y0));
         vector signed short qv1 = vec_add(vec_mule(q4x1, q8y1), vec_mulo(q4x1, q8y1));
 
-        qv0 = vec_add(qv0, qv1);
+        vector signed int vsumi0 = v0;
 
-        vector signed int vsumi0 = vec_add(vec_unpackh(qv0), vec_unpackl(qv0));
+        vsumi0 = vec_sum4s(qv0, vsumi0);
+        vsumi0 = vec_sum4s(qv1, vsumi0);
 
         vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
     }
@@ -4150,24 +4096,24 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r
     vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
     vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
 
-    *s = vec_extract(vsumf0, 0);
+    sumf = vec_extract(vsumf0, 0);
 
 #elif defined(__loongarch_asx)
     // Initialize accumulator with zeros
     __m256 acc = (__m256)__lasx_xvldi(0);
 
     // Main loop
-    for (int i = 0; i < nb; ++i) {
+    for (; ib < nb; ++ib) {
         /* Compute combined scale for the block */
-        const __m256 d = __lasx_xvreplfr2vr_s( GGML_FP16_TO_FP32(x[i].d) * GGML_FP16_TO_FP32(y[i].d) );
+        const __m256 d = __lasx_xvreplfr2vr_s( GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d) );
 
-        __m256i qx = bytes_from_nibbles_32(x[i].qs);
+        __m256i qx = bytes_from_nibbles_32(x[ib].qs);
 
         // Now we have a vector with bytes in [ 0 .. 15 ] interval. Offset them into [ -8 .. +7 ] interval.
         const __m256i off = __lasx_xvreplgr2vr_b( 8 );
         qx = __lasx_xvsub_b( qx, off );
 
-        __m256i qy = __lasx_xvld((const __m256i *)y[i].qs, 0);
+        __m256i qy = __lasx_xvld((const __m256i *)y[ib].qs, 0);
 
         const __m256 q = mul_sum_i8_pairs_float(qx, qy);
 
@@ -4175,7 +4121,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r
         acc = __lasx_xvfmadd_s( d, q, acc );
     }
 
-    *s = hsum_float_8(acc);
+    sumf = hsum_float_8(acc);
 #elif defined(__loongarch_sx)
     // set constants
     const __m128i low_mask = __lsx_vreplgr2vr_b(0xF);
@@ -4187,89 +4133,38 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r
     __m128 acc_2 = __lsx_vldi(0);
     __m128 acc_3 = __lsx_vldi(0);
 
-    // First round without accumulation
-    {
-        _mm_prefetch(&x[0] + sizeof(block_q4_0), _MM_HINT_T0);
-        _mm_prefetch(&y[0] + sizeof(block_q8_0), _MM_HINT_T0);
-
-        // Compute combined scale for the block 0 and 1
-        const __m128 d_0_1 = __lsx_vreplgr2vr_w( GGML_FP16_TO_FP32(x[0].d) * GGML_FP16_TO_FP32(y[0].d) );
-
-        const __m128i tmp_0_1 = __lsx_vld((const __m128i *)x[0].qs, 0);
-
-        __m128i bx_0 = __lsx_vand_v(low_mask, tmp_0_1);
-        __m128i by_0 = __lsx_vld((const __m128i *)y[0].qs, 0);
-        bx_0 = __lsx_vsub_b(bx_0, off);
-        const __m128i i32_0 = mul_sum_i8_pairs(bx_0, by_0);
-
-        __m128i bx_1 = __lsx_vand_v(low_mask, __lsx_vsrli_d(tmp_0_1, 4));
-        __m128i by_1 = __lsx_vld((const __m128i *)(y[0].qs + 16), 0);
-        bx_1 = __lsx_vsub_b(bx_1, off);
-        const __m128i i32_1 = mul_sum_i8_pairs(bx_1, by_1);
-
-        // Compute combined scale for the block 2 and 3
-        const __m128 d_2_3 = __lsx_vreplgr2vr_w( GGML_FP16_TO_FP32(x[1].d) * GGML_FP16_TO_FP32(y[1].d) );
-
-        const __m128i tmp_2_3 = __lsx_vld((const __m128i *)x[1].qs, 0);
-
-        __m128i bx_2 = __lsx_vand_v(low_mask, tmp_2_3);
-        __m128i by_2 = __lsx_vld((const __m128i *)y[1].qs, 0);
-        bx_2 = __lsx_vsub_b(bx_2, off);
-        const __m128i i32_2 = mul_sum_i8_pairs(bx_2, by_2);
-
-        __m128i bx_3 = __lsx_vand_v(low_mask, __lsx_vsrli_d(tmp_2_3, 4));
-        __m128i by_3 = __lsx_vld((const __m128i *)(y[1].qs + 16), 0);
-        bx_3 = __lsx_vsub_b(bx_3, off);
-        const __m128i i32_3 = mul_sum_i8_pairs(bx_3, by_3);
-
-        // Convert int32_t to float
-        __m128 p0 = __lsx_vffint_s_w(i32_0);
-        __m128 p1 = __lsx_vffint_s_w(i32_1);
-        __m128 p2 = __lsx_vffint_s_w(i32_2);
-        __m128 p3 = __lsx_vffint_s_w(i32_3);
-
-        // Apply the scale
-        acc_0 = __lsx_vfmul_s( d_0_1, p0 );
-        acc_1 = __lsx_vfmul_s( d_0_1, p1 );
-        acc_2 = __lsx_vfmul_s( d_2_3, p2 );
-        acc_3 = __lsx_vfmul_s( d_2_3, p3 );
-    }
-
-    assert(nb % 2 == 0); // TODO: handle odd nb
-
-    // Main loop
-    for (int i = 2; i < nb; i+=2) {
+    for (; ib + 1 < nb; ib += 2) {
 
         // Compute combined scale for the block 0 and 1
-        const __m128 d_0_1 = __lsx_vreplgr2vr_w( GGML_FP16_TO_FP32(x[i].d) * GGML_FP16_TO_FP32(y[i].d) );
+        const __m128 d_0_1 = __lsx_vreplgr2vr_w( GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d) );
 
-        const __m128i tmp_0_1 = __lsx_vld((const __m128i *)x[i].qs, 0);
+        const __m128i tmp_0_1 = __lsx_vld((const __m128i *)x[ib].qs, 0);
 
         __m128i bx_0 = __lsx_vand_v(low_mask, tmp_0_1);
-        __m128i by_0 = __lsx_vld((const __m128i *)y[i].qs, 0);
+        __m128i by_0 = __lsx_vld((const __m128i *)y[ib].qs, 0);
         bx_0 = __lsx_vsub_b(bx_0, off);
         const __m128i i32_0 = mul_sum_i8_pairs(bx_0, by_0);
 
         __m128i bx_1 = __lsx_vand_v(low_mask, __lsx_vsrli_d(tmp_0_1, 4));
-        __m128i by_1 = __lsx_vld((const __m128i *)(y[i].qs + 16), 0);
+        __m128i by_1 = __lsx_vld((const __m128i *)(y[ib].qs + 16), 0);
         bx_1 = __lsx_vsub_b(bx_1, off);
         const __m128i i32_1 = mul_sum_i8_pairs(bx_1, by_1);
 
-        //_mm_prefetch(&x[i] + 2 * sizeof(block_q4_0), _MM_HINT_T0);
-        //_mm_prefetch(&y[i] + 2 * sizeof(block_q8_0), _MM_HINT_T0);
+        //_mm_prefetch(&x[ib] + 2 * sizeof(block_q4_0), _MM_HINT_T0);
+        //_mm_prefetch(&y[ib] + 2 * sizeof(block_q8_0), _MM_HINT_T0);
 
         // Compute combined scale for the block 2 and 3
-        const __m128 d_2_3 = __lsx_vreplgr2vr_w( GGML_FP16_TO_FP32(x[i + 1].d) * GGML_FP16_TO_FP32(y[i + 1].d) );
+        const __m128 d_2_3 = __lsx_vreplgr2vr_w( GGML_FP16_TO_FP32(x[ib + 1].d) * GGML_FP16_TO_FP32(y[ib + 1].d) );
 
-        const __m128i tmp_2_3 = __lsx_vld((const __m128i *)x[i + 1].qs, 0);
+        const __m128i tmp_2_3 = __lsx_vld((const __m128i *)x[ib + 1].qs, 0);
 
         __m128i bx_2 = __lsx_vand_v(low_mask, tmp_2_3);
-        __m128i by_2 = __lsx_vld((const __m128i *)y[i + 1].qs, 0);
+        __m128i by_2 = __lsx_vld((const __m128i *)y[ib + 1].qs, 0);
         bx_2 = __lsx_vsub_b(bx_2, off);
         const __m128i i32_2 = mul_sum_i8_pairs(bx_2, by_2);
 
         __m128i bx_3 = __lsx_vand_v(low_mask, __lsx_vsrli_d(tmp_2_3, 4));
-        __m128i by_3 = __lsx_vld((const __m128i *)(y[i + 1].qs + 16), 0);
+        __m128i by_3 = __lsx_vld((const __m128i *)(y[ib + 1].qs + 16), 0);
         bx_3 = __lsx_vsub_b(bx_3, off);
         const __m128i i32_3 = mul_sum_i8_pairs(bx_3, by_3);
 
@@ -4292,27 +4187,25 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r
         acc_3 = __lsx_vfadd_s(p3_d, acc_3);
     }
 
-    *s = hsum_float_4x4(acc_0, acc_1, acc_2, acc_3);
-
-#else
-    // scalar
-    float sumf = 0.0;
-
-    for (int i = 0; i < nb; i++) {
-        int sumi = 0;
+    sumf = hsum_float_4x4(acc_0, acc_1, acc_2, acc_3);
+#endif
+    for (; ib < nb; ++ib) {
+        int sumi0 = 0;
+        int sumi1 = 0;
 
         for (int j = 0; j < qk/2; ++j) {
-            const int v0 = (x[i].qs[j] & 0x0F) - 8;
-            const int v1 = (x[i].qs[j] >>   4) - 8;
+            const int v0 = (x[ib].qs[j] & 0x0F) - 8;
+            const int v1 = (x[ib].qs[j] >>   4) - 8;
 
-            sumi += (v0 * y[i].qs[j]) + (v1 * y[i].qs[j + qk/2]);
+            sumi0 += (v0 * y[ib].qs[j]);
+            sumi1 += (v1 * y[ib].qs[j + qk/2]);
         }
 
-        sumf += sumi*GGML_FP16_TO_FP32(x[i].d)*GGML_FP16_TO_FP32(y[i].d);
+        int sumi = sumi0 + sumi1;
+        sumf += sumi*GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d);
     }
 
     *s = sumf;
-#endif
 }
 
 void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
@@ -4398,11 +4291,15 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * r
         float32x4_t sumv2 = vzip1q_f32(sumv0, sumv1);
         sumv2 = vaddq_f32(sumv2, summs0);
 
-        vst1_f32(s, vget_low_f32(sumv2));
+        vst1_f32(s,      vget_low_f32 (sumv2));
         vst1_f32(s + bs, vget_high_f32(sumv2));
         return;
     }
 #endif
+
+    int ib = 0;
+    float sumf = 0;
+
     // TODO: add WASM SIMD
 #if defined(__ARM_NEON)
     float32x4_t sumv0 = vdupq_n_f32(0.0f);
@@ -4410,13 +4307,11 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * r
 
     float summs = 0;
 
-    assert(nb % 2 == 0); // TODO: handle odd nb
-
-    for (int i = 0; i < nb; i += 2) {
-        const block_q4_1 * restrict x0 = &x[i + 0];
-        const block_q4_1 * restrict x1 = &x[i + 1];
-        const block_q8_1 * restrict y0 = &y[i + 0];
-        const block_q8_1 * restrict y1 = &y[i + 1];
+    for (; ib + 1 < nb; ib += 2) {
+        const block_q4_1 * restrict x0 = &x[ib + 0];
+        const block_q4_1 * restrict x1 = &x[ib + 1];
+        const block_q8_1 * restrict y0 = &y[ib + 0];
+        const block_q8_1 * restrict y1 = &y[ib + 1];
 
         summs += GGML_FP16_TO_FP32(x0->m) * GGML_FP16_TO_FP32(y0->s) + GGML_FP16_TO_FP32(x1->m) * GGML_FP16_TO_FP32(y1->s);
 
@@ -4445,7 +4340,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * r
         sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
     }
 
-    *s = vaddvq_f32(sumv0) + vaddvq_f32(sumv1) + summs;
+    sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1) + summs;
 #elif defined(__AVX2__) || defined(__AVX__)
     // Initialize accumulator with zeros
     __m256 acc = _mm256_setzero_ps();
@@ -4453,11 +4348,11 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * r
     float summs = 0;
 
     // Main loop
-    for (int i = 0; i < nb; ++i) {
-        const float d0 = GGML_FP16_TO_FP32(x[i].d);
-        const float d1 = GGML_FP16_TO_FP32(y[i].d);
+    for (; ib < nb; ++ib) {
+        const float d0 = GGML_FP16_TO_FP32(x[ib].d);
+        const float d1 = GGML_FP16_TO_FP32(y[ib].d);
 
-        summs += GGML_FP16_TO_FP32(x[i].m) * GGML_FP16_TO_FP32(y[i].s);
+        summs += GGML_FP16_TO_FP32(x[ib].m) * GGML_FP16_TO_FP32(y[ib].s);
 
         const __m256 d0v = _mm256_set1_ps( d0 );
         const __m256 d1v = _mm256_set1_ps( d1 );
@@ -4466,8 +4361,8 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * r
         const __m256 d0d1 = _mm256_mul_ps( d0v, d1v );
 
         // Load 16 bytes, and unpack 4 bit fields into bytes, making 32 bytes
-        const __m256i qx = bytes_from_nibbles_32(x[i].qs);
-        const __m256i qy = _mm256_loadu_si256( (const __m256i *)y[i].qs );
+        const __m256i qx = bytes_from_nibbles_32(x[ib].qs);
+        const __m256i qy = _mm256_loadu_si256( (const __m256i *)y[ib].qs );
 
         const __m256 xy = mul_sum_us8_pairs_float(qx, qy);
 
@@ -4479,18 +4374,16 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * r
 #endif
     }
 
-    *s = hsum_float_8(acc) + summs;
+    sumf = hsum_float_8(acc) + summs;
 #elif defined(__riscv_v_intrinsic)
-    float sumf = 0.0;
-
     size_t vl = __riscv_vsetvl_e8m1(qk/2);
 
-    for (int i = 0; i < nb; i++) {
+    for (; ib < nb; ++ib) {
         // load elements
-        vuint8mf2_t tx = __riscv_vle8_v_u8mf2(x[i].qs, vl);
+        vuint8mf2_t tx = __riscv_vle8_v_u8mf2(x[ib].qs, vl);
 
-        vint8mf2_t y0 = __riscv_vle8_v_i8mf2(y[i].qs, vl);
-        vint8mf2_t y1 = __riscv_vle8_v_i8mf2(y[i].qs+16, vl);
+        vint8mf2_t y0 = __riscv_vle8_v_i8mf2(y[ib].qs, vl);
+        vint8mf2_t y1 = __riscv_vle8_v_i8mf2(y[ib].qs+16, vl);
 
         // mask and store lower part of x, and then upper part
         vuint8mf2_t x_a = __riscv_vand_vx_u8mf2(tx, 0x0F, vl);
@@ -4509,43 +4402,40 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * r
 
         int sumi = __riscv_vmv_x_s_i32m1_i32(vs2);
 
-        sumf += (GGML_FP16_TO_FP32(x[i].d)*GGML_FP16_TO_FP32(y[i].d))*sumi + GGML_FP16_TO_FP32(x[i].m)*GGML_FP16_TO_FP32(y[i].s);
+        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
     }
 
-    *s = sumf;
-
 #elif defined(__POWER9_VECTOR__)
     const vector signed char lowMask = vec_splats((signed char)0xF);
+    const vector signed int v0 = vec_splats((int32_t)0);
     const vector unsigned char v4 = vec_splats((unsigned char)0x4);
 
     vector float vsumf0 = vec_splats(0.0f);
 
 #pragma GCC unroll 4
-    for (int i = 0; i < nb; i++) {
-        __builtin_prefetch(x[i].qs, 0, 1);
-        __builtin_prefetch(y[i].qs, 0, 1);
+    for (; ib < nb; ++ib) {
+        __builtin_prefetch(x[ib].qs, 0, 1);
+        __builtin_prefetch(y[ib].qs, 0, 1);
 
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
-        vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[i].d));
+        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ib].d));
+        vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[ib].d));
         vector float vd = vec_mul(vxd, vyd);
 
-        vector float vxmin = vec_splats(GGML_FP16_TO_FP32(x[i].m));
-        vector float vys = {GGML_FP16_TO_FP32(y[i].s), 0.0f, 0.0f, 0.0f};
+        vector float vxmin = vec_splats(GGML_FP16_TO_FP32(x[ib].m));
+        vector float vys = {GGML_FP16_TO_FP32(y[ib].s), 0.0f, 0.0f, 0.0f};
         vsumf0 = vec_madd(vxmin, vys, vsumf0);
 
-        vector signed char qxs = (vector signed char)vec_xl( 0, x[i].qs);
-        vector signed char q8y0 = vec_xl( 0, y[i].qs);
-        vector signed char q8y1 = vec_xl(16, y[i].qs);
-
-        vector signed char q4x0 = vec_and(qxs, lowMask);
-        vector signed char q4x1 = vec_sr(qxs, v4);
+        vector signed char qxs = (vector signed char)vec_xl( 0, x[ib].qs);
+        vector signed char q8y0 = vec_xl( 0, y[ib].qs);
+        vector signed char q8y1 = vec_xl(16, y[ib].qs);
 
-        vector signed short qv0 = vec_add(vec_mule(q4x0, q8y0), vec_mulo(q4x0, q8y0));
-        vector signed short qv1 = vec_add(vec_mule(q4x1, q8y1), vec_mulo(q4x1, q8y1));
+        vector unsigned char q4x0 = (vector unsigned char)vec_and(qxs, lowMask);
+        vector unsigned char q4x1 = (vector unsigned char)vec_sr(qxs, v4);
 
-        qv0 = vec_add(qv0, qv1);
+        vector signed int vsumi0 = v0;
 
-        vector signed int vsumi0 = vec_add(vec_unpackh(qv0), vec_unpackl(qv0));
+        vsumi0 = vec_msum(q8y0, q4x0, vsumi0);
+        vsumi0 = vec_msum(q8y1, q4x1, vsumi0);
 
         vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
     }
@@ -4553,7 +4443,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * r
     vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
     vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
 
-    *s = vec_extract(vsumf0, 0);
+    sumf = vec_extract(vsumf0, 0);
 
 #elif defined(__loongarch_asx)
     // Initialize accumulator with zeros
@@ -4562,11 +4452,11 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * r
     float summs = 0;
 
     // Main loop
-    for (int i = 0; i < nb; ++i) {
-        const float d0 = GGML_FP16_TO_FP32(x[i].d);
-        const float d1 = GGML_FP16_TO_FP32(y[i].d);
+    for (; ib < nb; ++ib) {
+        const float d0 = GGML_FP16_TO_FP32(x[ib].d);
+        const float d1 = GGML_FP16_TO_FP32(y[ib].d);
 
-        summs += GGML_FP16_TO_FP32(x[i].m) * GGML_FP16_TO_FP32(y[i].s);
+        summs += GGML_FP16_TO_FP32(x[ib].m) * GGML_FP16_TO_FP32(y[ib].s);
 
         const __m256 d0v = __lasx_xvreplfr2vr_s( d0 );
         const __m256 d1v = __lasx_xvreplfr2vr_s( d1 );
@@ -4575,8 +4465,8 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * r
         const __m256 d0d1 = __lasx_xvfmul_s( d0v, d1v );
 
         // Load 16 bytes, and unpack 4 bit fields into bytes, making 32 bytes
-        const __m256i qx = bytes_from_nibbles_32(x[i].qs);
-        const __m256i qy = __lasx_xvld( (const __m256i *)y[i].qs, 0);
+        const __m256i qx = bytes_from_nibbles_32(x[ib].qs);
+        const __m256i qy = __lasx_xvld( (const __m256i *)y[ib].qs, 0);
 
         const __m256 xy = mul_sum_us8_pairs_float(qx, qy);
 
@@ -4584,33 +4474,34 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * r
         acc = __lasx_xvfmadd_s( d0d1, xy, acc );
     }
 
-    *s = hsum_float_8(acc) + summs;
-
-#else
-    // scalar
-    float sumf = 0.0;
-
-    for (int i = 0; i < nb; i++) {
-        int sumi = 0;
+    sumf = hsum_float_8(acc) + summs;
+#endif
+    for (; ib < nb; ++ib) {
+        int sumi0 = 0;
+        int sumi1 = 0;
 
         for (int j = 0; j < qk/2; ++j) {
-            const int v0 = (x[i].qs[j] & 0x0F);
-            const int v1 = (x[i].qs[j] >>   4);
+            const int v0 = (x[ib].qs[j] & 0x0F);
+            const int v1 = (x[ib].qs[j] >>   4);
 
-            sumi += (v0 * y[i].qs[j]) + (v1 * y[i].qs[j + qk/2]);
+            sumi0 += (v0 * y[ib].qs[j]);
+            sumi1 += (v1 * y[ib].qs[j + qk/2]);
         }
 
-        sumf += (GGML_FP16_TO_FP32(x[i].d)*GGML_FP16_TO_FP32(y[i].d))*sumi + GGML_FP16_TO_FP32(x[i].m)*GGML_FP16_TO_FP32(y[i].s);
+        int sumi = sumi0 + sumi1;
+        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
     }
 
     *s = sumf;
-#endif
 }
 
 void ggml_vec_dot_q5_0_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
     const int qk = QK8_0;
     const int nb = n / qk;
 
+    int ib = 0;
+    float sumf = 0;
+
     assert(n % qk == 0);
     assert(qk == QK5_0);
     assert(nrc == 1);
@@ -4632,13 +4523,11 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * restrict s, size_t bs, const void * r
     uint64_t tmp0[4];
     uint64_t tmp1[4];
 
-    assert(nb % 2 == 0); // TODO: handle odd nb
-
-    for (int i = 0; i < nb; i += 2) {
-        const block_q5_0 * restrict x0 = &x[i];
-        const block_q5_0 * restrict x1 = &x[i + 1];
-        const block_q8_0 * restrict y0 = &y[i];
-        const block_q8_0 * restrict y1 = &y[i + 1];
+    for (; ib + 1 < nb; ib += 2) {
+        const block_q5_0 * restrict x0 = &x[ib];
+        const block_q5_0 * restrict x1 = &x[ib + 1];
+        const block_q8_0 * restrict y0 = &y[ib];
+        const block_q8_0 * restrict y1 = &y[ib + 1];
 
         const uint8x16_t m4b = vdupq_n_u8(0x0F);
 
@@ -4690,7 +4579,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * restrict s, size_t bs, const void * r
                         ggml_vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
     }
 
-    *s = vaddvq_f32(sumv0) + vaddvq_f32(sumv1);
+    sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1);
 #elif defined(__wasm_simd128__)
     v128_t sumv = wasm_f32x4_splat(0.0f);
 
@@ -4698,9 +4587,9 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * restrict s, size_t bs, const void * r
     uint64_t tmp[4];
 
     // TODO: check if unrolling this is better
-    for (int i = 0; i < nb; ++i) {
-        const block_q5_0 * restrict x0 = &x[i];
-        const block_q8_0 * restrict y0 = &y[i];
+    for (; ib < nb; ++ib) {
+        const block_q5_0 * restrict x0 = &x[ib];
+        const block_q8_0 * restrict y0 = &y[ib];
 
         const v128_t m4b  = wasm_i8x16_splat(0x0F);
 
@@ -4750,23 +4639,23 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * restrict s, size_t bs, const void * r
                     wasm_f32x4_splat(GGML_FP16_TO_FP32(x0->d) * GGML_FP16_TO_FP32(y0->d))));
     }
 
-    *s = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) +
-         wasm_f32x4_extract_lane(sumv, 2) + wasm_f32x4_extract_lane(sumv, 3);
+    sumf = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) +
+           wasm_f32x4_extract_lane(sumv, 2) + wasm_f32x4_extract_lane(sumv, 3);
 #elif defined(__AVX2__)
     // Initialize accumulator with zeros
     __m256 acc = _mm256_setzero_ps();
 
     // Main loop
-    for (int i = 0; i < nb; i++) {
+    for (; ib < nb; ++ib) {
         /* Compute combined scale for the block */
-        const __m256 d = _mm256_set1_ps(GGML_FP16_TO_FP32(x[i].d) * GGML_FP16_TO_FP32(y[i].d));
+        const __m256 d = _mm256_set1_ps(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d));
 
-        __m256i qx = bytes_from_nibbles_32(x[i].qs);
-        __m256i bxhi = bytes_from_bits_32(x[i].qh);
+        __m256i qx = bytes_from_nibbles_32(x[ib].qs);
+        __m256i bxhi = bytes_from_bits_32(x[ib].qh);
         bxhi = _mm256_andnot_si256(bxhi, _mm256_set1_epi8((char)0xF0));
         qx = _mm256_or_si256(qx, bxhi);
 
-        __m256i qy = _mm256_loadu_si256((const __m256i *)y[i].qs);
+        __m256i qy = _mm256_loadu_si256((const __m256i *)y[ib].qs);
 
         const __m256 q = mul_sum_i8_pairs_float(qx, qy);
 
@@ -4774,19 +4663,19 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * restrict s, size_t bs, const void * r
         acc = _mm256_fmadd_ps(d, q, acc);
     }
 
-    *s = hsum_float_8(acc);
+    sumf = hsum_float_8(acc);
 #elif defined(__AVX__)
     // Initialize accumulator with zeros
     __m256 acc = _mm256_setzero_ps();
     __m128i mask = _mm_set1_epi8((char)0xF0);
 
     // Main loop
-    for (int i = 0; i < nb; i++) {
+    for (; ib < nb; ++ib) {
         /* Compute combined scale for the block */
-        const __m256 d = _mm256_set1_ps(GGML_FP16_TO_FP32(x[i].d) * GGML_FP16_TO_FP32(y[i].d));
+        const __m256 d = _mm256_set1_ps(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d));
 
-        __m256i bx_0 = bytes_from_nibbles_32(x[i].qs);
-        const __m256i bxhi = bytes_from_bits_32(x[i].qh);
+        __m256i bx_0 = bytes_from_nibbles_32(x[ib].qs);
+        const __m256i bxhi = bytes_from_bits_32(x[ib].qh);
         __m128i bxhil = _mm256_castsi256_si128(bxhi);
         __m128i bxhih = _mm256_extractf128_si256(bxhi, 1);
         bxhil = _mm_andnot_si128(bxhil, mask);
@@ -4797,7 +4686,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * restrict s, size_t bs, const void * r
         bxh = _mm_or_si128(bxh, bxhih);
         bx_0 = MM256_SET_M128I(bxh, bxl);
 
-        const __m256i by_0 = _mm256_loadu_si256((const __m256i *)y[i].qs);
+        const __m256i by_0 = _mm256_loadu_si256((const __m256i *)y[ib].qs);
 
         const __m256 q = mul_sum_i8_pairs_float(bx_0, by_0);
 
@@ -4805,10 +4694,8 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * restrict s, size_t bs, const void * r
         acc = _mm256_add_ps(_mm256_mul_ps(d, q), acc);
     }
 
-    *s = hsum_float_8(acc);
+    sumf = hsum_float_8(acc);
 #elif defined(__riscv_v_intrinsic)
-    float sumf = 0.0;
-
     uint32_t qh;
 
     size_t vl = __riscv_vsetvl_e8m1(qk/2);
@@ -4820,8 +4707,8 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * restrict s, size_t bs, const void * r
     vuint32m2_t vt_3 = __riscv_vsll_vx_u32m2(vt_2, 16, vl);
     vuint32m2_t vt_4 = __riscv_vadd_vx_u32m2(vt_1, 12, vl);
 
-    for (int i = 0; i < nb; i++) {
-        memcpy(&qh, x[i].qh, sizeof(uint32_t));
+    for (; ib < nb; ++ib) {
+        memcpy(&qh, x[ib].qh, sizeof(uint32_t));
 
         // ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4;
         vuint32m2_t xha_0 = __riscv_vand_vx_u32m2(vt_2, qh, vl);
@@ -4840,10 +4727,10 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * restrict s, size_t bs, const void * r
         vuint8mf2_t xh_1 = __riscv_vncvt_x_x_w_u8mf2(xhc_1, vl);
 
         // load
-        vuint8mf2_t tx = __riscv_vle8_v_u8mf2(x[i].qs, vl);
+        vuint8mf2_t tx = __riscv_vle8_v_u8mf2(x[ib].qs, vl);
 
-        vint8mf2_t y0 = __riscv_vle8_v_i8mf2(y[i].qs, vl);
-        vint8mf2_t y1 = __riscv_vle8_v_i8mf2(y[i].qs+16, vl);
+        vint8mf2_t y0 = __riscv_vle8_v_i8mf2(y[ib].qs, vl);
+        vint8mf2_t y1 = __riscv_vle8_v_i8mf2(y[ib].qs+16, vl);
 
         vuint8mf2_t x_at = __riscv_vand_vx_u8mf2(tx, 0x0F, vl);
         vuint8mf2_t x_lt = __riscv_vsrl_vx_u8mf2(tx, 0x04, vl);
@@ -4867,11 +4754,9 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * restrict s, size_t bs, const void * r
 
         int sumi = __riscv_vmv_x_s_i32m1_i32(vs2);
 
-        sumf += (GGML_FP16_TO_FP32(x[i].d)*GGML_FP16_TO_FP32(y[i].d)) * sumi;
+        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)) * sumi;
     }
 
-    *s = sumf;
-
 #elif defined(__POWER9_VECTOR__)
     const vector signed char lowMask = vec_splats((signed char)0xF);
     const vector unsigned char v4 = vec_splats((unsigned char)4);
@@ -4879,27 +4764,27 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * restrict s, size_t bs, const void * r
     vector float vsumf0 = vec_splats(0.0f);
 
 #pragma GCC unroll 4
-    for (int i = 0; i < nb; ++i) {
-        __builtin_prefetch(x[i].qs, 0, 1);
-        __builtin_prefetch(y[i].qs, 0, 1);
+    for (; ib < nb; ++ib) {
+        __builtin_prefetch(x[ib].qs, 0, 1);
+        __builtin_prefetch(y[ib].qs, 0, 1);
 
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
-        vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[i].d));
+        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ib].d));
+        vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[ib].d));
         vector float vd = vec_mul(vxd, vyd);
 
-        vector signed long long aux64x2_0 = {(uint64_t)(table_b2b_1[x[i].qh[0]]), (uint64_t)(table_b2b_1[x[i].qh[1]])};
-        vector signed long long aux64x2_1 = {(uint64_t)(table_b2b_1[x[i].qh[2]]), (uint64_t)(table_b2b_1[x[i].qh[3]])};
+        vector signed long long aux64x2_0 = {(uint64_t)(table_b2b_1[x[ib].qh[0]]), (uint64_t)(table_b2b_1[x[ib].qh[1]])};
+        vector signed long long aux64x2_1 = {(uint64_t)(table_b2b_1[x[ib].qh[2]]), (uint64_t)(table_b2b_1[x[ib].qh[3]])};
 
         vector signed char qh0 = (vector signed char)aux64x2_0;
         vector signed char qh1 = (vector signed char)aux64x2_1;
 
-        vector signed char qxs = (vector signed char)vec_xl( 0, x[i].qs);
+        vector signed char qxs = (vector signed char)vec_xl( 0, x[ib].qs);
 
         vector signed char q5x0 = vec_sub(vec_and (qxs, lowMask), qh0);
         vector signed char q5x1 = vec_sub(vec_sr(qxs, v4), qh1);
 
-        vector signed char q8y0 = vec_xl(  0, y[i].qs);
-        vector signed char q8y1 = vec_xl( 16, y[i].qs);
+        vector signed char q8y0 = vec_xl(  0, y[ib].qs);
+        vector signed char q8y1 = vec_xl( 16, y[ib].qs);
 
         vector signed short qv0 = vec_add(vec_mule(q5x0, q8y0), vec_mulo(q5x0, q8y0));
         vector signed short qv1 = vec_add(vec_mule(q5x1, q8y1), vec_mulo(q5x1, q8y1));
@@ -4914,23 +4799,23 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * restrict s, size_t bs, const void * r
     vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
     vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
 
-    *s = vec_extract(vsumf0, 0);
+    sumf = vec_extract(vsumf0, 0);
 
 #elif defined(__loongarch_asx)
     // Initialize accumulator with zeros
     __m256 acc = (__m256)__lasx_xvldi(0);
 
     // Main loop
-    for (int i = 0; i < nb; i++) {
+    for (; ib < nb; ++ib) {
         /* Compute combined scale for the block */
-        const __m256 d = __lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(x[i].d) * GGML_FP16_TO_FP32(y[i].d)); //FIXME
+        const __m256 d = __lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d)); //FIXME
 
-        __m256i qx = bytes_from_nibbles_32(x[i].qs);
-        __m256i bxhi = bytes_from_bits_32(x[i].qh);
+        __m256i qx = bytes_from_nibbles_32(x[ib].qs);
+        __m256i bxhi = bytes_from_bits_32(x[ib].qh);
         bxhi = __lasx_xvandn_v(bxhi, __lasx_xvreplgr2vr_b((char)0xF0));
         qx = __lasx_xvor_v(qx, bxhi);
 
-        __m256i qy = __lasx_xvld((const __m256i *)y[i].qs, 0);
+        __m256i qy = __lasx_xvld((const __m256i *)y[ib].qs, 0);
 
         const __m256 q = mul_sum_i8_pairs_float(qx, qy);
 
@@ -4938,39 +4823,40 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * restrict s, size_t bs, const void * r
         acc = __lasx_xvfmadd_s(d, q, acc);
     }
 
-    *s = hsum_float_8(acc);
-
-#else
-    // scalar
-    float sumf = 0.0;
-
-    for (int i = 0; i < nb; i++) {
+    sumf = hsum_float_8(acc);
+#endif
+    for (; ib < nb; ++ib) {
         uint32_t qh;
-        memcpy(&qh, x[i].qh, sizeof(qh));
+        memcpy(&qh, x[ib].qh, sizeof(qh));
 
-        int sumi = 0;
+        int sumi0 = 0;
+        int sumi1 = 0;
 
         for (int j = 0; j < qk/2; ++j) {
             const uint8_t xh_0 = ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4;
             const uint8_t xh_1 = ((qh & (1u << (j + 16))) >> (j + 12));
 
-            const int32_t x0 = ((x[i].qs[j] & 0x0F) | xh_0) - 16;
-            const int32_t x1 = ((x[i].qs[j] >>   4) | xh_1) - 16;
+            const int32_t x0 = (int8_t)(((x[ib].qs[j] & 0x0F) | xh_0) - 16);
+            const int32_t x1 = (int8_t)(((x[ib].qs[j] >>   4) | xh_1) - 16);
 
-            sumi += (x0 * y[i].qs[j]) + (x1 * y[i].qs[j + qk/2]);
+            sumi0 += (x0 * y[ib].qs[j]);
+            sumi1 += (x1 * y[ib].qs[j + qk/2]);
         }
 
-        sumf += (GGML_FP16_TO_FP32(x[i].d)*GGML_FP16_TO_FP32(y[i].d)) * sumi;
+        int sumi = sumi0 + sumi1;
+        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)) * sumi;
     }
 
     *s = sumf;
-#endif
 }
 
 void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
     const int qk = QK8_1;
     const int nb = n / qk;
 
+    int ib = 0;
+    float sumf = 0;
+
     assert(n % qk == 0);
     assert(qk == QK5_1);
     assert(nrc == 1);
@@ -4995,13 +4881,11 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r
     uint64_t tmp0[4];
     uint64_t tmp1[4];
 
-    assert(nb % 2 == 0); // TODO: handle odd nb
-
-    for (int i = 0; i < nb; i += 2) {
-        const block_q5_1 * restrict x0 = &x[i];
-        const block_q5_1 * restrict x1 = &x[i + 1];
-        const block_q8_1 * restrict y0 = &y[i];
-        const block_q8_1 * restrict y1 = &y[i + 1];
+    for (; ib + 1 < nb; ib += 2) {
+        const block_q5_1 * restrict x0 = &x[ib];
+        const block_q5_1 * restrict x1 = &x[ib + 1];
+        const block_q8_1 * restrict y0 = &y[ib];
+        const block_q8_1 * restrict y1 = &y[ib + 1];
 
         const uint8x16_t m4b = vdupq_n_u8(0x0F);
 
@@ -5056,7 +4940,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r
                         ggml_vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
     }
 
-    *s = vaddvq_f32(sumv0) + vaddvq_f32(sumv1) + summs0 + summs1;
+    sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1) + summs0 + summs1;
 #elif defined(__wasm_simd128__)
     v128_t sumv = wasm_f32x4_splat(0.0f);
 
@@ -5066,9 +4950,9 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r
     uint64_t tmp[4];
 
     // TODO: check if unrolling this is better
-    for (int i = 0; i < nb; ++i) {
-        const block_q5_1 * restrict x0 = &x[i];
-        const block_q8_1 * restrict y0 = &y[i];
+    for (; ib < nb; ++ib) {
+        const block_q5_1 * restrict x0 = &x[ib];
+        const block_q8_1 * restrict y0 = &y[ib];
 
         summs += GGML_FP16_TO_FP32(x0->m) * GGML_FP16_TO_FP32(y0->s);
 
@@ -5120,8 +5004,8 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r
                     wasm_f32x4_splat(GGML_FP16_TO_FP32(x0->d) * GGML_FP16_TO_FP32(y0->d))));
     }
 
-    *s = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) +
-         wasm_f32x4_extract_lane(sumv, 2) + wasm_f32x4_extract_lane(sumv, 3) + summs;
+    sumf = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) +
+           wasm_f32x4_extract_lane(sumv, 2) + wasm_f32x4_extract_lane(sumv, 3) + summs;
 #elif defined(__AVX2__)
     // Initialize accumulator with zeros
     __m256 acc = _mm256_setzero_ps();
@@ -5129,25 +5013,25 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r
     float summs = 0.0f;
 
     // Main loop
-    for (int i = 0; i < nb; i++) {
-        const __m256 dx = _mm256_set1_ps(GGML_FP16_TO_FP32(x[i].d));
+    for (; ib < nb; ++ib) {
+        const __m256 dx = _mm256_set1_ps(GGML_FP16_TO_FP32(x[ib].d));
 
-        summs += GGML_FP16_TO_FP32(x[i].m) * GGML_FP16_TO_FP32(y[i].s);
+        summs += GGML_FP16_TO_FP32(x[ib].m) * GGML_FP16_TO_FP32(y[ib].s);
 
-        __m256i qx = bytes_from_nibbles_32(x[i].qs);
-        __m256i bxhi = bytes_from_bits_32(x[i].qh);
+        __m256i qx = bytes_from_nibbles_32(x[ib].qs);
+        __m256i bxhi = bytes_from_bits_32(x[ib].qh);
         bxhi = _mm256_and_si256(bxhi, _mm256_set1_epi8(0x10));
         qx = _mm256_or_si256(qx, bxhi);
 
-        const __m256 dy = _mm256_set1_ps(GGML_FP16_TO_FP32(y[i].d));
-        const __m256i qy = _mm256_loadu_si256((const __m256i *)y[i].qs);
+        const __m256 dy = _mm256_set1_ps(GGML_FP16_TO_FP32(y[ib].d));
+        const __m256i qy = _mm256_loadu_si256((const __m256i *)y[ib].qs);
 
         const __m256 q = mul_sum_us8_pairs_float(qx, qy);
 
         acc = _mm256_fmadd_ps(q, _mm256_mul_ps(dx, dy), acc);
     }
 
-    *s = hsum_float_8(acc) + summs;
+    sumf = hsum_float_8(acc) + summs;
 #elif defined(__AVX__)
     // Initialize accumulator with zeros
     __m256 acc = _mm256_setzero_ps();
@@ -5156,13 +5040,13 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r
     float summs = 0.0f;
 
     // Main loop
-    for (int i = 0; i < nb; i++) {
-        const __m256 dx = _mm256_set1_ps(GGML_FP16_TO_FP32(x[i].d));
+    for (; ib < nb; ++ib) {
+        const __m256 dx = _mm256_set1_ps(GGML_FP16_TO_FP32(x[ib].d));
 
-        summs += GGML_FP16_TO_FP32(x[i].m) * GGML_FP16_TO_FP32(y[i].s);
+        summs += GGML_FP16_TO_FP32(x[ib].m) * GGML_FP16_TO_FP32(y[ib].s);
 
-        __m256i bx_0 = bytes_from_nibbles_32(x[i].qs);
-        const __m256i bxhi = bytes_from_bits_32(x[i].qh);
+        __m256i bx_0 = bytes_from_nibbles_32(x[ib].qs);
+        const __m256i bxhi = bytes_from_bits_32(x[ib].qh);
         __m128i bxhil = _mm256_castsi256_si128(bxhi);
         __m128i bxhih = _mm256_extractf128_si256(bxhi, 1);
         bxhil = _mm_and_si128(bxhil, mask);
@@ -5173,18 +5057,16 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r
         bxh = _mm_or_si128(bxh, bxhih);
         bx_0 = MM256_SET_M128I(bxh, bxl);
 
-        const __m256 dy = _mm256_set1_ps(GGML_FP16_TO_FP32(y[i].d));
-        const __m256i by_0 = _mm256_loadu_si256((const __m256i *)y[i].qs);
+        const __m256 dy = _mm256_set1_ps(GGML_FP16_TO_FP32(y[ib].d));
+        const __m256i by_0 = _mm256_loadu_si256((const __m256i *)y[ib].qs);
 
         const __m256 q = mul_sum_us8_pairs_float(bx_0, by_0);
 
         acc = _mm256_add_ps(_mm256_mul_ps(q, _mm256_mul_ps(dx, dy)), acc);
     }
 
-    *s = hsum_float_8(acc) + summs;
+    sumf = hsum_float_8(acc) + summs;
 #elif defined(__riscv_v_intrinsic)
-    float sumf = 0.0;
-
     uint32_t qh;
 
     size_t vl = __riscv_vsetvl_e8m1(qk/2);
@@ -5193,8 +5075,8 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r
     vuint32m2_t vt_1 = __riscv_vid_v_u32m2(vl);
     vuint32m2_t vt_2 = __riscv_vadd_vx_u32m2(vt_1, 12, vl);
 
-    for (int i = 0; i < nb; i++) {
-        memcpy(&qh, x[i].qh, sizeof(uint32_t));
+    for (; ib < nb; ++ib) {
+        memcpy(&qh, x[ib].qh, sizeof(uint32_t));
 
         // load qh
         vuint32m2_t vqh = __riscv_vmv_v_x_u32m2(qh, vl);
@@ -5216,10 +5098,10 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r
         vuint8mf2_t xh_1 = __riscv_vncvt_x_x_w_u8mf2(xhc_1, vl);
 
         // load
-        vuint8mf2_t tx = __riscv_vle8_v_u8mf2(x[i].qs, vl);
+        vuint8mf2_t tx = __riscv_vle8_v_u8mf2(x[ib].qs, vl);
 
-        vint8mf2_t y0 = __riscv_vle8_v_i8mf2(y[i].qs, vl);
-        vint8mf2_t y1 = __riscv_vle8_v_i8mf2(y[i].qs+16, vl);
+        vint8mf2_t y0 = __riscv_vle8_v_i8mf2(y[ib].qs, vl);
+        vint8mf2_t y1 = __riscv_vle8_v_i8mf2(y[ib].qs+16, vl);
 
         vuint8mf2_t x_at = __riscv_vand_vx_u8mf2(tx, 0x0F, vl);
         vuint8mf2_t x_lt = __riscv_vsrl_vx_u8mf2(tx, 0x04, vl);
@@ -5240,50 +5122,47 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r
 
         int sumi = __riscv_vmv_x_s_i32m1_i32(vs2);
 
-        sumf += (GGML_FP16_TO_FP32(x[i].d)*GGML_FP16_TO_FP32(y[i].d))*sumi + GGML_FP16_TO_FP32(x[i].m)*GGML_FP16_TO_FP32(y[i].s);
+        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
     }
 
-    *s = sumf;
-
 #elif defined(__POWER9_VECTOR__)
     const vector signed char lowMask = vec_splats((signed char)0xF);
+    const vector signed int v0 = vec_splats((int32_t)0);
     const vector unsigned char v4 = vec_splats((unsigned char)0x4);
 
     vector float vsumf0 = vec_splats(0.0f);
 
 #pragma GCC unroll 4
-    for (int i = 0; i < nb; ++i) {
-        __builtin_prefetch(x[i].qs, 0, 1);
-        __builtin_prefetch(y[i].qs, 0, 1);
+    for (; ib < nb; ++ib) {
+        __builtin_prefetch(x[ib].qs, 0, 1);
+        __builtin_prefetch(y[ib].qs, 0, 1);
 
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
-        vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[i].d));
+        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ib].d));
+        vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[ib].d));
         vector float vd = vec_mul(vxd, vyd);
 
-        vector float vxmin = vec_splats(GGML_FP16_TO_FP32(x[i].m));
-        vector float vys = {GGML_FP16_TO_FP32(y[i].s), 0.f, 0.f, 0.f};
+        vector float vxmin = vec_splats(GGML_FP16_TO_FP32(x[ib].m));
+        vector float vys = {GGML_FP16_TO_FP32(y[ib].s), 0.f, 0.f, 0.f};
         vsumf0 = vec_madd(vxmin, vys, vsumf0);
 
-        vector unsigned long long aux64x2_0 = {(uint64_t)(table_b2b_0[x[i].qh[0]]), (uint64_t)(table_b2b_0[x[i].qh[1]])};
-        vector unsigned long long aux64x2_1 = {(uint64_t)(table_b2b_0[x[i].qh[2]]), (uint64_t)(table_b2b_0[x[i].qh[3]])};
+        vector unsigned long long aux64x2_0 = {(uint64_t)(table_b2b_0[x[ib].qh[0]]), (uint64_t)(table_b2b_0[x[ib].qh[1]])};
+        vector unsigned long long aux64x2_1 = {(uint64_t)(table_b2b_0[x[ib].qh[2]]), (uint64_t)(table_b2b_0[x[ib].qh[3]])};
 
         vector signed char qh0 = (vector signed char)aux64x2_0;
         vector signed char qh1 = (vector signed char)aux64x2_1;
 
-        vector signed char qxs = (vector signed char)vec_xl( 0, x[i].qs);
-
-        vector signed char q5x0 = vec_or(vec_and(qxs, lowMask), qh0);
-        vector signed char q5x1 = vec_or(vec_sr(qxs, v4), qh1);
+        vector signed char qxs = (vector signed char)vec_xl( 0, x[ib].qs);
 
-        vector signed char q8y0 = vec_xl(  0, y[i].qs);
-        vector signed char q8y1 = vec_xl( 16, y[i].qs);
+        vector unsigned char q5x0 = (vector unsigned char)vec_or(vec_and(qxs, lowMask), qh0);
+        vector unsigned char q5x1 = (vector unsigned char)vec_or(vec_sr(qxs, v4), qh1);
 
-        vector signed short qv0 = vec_add(vec_mule(q5x0, q8y0), vec_mulo(q5x0, q8y0));
-        vector signed short qv1 = vec_add(vec_mule(q5x1, q8y1), vec_mulo(q5x1, q8y1));
+        vector signed char q8y0 = vec_xl(  0, y[ib].qs);
+        vector signed char q8y1 = vec_xl( 16, y[ib].qs);
 
-        qv0 = vec_add(qv0, qv1);
+        vector signed int vsumi0 = v0;
 
-        vector signed int vsumi0 = vec_add(vec_unpackh(qv0), vec_unpackl(qv0));
+        vsumi0 = vec_msum(q8y0, q5x0, vsumi0);
+        vsumi0 = vec_msum(q8y1, q5x1, vsumi0);
 
         vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
     }
@@ -5291,7 +5170,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r
     vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
     vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
 
-    *s = vec_extract(vsumf0, 0);
+    sumf = vec_extract(vsumf0, 0);
 
 #elif defined(__loongarch_asx)
     // Initialize accumulator with zeros
@@ -5300,51 +5179,49 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r
     float summs = 0.0f;
 
     // Main loop
-    for (int i = 0; i < nb; i++) {
-        const __m256 dx = __lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(x[i].d));
+    for (; ib < nb; ++ib) {
+        const __m256 dx = __lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(x[ib].d));
 
-        summs += GGML_FP16_TO_FP32(x[i].m) * GGML_FP16_TO_FP32(y[i].s);
+        summs += GGML_FP16_TO_FP32(x[ib].m) * GGML_FP16_TO_FP32(y[ib].s);
 
-        __m256i qx = bytes_from_nibbles_32(x[i].qs);
-        __m256i bxhi = bytes_from_bits_32(x[i].qh);
+        __m256i qx = bytes_from_nibbles_32(x[ib].qs);
+        __m256i bxhi = bytes_from_bits_32(x[ib].qh);
         bxhi = __lasx_xvand_v(bxhi, __lasx_xvreplgr2vr_b(0x10));
         qx = __lasx_xvor_v(qx, bxhi);
 
-        const __m256 dy = __lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(y[i].d));
-        const __m256i qy = __lasx_xvld((const __m256i *)y[i].qs, 0);
+        const __m256 dy = __lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(y[ib].d));
+        const __m256i qy = __lasx_xvld((const __m256i *)y[ib].qs, 0);
 
         const __m256 q = mul_sum_us8_pairs_float(qx, qy);
 
         acc = __lasx_xvfmadd_s(q, __lasx_xvfmul_s(dx, dy), acc);
     }
 
-    *s = hsum_float_8(acc) + summs;
-
-#else
-    // scalar
-    float sumf = 0.0;
-
-    for (int i = 0; i < nb; i++) {
+    sumf = hsum_float_8(acc) + summs;
+#endif
+    for (; ib < nb; ++ib) {
         uint32_t qh;
-        memcpy(&qh, x[i].qh, sizeof(qh));
+        memcpy(&qh, x[ib].qh, sizeof(qh));
 
-        int sumi = 0;
+        int sumi0 = 0;
+        int sumi1 = 0;
 
         for (int j = 0; j < qk/2; ++j) {
             const uint8_t xh_0 = ((qh >> (j +  0)) << 4) & 0x10;
             const uint8_t xh_1 = ((qh >> (j + 12))     ) & 0x10;
 
-            const int32_t x0 = (x[i].qs[j] & 0xF) | xh_0;
-            const int32_t x1 = (x[i].qs[j] >>  4) | xh_1;
+            const int32_t x0 = (x[ib].qs[j] & 0xF) | xh_0;
+            const int32_t x1 = (x[ib].qs[j] >>  4) | xh_1;
 
-            sumi += (x0 * y[i].qs[j]) + (x1 * y[i].qs[j + qk/2]);
+            sumi0 += (x0 * y[ib].qs[j]);
+            sumi1 += (x1 * y[ib].qs[j + qk/2]);
         }
 
-        sumf += (GGML_FP16_TO_FP32(x[i].d)*GGML_FP16_TO_FP32(y[i].d))*sumi + GGML_FP16_TO_FP32(x[i].m)*GGML_FP16_TO_FP32(y[i].s);
+        int sumi = sumi0 + sumi1;
+        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
     }
 
     *s = sumf;
-#endif
 }
 
 void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
@@ -5421,42 +5298,44 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, size_t bs, const void * r
         return;
     }
 #endif
+
+    int ib = 0;
+    float sumf = 0;
+
 #if defined(__ARM_FEATURE_SVE)
-    svfloat32_t sumv0 = svdup_n_f32(0.0f);
-    svfloat32_t sumv1 = svdup_n_f32(0.0f);
+    if (svcntb() == QK8_0) {
+        svfloat32_t sumv0 = svdup_n_f32(0.0f);
+        svfloat32_t sumv1 = svdup_n_f32(0.0f);
 
-    assert(nb % 2 == 0); // TODO: handle odd nb
+        for (; ib + 1 < nb; ib += 2) {
+            const block_q8_0 * restrict x0 = &x[ib + 0];
+            const block_q8_0 * restrict x1 = &x[ib + 1];
+            const block_q8_0 * restrict y0 = &y[ib + 0];
+            const block_q8_0 * restrict y1 = &y[ib + 1];
 
-    for (int i = 0; i < nb; i += 2) {
-        const block_q8_0 * restrict x0 = &x[i + 0];
-        const block_q8_0 * restrict x1 = &x[i + 1];
-        const block_q8_0 * restrict y0 = &y[i + 0];
-        const block_q8_0 * restrict y1 = &y[i + 1];
+            // load x
+            const svint8_t qx0 = svld1_s8(svptrue_b8(), x0->qs);
+            const svint8_t qx1 = svld1_s8(svptrue_b8(), x1->qs);
 
-        // load x
-        const svint8_t qx0 = svld1_s8(svptrue_b8(), x0->qs);
-        const svint8_t qx1 = svld1_s8(svptrue_b8(), x1->qs);
+            // load y
+            const svint8_t qy0 = svld1_s8(svptrue_b8(), y0->qs);
+            const svint8_t qy1 = svld1_s8(svptrue_b8(), y1->qs);
 
-        // load y
-        const svint8_t qy0 = svld1_s8(svptrue_b8(), y0->qs);
-        const svint8_t qy1 = svld1_s8(svptrue_b8(), y1->qs);
+            sumv0 = svmla_n_f32_x(svptrue_b32(), sumv0, svcvt_f32_s32_x(svptrue_b32(), svdot_s32(svdup_n_s32(0), qx0, qy0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
+            sumv1 = svmla_n_f32_x(svptrue_b32(), sumv1, svcvt_f32_s32_x(svptrue_b32(), svdot_s32(svdup_n_s32(0), qx1, qy1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
+        }
 
-        sumv0 = svmla_n_f32_x(svptrue_b32(), sumv0, svcvt_f32_s32_x(svptrue_b32(), svdot_s32(svdup_n_s32(0), qx0, qy0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
-        sumv1 = svmla_n_f32_x(svptrue_b32(), sumv1, svcvt_f32_s32_x(svptrue_b32(), svdot_s32(svdup_n_s32(0), qx1, qy1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
+        sumf = svaddv_f32(svptrue_b32(), svadd_f32_x(svptrue_b32(), sumv0, sumv1));
     }
-
-    *s = svaddv_f32(svptrue_b32(), svadd_f32_x(svptrue_b32(), sumv0, sumv1));
 #elif defined(__ARM_NEON)
     float32x4_t sumv0 = vdupq_n_f32(0.0f);
     float32x4_t sumv1 = vdupq_n_f32(0.0f);
 
-    assert(nb % 2 == 0); // TODO: handle odd nb
-
-    for (int i = 0; i < nb; i += 2) {
-        const block_q8_0 * restrict x0 = &x[i + 0];
-        const block_q8_0 * restrict x1 = &x[i + 1];
-        const block_q8_0 * restrict y0 = &y[i + 0];
-        const block_q8_0 * restrict y1 = &y[i + 1];
+    for (; ib + 1 < nb; ib += 2) {
+        const block_q8_0 * restrict x0 = &x[ib + 0];
+        const block_q8_0 * restrict x1 = &x[ib + 1];
+        const block_q8_0 * restrict y0 = &y[ib + 0];
+        const block_q8_0 * restrict y1 = &y[ib + 1];
 
         const int8x16_t x0_0 = vld1q_s8(x0->qs);
         const int8x16_t x0_1 = vld1q_s8(x0->qs + 16);
@@ -5478,17 +5357,17 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, size_t bs, const void * r
                         ggml_vdotq_s32(vdupq_n_s32(0), x1_1, y1_1))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
     }
 
-    *s = vaddvq_f32(sumv0) + vaddvq_f32(sumv1);
+    sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1);
 #elif defined(__AVX2__) || defined(__AVX__)
     // Initialize accumulator with zeros
     __m256 acc = _mm256_setzero_ps();
 
     // Main loop
-    for (int i = 0; i < nb; ++i) {
+    for (; ib < nb; ++ib) {
         // Compute combined scale for the block
-        const __m256 d = _mm256_set1_ps(GGML_FP16_TO_FP32(x[i].d) * GGML_FP16_TO_FP32(y[i].d));
-        __m256i qx = _mm256_loadu_si256((const __m256i *)x[i].qs);
-        __m256i qy = _mm256_loadu_si256((const __m256i *)y[i].qs);
+        const __m256 d = _mm256_set1_ps(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d));
+        __m256i qx = _mm256_loadu_si256((const __m256i *)x[ib].qs);
+        __m256i qy = _mm256_loadu_si256((const __m256i *)y[ib].qs);
 
         const __m256 q = mul_sum_i8_pairs_float(qx, qy);
 
@@ -5500,15 +5379,14 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, size_t bs, const void * r
 #endif
     }
 
-    *s = hsum_float_8(acc);
+    sumf = hsum_float_8(acc);
 #elif defined(__riscv_v_intrinsic)
-    float sumf = 0.0;
     size_t vl = __riscv_vsetvl_e8m1(qk);
 
-    for (int i = 0; i < nb; i++) {
+    for (; ib < nb; ++ib) {
         // load elements
-        vint8m1_t bx_0 = __riscv_vle8_v_i8m1(x[i].qs, vl);
-        vint8m1_t by_0 = __riscv_vle8_v_i8m1(y[i].qs, vl);
+        vint8m1_t bx_0 = __riscv_vle8_v_i8m1(x[ib].qs, vl);
+        vint8m1_t by_0 = __riscv_vle8_v_i8m1(y[ib].qs, vl);
 
         vint16m2_t vw_mul = __riscv_vwmul_vv_i16m2(bx_0, by_0, vl);
 
@@ -5517,40 +5395,38 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, size_t bs, const void * r
 
         int sumi = __riscv_vmv_x_s_i32m1_i32(v_sum);
 
-        sumf += sumi*(GGML_FP16_TO_FP32(x[i].d)*GGML_FP16_TO_FP32(y[i].d));
+        sumf += sumi*(GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d));
     }
-
-    *s = sumf;
-
 #elif defined(__POWER9_VECTOR__)
+    const vector signed int v0 = vec_splats((int32_t)0);
     vector float vsumf0 = vec_splats(0.0f);
 
-#pragma GCC unroll 4
-    for (int i = 0; i < nb; i++) {
-        __builtin_prefetch(x[i].qs, 0, 1);
-        __builtin_prefetch(y[i].qs, 0, 1);
+#pragma GCC unroll 8
+    for (; ib < nb; ++ib) {
+        __builtin_prefetch(x[ib].qs, 0, 1);
+        __builtin_prefetch(y[ib].qs, 0, 1);
 
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
-        vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[i].d));
+        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ib].d));
+        vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[ib].d));
         vector float vd = vec_mul(vxd, vyd);
 
-        vector signed char q8x0 = vec_xl( 0, x[i].qs);
-        vector signed char q8x1 = vec_xl(16, x[i].qs);
-        vector signed char q8y0 = vec_xl( 0, y[i].qs);
-        vector signed char q8y1 = vec_xl(16, y[i].qs);
+        vector signed char q8x0 = vec_xl( 0, x[ib].qs);
+        vector signed char q8x1 = vec_xl(16, x[ib].qs);
+        vector signed char q8y0 = vec_xl( 0, y[ib].qs);
+        vector signed char q8y1 = vec_xl(16, y[ib].qs);
 
         vector signed short qv0 = vec_mule(q8x0, q8y0);
         vector signed short qv1 = vec_mulo(q8x0, q8y0);
         vector signed short qv2 = vec_mule(q8x1, q8y1);
         vector signed short qv3 = vec_mulo(q8x1, q8y1);
 
-        vector signed int vsumi0 = vec_add(vec_unpackh(qv0), vec_unpackh(qv1));
-        vector signed int vsumi1 = vec_add(vec_unpackl(qv0), vec_unpackl(qv1));
-        vector signed int vsumi2 = vec_add(vec_unpackh(qv2), vec_unpackh(qv3));
-        vector signed int vsumi3 = vec_add(vec_unpackl(qv2), vec_unpackl(qv3));
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
 
-        vsumi0 = vec_add(vsumi0, vsumi2);
-        vsumi1 = vec_add(vsumi1, vsumi3);
+        vsumi0 = vec_sum4s(qv0, vsumi0);
+        vsumi1 = vec_sum4s(qv1, vsumi1);
+        vsumi0 = vec_sum4s(qv2, vsumi0);
+        vsumi1 = vec_sum4s(qv3, vsumi1);
 
         vsumi0 = vec_add(vsumi0, vsumi1);
 
@@ -5560,18 +5436,18 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, size_t bs, const void * r
     vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
     vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
 
-    *s = vec_extract(vsumf0, 0);
+    sumf = vec_extract(vsumf0, 0);
 
 #elif defined(__loongarch_asx)
     // Initialize accumulator with zeros
     __m256 acc = (__m256)__lasx_xvldi(0);
 
     // Main loop
-    for (int i = 0; i < nb; ++i) {
+    for (; ib < nb; ++ib) {
         // Compute combined scale for the block
-        const __m256 d = __lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(x[i].d) * GGML_FP16_TO_FP32(y[i].d));
-        __m256i qx = __lasx_xvld((const __m256i *)x[i].qs, 0);
-        __m256i qy = __lasx_xvld((const __m256i *)y[i].qs, 0);
+        const __m256 d = __lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d));
+        __m256i qx = __lasx_xvld((const __m256i *)x[ib].qs, 0);
+        __m256i qy = __lasx_xvld((const __m256i *)y[ib].qs, 0);
 
         const __m256 q = mul_sum_i8_pairs_float(qx, qy);
 
@@ -5579,24 +5455,19 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, size_t bs, const void * r
         acc = __lasx_xvfmadd_s( d, q, acc );
     }
 
-    *s = hsum_float_8(acc);
-
-#else
-    // scalar
-    float sumf = 0.0;
-
-    for (int i = 0; i < nb; i++) {
+    sumf = hsum_float_8(acc);
+#endif
+    for (; ib < nb; ++ib) {
         int sumi = 0;
 
         for (int j = 0; j < qk; j++) {
-            sumi += x[i].qs[j]*y[i].qs[j];
+            sumi += x[ib].qs[j]*y[ib].qs[j];
         }
 
-        sumf += sumi*(GGML_FP16_TO_FP32(x[i].d)*GGML_FP16_TO_FP32(y[i].d));
+        sumf += sumi*(GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d));
     }
 
     *s = sumf;
-#endif
 }
 
 void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
@@ -5938,6 +5809,7 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, size_t bs, const void * r
 #elif defined(__POWER9_VECTOR__)
     const vector signed char lowMask = vec_splats((signed char)0x3);
     const vector signed char lowScaleMask = vec_splats((signed char)0xF);
+    const vector int v0 = vec_splats((int32_t)0);
     const vector unsigned char v2 = vec_splats((unsigned char)0x2);
     const vector unsigned char v6 = vec_splats((unsigned char)0x6);
     const vector unsigned char v4 = vec_splats((unsigned char)0x4);
@@ -5975,15 +5847,17 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, size_t bs, const void * r
         vsumf2 = vec_nmsub(vec_ctf(prod2, 0), vdmin, vsumf2);
         vsumf3 = vec_nmsub(vec_ctf(prod3, 0), vdmin, vsumf3);
 
-        vector signed int vsumi0 = vec_splats((int32_t)0);
-        vector signed int vsumi1 = vec_splats((int32_t)0);
-        vector signed int vsumi2 = vec_splats((int32_t)0);
-        vector signed int vsumi3 = vec_splats((int32_t)0);
-        vector signed int vsumi4 = vec_splats((int32_t)0);
-        vector signed int vsumi5 = vec_splats((int32_t)0);
-        vector signed int vsumi6 = vec_splats((int32_t)0);
-        vector signed int vsumi7 = vec_splats((int32_t)0);
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
+        vector signed int vsumi2 = v0;
+        vector signed int vsumi3 = v0;
+        vector signed int vsumi4 = v0;
+        vector signed int vsumi5 = v0;
+        vector signed int vsumi6 = v0;
+        vector signed int vsumi7 = v0;
 
+        const uint8_t * restrict q2 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
 
         for (int j = 0; j < QK_K/128; ++j) {
             __builtin_prefetch(q2, 0, 1);
@@ -5993,14 +5867,14 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, size_t bs, const void * r
             vector signed char qxs1 = (vector signed char)vec_xl(16, q2);
             q2 += 32;
 
-            vector signed char q2x00 = vec_and(qxs0, lowMask);
-            vector signed char q2x01 = vec_and(vec_sr(qxs0, v2), lowMask);
-            vector signed char q2x02 = vec_and(vec_sr(qxs0, v4), lowMask);
-            vector signed char q2x03 = vec_and(vec_sr(qxs0, v6), lowMask);
-            vector signed char q2x10 = vec_and(qxs1, lowMask);
-            vector signed char q2x11 = vec_and(vec_sr(qxs1, v2), lowMask);
-            vector signed char q2x12 = vec_and(vec_sr(qxs1, v4), lowMask);
-            vector signed char q2x13 = vec_and(vec_sr(qxs1, v6), lowMask);
+            vector unsigned char q2x00 = (vector unsigned char)vec_and(qxs0, lowMask);
+            vector unsigned char q2x01 = (vector unsigned char)vec_and(vec_sr(qxs0, v2), lowMask);
+            vector unsigned char q2x02 = (vector unsigned char)vec_and(vec_sr(qxs0, v4), lowMask);
+            vector unsigned char q2x03 = (vector unsigned char)vec_and(vec_sr(qxs0, v6), lowMask);
+            vector unsigned char q2x10 = (vector unsigned char)vec_and(qxs1, lowMask);
+            vector unsigned char q2x11 = (vector unsigned char)vec_and(vec_sr(qxs1, v2), lowMask);
+            vector unsigned char q2x12 = (vector unsigned char)vec_and(vec_sr(qxs1, v4), lowMask);
+            vector unsigned char q2x13 = (vector unsigned char)vec_and(vec_sr(qxs1, v6), lowMask);
 
             vector signed char q8y00 = vec_xl(  0, q8);
             vector signed char q8y10 = vec_xl( 16, q8);
@@ -6012,45 +5886,36 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, size_t bs, const void * r
             vector signed char q8y13 = vec_xl(112, q8);
             q8 += 128;
 
-            vector signed short qv0 = vec_add(vec_mule(q2x00, q8y00), vec_mulo(q2x00, q8y00));
-            vector signed short qv1 = vec_add(vec_mule(q2x01, q8y01), vec_mulo(q2x01, q8y01));
-            vector signed short qv2 = vec_add(vec_mule(q2x02, q8y02), vec_mulo(q2x02, q8y02));
-            vector signed short qv3 = vec_add(vec_mule(q2x03, q8y03), vec_mulo(q2x03, q8y03));
-            vector signed short qv4 = vec_add(vec_mule(q2x10, q8y10), vec_mulo(q2x10, q8y10));
-            vector signed short qv5 = vec_add(vec_mule(q2x11, q8y11), vec_mulo(q2x11, q8y11));
-            vector signed short qv6 = vec_add(vec_mule(q2x12, q8y12), vec_mulo(q2x12, q8y12));
-            vector signed short qv7 = vec_add(vec_mule(q2x13, q8y13), vec_mulo(q2x13, q8y13));
-
-            vector signed short vscales_h = vec_unpackh(vscales);
-            vector signed short vs0 = vec_splat(vscales_h, 0);
-            vector signed short vs1 = vec_splat(vscales_h, 1);
-            vector signed short vs2 = vec_splat(vscales_h, 2);
-            vector signed short vs3 = vec_splat(vscales_h, 3);
-            vector signed short vs4 = vec_splat(vscales_h, 4);
-            vector signed short vs5 = vec_splat(vscales_h, 5);
-            vector signed short vs6 = vec_splat(vscales_h, 6);
-            vector signed short vs7 = vec_splat(vscales_h, 7);
+            vector signed int qv0 = vec_msum(q8y00, q2x00, v0);
+            vector signed int qv1 = vec_msum(q8y01, q2x01, v0);
+            vector signed int qv2 = vec_msum(q8y02, q2x02, v0);
+            vector signed int qv3 = vec_msum(q8y03, q2x03, v0);
+            vector signed int qv4 = vec_msum(q8y10, q2x10, v0);
+            vector signed int qv5 = vec_msum(q8y11, q2x11, v0);
+            vector signed int qv6 = vec_msum(q8y12, q2x12, v0);
+            vector signed int qv7 = vec_msum(q8y13, q2x13, v0);
+
+            vector signed short vscales_07 = vec_unpackh(vscales);
+            vector signed int vscales_03 = vec_unpackh(vscales_07);
+            vector signed int vscales_47 = vec_unpackl(vscales_07);
+            vector signed int vs0 = vec_splat(vscales_03, 0);
+            vector signed int vs1 = vec_splat(vscales_03, 1);
+            vector signed int vs2 = vec_splat(vscales_03, 2);
+            vector signed int vs3 = vec_splat(vscales_03, 3);
+            vector signed int vs4 = vec_splat(vscales_47, 0);
+            vector signed int vs5 = vec_splat(vscales_47, 1);
+            vector signed int vs6 = vec_splat(vscales_47, 2);
+            vector signed int vs7 = vec_splat(vscales_47, 3);
             vscales = vec_sld(vscales, vscales, 8);
 
-            qv0 = vec_mul(qv0, vs0);
-            qv1 = vec_mul(qv1, vs2);
-            qv2 = vec_mul(qv2, vs4);
-            qv3 = vec_mul(qv3, vs6);
-
-            qv0 = vec_madd(qv4, vs1, qv0);
-            qv1 = vec_madd(qv5, vs3, qv1);
-            qv2 = vec_madd(qv6, vs5, qv2);
-            qv3 = vec_madd(qv7, vs7, qv3);
-
-            vsumi0 = vec_add(vec_unpackh(qv0), vsumi0);
-            vsumi1 = vec_add(vec_unpackh(qv1), vsumi1);
-            vsumi2 = vec_add(vec_unpackh(qv2), vsumi2);
-            vsumi3 = vec_add(vec_unpackh(qv3), vsumi3);
-
-            vsumi4 = vec_add(vec_unpackl(qv0), vsumi4);
-            vsumi5 = vec_add(vec_unpackl(qv1), vsumi5);
-            vsumi6 = vec_add(vec_unpackl(qv2), vsumi6);
-            vsumi7 = vec_add(vec_unpackl(qv3), vsumi7);
+            vsumi0 = vec_add(vec_mul(qv0, vs0), vsumi0);
+            vsumi1 = vec_add(vec_mul(qv1, vs2), vsumi1);
+            vsumi2 = vec_add(vec_mul(qv2, vs4), vsumi2);
+            vsumi3 = vec_add(vec_mul(qv3, vs6), vsumi3);
+            vsumi4 = vec_add(vec_mul(qv4, vs1), vsumi4);
+            vsumi5 = vec_add(vec_mul(qv5, vs3), vsumi5);
+            vsumi6 = vec_add(vec_mul(qv6, vs5), vsumi6);
+            vsumi7 = vec_add(vec_mul(qv7, vs7), vsumi7);
         }
 
         vsumi0 = vec_add(vsumi0, vsumi4);
@@ -6088,6 +5953,7 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, size_t bs, const void * r
 
         const uint8_t * restrict q2 = x[i].qs;
         const int8_t  * restrict q8 = y[i].qs;
+
         const __m128i mins_and_scales = __lsx_vld((const __m128i*)x[i].scales, 0);
         const __m128i scales8 = __lsx_vand_v(mins_and_scales, m4);
         const __m128i mins8 = __lsx_vand_v(__lsx_vsrli_h(mins_and_scales, 4), m4);
@@ -6640,6 +6506,9 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * r
 
 #elif defined(__POWER9_VECTOR__)
     const vector signed char lowMask = vec_splats((signed char)0x3);
+    const vector signed char lowMask1 = vec_splats((int8_t)0xf);
+    const vector signed char lowMask2 = vec_splats((int8_t)0x30);
+    const vector int v0 = vec_splats((int32_t)0);
     const vector signed char v1 = vec_splats((signed char)0x1);
     const vector unsigned char v2 = vec_splats((unsigned char)0x2);
     const vector unsigned char v3 = vec_splats((unsigned char)0x3);
@@ -6657,30 +6526,33 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * r
         vector float vyd = vec_splats(y[i].d);
         vector float vd = vec_mul(vxd, vyd);
 
-        uint32_t aux[3];
-        uint32_t utmp[4];
+        UNUSED(kmask1);
+        UNUSED(kmask2);
 
-        memcpy(aux, x[i].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);
+        vector signed char u0 = (vector signed char)vec_xl_len(x[i].scales, 8);
+        vector signed char u1 = vec_and(u0, lowMask1);
+        vector signed char u2 = (vector signed char)vec_xl_len(x[i].scales + 8, 4);
+        vector signed char u3 = (vector signed char)vec_mergeh((vector signed int)u2, (vector signed int)vec_sr(u2, v2));
+        vector signed char u30 = vec_sl(vec_and(u3, lowMask), v4);
+        vector signed char u31 = vec_and(u3, lowMask2);
+
+        u1 = vec_or(u1, u30);
+        u2 = vec_or(vec_sr(u0, v4), u31);
 
-        vector signed char vscales = (vector signed char)vec_xl( 0, utmp);
+        vector signed char vscales = (vector signed char)vec_mergeh((vector signed long long)u1, (vector signed long long)u2);
         vector signed char qxhs0 = (vector signed char)vec_xl( 0, x[i].hmask);
         vector signed char qxhs1 = (vector signed char)vec_xl(16, x[i].hmask);
 
         vscales = vec_sub(vscales, off);
 
-        vector signed int vsumi0 = vec_splats((int32_t)0);
-        vector signed int vsumi1 = vec_splats((int32_t)0);
-        vector signed int vsumi2 = vec_splats((int32_t)0);
-        vector signed int vsumi3 = vec_splats((int32_t)0);
-        vector signed int vsumi4 = vec_splats((int32_t)0);
-        vector signed int vsumi5 = vec_splats((int32_t)0);
-        vector signed int vsumi6 = vec_splats((int32_t)0);
-        vector signed int vsumi7 = vec_splats((int32_t)0);
-
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
+        vector signed int vsumi2 = v0;
+        vector signed int vsumi3 = v0;
+        vector signed int vsumi4 = v0;
+        vector signed int vsumi5 = v0;
+        vector signed int vsumi6 = v0;
+        vector signed int vsumi7 = v0;
 
         const uint8_t * restrict q3 = x[i].qs;
         const int8_t  * restrict q8 = y[i].qs;
@@ -6754,23 +6626,14 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * r
             vector signed short qv12 = vec_add(vec_mule(q3x12, q8y12), vec_mulo(q3x12, q8y12));
             vector signed short qv13 = vec_add(vec_mule(q3x13, q8y13), vec_mulo(q3x13, q8y13));
 
-            vector signed int vsum0 = vec_add(vec_mule(qv00, vs0), vec_mulo(qv00, vs0));
-            vector signed int vsum1 = vec_add(vec_mule(qv01, vs2), vec_mulo(qv01, vs2));
-            vector signed int vsum2 = vec_add(vec_mule(qv02, vs4), vec_mulo(qv02, vs4));
-            vector signed int vsum3 = vec_add(vec_mule(qv03, vs6), vec_mulo(qv03, vs6));
-            vector signed int vsum4 = vec_add(vec_mule(qv10, vs1), vec_mulo(qv10, vs1));
-            vector signed int vsum5 = vec_add(vec_mule(qv11, vs3), vec_mulo(qv11, vs3));
-            vector signed int vsum6 = vec_add(vec_mule(qv12, vs5), vec_mulo(qv12, vs5));
-            vector signed int vsum7 = vec_add(vec_mule(qv13, vs7), vec_mulo(qv13, vs7));
-
-            vsumi0 = vec_add(vsum0, vsumi0);
-            vsumi1 = vec_add(vsum1, vsumi1);
-            vsumi2 = vec_add(vsum2, vsumi2);
-            vsumi3 = vec_add(vsum3, vsumi3);
-            vsumi4 = vec_add(vsum4, vsumi4);
-            vsumi5 = vec_add(vsum5, vsumi5);
-            vsumi6 = vec_add(vsum6, vsumi6);
-            vsumi7 = vec_add(vsum7, vsumi7);
+            vsumi0 = vec_msum(qv00, vs0, vsumi0);
+            vsumi1 = vec_msum(qv01, vs2, vsumi1);
+            vsumi2 = vec_msum(qv02, vs4, vsumi2);
+            vsumi3 = vec_msum(qv03, vs6, vsumi3);
+            vsumi4 = vec_msum(qv10, vs1, vsumi4);
+            vsumi5 = vec_msum(qv11, vs3, vsumi5);
+            vsumi6 = vec_msum(qv12, vs5, vsumi6);
+            vsumi7 = vec_msum(qv13, vs7, vsumi7);
         }
 
         vsumi0 = vec_add(vsumi0, vsumi4);
@@ -6807,6 +6670,8 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * r
     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;
         // Set up scales
         memcpy(aux, x[i].scales, 12);
         __m128i scales128 = lsx_set_w(
@@ -6828,29 +6693,32 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * r
 
         int bit = 0;
         int is  = 0;
+        __m256i xvbit;
 
-        const uint8_t * restrict q3 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
 
         for (int j = 0; j < QK_K/128; ++j) {
             // load low 2 bits
             const __m256i q3bits = __lasx_xvld((const __m256i*)q3, 0); q3 += 32;
 
+            xvbit = __lasx_xvreplgr2vr_h(bit);
             // prepare low and high bits
             const __m256i q3l_0 = __lasx_xvand_v(q3bits, m3);
-            const __m256i q3h_0 = __lasx_xvslli_h(__lasx_xvsrli_h(__lasx_xvandn_v(hbits, __lasx_xvslli_h(mone, bit)), bit), 2);
+            const __m256i q3h_0 = __lasx_xvslli_h(__lasx_xvsrl_h(__lasx_xvandn_v(hbits, __lasx_xvsll_h(mone, xvbit)), xvbit), 2);
             ++bit;
 
+            xvbit = __lasx_xvreplgr2vr_h(bit);
             const __m256i q3l_1 = __lasx_xvand_v(__lasx_xvsrli_h(q3bits, 2), m3);
-            const __m256i q3h_1 = __lasx_xvslli_h(__lasx_xvsrli_h(__lasx_xvandn_v(hbits, __lasx_xvslli_h(mone, bit)), bit), 2);
+            const __m256i q3h_1 = __lasx_xvslli_h(__lasx_xvsrl_h(__lasx_xvandn_v(hbits, __lasx_xvsll_h(mone, xvbit)), xvbit), 2);
             ++bit;
 
+            xvbit = __lasx_xvreplgr2vr_h(bit);
             const __m256i q3l_2 = __lasx_xvand_v(__lasx_xvsrli_h(q3bits, 4), m3);
-            const __m256i q3h_2 = __lasx_xvslli_h(__lasx_xvsrli_h(__lasx_xvandn_v(hbits, __lasx_xvslli_h(mone, bit)), bit), 2);
+            const __m256i q3h_2 = __lasx_xvslli_h(__lasx_xvsrl_h(__lasx_xvandn_v(hbits, __lasx_xvsll_h(mone, xvbit)), xvbit), 2);
             ++bit;
 
+            xvbit = __lasx_xvreplgr2vr_h(bit);
             const __m256i q3l_3 = __lasx_xvand_v(__lasx_xvsrli_h(q3bits, 6), m3);
-            const __m256i q3h_3 = __lasx_xvslli_h(__lasx_xvsrli_h(__lasx_xvandn_v(hbits, __lasx_xvslli_h(mone, bit)), bit), 2);
+            const __m256i q3h_3 = __lasx_xvslli_h(__lasx_xvsrl_h(__lasx_xvandn_v(hbits, __lasx_xvsll_h(mone, xvbit)), xvbit), 2);
             ++bit;
 
             // load Q8 quants
@@ -7264,6 +7132,10 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * r
 
 #elif defined(__POWER9_VECTOR__)
     const vector signed char lowMask = vec_splats((signed char)0xF);
+    const vector signed char lowMask1 = vec_splats((int8_t)0x3f);
+    const vector signed char lowMask2 = vec_splats((int8_t)0x30);
+    const vector int v0 = vec_splats((int32_t)0);
+    const vector unsigned char v2 = vec_splats((uint8_t)2);
     const vector unsigned char v4 = vec_splats((unsigned char)0x4);
 
     vector float vsumf0 = vec_splats(0.0f);
@@ -7282,15 +7154,24 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * r
         vector signed short q8ysums0 = vec_xl( 0, y[i].bsums);
         vector signed short q8ysums1 = vec_xl(16, y[i].bsums);
 
-        memcpy(utmp, x[i].scales, 12);
+        UNUSED(kmask1);
+        UNUSED(kmask2);
+        UNUSED(kmask3);
+        UNUSED(utmp);
 
-        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;
+        vector signed char u0 = (vector signed char)vec_xl_len(x[i].scales, 8);
+        vector signed char u1 = vec_and(vec_sr(u0, v2), lowMask2);
+        vector signed char u2 = (vector signed char)vec_xl_len(x[i].scales + 8, 4);
+        vector signed char u3 = vec_sr(u2, v4);
+
+        vector signed char u30 = u1;
+        vector signed char u31 = (vector signed char)vec_mergeh((vector signed int)vec_and(u2, lowMask), (vector signed int)u3);
+
+        u1 = vec_and(u0, lowMask1);
+        u2 = vec_or(u30, u31);
+
+        vector signed char utmps = (vector signed char)vec_mergeh((vector signed int)u1, (vector signed int)u2);
 
-        vector signed char utmps = (vector signed char)vec_xl( 0, utmp);
         vector signed short vscales = vec_unpackh(utmps);
         vector signed short q4xmins = vec_unpackl(utmps);
         vector signed short q4xmins0 = vec_mergeh(q4xmins, q4xmins);
@@ -7306,14 +7187,10 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * r
         vsumf2 = vec_nmsub(vec_ctf(prod2, 0), vdmin, vsumf2);
         vsumf3 = vec_nmsub(vec_ctf(prod3, 0), vdmin, vsumf3);
 
-        vector signed int vsumi0 = vec_splats((int32_t)0);
-        vector signed int vsumi1 = vec_splats((int32_t)0);
-        vector signed int vsumi2 = vec_splats((int32_t)0);
-        vector signed int vsumi3 = vec_splats((int32_t)0);
-        vector signed int vsumi4 = vec_splats((int32_t)0);
-        vector signed int vsumi5 = vec_splats((int32_t)0);
-        vector signed int vsumi6 = vec_splats((int32_t)0);
-        vector signed int vsumi7 = vec_splats((int32_t)0);
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
+        vector signed int vsumi2 = v0;
+        vector signed int vsumi3 = v0;
 
         const uint8_t * restrict q4 = x[i].qs;
         const int8_t  * restrict q8 = y[i].qs;
@@ -7328,14 +7205,14 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * r
             vector signed char qxs3 = (vector signed char)vec_xl(48, q4);
             q4 += 64;
 
-            vector signed char q4x00 = vec_and(qxs0, lowMask);
-            vector signed char q4x01 = vec_sr(qxs0, v4);
-            vector signed char q4x10 = vec_and(qxs1, lowMask);
-            vector signed char q4x11 = vec_sr(qxs1, v4);
-            vector signed char q4x20 = vec_and(qxs2, lowMask);
-            vector signed char q4x21 = vec_sr(qxs2, v4);
-            vector signed char q4x30 = vec_and(qxs3, lowMask);
-            vector signed char q4x31 = vec_sr(qxs3, v4);
+            vector unsigned char q4x00 = (vector unsigned char)vec_and(qxs0, lowMask);
+            vector unsigned char q4x01 = (vector unsigned char)vec_sr(qxs0, v4);
+            vector unsigned char q4x10 = (vector unsigned char)vec_and(qxs1, lowMask);
+            vector unsigned char q4x11 = (vector unsigned char)vec_sr(qxs1, v4);
+            vector unsigned char q4x20 = (vector unsigned char)vec_and(qxs2, lowMask);
+            vector unsigned char q4x21 = (vector unsigned char)vec_sr(qxs2, v4);
+            vector unsigned char q4x30 = (vector unsigned char)vec_and(qxs3, lowMask);
+            vector unsigned char q4x31 = (vector unsigned char)vec_sr(qxs3, v4);
 
             vector signed char q8y00 = vec_xl(  0, q8);
             vector signed char q8y10 = vec_xl( 16, q8);
@@ -7347,41 +7224,33 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * r
             vector signed char q8y31 = vec_xl(112, q8);
             q8 += 128;
 
-            vector signed short qv00 = vec_add(vec_mule(q4x00, q8y00), vec_mulo(q4x00, q8y00));
-            vector signed short qv01 = vec_add(vec_mule(q4x01, q8y01), vec_mulo(q4x01, q8y01));
-            vector signed short qv10 = vec_add(vec_mule(q4x10, q8y10), vec_mulo(q4x10, q8y10));
-            vector signed short qv11 = vec_add(vec_mule(q4x11, q8y11), vec_mulo(q4x11, q8y11));
-            vector signed short qv20 = vec_add(vec_mule(q4x20, q8y20), vec_mulo(q4x20, q8y20));
-            vector signed short qv21 = vec_add(vec_mule(q4x21, q8y21), vec_mulo(q4x21, q8y21));
-            vector signed short qv30 = vec_add(vec_mule(q4x30, q8y30), vec_mulo(q4x30, q8y30));
-            vector signed short qv31 = vec_add(vec_mule(q4x31, q8y31), vec_mulo(q4x31, q8y31));
-
-            vector signed short vs0 = vec_splat(vscales, 0);
-            vector signed short vs1 = vec_splat(vscales, 1);
-            vector signed short vs2 = vec_splat(vscales, 2);
-            vector signed short vs3 = vec_splat(vscales, 3);
+            vector signed int qv00 = vec_msum(q8y00, q4x00, v0);
+            vector signed int qv01 = vec_msum(q8y01, q4x01, v0);
+            vector signed int qv10 = vec_msum(q8y10, q4x10, v0);
+            vector signed int qv11 = vec_msum(q8y11, q4x11, v0);
+            vector signed int qv20 = vec_msum(q8y20, q4x20, v0);
+            vector signed int qv21 = vec_msum(q8y21, q4x21, v0);
+            vector signed int qv30 = vec_msum(q8y30, q4x30, v0);
+            vector signed int qv31 = vec_msum(q8y31, q4x31, v0);
+
+            vector signed int vscales_h = vec_unpackh(vscales);
+            vector signed int vs0 = vec_splat(vscales_h, 0);
+            vector signed int vs1 = vec_splat(vscales_h, 1);
+            vector signed int vs2 = vec_splat(vscales_h, 2);
+            vector signed int vs3 = vec_splat(vscales_h, 3);
             vscales = vec_sld(vscales, vscales, 8);
 
-            qv00 = vec_add(qv00, qv10);
-            qv10 = vec_add(qv01, qv11);
-            qv20 = vec_add(qv20, qv30);
-            qv30 = vec_add(qv21, qv31);
+            vsumi0 = vec_add(vec_mul(qv00, vs0), vsumi0);
+            vsumi1 = vec_add(vec_mul(qv01, vs1), vsumi1);
+            vsumi2 = vec_add(vec_mul(qv20, vs2), vsumi2);
+            vsumi3 = vec_add(vec_mul(qv21, vs3), vsumi3);
 
-            vsumi0 = vec_add(vec_mule(qv00, vs0), vsumi0);
-            vsumi1 = vec_add(vec_mulo(qv00, vs0), vsumi1);
-            vsumi2 = vec_add(vec_mule(qv10, vs1), vsumi2);
-            vsumi3 = vec_add(vec_mulo(qv10, vs1), vsumi3);
-            vsumi4 = vec_add(vec_mule(qv20, vs2), vsumi4);
-            vsumi5 = vec_add(vec_mulo(qv20, vs2), vsumi5);
-            vsumi6 = vec_add(vec_mule(qv30, vs3), vsumi6);
-            vsumi7 = vec_add(vec_mulo(qv30, vs3), vsumi7);
+            vsumi0 = vec_add(vec_mul(qv10, vs0), vsumi0);
+            vsumi1 = vec_add(vec_mul(qv11, vs1), vsumi1);
+            vsumi2 = vec_add(vec_mul(qv30, vs2), vsumi2);
+            vsumi3 = vec_add(vec_mul(qv31, vs3), vsumi3);
         }
 
-        vsumi0 = vec_add(vsumi0, vsumi4);
-        vsumi1 = vec_add(vsumi1, vsumi5);
-        vsumi2 = vec_add(vsumi2, vsumi6);
-        vsumi3 = vec_add(vsumi3, vsumi7);
-
         vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
         vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
         vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
@@ -7399,6 +7268,9 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * r
     *s = vec_extract(vsumf0, 0);
 
 #elif defined __loongarch_asx
+    GGML_UNUSED(kmask1);
+    GGML_UNUSED(kmask2);
+    GGML_UNUSED(kmask3);
 
     const __m256i m4 = __lasx_xvreplgr2vr_b(0xF);
 
@@ -7411,6 +7283,11 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * r
         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;
 
         const uint8_t * restrict q4 = x[i].qs;
         const int8_t  * restrict q8 = y[i].qs;
@@ -7450,16 +7327,17 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * r
 
         __m256 vd = __lasx_xvreplfr2vr_s(d);
         acc = __lasx_xvfmadd_s(vd, __lasx_xvffint_s_w(sumi), acc);
+
     }
 
     acc_m = __lsx_vfadd_s(acc_m, (__m128)__lsx_vpermi_w((__m128i)acc_m, (__m128i)acc_m, 0xee));
     __m128i tmp1 = __lsx_vinsgr2vr_w(__lsx_vldi(0), __lsx_vpickve2gr_w((__m128i)acc_m, 1), 0);
     acc_m = __lsx_vfadd_s(acc_m, (__m128)tmp1);
 
+
     ft_union fi;
     fi.i = __lsx_vpickve2gr_w(acc_m, 0);
     *s = hsum_float_8(acc) + fi.f ;
-
 #else
 
     const uint8_t * scales = (const uint8_t*)&utmp[0];
@@ -7874,6 +7752,9 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r
 
 #elif defined(__POWER9_VECTOR__)
     const vector signed char lowMask = vec_splats((signed char)0xF);
+    const vector signed char lowMask1 = vec_splats((int8_t)0x3f);
+    const vector signed char lowMask2 = vec_splats((int8_t)0x30);
+    const vector int v0 = vec_splats((int32_t)0);
     const vector unsigned char v1 = vec_splats((unsigned char)0x1);
     const vector unsigned char v2 = vec_splats((unsigned char)0x2);
     const vector unsigned char v3 = vec_splats((unsigned char)0x3);
@@ -7892,18 +7773,27 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r
         vector float vxmin = vec_splats(GGML_FP16_TO_FP32(x[i].dmin));
         vector float vdmin = vec_mul(vxmin, vyd);
 
-        memcpy(utmp, x[i].scales, 12);
+        UNUSED(kmask1);
+        UNUSED(kmask2);
+        UNUSED(kmask3);
+        UNUSED(utmp);
 
-        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;
+        vector signed char u0 = (vector signed char)vec_xl_len(x[i].scales, 8);
+        vector signed char u1 = vec_and(vec_sr(u0, v2), lowMask2);
+        vector signed char u2 = (vector signed char)vec_xl_len(x[i].scales + 8, 4);
+        vector signed char u3 = vec_sr(u2, v4);
+
+        vector signed char u30 = u1;
+        vector signed char u31 = (vector signed char)vec_mergeh((vector signed int)vec_and(u2, lowMask), (vector signed int)u3);
+
+        u1 = vec_and(u0, lowMask1);
+        u2 = vec_or(u30, u31);
+
+        vector signed char utmps = (vector signed char)vec_mergeh((vector signed int)u1, (vector signed int)u2);
 
         vector signed short q8ysums0 = vec_xl( 0, y[i].bsums);
         vector signed short q8ysums1 = vec_xl(16, y[i].bsums);
 
-        vector signed char utmps = (vector signed char)vec_xl( 0, utmp);
         vector signed short vscales = vec_unpackh(utmps);
 
         vector signed short q5xmins = vec_unpackl(utmps);
@@ -7923,10 +7813,10 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r
         vector signed char qxhs0 = (vector signed char)vec_xl( 0, x[i].qh);
         vector signed char qxhs1 = (vector signed char)vec_xl(16, x[i].qh);
 
-        vector signed int vsumi0 = vec_splats((int32_t)0);
-        vector signed int vsumi1 = vec_splats((int32_t)0);
-        vector signed int vsumi2 = vec_splats((int32_t)0);
-        vector signed int vsumi3 = vec_splats((int32_t)0);
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
+        vector signed int vsumi2 = v0;
+        vector signed int vsumi3 = v0;
 
         const uint8_t * restrict q5 = x[i].qs;
         const int8_t  * restrict q8 = y[i].qs;
@@ -7951,10 +7841,10 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r
             qxhs0 = vec_sr(qxhs0, v2);
             qxhs1 = vec_sr(qxhs1, v2);
 
-            vector signed char q5x00 = vec_or(q5h00, qxs00);
-            vector signed char q5x01 = vec_or(q5h01, qxs01);
-            vector signed char q5x10 = vec_or(q5h10, qxs10);
-            vector signed char q5x11 = vec_or(q5h11, qxs11);
+            vector unsigned char q5x00 = (vector unsigned char)vec_or(q5h00, qxs00);
+            vector unsigned char q5x01 = (vector unsigned char)vec_or(q5h01, qxs01);
+            vector unsigned char q5x10 = (vector unsigned char)vec_or(q5h10, qxs10);
+            vector unsigned char q5x11 = (vector unsigned char)vec_or(q5h11, qxs11);
 
             vector signed char q8y00 = vec_xl( 0, q8);
             vector signed char q8y10 = vec_xl(16, q8);
@@ -7962,22 +7852,20 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r
             vector signed char q8y11 = vec_xl(48, q8);
             q8 += 64;
 
-            vector signed short qv00 = vec_add(vec_mule(q5x00, q8y00), vec_mulo(q5x00, q8y00));
-            vector signed short qv01 = vec_add(vec_mule(q5x01, q8y01), vec_mulo(q5x01, q8y01));
-            vector signed short qv10 = vec_add(vec_mule(q5x10, q8y10), vec_mulo(q5x10, q8y10));
-            vector signed short qv11 = vec_add(vec_mule(q5x11, q8y11), vec_mulo(q5x11, q8y11));
+            vector signed int qv00 = vec_msum(q8y00, q5x00, v0);
+            vector signed int qv01 = vec_msum(q8y01, q5x01, v0);
+            vector signed int qv10 = vec_msum(q8y10, q5x10, v0);
+            vector signed int qv11 = vec_msum(q8y11, q5x11, v0);
 
-            vector signed short vs0 = vec_splat(vscales, 0);
-            vector signed short vs1 = vec_splat(vscales, 1);
+            vector signed int vscales_h = vec_unpackh(vscales);
+            vector signed int vs0 = vec_splat(vscales_h, 0);
+            vector signed int vs1 = vec_splat(vscales_h, 1);
             vscales = vec_sld(vscales, vscales, 12);
 
-            qv00 = vec_add(qv00, qv10);
-            qv01 = vec_add(qv01, qv11);
-
-            vsumi0 = vec_add(vec_mule(qv00, vs0), vsumi0);
-            vsumi1 = vec_add(vec_mulo(qv00, vs0), vsumi1);
-            vsumi2 = vec_add(vec_mule(qv01, vs1), vsumi2);
-            vsumi3 = vec_add(vec_mulo(qv01, vs1), vsumi3);
+            vsumi0 = vec_add(vec_mul(qv00, vs0), vsumi0);
+            vsumi1 = vec_add(vec_mul(qv10, vs0), vsumi1);
+            vsumi2 = vec_add(vec_mul(qv01, vs1), vsumi2);
+            vsumi3 = vec_add(vec_mul(qv11, vs1), vsumi3);
         }
 
         vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
@@ -7997,6 +7885,9 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r
     *s = vec_extract(vsumf0, 0);
 
 #elif defined __loongarch_asx
+    GGML_UNUSED(kmask1);
+    GGML_UNUSED(kmask2);
+    GGML_UNUSED(kmask3);
 
     const __m256i m4 = __lasx_xvreplgr2vr_b(0xF);
     const __m128i mzero = __lsx_vldi(0);
@@ -8015,6 +7906,11 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r
         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;
 
         const __m256i mins_and_scales = lasx_extu8_16(lsx_set_w(utmp[3], utmp[2], utmp[1], utmp[0]));
 
@@ -8033,6 +7929,7 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r
         __m256i sumi = __lasx_xvldi(0);
 
         int bit = 0;
+        __m256i xvbit;
 
         for (int j = 0; j < QK_K/64; ++j) {
 
@@ -8041,13 +7938,15 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r
 
             const __m256i q5bits = __lasx_xvld((const __m256i*)q5, 0); q5 += 32;
 
+            xvbit = __lasx_xvreplgr2vr_h(bit++);
             const __m256i q5l_0 = __lasx_xvand_v(q5bits, m4);
-            const __m256i q5h_0 = __lasx_xvslli_h(__lasx_xvsrli_h(__lasx_xvand_v(hbits, hmask), bit++), 4);
+            const __m256i q5h_0 = __lasx_xvslli_h(__lasx_xvsrl_h(__lasx_xvand_v(hbits, hmask), xvbit), 4);
             const __m256i q5_0  = __lasx_xvadd_b(q5l_0, q5h_0);
             hmask = __lasx_xvslli_h(hmask, 1);
 
+            xvbit = __lasx_xvreplgr2vr_h(bit++);
             const __m256i q5l_1 = __lasx_xvand_v(__lasx_xvsrli_h(q5bits, 4), m4);
-            const __m256i q5h_1 = __lasx_xvslli_h(__lasx_xvsrli_h(__lasx_xvand_v(hbits, hmask), bit++), 4);
+            const __m256i q5h_1 = __lasx_xvslli_h(__lasx_xvsrl_h(__lasx_xvand_v(hbits, hmask), xvbit), 4);
             const __m256i q5_1  = __lasx_xvadd_b(q5l_1, q5h_1);
             hmask = __lasx_xvslli_h(hmask, 1);
 
@@ -8061,10 +7960,12 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r
             p16_1 = lasx_madd_h(scale_1, p16_1);
 
             sumi = __lasx_xvadd_w(sumi, __lasx_xvadd_w(p16_0, p16_1));
+
         }
 
         __m256 vd = __lasx_xvreplfr2vr_s(d);
         acc = __lasx_xvfmadd_s(vd, __lasx_xvffint_s_w(sumi), acc);
+
     }
 
     *s = hsum_float_8(acc) + summs;
@@ -8525,6 +8426,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, size_t bs, const void * r
 
 #elif defined(__POWER9_VECTOR__)
     const vector signed char lowMask = vec_splats((signed char)0xF);
+    const vector int v0 = vec_splats((int32_t)0);
     const vector unsigned char v2 = vec_splats((unsigned char)0x2);
     const vector unsigned char v3 = vec_splats((unsigned char)0x3);
     const vector unsigned char v4 = vec_splats((unsigned char)0x4);
@@ -8541,14 +8443,14 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, size_t bs, const void * r
         vector float vyd = vec_splats(y[i].d);
         vector float vd = vec_mul(vxd, vyd);
 
-        vector signed int vsumi0 = vec_splats((int32_t)0);
-        vector signed int vsumi1 = vec_splats((int32_t)0);
-        vector signed int vsumi2 = vec_splats((int32_t)0);
-        vector signed int vsumi3 = vec_splats((int32_t)0);
-        vector signed int vsumi4 = vec_splats((int32_t)0);
-        vector signed int vsumi5 = vec_splats((int32_t)0);
-        vector signed int vsumi6 = vec_splats((int32_t)0);
-        vector signed int vsumi7 = vec_splats((int32_t)0);
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
+        vector signed int vsumi2 = v0;
+        vector signed int vsumi3 = v0;
+        vector signed int vsumi4 = v0;
+        vector signed int vsumi5 = v0;
+        vector signed int vsumi6 = v0;
+        vector signed int vsumi7 = v0;
 
         const uint8_t * restrict q6 = x[i].ql;
         const uint8_t * restrict qh = x[i].qh;
@@ -8628,23 +8530,14 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, size_t bs, const void * r
             vector signed short vs6 = vec_splat(vscales, 6);
             vector signed short vs7 = vec_splat(vscales, 7);
 
-            vsumi0 = vec_add(vec_mule(qv00, vs0), vsumi0);
-            vsumi1 = vec_add(vec_mulo(qv00, vs0), vsumi1);
-            vsumi2 = vec_add(vec_mule(qv01, vs4), vsumi2);
-            vsumi3 = vec_add(vec_mulo(qv01, vs4), vsumi3);
-            vsumi4 = vec_add(vec_mule(qv10, vs1), vsumi4);
-            vsumi5 = vec_add(vec_mulo(qv10, vs1), vsumi5);
-            vsumi6 = vec_add(vec_mule(qv11, vs5), vsumi6);
-            vsumi7 = vec_add(vec_mulo(qv11, vs5), vsumi7);
-
-            vsumi0 = vec_add(vec_mule(qv20, vs2), vsumi0);
-            vsumi1 = vec_add(vec_mulo(qv20, vs2), vsumi1);
-            vsumi2 = vec_add(vec_mule(qv21, vs6), vsumi2);
-            vsumi3 = vec_add(vec_mulo(qv21, vs6), vsumi3);
-            vsumi4 = vec_add(vec_mule(qv30, vs3), vsumi4);
-            vsumi5 = vec_add(vec_mulo(qv30, vs3), vsumi5);
-            vsumi6 = vec_add(vec_mule(qv31, vs7), vsumi6);
-            vsumi7 = vec_add(vec_mulo(qv31, vs7), vsumi7);
+            vsumi0 = vec_msum(qv00, vs0, vsumi0);
+            vsumi1 = vec_msum(qv01, vs4, vsumi1);
+            vsumi2 = vec_msum(qv10, vs1, vsumi2);
+            vsumi3 = vec_msum(qv11, vs5, vsumi3);
+            vsumi4 = vec_msum(qv20, vs2, vsumi4);
+            vsumi5 = vec_msum(qv21, vs6, vsumi5);
+            vsumi6 = vec_msum(qv30, vs3, vsumi6);
+            vsumi7 = vec_msum(qv31, vs7, vsumi7);
         }
 
         vsumi0 = vec_add(vsumi0, vsumi4);
@@ -8791,7 +8684,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, size_t bs, const void * r
 #endif
 }
 
-#if defined (__AVX2__) || defined (__ARM_NEON) || defined (__POWER9_VECTOR__) || defined(__loongarch_asx)
+#if defined (__AVX__) || defined (__AVX2__) || defined (__ARM_NEON) || defined (__POWER9_VECTOR__) || defined(__loongarch_asx)
 static const int8_t keven_signs_q2xs[1024] = {
      1,  1,  1,  1,  1,  1,  1,  1, -1,  1,  1,  1,  1,  1,  1, -1,  1, -1,  1,  1,  1,  1,  1, -1, -1, -1,  1,  1,  1,  1,  1,  1,
      1,  1, -1,  1,  1,  1,  1, -1, -1,  1, -1,  1,  1,  1,  1,  1,  1, -1, -1,  1,  1,  1,  1,  1, -1, -1, -1,  1,  1,  1,  1, -1,
@@ -8924,7 +8817,63 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * restrict s, size_t bs, const void
 
     *s = 0.125f * hsum_float_8(accumf);
 
+#elif defined(__AVX__)
+    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
+
+    uint32_t aux32[4];
+    const uint8_t * aux8 = (const uint8_t *)aux32;
+
+    __m256 accumf = _mm256_setzero_ps();
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint16_t * restrict q2 = x[i].qs;
+        const int8_t   * restrict q8 = y[i].qs;
+        __m128i sumi1_0 = _mm_setzero_si128();
+        __m128i sumi1_1 = _mm_setzero_si128();
+        __m128i sumi2_0 = _mm_setzero_si128();
+        __m128i sumi2_1 = _mm_setzero_si128();
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+            const __m128i q8_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            memcpy(aux32, q2, 4*sizeof(uint32_t)); q2 += 8;
+            const __m128i q2_1_0 = _mm_set_epi64x(iq2xxs_grid[aux8[1]], iq2xxs_grid[aux8[0]]);
+            const __m128i q2_1_1 = _mm_set_epi64x(iq2xxs_grid[aux8[3]], iq2xxs_grid[aux8[2]]);
+            const __m128i q2_2_0 = _mm_set_epi64x(iq2xxs_grid[aux8[9]], iq2xxs_grid[aux8[8]]);
+            const __m128i q2_2_1 = _mm_set_epi64x(iq2xxs_grid[aux8[11]], iq2xxs_grid[aux8[10]]);
+            const __m128i s2_1_0 = _mm_set_epi64x(signs64[(aux32[1] >>  7) & 127], signs64[(aux32[1] >>  0) & 127]);
+            const __m128i s2_1_1 = _mm_set_epi64x(signs64[(aux32[1] >> 21) & 127], signs64[(aux32[1] >> 14) & 127]);
+            const __m128i s2_2_0 = _mm_set_epi64x(signs64[(aux32[3] >>  7) & 127], signs64[(aux32[3] >>  0) & 127]);
+            const __m128i s2_2_1 = _mm_set_epi64x(signs64[(aux32[3] >> 21) & 127], signs64[(aux32[3] >> 14) & 127]);
+            const __m128i q8s_1_0 = _mm_sign_epi8(q8_1_0, s2_1_0);
+            const __m128i q8s_1_1 = _mm_sign_epi8(q8_1_1, s2_1_1);
+            const __m128i q8s_2_0 = _mm_sign_epi8(q8_2_0, s2_2_0);
+            const __m128i q8s_2_1 = _mm_sign_epi8(q8_2_1, s2_2_1);
+            const __m128i dot1_0  = _mm_maddubs_epi16(q2_1_0, q8s_1_0);
+            const __m128i dot1_1  = _mm_maddubs_epi16(q2_1_1, q8s_1_1);
+            const __m128i dot2_0  = _mm_maddubs_epi16(q2_2_0, q8s_2_0);
+            const __m128i dot2_1  = _mm_maddubs_epi16(q2_2_1, q8s_2_1);
+            const uint16_t ls1 = aux32[1] >> 28;
+            const uint16_t ls2 = aux32[3] >> 28;
+            const __m128i p1_0 = _mm_madd_epi16(dot1_0, _mm_set1_epi16(2*ls1+1));
+            const __m128i p1_1 = _mm_madd_epi16(dot1_1, _mm_set1_epi16(2*ls1+1));
+            const __m128i p2_0 = _mm_madd_epi16(dot2_0, _mm_set1_epi16(2*ls2+1));
+            const __m128i p2_1 = _mm_madd_epi16(dot2_1, _mm_set1_epi16(2*ls2+1));
+            sumi1_0 = _mm_add_epi32(sumi1_0, p1_0);
+            sumi1_1 = _mm_add_epi32(sumi1_1, p1_1);
+            sumi2_0 = _mm_add_epi32(sumi2_0, p2_0);
+            sumi2_1 = _mm_add_epi32(sumi2_1, p2_1);
+        }
+
+        accumf = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(MM256_SET_M128I(_mm_add_epi32(sumi1_1, sumi2_1), _mm_add_epi32(sumi1_0, sumi2_0)))), accumf);
+
+    }
+
+    *s = 0.125f * hsum_float_8(accumf);
+
 #elif defined(__POWER9_VECTOR__)
+    const vector int v0 = vec_splats((int32_t)0);
     vector float vsumf0 = vec_splats(0.0f);
     vector float vsumf1 = vec_splats(0.0f);
     vector float vsumf2 = vec_splats(0.0f);
@@ -8937,14 +8886,10 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * restrict s, size_t bs, const void
         vector float vyd = vec_splats(y[i].d);
         vector float vd = vec_mul(vxd, vyd);
 
-        vector signed int vsumi0 = vec_splats((int32_t)0);
-        vector signed int vsumi1 = vec_splats((int32_t)0);
-        vector signed int vsumi2 = vec_splats((int32_t)0);
-        vector signed int vsumi3 = vec_splats((int32_t)0);
-        vector signed int vsumi4 = vec_splats((int32_t)0);
-        vector signed int vsumi5 = vec_splats((int32_t)0);
-        vector signed int vsumi6 = vec_splats((int32_t)0);
-        vector signed int vsumi7 = vec_splats((int32_t)0);
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
+        vector signed int vsumi2 = v0;
+        vector signed int vsumi3 = v0;
 
         const uint16_t * restrict q2 = x[i].qs;
         const int8_t  *  restrict q8 = y[i].qs;
@@ -8991,21 +8936,12 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * restrict s, size_t bs, const void
             vector signed short vscales01 = vec_splats((int16_t)(2*ls0+1));
             vector signed short vscales23 = vec_splats((int16_t)(2*ls1+1));
 
-            vsumi0 = vec_add(vec_mule(qv0, vscales01), vsumi0);
-            vsumi1 = vec_add(vec_mule(qv1, vscales01), vsumi1);
-            vsumi2 = vec_add(vec_mule(qv2, vscales23), vsumi2);
-            vsumi3 = vec_add(vec_mule(qv3, vscales23), vsumi3);
-            vsumi4 = vec_add(vec_mulo(qv0, vscales01), vsumi4);
-            vsumi5 = vec_add(vec_mulo(qv1, vscales01), vsumi5);
-            vsumi6 = vec_add(vec_mulo(qv2, vscales23), vsumi6);
-            vsumi7 = vec_add(vec_mulo(qv3, vscales23), vsumi7);
+            vsumi0 = vec_msum(qv0, vscales01, vsumi0);
+            vsumi1 = vec_msum(qv1, vscales01, vsumi1);
+            vsumi2 = vec_msum(qv2, vscales23, vsumi2);
+            vsumi3 = vec_msum(qv3, vscales23, vsumi3);
         }
 
-        vsumi0 = vec_add(vsumi0, vsumi4);
-        vsumi1 = vec_add(vsumi1, vsumi5);
-        vsumi2 = vec_add(vsumi2, vsumi6);
-        vsumi3 = vec_add(vsumi3, vsumi7);
-
         vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
         vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
         vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
@@ -9279,6 +9215,165 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * restrict s, size_t bs, const void *
     }
 
     *s = 0.125f * hsum_float_8(accumf);
+
+#elif defined(__AVX__)
+    const __m128i mone = _mm_set1_epi8(1);
+    static const char block_sign_shuffle_mask_1[32] = {
+        0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02,
+        0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06,
+    };
+    static const char block_sign_shuffle_mask_2[32] = {
+        0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a,
+        0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e,
+    };
+    static const uint8_t bit_selector_mask_bytes[32] = {
+        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+    };
+
+    const __m128i bit_selector_mask_0 = _mm_loadu_si128((const __m128i*)bit_selector_mask_bytes);
+    const __m128i bit_selector_mask_1 = _mm_loadu_si128((const __m128i*)bit_selector_mask_bytes + 1);
+    const __m128i block_sign_shuffle_1_0 = _mm_loadu_si128((const __m128i*)block_sign_shuffle_mask_1);
+    const __m128i block_sign_shuffle_1_1 = _mm_loadu_si128((const __m128i*)block_sign_shuffle_mask_1 + 1);
+    const __m128i block_sign_shuffle_2_0 = _mm_loadu_si128((const __m128i*)block_sign_shuffle_mask_2);
+    const __m128i block_sign_shuffle_2_1 = _mm_loadu_si128((const __m128i*)block_sign_shuffle_mask_2 + 1);
+
+    static const uint8_t k_bit_helper[32] = {
+        0x00, 0x80, 0x80, 0x00, 0x80, 0x00, 0x00, 0x80, 0x80, 0x00, 0x00, 0x80, 0x00, 0x80, 0x80, 0x00,
+        0x00, 0x80, 0x80, 0x00, 0x80, 0x00, 0x00, 0x80, 0x80, 0x00, 0x00, 0x80, 0x00, 0x80, 0x80, 0x00,
+    };
+    const __m128i bit_helper_0 = _mm_loadu_si128((const __m128i*)k_bit_helper);
+    const __m128i bit_helper_1 = _mm_loadu_si128((const __m128i*)k_bit_helper + 1);
+    const __m128i m511 = _mm_set1_epi16(511);
+    const __m128i m4 = _mm_set1_epi8(0xf);
+    const __m128i m1 = _mm_set1_epi8(1);
+
+    uint64_t aux64;
+
+    // somewhat hacky, but gives a significant boost in performance
+    __m256i aux_gindex;
+    const uint16_t * gindex = (const uint16_t *)&aux_gindex;
+
+    __m256 accumf = _mm256_setzero_ps();
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint16_t * restrict q2 = x[i].qs;
+        const int8_t   * restrict q8 = y[i].qs;
+
+        memcpy(&aux64, x[i].scales, 8);
+        __m128i stmp = _mm_set1_epi64x(aux64);
+        stmp = _mm_unpacklo_epi8(_mm_and_si128(stmp, m4), _mm_and_si128(_mm_srli_epi16(stmp, 4), m4));
+        const __m128i scales = _mm_add_epi8(_mm_slli_epi16(stmp, 1), m1);
+
+        __m128i sumi1_0 = _mm_setzero_si128();
+        __m128i sumi1_1 = _mm_setzero_si128();
+        __m128i sumi2_0 = _mm_setzero_si128();
+        __m128i sumi2_1 = _mm_setzero_si128();
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 4) {
+
+            const __m128i q2_data_0 = _mm_loadu_si128((const __m128i*)q2);
+            const __m128i q2_data_1 = _mm_loadu_si128((const __m128i*)q2 + 1);  q2 += 16;
+            aux_gindex = MM256_SET_M128I(_mm_and_si128(q2_data_1, m511), _mm_and_si128(q2_data_0, m511));
+
+            const __m128i partial_sign_bits_0 = _mm_srli_epi16(q2_data_0, 9);
+            const __m128i partial_sign_bits_1 = _mm_srli_epi16(q2_data_1, 9);
+            const __m128i partial_sign_bits_upper_0 = _mm_srli_epi16(q2_data_0, 13);
+            const __m128i partial_sign_bits_upper_1 = _mm_srli_epi16(q2_data_1, 13);
+            const __m128i partial_sign_bits_for_counting_0 = _mm_xor_si128(partial_sign_bits_0, partial_sign_bits_upper_0);
+            const __m128i partial_sign_bits_for_counting_1 = _mm_xor_si128(partial_sign_bits_1, partial_sign_bits_upper_1);
+
+            const __m128i odd_bits_0 = _mm_shuffle_epi8(bit_helper_0, partial_sign_bits_for_counting_0);
+            const __m128i odd_bits_1 = _mm_shuffle_epi8(bit_helper_1, partial_sign_bits_for_counting_1);
+            const __m128i full_sign_bits_0 = _mm_or_si128(partial_sign_bits_0, odd_bits_0);
+            const __m128i full_sign_bits_1 = _mm_or_si128(partial_sign_bits_1, odd_bits_1);
+
+            const __m128i q8_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_3_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_3_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_4_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_4_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+
+            const __m128i q2_1_0 = _mm_set_epi64x(iq2xs_grid[gindex[1]], iq2xs_grid[gindex[0]]);
+            const __m128i q2_1_1 = _mm_set_epi64x(iq2xs_grid[gindex[3]], iq2xs_grid[gindex[2]]);
+            const __m128i q2_2_0 = _mm_set_epi64x(iq2xs_grid[gindex[5]], iq2xs_grid[gindex[4]]);
+            const __m128i q2_2_1 = _mm_set_epi64x(iq2xs_grid[gindex[7]], iq2xs_grid[gindex[6]]);
+            const __m128i q2_3_0 = _mm_set_epi64x(iq2xs_grid[gindex[9]], iq2xs_grid[gindex[8]]);
+            const __m128i q2_3_1 = _mm_set_epi64x(iq2xs_grid[gindex[11]], iq2xs_grid[gindex[10]]);
+            const __m128i q2_4_0 = _mm_set_epi64x(iq2xs_grid[gindex[13]], iq2xs_grid[gindex[12]]);
+            const __m128i q2_4_1 = _mm_set_epi64x(iq2xs_grid[gindex[15]], iq2xs_grid[gindex[14]]);
+
+            // AVX2 full_signs_1 is full_sign_bits_0 here
+            // AVX2 full_signs_2 is full_sign_bits_1 here
+            __m128i signs_0, signs_1;
+            signs_0 = _mm_shuffle_epi8(full_sign_bits_0, block_sign_shuffle_1_0);
+            signs_1 = _mm_shuffle_epi8(full_sign_bits_0, block_sign_shuffle_1_1);
+            signs_0 = _mm_cmpeq_epi8(_mm_and_si128(signs_0, bit_selector_mask_0), bit_selector_mask_0);
+            signs_1 = _mm_cmpeq_epi8(_mm_and_si128(signs_1, bit_selector_mask_1), bit_selector_mask_1);
+            const __m128i q8s_1_0 = _mm_sign_epi8(q8_1_0, _mm_or_si128(signs_0, mone));
+            const __m128i q8s_1_1 = _mm_sign_epi8(q8_1_1, _mm_or_si128(signs_1, mone));
+
+            signs_0 = _mm_shuffle_epi8(full_sign_bits_0, block_sign_shuffle_2_0);
+            signs_1 = _mm_shuffle_epi8(full_sign_bits_0, block_sign_shuffle_2_1);
+            signs_0 = _mm_cmpeq_epi8(_mm_and_si128(signs_0, bit_selector_mask_0), bit_selector_mask_0);
+            signs_1 = _mm_cmpeq_epi8(_mm_and_si128(signs_1, bit_selector_mask_1), bit_selector_mask_1);
+            const __m128i q8s_2_0 = _mm_sign_epi8(q8_2_0, _mm_or_si128(signs_0, mone));
+            const __m128i q8s_2_1 = _mm_sign_epi8(q8_2_1, _mm_or_si128(signs_1, mone));
+
+            signs_0 = _mm_shuffle_epi8(full_sign_bits_1, block_sign_shuffle_1_0);
+            signs_1 = _mm_shuffle_epi8(full_sign_bits_1, block_sign_shuffle_1_1);
+            signs_0 = _mm_cmpeq_epi8(_mm_and_si128(signs_0, bit_selector_mask_0), bit_selector_mask_0);
+            signs_1 = _mm_cmpeq_epi8(_mm_and_si128(signs_1, bit_selector_mask_1), bit_selector_mask_1);
+            const __m128i q8s_3_0 = _mm_sign_epi8(q8_3_0, _mm_or_si128(signs_0, mone));
+            const __m128i q8s_3_1 = _mm_sign_epi8(q8_3_1, _mm_or_si128(signs_1, mone));
+
+            signs_0 = _mm_shuffle_epi8(full_sign_bits_1, block_sign_shuffle_2_0);
+            signs_1 = _mm_shuffle_epi8(full_sign_bits_1, block_sign_shuffle_2_1);
+            signs_0 = _mm_cmpeq_epi8(_mm_and_si128(signs_0, bit_selector_mask_0), bit_selector_mask_0);
+            signs_1 = _mm_cmpeq_epi8(_mm_and_si128(signs_1, bit_selector_mask_1), bit_selector_mask_1);
+            const __m128i q8s_4_0 = _mm_sign_epi8(q8_4_0, _mm_or_si128(signs_0, mone));
+            const __m128i q8s_4_1 = _mm_sign_epi8(q8_4_1, _mm_or_si128(signs_1, mone));
+
+            const __m128i dot1_0  = _mm_maddubs_epi16(q2_1_0, q8s_1_0);
+            const __m128i dot1_1  = _mm_maddubs_epi16(q2_1_1, q8s_1_1);
+            const __m128i dot2_0  = _mm_maddubs_epi16(q2_2_0, q8s_2_0);
+            const __m128i dot2_1  = _mm_maddubs_epi16(q2_2_1, q8s_2_1);
+            const __m128i dot3_0  = _mm_maddubs_epi16(q2_3_0, q8s_3_0);
+            const __m128i dot3_1  = _mm_maddubs_epi16(q2_3_1, q8s_3_1);
+            const __m128i dot4_0  = _mm_maddubs_epi16(q2_4_0, q8s_4_0);
+            const __m128i dot4_1  = _mm_maddubs_epi16(q2_4_1, q8s_4_1);
+
+            __m128i sc_tmp = _mm_shuffle_epi8(scales, get_scale_shuffle(ib32+0));
+            const __m128i sc1_0 = _mm_cvtepi8_epi16(sc_tmp);
+            const __m128i sc1_1 = _mm_cvtepi8_epi16(_mm_srli_si128(sc_tmp, 8));
+            sc_tmp = _mm_shuffle_epi8(scales, get_scale_shuffle(ib32+1));
+            const __m128i sc2_0 = _mm_cvtepi8_epi16(sc_tmp);
+            const __m128i sc2_1 = _mm_cvtepi8_epi16(_mm_srli_si128(sc_tmp, 8));
+            sc_tmp = _mm_shuffle_epi8(scales, get_scale_shuffle(ib32+2));
+            const __m128i sc3_0 = _mm_cvtepi8_epi16(sc_tmp);
+            const __m128i sc3_1 = _mm_cvtepi8_epi16(_mm_srli_si128(sc_tmp, 8));
+            sc_tmp = _mm_shuffle_epi8(scales, get_scale_shuffle(ib32+3));
+            const __m128i sc4_0 = _mm_cvtepi8_epi16(sc_tmp);
+            const __m128i sc4_1 = _mm_cvtepi8_epi16(_mm_srli_si128(sc_tmp, 8));
+
+            sumi1_0 = _mm_add_epi32(sumi1_0, _mm_madd_epi16(dot1_0, sc1_0));
+            sumi1_1 = _mm_add_epi32(sumi1_1, _mm_madd_epi16(dot1_1, sc1_1));
+            sumi2_0 = _mm_add_epi32(sumi2_0, _mm_madd_epi16(dot2_0, sc2_0));
+            sumi2_1 = _mm_add_epi32(sumi2_1, _mm_madd_epi16(dot2_1, sc2_1));
+            sumi1_0 = _mm_add_epi32(sumi1_0, _mm_madd_epi16(dot3_0, sc3_0));
+            sumi1_1 = _mm_add_epi32(sumi1_1, _mm_madd_epi16(dot3_1, sc3_1));
+            sumi2_0 = _mm_add_epi32(sumi2_0, _mm_madd_epi16(dot4_0, sc4_0));
+            sumi2_1 = _mm_add_epi32(sumi2_1, _mm_madd_epi16(dot4_1, sc4_1));
+        }
+
+        accumf = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(MM256_SET_M128I(_mm_add_epi32(sumi1_1, sumi2_1), _mm_add_epi32(sumi1_0, sumi2_0)))), accumf);
+
+    }
+
+    *s = 0.125f * hsum_float_8(accumf);
+
 #elif defined(__loongarch_asx)
 
     const __m256i mone = __lasx_xvreplgr2vr_b(1);
@@ -9397,6 +9492,7 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * restrict s, size_t bs, const void *
 
     *s = 0.125f * hsum_float_8(accumf);
 #elif defined(__POWER9_VECTOR__)
+    const vector int v0 = vec_splats((int32_t)0);
     vector float vsumf0 = vec_splats(0.0f);
     vector float vsumf1 = vec_splats(0.0f);
     vector float vsumf2 = vec_splats(0.0f);
@@ -9409,14 +9505,10 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * restrict s, size_t bs, const void *
         vector float vyd = vec_splats(y[i].d);
         vector float vd = vec_mul(vxd, vyd);
 
-        vector signed int vsumi0 = vec_splats((int32_t)0);
-        vector signed int vsumi1 = vec_splats((int32_t)0);
-        vector signed int vsumi2 = vec_splats((int32_t)0);
-        vector signed int vsumi3 = vec_splats((int32_t)0);
-        vector signed int vsumi4 = vec_splats((int32_t)0);
-        vector signed int vsumi5 = vec_splats((int32_t)0);
-        vector signed int vsumi6 = vec_splats((int32_t)0);
-        vector signed int vsumi7 = vec_splats((int32_t)0);
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
+        vector signed int vsumi2 = v0;
+        vector signed int vsumi3 = v0;
 
         const uint16_t * restrict q2 = x[i].qs;
         const uint8_t  * restrict sc = x[i].scales;
@@ -9464,21 +9556,12 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * restrict s, size_t bs, const void *
             vector signed short vscales2 = vec_splats((int16_t)(2*ls2+1));
             vector signed short vscales3 = vec_splats((int16_t)(2*ls3+1));
 
-            vsumi0 = vec_add(vec_mule(qv0, vscales0), vsumi0);
-            vsumi1 = vec_add(vec_mule(qv1, vscales1), vsumi1);
-            vsumi2 = vec_add(vec_mule(qv2, vscales2), vsumi2);
-            vsumi3 = vec_add(vec_mule(qv3, vscales3), vsumi3);
-            vsumi4 = vec_add(vec_mulo(qv0, vscales0), vsumi4);
-            vsumi5 = vec_add(vec_mulo(qv1, vscales1), vsumi5);
-            vsumi6 = vec_add(vec_mulo(qv2, vscales2), vsumi6);
-            vsumi7 = vec_add(vec_mulo(qv3, vscales3), vsumi7);
+            vsumi0 = vec_msum(qv0, vscales0, vsumi0);
+            vsumi1 = vec_msum(qv1, vscales1, vsumi1);
+            vsumi2 = vec_msum(qv2, vscales2, vsumi2);
+            vsumi3 = vec_msum(qv3, vscales3, vsumi3);
         }
 
-        vsumi0 = vec_add(vsumi0, vsumi4);
-        vsumi1 = vec_add(vsumi1, vsumi5);
-        vsumi2 = vec_add(vsumi2, vsumi6);
-        vsumi3 = vec_add(vsumi3, vsumi7);
-
         vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
         vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
         vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
@@ -9694,6 +9777,98 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * restrict s, size_t bs, const void *
 
     *s = 0.125f * hsum_float_8(accumf);
 
+#elif defined(__AVX__)
+   static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
+                                       0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03
+   };
+
+    static const uint8_t k_mask2[32] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+                                        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+    };
+
+    const __m128i m4 = _mm_set1_epi8(0xf);
+    const __m128i m1 = _mm_set1_epi8(1);
+
+    const __m128i mask1_0 = _mm_loadu_si128((const __m128i*)k_mask1);
+    const __m128i mask1_1 = _mm_loadu_si128((const __m128i*)k_mask1 + 1);
+    const __m128i mask2_0 = _mm_loadu_si128((const __m128i*)k_mask2);
+    const __m128i mask2_1 = _mm_loadu_si128((const __m128i*)k_mask2 + 1);
+
+    uint64_t aux64;
+
+    __m256 accumf = _mm256_setzero_ps();
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint8_t * restrict qs = x[i].qs;
+        const uint8_t * restrict qh = x[i].qh;
+        const uint16_t * restrict signs = (const uint16_t *)(x[i].qs + QK_K/8);
+        const int8_t  * restrict q8 = y[i].qs;
+
+        memcpy(&aux64, x[i].scales, 8);
+        const __m128i scales8 = _mm_add_epi8(_mm_slli_epi16(_mm_and_si128(_mm_set_epi64x(aux64 >> 4, aux64), m4), 1), m1);
+        const __m128i scales16_0 = _mm_cvtepi8_epi16(scales8);
+        const __m128i scales16_1 = _mm_cvtepi8_epi16(_mm_srli_si128(scales8, 8));
+
+        __m128i sumi1_0 = _mm_setzero_si128();
+        __m128i sumi1_1 = _mm_setzero_si128();
+        __m128i sumi2_0 = _mm_setzero_si128();
+        __m128i sumi2_1 = _mm_setzero_si128();
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+            const __m128i q8_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q2_1_0 = _mm_set_epi64x(iq2s_grid[qs[1] | ((qh[ib32+0] << 6) & 0x300)],
+                                                  iq2s_grid[qs[0] | ((qh[ib32+0] << 8) & 0x300)]);
+            const __m128i q2_1_1 = _mm_set_epi64x(iq2s_grid[qs[3] | ((qh[ib32+0] << 2) & 0x300)],
+                                                  iq2s_grid[qs[2] | ((qh[ib32+0] << 4) & 0x300)]);
+            const __m128i q2_2_0 = _mm_set_epi64x(iq2s_grid[qs[5] | ((qh[ib32+1] << 6) & 0x300)],
+                                                  iq2s_grid[qs[4] | ((qh[ib32+1] << 8) & 0x300)]);
+            const __m128i q2_2_1 = _mm_set_epi64x(iq2s_grid[qs[7] | ((qh[ib32+1] << 2) & 0x300)],
+                                                  iq2s_grid[qs[6] | ((qh[ib32+1] << 4) & 0x300)]);
+            qs += 8;
+
+            __m128i aux128_0 = _mm_set1_epi32(signs[0] | ((uint32_t) signs[1] << 16));
+            __m128i aux128_1 = aux128_0;
+            aux128_0 = _mm_and_si128(_mm_shuffle_epi8(aux128_0,mask1_0), mask2_0);
+            aux128_1 = _mm_and_si128(_mm_shuffle_epi8(aux128_1,mask1_1), mask2_1);
+            const __m128i s2_1_0 = _mm_cmpeq_epi8(aux128_0, mask2_0);
+            const __m128i s2_1_1 = _mm_cmpeq_epi8(aux128_1, mask2_1);
+            const __m128i q8s_1_0 = _mm_sub_epi8(_mm_xor_si128(s2_1_0, q8_1_0), s2_1_0);
+            const __m128i q8s_1_1 = _mm_sub_epi8(_mm_xor_si128(s2_1_1, q8_1_1), s2_1_1);
+
+            aux128_0 = _mm_set1_epi32(signs[2] | ((uint32_t) signs[3] << 16));
+            aux128_1 = aux128_0;
+            aux128_0 = _mm_and_si128(_mm_shuffle_epi8(aux128_0,mask1_0), mask2_0);
+            aux128_1 = _mm_and_si128(_mm_shuffle_epi8(aux128_1,mask1_1), mask2_1);
+            const __m128i s2_2_0 = _mm_cmpeq_epi8(aux128_0, mask2_0);
+            const __m128i s2_2_1 = _mm_cmpeq_epi8(aux128_1, mask2_1);
+            const __m128i q8s_2_0 = _mm_sub_epi8(_mm_xor_si128(s2_2_0, q8_2_0), s2_2_0);
+            const __m128i q8s_2_1 = _mm_sub_epi8(_mm_xor_si128(s2_2_1, q8_2_1), s2_2_1);
+
+            signs += 4;
+
+            const __m128i dot1_0  = _mm_maddubs_epi16(q2_1_0, q8s_1_0);
+            const __m128i dot1_1  = _mm_maddubs_epi16(q2_1_1, q8s_1_1);
+            const __m128i dot2_0  = _mm_maddubs_epi16(q2_2_0, q8s_2_0);
+            const __m128i dot2_1  = _mm_maddubs_epi16(q2_2_1, q8s_2_1);
+
+            const __m128i p1_0 = _mm_madd_epi16(dot1_0, _mm_shuffle_epi8(scales16_0, _mm256_extractf128_si256(get_scale_shuffle_k4(ib32+0), 0)));
+            const __m128i p1_1 = _mm_madd_epi16(dot1_1, _mm_shuffle_epi8(scales16_1, _mm256_extractf128_si256(get_scale_shuffle_k4(ib32+0), 1)));
+            const __m128i p2_0 = _mm_madd_epi16(dot2_0, _mm_shuffle_epi8(scales16_0, _mm256_extractf128_si256(get_scale_shuffle_k4(ib32+1), 0)));
+            const __m128i p2_1 = _mm_madd_epi16(dot2_1, _mm_shuffle_epi8(scales16_1, _mm256_extractf128_si256(get_scale_shuffle_k4(ib32+1), 1)));
+            sumi1_0 = _mm_add_epi32(sumi1_0, p1_0);
+            sumi1_1 = _mm_add_epi32(sumi1_1, p1_1);
+            sumi2_0 = _mm_add_epi32(sumi2_0, p2_0);
+            sumi2_1 = _mm_add_epi32(sumi2_1, p2_1);
+        }
+
+        accumf = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(MM256_SET_M128I(_mm_add_epi32(sumi1_1, sumi2_1), _mm_add_epi32(sumi1_0, sumi2_0)))), accumf);
+
+    }
+
+    *s = 0.125f * hsum_float_8(accumf);
+
 #elif defined(__POWER9_VECTOR__)
     static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
                                         0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03
@@ -9701,6 +9876,8 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * restrict s, size_t bs, const void *
 
     static const uint8_t k_mask2[16] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,};
 
+    const vector int v0 = vec_splats((int32_t)0);
+
     vector float vsumf0 = vec_splats(0.0f);
     vector float vsumf1 = vec_splats(0.0f);
     vector float vsumf2 = vec_splats(0.0f);
@@ -9715,14 +9892,10 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * restrict s, size_t bs, const void *
         vector float vyd = vec_splats(y[i].d);
         vector float vd = vec_mul(vxd, vyd);
 
-        vector signed int vsumi0 = vec_splats((int32_t)0);
-        vector signed int vsumi1 = vec_splats((int32_t)0);
-        vector signed int vsumi2 = vec_splats((int32_t)0);
-        vector signed int vsumi3 = vec_splats((int32_t)0);
-        vector signed int vsumi4 = vec_splats((int32_t)0);
-        vector signed int vsumi5 = vec_splats((int32_t)0);
-        vector signed int vsumi6 = vec_splats((int32_t)0);
-        vector signed int vsumi7 = vec_splats((int32_t)0);
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
+        vector signed int vsumi2 = v0;
+        vector signed int vsumi3 = v0;
 
         const uint8_t *  restrict q2 = x[i].qs;
         const uint8_t *  restrict qh = x[i].qh;
@@ -9782,21 +9955,12 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * restrict s, size_t bs, const void *
             vector signed short vscales2 = vec_splats((int16_t)(2*ls2+1));
             vector signed short vscales3 = vec_splats((int16_t)(2*ls3+1));
 
-            vsumi0 = vec_add(vec_mule(qv0, vscales0), vsumi0);
-            vsumi1 = vec_add(vec_mule(qv1, vscales1), vsumi1);
-            vsumi2 = vec_add(vec_mule(qv2, vscales2), vsumi2);
-            vsumi3 = vec_add(vec_mule(qv3, vscales3), vsumi3);
-            vsumi4 = vec_add(vec_mulo(qv0, vscales0), vsumi4);
-            vsumi5 = vec_add(vec_mulo(qv1, vscales1), vsumi5);
-            vsumi6 = vec_add(vec_mulo(qv2, vscales2), vsumi6);
-            vsumi7 = vec_add(vec_mulo(qv3, vscales3), vsumi7);
+            vsumi0 = vec_msum(qv0, vscales0, vsumi0);
+            vsumi1 = vec_msum(qv1, vscales1, vsumi1);
+            vsumi2 = vec_msum(qv2, vscales2, vsumi2);
+            vsumi3 = vec_msum(qv3, vscales3, vsumi3);
         }
 
-        vsumi0 = vec_add(vsumi0, vsumi4);
-        vsumi1 = vec_add(vsumi1, vsumi5);
-        vsumi2 = vec_add(vsumi2, vsumi6);
-        vsumi3 = vec_add(vsumi3, vsumi7);
-
         vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
         vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
         vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
@@ -10031,9 +10195,68 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * restrict s, size_t bs, const void
 
     *s = 0.25f * hsum_float_8(accumf);
 
+#elif defined(__AVX__)
+    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
+
+    uint32_t aux32[2];
+
+    __m256 accumf = _mm256_setzero_ps();
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint8_t * restrict q3 = x[i].qs;
+        const uint8_t * restrict gas = x[i].qs + QK_K/4;
+        const int8_t  * restrict q8 = y[i].qs;
+        __m128i sumi1_0 = _mm_setzero_si128();
+        __m128i sumi1_1 = _mm_setzero_si128();
+        __m128i sumi2_0 = _mm_setzero_si128();
+        __m128i sumi2_1 = _mm_setzero_si128();
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+            const __m128i q8_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q2_1_0 = _mm_set_epi32(iq3xxs_grid[q3[3]], iq3xxs_grid[q3[2]], iq3xxs_grid[q3[1]], iq3xxs_grid[q3[0]]);
+            const __m128i q2_1_1 = _mm_set_epi32(iq3xxs_grid[q3[7]], iq3xxs_grid[q3[6]], iq3xxs_grid[q3[5]], iq3xxs_grid[q3[4]]);
+            q3 += 8;
+            const __m128i q2_2_0 = _mm_set_epi32(iq3xxs_grid[q3[3]], iq3xxs_grid[q3[2]], iq3xxs_grid[q3[1]], iq3xxs_grid[q3[0]]);
+            const __m128i q2_2_1 = _mm_set_epi32(iq3xxs_grid[q3[7]], iq3xxs_grid[q3[6]], iq3xxs_grid[q3[5]], iq3xxs_grid[q3[4]]);
+            q3 += 8;
+            memcpy(aux32, gas, 8); gas += 8;
+            const __m128i s2_1_0 = _mm_set_epi64x(signs64[(aux32[0] >>  7) & 127], signs64[(aux32[0] >>  0) & 127]);
+            const __m128i s2_1_1 = _mm_set_epi64x(signs64[(aux32[0] >> 21) & 127], signs64[(aux32[0] >> 14) & 127]);
+            const __m128i s2_2_0 = _mm_set_epi64x(signs64[(aux32[1] >>  7) & 127], signs64[(aux32[1] >>  0) & 127]);
+            const __m128i s2_2_1 = _mm_set_epi64x(signs64[(aux32[1] >> 21) & 127], signs64[(aux32[1] >> 14) & 127]);
+            const __m128i q8s_1_0 = _mm_sign_epi8(q8_1_0, s2_1_0);
+            const __m128i q8s_1_1 = _mm_sign_epi8(q8_1_1, s2_1_1);
+            const __m128i q8s_2_0 = _mm_sign_epi8(q8_2_0, s2_2_0);
+            const __m128i q8s_2_1 = _mm_sign_epi8(q8_2_1, s2_2_1);
+            const __m128i dot1_0  = _mm_maddubs_epi16(q2_1_0, q8s_1_0);
+            const __m128i dot1_1  = _mm_maddubs_epi16(q2_1_1, q8s_1_1);
+            const __m128i dot2_0  = _mm_maddubs_epi16(q2_2_0, q8s_2_0);
+            const __m128i dot2_1  = _mm_maddubs_epi16(q2_2_1, q8s_2_1);
+            const uint16_t ls1 = aux32[0] >> 28;
+            const uint16_t ls2 = aux32[1] >> 28;
+            const __m128i p1_0 = _mm_madd_epi16(dot1_0, _mm_set1_epi16(2*ls1+1));
+            const __m128i p1_1 = _mm_madd_epi16(dot1_1, _mm_set1_epi16(2*ls1+1));
+            const __m128i p2_0 = _mm_madd_epi16(dot2_0, _mm_set1_epi16(2*ls2+1));
+            const __m128i p2_1 = _mm_madd_epi16(dot2_1, _mm_set1_epi16(2*ls2+1));
+            sumi1_0 = _mm_add_epi32(sumi1_0, p1_0);
+            sumi1_1 = _mm_add_epi32(sumi1_1, p1_1);
+            sumi2_0 = _mm_add_epi32(sumi2_0, p2_0);
+            sumi2_1 = _mm_add_epi32(sumi2_1, p2_1);
+        }
+
+        accumf = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(MM256_SET_M128I(_mm_add_epi32(sumi1_1, sumi2_1), _mm_add_epi32(sumi1_0, sumi2_0)))), accumf);
+
+    }
+
+    *s = 0.25f * hsum_float_8(accumf);
+
 #elif defined(__POWER9_VECTOR__)
     const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
 
+    const vector int v0 = vec_splats((int32_t)0);
+
     vector float vsumf0 = vec_splats(0.0f);
     vector float vsumf1 = vec_splats(0.0f);
     vector float vsumf2 = vec_splats(0.0f);
@@ -10044,14 +10267,10 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * restrict s, size_t bs, const void
         vector float vyd = vec_splats(y[i].d);
         vector float vd = vec_mul(vxd, vyd);
 
-        vector signed int vsumi0 = vec_splats((int32_t)0);
-        vector signed int vsumi1 = vec_splats((int32_t)0);
-        vector signed int vsumi2 = vec_splats((int32_t)0);
-        vector signed int vsumi3 = vec_splats((int32_t)0);
-        vector signed int vsumi4 = vec_splats((int32_t)0);
-        vector signed int vsumi5 = vec_splats((int32_t)0);
-        vector signed int vsumi6 = vec_splats((int32_t)0);
-        vector signed int vsumi7 = vec_splats((int32_t)0);
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
+        vector signed int vsumi2 = v0;
+        vector signed int vsumi3 = v0;
 
         const uint8_t * restrict q3 = x[i].qs;
         const uint32_t * restrict signs = (const uint32_t *)(x[i].qs + QK_K/4);
@@ -10096,21 +10315,12 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * restrict s, size_t bs, const void
             vector signed short vscales01 = (vector signed short)vec_splats((uint16_t)(2*ls0+1));
             vector signed short vscales23 = (vector signed short)vec_splats((uint16_t)(2*ls1+1));
 
-            vsumi0 = vec_add(vec_mule(qv0, vscales01), vsumi0);
-            vsumi1 = vec_add(vec_mule(qv1, vscales01), vsumi1);
-            vsumi2 = vec_add(vec_mule(qv2, vscales23), vsumi2);
-            vsumi3 = vec_add(vec_mule(qv3, vscales23), vsumi3);
-            vsumi4 = vec_add(vec_mulo(qv0, vscales01), vsumi4);
-            vsumi5 = vec_add(vec_mulo(qv1, vscales01), vsumi5);
-            vsumi6 = vec_add(vec_mulo(qv2, vscales23), vsumi6);
-            vsumi7 = vec_add(vec_mulo(qv3, vscales23), vsumi7);
+            vsumi0 = vec_msum(qv0, vscales01, vsumi0);
+            vsumi1 = vec_msum(qv1, vscales01, vsumi1);
+            vsumi2 = vec_msum(qv2, vscales23, vsumi2);
+            vsumi3 = vec_msum(qv3, vscales23, vsumi3);
         }
 
-        vsumi0 = vec_add(vsumi0, vsumi4);
-        vsumi1 = vec_add(vsumi1, vsumi5);
-        vsumi2 = vec_add(vsumi2, vsumi6);
-        vsumi3 = vec_add(vsumi3, vsumi7);
-
         vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
         vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
         vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
@@ -10393,6 +10603,112 @@ void ggml_vec_dot_iq3_s_q8_K (int n, float * restrict s, size_t bs, const void *
 
     *s = hsum_float_8(accumf);
 
+#elif defined(__AVX__)
+   static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
+                                       0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03
+   };
+
+    static const uint8_t k_mask2[32] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+                                        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+    };
+
+    const __m128i mask1_0 = _mm_loadu_si128((const __m128i*)k_mask1);
+    const __m128i mask1_1 = _mm_loadu_si128((const __m128i*)k_mask1 + 1);
+    const __m128i mask2_0 = _mm_loadu_si128((const __m128i*)k_mask2);
+    const __m128i mask2_1 = _mm_loadu_si128((const __m128i*)k_mask2 + 1);
+
+    const __m128i idx_mul_0 = _mm_set_epi32(32, 64, 128, 256);
+    const __m128i idx_mul_1 = _mm_set_epi32(2, 4, 8, 16);
+    const __m128i idx_mask  = _mm_set1_epi32(256);
+
+    typedef union {
+        __m128i  vec[4];
+        uint32_t index[16];
+    } index_t;
+
+    index_t idx;
+
+    __m256 accumf = _mm256_setzero_ps();
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint8_t * restrict qs = x[i].qs;
+        const uint8_t * restrict qh = x[i].qh;
+        const uint16_t * restrict signs = (const uint16_t *)x[i].signs;
+        const int8_t  * restrict q8 = y[i].qs;
+        __m128i sumi1_0 = _mm_setzero_si128();
+        __m128i sumi1_1 = _mm_setzero_si128();
+        __m128i sumi2_0 = _mm_setzero_si128();
+        __m128i sumi2_1 = _mm_setzero_si128();
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+            const __m128i q8_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i qs_tmp = _mm_loadu_si128((const __m128i *)qs);
+            const __m128i idx_l_0 = _mm_cvtepu8_epi16(qs_tmp);
+            const __m128i idx_l_1 = _mm_cvtepu8_epi16(_mm_srli_si128(qs_tmp, 8)); qs += 16;
+            idx.vec[0] = _mm_set1_epi32(qh[ib32+0]);
+            idx.vec[1] = idx.vec[0];
+            idx.vec[2] = _mm_set1_epi32(qh[ib32+1]);
+            idx.vec[3] = idx.vec[2];
+
+            idx.vec[0] = _mm_and_si128(_mm_mullo_epi32(idx.vec[0], idx_mul_0), idx_mask);
+            idx.vec[1] = _mm_and_si128(_mm_mullo_epi32(idx.vec[1], idx_mul_1), idx_mask);
+            idx.vec[2] = _mm_and_si128(_mm_mullo_epi32(idx.vec[2], idx_mul_0), idx_mask);
+            idx.vec[3] = _mm_and_si128(_mm_mullo_epi32(idx.vec[3], idx_mul_1), idx_mask);
+
+            idx.vec[0] = _mm_or_si128(idx.vec[0], _mm_cvtepi16_epi32(idx_l_0));
+            idx.vec[1] = _mm_or_si128(idx.vec[1], _mm_cvtepi16_epi32(_mm_srli_si128(idx_l_0, 8)));
+            idx.vec[2] = _mm_or_si128(idx.vec[2], _mm_cvtepi16_epi32(idx_l_1));
+            idx.vec[3] = _mm_or_si128(idx.vec[3], _mm_cvtepi16_epi32(_mm_srli_si128(idx_l_1, 8)));
+
+            const __m128i q2_1_0 = _mm_set_epi32(iq3s_grid[idx.index[3]], iq3s_grid[idx.index[2]], iq3s_grid[idx.index[1]], iq3s_grid[idx.index[0]]);
+            const __m128i q2_1_1 = _mm_set_epi32(iq3s_grid[idx.index[7]], iq3s_grid[idx.index[6]], iq3s_grid[idx.index[5]], iq3s_grid[idx.index[4]]);
+            const __m128i q2_2_0 = _mm_set_epi32(iq3s_grid[idx.index[11]], iq3s_grid[idx.index[10]], iq3s_grid[idx.index[9]], iq3s_grid[idx.index[8]]);
+            const __m128i q2_2_1 = _mm_set_epi32(iq3s_grid[idx.index[15]], iq3s_grid[idx.index[14]], iq3s_grid[idx.index[13]], iq3s_grid[idx.index[12]]);
+
+            __m128i aux128_0 = _mm_set1_epi32(signs[0] | (signs[1] << 16));
+            __m128i aux128_1 = aux128_0;
+            aux128_0 = _mm_and_si128(_mm_shuffle_epi8(aux128_0,mask1_0), mask2_0);
+            aux128_1 = _mm_and_si128(_mm_shuffle_epi8(aux128_1,mask1_1), mask2_1);
+            const __m128i s2_1_0 = _mm_cmpeq_epi8(aux128_0, mask2_0);
+            const __m128i s2_1_1 = _mm_cmpeq_epi8(aux128_1, mask2_1);
+            const __m128i q8s_1_0 = _mm_sub_epi8(_mm_xor_si128(s2_1_0, q8_1_0), s2_1_0);
+            const __m128i q8s_1_1 = _mm_sub_epi8(_mm_xor_si128(s2_1_1, q8_1_1), s2_1_1);
+
+            aux128_0 = _mm_set1_epi32(signs[2] | (signs[3] << 16));
+            aux128_1 = aux128_0;
+            aux128_0 = _mm_and_si128(_mm_shuffle_epi8(aux128_0,mask1_0), mask2_0);
+            aux128_1 = _mm_and_si128(_mm_shuffle_epi8(aux128_1,mask1_1), mask2_1);
+            const __m128i s2_2_0 = _mm_cmpeq_epi8(aux128_0, mask2_0);
+            const __m128i s2_2_1 = _mm_cmpeq_epi8(aux128_1, mask2_1);
+            const __m128i q8s_2_0 = _mm_sub_epi8(_mm_xor_si128(s2_2_0, q8_2_0), s2_2_0);
+            const __m128i q8s_2_1 = _mm_sub_epi8(_mm_xor_si128(s2_2_1, q8_2_1), s2_2_1);
+
+            signs += 4;
+
+            const __m128i dot1_0  = _mm_maddubs_epi16(q2_1_0, q8s_1_0);
+            const __m128i dot1_1  = _mm_maddubs_epi16(q2_1_1, q8s_1_1);
+            const __m128i dot2_0  = _mm_maddubs_epi16(q2_2_0, q8s_2_0);
+            const __m128i dot2_1  = _mm_maddubs_epi16(q2_2_1, q8s_2_1);
+            const uint16_t ls1 = x[i].scales[ib32/2] & 0xf;
+            const uint16_t ls2 = x[i].scales[ib32/2] >>  4;
+            const __m128i p1_0 = _mm_madd_epi16(dot1_0, _mm_set1_epi16(2*ls1+1));
+            const __m128i p1_1 = _mm_madd_epi16(dot1_1, _mm_set1_epi16(2*ls1+1));
+            const __m128i p2_0 = _mm_madd_epi16(dot2_0, _mm_set1_epi16(2*ls2+1));
+            const __m128i p2_1 = _mm_madd_epi16(dot2_1, _mm_set1_epi16(2*ls2+1));
+            sumi1_0 = _mm_add_epi32(sumi1_0, p1_0);
+            sumi1_1 = _mm_add_epi32(sumi1_1, p1_1);
+            sumi2_0 = _mm_add_epi32(sumi2_0, p2_0);
+            sumi2_1 = _mm_add_epi32(sumi2_1, p2_1);
+        }
+
+        accumf = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(MM256_SET_M128I(_mm_add_epi32(sumi1_1, sumi2_1), _mm_add_epi32(sumi1_0, sumi2_0)))), accumf);
+
+    }
+
+    *s = hsum_float_8(accumf);
+
 #elif defined(__POWER9_VECTOR__)
     static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
                                         0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03
@@ -10400,6 +10716,8 @@ void ggml_vec_dot_iq3_s_q8_K (int n, float * restrict s, size_t bs, const void *
 
     static const uint8_t k_mask2[16] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,};
 
+    const vector int v0 = vec_splats((int32_t)0);
+
     vector float vsumf0 = vec_splats(0.0f);
     vector float vsumf1 = vec_splats(0.0f);
     vector float vsumf2 = vec_splats(0.0f);
@@ -10420,14 +10738,10 @@ void ggml_vec_dot_iq3_s_q8_K (int n, float * restrict s, size_t bs, const void *
         const uint8_t *  restrict sc = x[i].scales;
         const int8_t  *  restrict q8 = y[i].qs;
 
-        vector signed int vsumi0 = vec_splats((int32_t)0);
-        vector signed int vsumi1 = vec_splats((int32_t)0);
-        vector signed int vsumi2 = vec_splats((int32_t)0);
-        vector signed int vsumi3 = vec_splats((int32_t)0);
-        vector signed int vsumi4 = vec_splats((int32_t)0);
-        vector signed int vsumi5 = vec_splats((int32_t)0);
-        vector signed int vsumi6 = vec_splats((int32_t)0);
-        vector signed int vsumi7 = vec_splats((int32_t)0);
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
+        vector signed int vsumi2 = v0;
+        vector signed int vsumi3 = v0;
 
         for (int j = 0; j < QK_K/32; j += 2) {
             __builtin_prefetch(q3, 0, 1);
@@ -10481,21 +10795,12 @@ void ggml_vec_dot_iq3_s_q8_K (int n, float * restrict s, size_t bs, const void *
             vector signed short vscales01 = (vector signed short)vec_splats((uint16_t)(2*ls0+1));
             vector signed short vscales23 = (vector signed short)vec_splats((uint16_t)(2*ls1+1));
 
-            vsumi0 = vec_add(vec_mule(qv0, vscales01), vsumi0);
-            vsumi1 = vec_add(vec_mule(qv1, vscales01), vsumi1);
-            vsumi2 = vec_add(vec_mule(qv2, vscales23), vsumi2);
-            vsumi3 = vec_add(vec_mule(qv3, vscales23), vsumi3);
-            vsumi4 = vec_add(vec_mulo(qv0, vscales01), vsumi4);
-            vsumi5 = vec_add(vec_mulo(qv1, vscales01), vsumi5);
-            vsumi6 = vec_add(vec_mulo(qv2, vscales23), vsumi6);
-            vsumi7 = vec_add(vec_mulo(qv3, vscales23), vsumi7);
+            vsumi0 = vec_msum(qv0, vscales01, vsumi0);
+            vsumi1 = vec_msum(qv1, vscales01, vsumi1);
+            vsumi2 = vec_msum(qv2, vscales23, vsumi2);
+            vsumi3 = vec_msum(qv3, vscales23, vsumi3);
         }
 
-        vsumi0 = vec_add(vsumi0, vsumi4);
-        vsumi1 = vec_add(vsumi1, vsumi5);
-        vsumi2 = vec_add(vsumi2, vsumi6);
-        vsumi3 = vec_add(vsumi3, vsumi7);
-
         vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
         vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
         vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
@@ -10641,6 +10946,14 @@ void ggml_vec_dot_iq3_s_q8_K (int n, float * restrict s, size_t bs, const void *
 }
 
 
+#if defined(__AVX__)
+static inline __m128i mul_add_epi8_sse(const __m128i x, const __m128i y) {
+    const __m128i ax = _mm_sign_epi8(x, x);
+    const __m128i sy = _mm_sign_epi8(y, x);
+    return _mm_maddubs_epi16(ax, sy);
+}
+#endif
+
 #if defined(__AVX2__)
 static inline __m256i mul_add_epi8(const __m256i x, const __m256i y) {
     const __m256i ax = _mm256_sign_epi8(x, x);
@@ -10758,6 +11071,54 @@ void ggml_vec_dot_iq1_s_q8_K  (int n, float * restrict s, size_t bs, const void
 
     *s = hsum_float_8(accum) + IQ1S_DELTA * accum1;
 
+#elif defined __AVX__
+    __m256 accum = _mm256_setzero_ps();
+    float accum1 = 0;
+    for (int i = 0; i < nb; ++i) {
+
+        const int8_t   * q8 = y[i].qs;
+        const uint8_t  * qs = x[i].qs;
+        const uint16_t * qh = x[i].qh;
+
+        __m128i sumi1_0 = _mm_setzero_si128();
+        __m128i sumi1_1 = _mm_setzero_si128();
+        int sumi1 = 0;
+        for (int ib = 0; ib < QK_K/32; ib += 2) {
+            const __m128i q1b_1_0 = _mm_set_epi64x(iq1s_grid[qs[1] | ((qh[ib+0] << 5) & 0x700)], iq1s_grid[qs[0] | ((qh[ib+0] << 8) & 0x700)]);
+            const __m128i q1b_1_1 = _mm_set_epi64x(iq1s_grid[qs[3] | ((qh[ib+0] >> 1) & 0x700)], iq1s_grid[qs[2] | ((qh[ib+0] << 2) & 0x700)]);
+            const __m128i q1b_2_0 = _mm_set_epi64x(iq1s_grid[qs[5] | ((qh[ib+1] << 5) & 0x700)], iq1s_grid[qs[4] | ((qh[ib+1] << 8) & 0x700)]);
+            const __m128i q1b_2_1 = _mm_set_epi64x(iq1s_grid[qs[7] | ((qh[ib+1] >> 1) & 0x700)], iq1s_grid[qs[6] | ((qh[ib+1] << 2) & 0x700)]);
+            qs += 8;
+            const __m128i q8b_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8b_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8b_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8b_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+
+            const __m128i dot1_0 = mul_add_epi8_sse(q1b_1_0, q8b_1_0);
+            const __m128i dot1_1 = mul_add_epi8_sse(q1b_1_1, q8b_1_1);
+            const __m128i dot2_0 = mul_add_epi8_sse(q1b_2_0, q8b_2_0);
+            const __m128i dot2_1 = mul_add_epi8_sse(q1b_2_1, q8b_2_1);
+            const int16_t ls1 = 2*((qh[ib+0] >> 12) & 7) + 1;
+            const int16_t ls2 = 2*((qh[ib+1] >> 12) & 7) + 1;
+            const __m128i p1_0 = _mm_madd_epi16(dot1_0, _mm_set1_epi16(ls1));
+            const __m128i p1_1 = _mm_madd_epi16(dot1_1, _mm_set1_epi16(ls1));
+            const __m128i p2_0 = _mm_madd_epi16(dot2_0, _mm_set1_epi16(ls2));
+            const __m128i p2_1 = _mm_madd_epi16(dot2_1, _mm_set1_epi16(ls2));
+
+            sumi1_0 = _mm_add_epi32(sumi1_0, _mm_add_epi32(p1_0, p2_0));
+            sumi1_1 = _mm_add_epi32(sumi1_1, _mm_add_epi32(p1_1, p2_1));
+            sumi1 += (y[i].bsums[2*ib+0] + y[i].bsums[2*ib+1]) * (qh[ib+0] & 0x8000 ? -1 : 1) * ls1
+                   + (y[i].bsums[2*ib+2] + y[i].bsums[2*ib+3]) * (qh[ib+1] & 0x8000 ? -1 : 1) * ls2;
+        }
+
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        accum = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(MM256_SET_M128I(sumi1_1, sumi1_0))), accum);
+        accum1 += d * sumi1;
+
+    }
+
+    *s = hsum_float_8(accum) + IQ1S_DELTA * accum1;
+
 #elif defined(__POWER9_VECTOR__)
     const vector unsigned char v0 = vec_splats((unsigned char)0x0);
     const vector unsigned short vsign = vec_splats((unsigned short)0x8000);
@@ -10776,10 +11137,6 @@ void ggml_vec_dot_iq1_s_q8_K  (int n, float * restrict s, size_t bs, const void
         vector signed int vsumi1 = vec_splats((int32_t)0);
         vector signed int vsumi2 = vec_splats((int32_t)0);
         vector signed int vsumi3 = vec_splats((int32_t)0);
-        vector signed int vsumi4 = vec_splats((int32_t)0);
-        vector signed int vsumi5 = vec_splats((int32_t)0);
-        vector signed int vsumi6 = vec_splats((int32_t)0);
-        vector signed int vsumi7 = vec_splats((int32_t)0);
         vector signed int vsumi8 = vec_splats((int32_t)0);
 
         const uint8_t  * restrict q1 = x[i].qs;
@@ -10821,14 +11178,10 @@ void ggml_vec_dot_iq1_s_q8_K  (int n, float * restrict s, size_t bs, const void
             vector signed short vscales23 = (vector signed short)vec_splats((uint16_t)(2*ls1+1));
             vector signed short vscales = vec_sld(vscales23, vscales01, 8);
 
-            vsumi0 = vec_add(vec_mule(qv0, vscales01), vsumi0);
-            vsumi1 = vec_add(vec_mule(qv1, vscales01), vsumi1);
-            vsumi2 = vec_add(vec_mule(qv2, vscales23), vsumi2);
-            vsumi3 = vec_add(vec_mule(qv3, vscales23), vsumi3);
-            vsumi4 = vec_add(vec_mulo(qv0, vscales01), vsumi4);
-            vsumi5 = vec_add(vec_mulo(qv1, vscales01), vsumi5);
-            vsumi6 = vec_add(vec_mulo(qv2, vscales23), vsumi6);
-            vsumi7 = vec_add(vec_mulo(qv3, vscales23), vsumi7);
+            vsumi0 = vec_msum(qv0, vscales01, vsumi0);
+            vsumi1 = vec_msum(qv1, vscales01, vsumi1);
+            vsumi2 = vec_msum(qv2, vscales23, vsumi2);
+            vsumi3 = vec_msum(qv3, vscales23, vsumi3);
 
             vector signed short q8ysums = vec_xl_len(qs, 8);
             qs += 4;
@@ -10843,11 +11196,6 @@ void ggml_vec_dot_iq1_s_q8_K  (int n, float * restrict s, size_t bs, const void
             vsumi8 = vec_add(vec_mule(q8ysum, vscales), vsumi8);
         }
 
-        vsumi0 = vec_add(vsumi0, vsumi4);
-        vsumi1 = vec_add(vsumi1, vsumi5);
-        vsumi2 = vec_add(vsumi2, vsumi6);
-        vsumi3 = vec_add(vsumi3, vsumi7);
-
         vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
         vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
         vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
@@ -11109,6 +11457,92 @@ void ggml_vec_dot_iq1_m_q8_K  (int n, float * restrict s, size_t bs, const void
 
     *s = hsum_float_8(accum1) + IQ1M_DELTA * hsum_float_8(accum2);
 
+#elif defined __AVX__
+    const __m128i mask = _mm_set1_epi16(0x7);
+    const __m128i mone = _mm_set1_epi16(1);
+
+    __m256 accum1 = _mm256_setzero_ps();
+    __m256 accum2 = _mm256_setzero_ps();
+    for (int i = 0; i < nb; ++i) {
+
+        const int8_t   * q8 = y[i].qs;
+        const uint8_t  * qs = x[i].qs;
+        const uint8_t  * qh = x[i].qh;
+        const uint16_t * sc = (const uint16_t *)x[i].scales;
+
+        scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
+
+        __m128i sumi1_0 = _mm_setzero_si128();
+        __m128i sumi1_1 = _mm_setzero_si128();
+        __m128i sumi2_0 = _mm_setzero_si128();
+        __m128i sumi2_1 = _mm_setzero_si128();
+        for (int ib = 0; ib < QK_K/32; ib += 2) {
+            const __m128i q1b_1_0 = _mm_set_epi64x(
+                    iq1s_grid[qs[1] | (((uint16_t)qh[0] << 4) & 0x700)], iq1s_grid[qs[0] | (((uint16_t)qh[0] << 8) & 0x700)]);
+            const __m128i q1b_1_1 = _mm_set_epi64x(
+                    iq1s_grid[qs[3] | (((uint16_t)qh[1] << 4) & 0x700)], iq1s_grid[qs[2] | (((uint16_t)qh[1] << 8) & 0x700)]);
+            const __m128i q1b_2_0 = _mm_set_epi64x(
+                    iq1s_grid[qs[5] | (((uint16_t)qh[2] << 4) & 0x700)], iq1s_grid[qs[4] | (((uint16_t)qh[2] << 8) & 0x700)]);
+            const __m128i q1b_2_1 = _mm_set_epi64x(
+                    iq1s_grid[qs[7] | (((uint16_t)qh[3] << 4) & 0x700)], iq1s_grid[qs[6] | (((uint16_t)qh[3] << 8) & 0x700)]);
+            const __m128i q8b_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8b_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8b_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8b_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+
+            const __m128i dot1_0 = mul_add_epi8_sse(q1b_1_0, q8b_1_0);
+            const __m128i dot1_1 = mul_add_epi8_sse(q1b_1_1, q8b_1_1);
+            const __m128i dot2_0 = mul_add_epi8_sse(q1b_2_0, q8b_2_0);
+            const __m128i dot2_1 = mul_add_epi8_sse(q1b_2_1, q8b_2_1);
+
+            const __m128i delta1_0 = _mm_set_epi64x(qh[0] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101,
+                                                     qh[0] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101);
+            const __m128i delta1_1 = _mm_set_epi64x(qh[1] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101,
+                                                     qh[1] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101);
+            const __m128i delta2_0 = _mm_set_epi64x(qh[2] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101,
+                                                     qh[2] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101);
+            const __m128i delta2_1 = _mm_set_epi64x(qh[3] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101,
+                                                     qh[3] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101);
+
+            const __m128i dot3_0 = mul_add_epi8_sse(delta1_0, q8b_1_0);
+            const __m128i dot3_1 = mul_add_epi8_sse(delta1_1, q8b_1_1);
+            const __m128i dot4_0 = mul_add_epi8_sse(delta2_0, q8b_2_0);
+            const __m128i dot4_1 = mul_add_epi8_sse(delta2_1, q8b_2_1);
+
+            __m128i scale1_0 = _mm_set1_epi16(sc[ib/2] >> 0);
+            __m128i scale1_1 = _mm_set1_epi16(sc[ib/2] >> 3);
+            __m128i scale2_0 = _mm_set1_epi16(sc[ib/2] >> 6);
+            __m128i scale2_1 = _mm_set1_epi16(sc[ib/2] >> 9);
+
+            scale1_0 = _mm_add_epi16(_mm_slli_epi16(_mm_and_si128(scale1_0, mask), 1), mone);
+            scale1_1 = _mm_add_epi16(_mm_slli_epi16(_mm_and_si128(scale1_1, mask), 1), mone);
+            scale2_0 = _mm_add_epi16(_mm_slli_epi16(_mm_and_si128(scale2_0, mask), 1), mone);
+            scale2_1 = _mm_add_epi16(_mm_slli_epi16(_mm_and_si128(scale2_1, mask), 1), mone);
+            const __m128i p1_0 = _mm_madd_epi16(dot1_0, scale1_0);
+            const __m128i p1_1 = _mm_madd_epi16(dot1_1, scale1_1);
+            const __m128i p2_0 = _mm_madd_epi16(dot2_0, scale2_0);
+            const __m128i p2_1 = _mm_madd_epi16(dot2_1, scale2_1);
+            const __m128i p3_0 = _mm_madd_epi16(dot3_0, scale1_0);
+            const __m128i p3_1 = _mm_madd_epi16(dot3_1, scale1_1);
+            const __m128i p4_0 = _mm_madd_epi16(dot4_0, scale2_0);
+            const __m128i p4_1 = _mm_madd_epi16(dot4_1, scale2_1);
+
+            sumi1_0 = _mm_add_epi32(sumi1_0, _mm_add_epi32(p1_0, p2_0));
+            sumi1_1 = _mm_add_epi32(sumi1_1, _mm_add_epi32(p1_1, p2_1));
+            sumi2_0 = _mm_add_epi32(sumi2_0, _mm_add_epi32(p3_0, p4_0));
+            sumi2_1 = _mm_add_epi32(sumi2_1, _mm_add_epi32(p3_1, p4_1));
+
+            qs += 8; qh += 4;
+        }
+
+        const __m256 d = _mm256_set1_ps(y[i].d * GGML_FP16_TO_FP32(scale.f16));
+
+        accum1 = _mm256_add_ps(_mm256_mul_ps(d, _mm256_cvtepi32_ps(MM256_SET_M128I(sumi1_1, sumi1_0))), accum1);
+        accum2 = _mm256_add_ps(_mm256_mul_ps(d, _mm256_cvtepi32_ps(MM256_SET_M128I(sumi2_1, sumi2_0))), accum2);
+    }
+
+    *s = hsum_float_8(accum1) + IQ1M_DELTA * hsum_float_8(accum2);
+
 #else
 
     int sum1[2], sum2[2], delta[4];
@@ -11173,6 +11607,9 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * restrict s, size_t bs, const void *
 
     const int nb = n / QK4_NL;
 
+    int ib = 0;
+    float sumf = 0;
+
 #if defined __ARM_NEON
     const int8x16_t values = vld1q_s8(kvalues_iq4nl);
     const uint8x16_t m4b = vdupq_n_u8(0x0f);
@@ -11181,16 +11618,14 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * restrict s, size_t bs, const void *
     int8x16x4_t q8b;
     int32x4_t prod_1, prod_2;
 
-    float sumf = 0;
-
-    for (int ib = 0; ib < nb; ib += 2) {
+    for (; ib + 1 < nb; ib += 2) {
 
-        q4bits.val[0] = vld1q_u8(x[ib+0].qs);
-        q4bits.val[1] = vld1q_u8(x[ib+1].qs);
-        q8b.val[0]    = vld1q_s8(y[ib+0].qs);
-        q8b.val[1]    = vld1q_s8(y[ib+0].qs + 16);
-        q8b.val[2]    = vld1q_s8(y[ib+1].qs);
-        q8b.val[3]    = vld1q_s8(y[ib+1].qs + 16);
+        q4bits.val[0] = vld1q_u8(x[ib + 0].qs);
+        q4bits.val[1] = vld1q_u8(x[ib + 1].qs);
+        q8b.val[0]    = vld1q_s8(y[ib + 0].qs);
+        q8b.val[1]    = vld1q_s8(y[ib + 0].qs + 16);
+        q8b.val[2]    = vld1q_s8(y[ib + 1].qs);
+        q8b.val[3]    = vld1q_s8(y[ib + 1].qs + 16);
 
         q4b.val[0] = ggml_vqtbl1q_s8(values, vandq_u8  (q4bits.val[0], m4b));
         q4b.val[1] = ggml_vqtbl1q_s8(values, vshrq_n_u8(q4bits.val[0], 4));
@@ -11201,12 +11636,10 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * restrict s, size_t bs, const void *
         prod_2 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q4b.val[2], q8b.val[2]), q4b.val[3], q8b.val[3]);
 
         sumf +=
-            GGML_FP16_TO_FP32(x[ib+0].d) * GGML_FP16_TO_FP32(y[ib+0].d) * vaddvq_s32(prod_1) +
-            GGML_FP16_TO_FP32(x[ib+1].d) * GGML_FP16_TO_FP32(y[ib+1].d) * vaddvq_s32(prod_2);
+            GGML_FP16_TO_FP32(x[ib+0].d) * GGML_FP16_TO_FP32(y[ib + 0].d) * vaddvq_s32(prod_1) +
+            GGML_FP16_TO_FP32(x[ib+1].d) * GGML_FP16_TO_FP32(y[ib + 1].d) * vaddvq_s32(prod_2);
     }
 
-    *s = sumf;
-
 #elif defined __AVX2__
 
     const __m128i values128 = _mm_loadu_si128((const __m128i*)kvalues_iq4nl);
@@ -11215,11 +11648,11 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * restrict s, size_t bs, const void *
 
     __m256 accum1 = _mm256_setzero_ps();
     __m256 accum2 = _mm256_setzero_ps();
-    for (int ib = 0; ib < nb; ib += 2) {
-        const __m128i q4bits_1 = _mm_loadu_si128((const __m128i*)x[0].qs);
-        const __m128i q4bits_2 = _mm_loadu_si128((const __m128i*)x[1].qs);
-        const __m256i q8b_1 = _mm256_loadu_si256((const __m256i *)y[0].qs);
-        const __m256i q8b_2 = _mm256_loadu_si256((const __m256i *)y[1].qs);
+    for (; ib + 1 < nb; ib += 2) {
+        const __m128i q4bits_1 = _mm_loadu_si128((const __m128i*)x[ib + 0].qs);
+        const __m128i q4bits_2 = _mm_loadu_si128((const __m128i*)x[ib + 1].qs);
+        const __m256i q8b_1 = _mm256_loadu_si256((const __m256i *)y[ib + 0].qs);
+        const __m256i q8b_2 = _mm256_loadu_si256((const __m256i *)y[ib + 1].qs);
         const __m256i q4b_1 = MM256_SET_M128I(_mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_1, 4), m4b)),
                                               _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_1, m4b)));
         const __m256i q4b_2 = MM256_SET_M128I(_mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_2, 4), m4b)),
@@ -11228,19 +11661,52 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * restrict s, size_t bs, const void *
         const __m256i p16_2 = mul_add_epi8(q4b_2, q8b_2);
         const __m256i p_1 = _mm256_madd_epi16(p16_1, mone);
         const __m256i p_2 = _mm256_madd_epi16(p16_2, mone);
-        accum1 = _mm256_fmadd_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(y[0].d)*GGML_FP16_TO_FP32(x[0].d)),
+        accum1 = _mm256_fmadd_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(y[ib + 0].d)*GGML_FP16_TO_FP32(x[ib + 0].d)),
                 _mm256_cvtepi32_ps(p_1), accum1);
-        accum2 = _mm256_fmadd_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(y[1].d)*GGML_FP16_TO_FP32(x[1].d)),
+        accum2 = _mm256_fmadd_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(y[ib + 1].d)*GGML_FP16_TO_FP32(x[ib + 1].d)),
                 _mm256_cvtepi32_ps(p_2), accum2);
-
-        y += 2;
-        x += 2;
     }
 
-    *s = hsum_float_8(_mm256_add_ps(accum1, accum2));
+    sumf = hsum_float_8(_mm256_add_ps(accum1, accum2));
+
+#elif defined __AVX__
+    const __m128i values128 = _mm_loadu_si128((const __m128i*)kvalues_iq4nl);
+    const __m128i m4b  = _mm_set1_epi8(0x0f);
+    const __m128i mone = _mm_set1_epi16(1);
+
+    __m256 accum1 = _mm256_setzero_ps();
+    __m256 accum2 = _mm256_setzero_ps();
+    for (; ib + 1 < nb; ib += 2) {
+        const __m128i q4bits_1 = _mm_loadu_si128((const __m128i *)x[ib + 0].qs);
+        const __m128i q4bits_2 = _mm_loadu_si128((const __m128i *)x[ib + 1].qs);
+        const __m128i q8b_1_0 = _mm_loadu_si128((const __m128i *)y[ib + 0].qs);
+        const __m128i q8b_1_1 = _mm_loadu_si128((const __m128i *)y[ib + 0].qs + 1);
+        const __m128i q8b_2_0 = _mm_loadu_si128((const __m128i *)y[ib + 1].qs);
+        const __m128i q8b_2_1 = _mm_loadu_si128((const __m128i *)y[ib + 1].qs + 1);
+
+        const __m128i q4b_1_0 = _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_1, m4b));
+        const __m128i q4b_1_1 = _mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_1, 4), m4b));
+        const __m128i q4b_2_0 = _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_2, m4b));
+        const __m128i q4b_2_1 = _mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_2, 4), m4b));
+        const __m128i p16_1_0 = mul_add_epi8_sse(q4b_1_0, q8b_1_0);
+        const __m128i p16_1_1 = mul_add_epi8_sse(q4b_1_1, q8b_1_1);
+        const __m128i p16_2_0 = mul_add_epi8_sse(q4b_2_0, q8b_2_0);
+        const __m128i p16_2_1 = mul_add_epi8_sse(q4b_2_1, q8b_2_1);
+        const __m128i p_1_0 = _mm_madd_epi16(p16_1_0, mone);
+        const __m128i p_1_1 = _mm_madd_epi16(p16_1_1, mone);
+        const __m128i p_2_0 = _mm_madd_epi16(p16_2_0, mone);
+        const __m128i p_2_1 = _mm_madd_epi16(p16_2_1, mone);
+        accum1 = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(y[ib + 0].d)*GGML_FP16_TO_FP32(x[ib + 0].d)),
+                _mm256_cvtepi32_ps(MM256_SET_M128I(p_1_1, p_1_0))), accum1);
+        accum2 = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(y[ib + 1].d)*GGML_FP16_TO_FP32(x[ib + 1].d)),
+                _mm256_cvtepi32_ps(MM256_SET_M128I(p_2_1, p_2_0))), accum2);
+    }
+
+    sumf = hsum_float_8(_mm256_add_ps(accum1, accum2));
 
 #elif defined(__POWER9_VECTOR__)
     const vector signed char lowMask = vec_splats((signed char)0xF);
+    const vector signed int v0 = vec_splats((int32_t)0);
     const vector unsigned char v4 = vec_splats((unsigned char)0x4);
 
     vector float vsumf0 = vec_splats(0.0f);
@@ -11249,7 +11715,7 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * restrict s, size_t bs, const void *
     const vector signed char values = vec_xl( 0, kvalues_iq4nl);
 
 #pragma GCC unroll 4
-    for (int ib = 0; ib < nb; ++ib) {
+    for (; ib < nb; ++ib) {
         __builtin_prefetch(x[ib].qs, 0, 1);
         __builtin_prefetch(y[ib].qs, 0, 1);
 
@@ -11271,8 +11737,11 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * restrict s, size_t bs, const void *
         vector signed short qv0 = vec_add(vec_mule(q4x0, q8y0), vec_mulo(q4x0, q8y0));
         vector signed short qv1 = vec_add(vec_mule(q4x1, q8y1), vec_mulo(q4x1, q8y1));
 
-        vector signed int vsumi0 = vec_add(vec_unpackh(qv0), vec_unpackl(qv0));
-        vector signed int vsumi1 = vec_add(vec_unpackh(qv1), vec_unpackl(qv1));
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
+
+        vsumi0 = vec_sum4s(qv0, vsumi0);
+        vsumi1 = vec_sum4s(qv1, vsumi1);
 
         vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
         vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
@@ -11283,7 +11752,7 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * restrict s, size_t bs, const void *
     vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
     vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
 
-    *s = vec_extract(vsumf0, 0);
+    sumf = vec_extract(vsumf0, 0);
 
 #elif defined (__loongarch_asx)
 
@@ -11293,11 +11762,11 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * restrict s, size_t bs, const void *
 
     __m256 accum1 = (__m256)__lasx_xvldi(0);
     __m256 accum2 = (__m256)__lasx_xvldi(0);
-    for (int ib = 0; ib < nb; ib += 2) {
-        const __m128i q4bits_1 = __lsx_vld((const __m128i*)x[0].qs, 0);
-        const __m128i q4bits_2 = __lsx_vld((const __m128i*)x[1].qs, 0);
-        const __m256i q8b_1 = __lasx_xvld((const __m256i *)y[0].qs, 0);
-        const __m256i q8b_2 = __lasx_xvld((const __m256i *)y[1].qs, 0);
+    for (; ib + 1 < nb; ib += 2) {
+        const __m128i q4bits_1 = __lsx_vld((const __m128i*)x[ib + 0].qs, 0);
+        const __m128i q4bits_2 = __lsx_vld((const __m128i*)x[ib + 1].qs, 0);
+        const __m256i q8b_1 = __lasx_xvld((const __m256i *)y[ib + 0].qs, 0);
+        const __m256i q8b_2 = __lasx_xvld((const __m256i *)y[ib + 1].qs, 0);
         const __m256i q4b_1 = lasx_insertf128(lsx_shuffle_b(values128, __lsx_vand_v(__lsx_vsrli_h(q4bits_1, 4), m4b)),
                                               lsx_shuffle_b(values128, __lsx_vand_v(q4bits_1, m4b)));
         const __m256i q4b_2 = lasx_insertf128(lsx_shuffle_b(values128, __lsx_vand_v(__lsx_vsrli_h(q4bits_2, 4), m4b)),
@@ -11306,20 +11775,16 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * restrict s, size_t bs, const void *
         const __m256i p16_2 = mul_add_epi8(q4b_2, q8b_2);
         const __m256i p_1 = lasx_madd_h(p16_1, mone);
         const __m256i p_2 = lasx_madd_h(p16_2, mone);
-        accum1 = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(y[0].d)*GGML_FP16_TO_FP32(x[0].d)),
+        accum1 = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(y[ib + 0].d)*GGML_FP16_TO_FP32(x[ib + 0].d)),
                 __lasx_xvffint_s_w(p_1), accum1);
-        accum2 = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(y[1].d)*GGML_FP16_TO_FP32(x[1].d)),
+        accum2 = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(y[ib + 1].d)*GGML_FP16_TO_FP32(x[ib + 1].d)),
                 __lasx_xvffint_s_w(p_2), accum2);
-
-        y += 2;
-        x += 2;
     }
 
-    *s = hsum_float_8(__lasx_xvfadd_s(accum1, accum2));
+    sumf = hsum_float_8(__lasx_xvfadd_s(accum1, accum2));
 
-#else
-    float sumf = 0;
-    for (int ib = 0; ib < nb; ++ib) {
+#endif
+    for (; ib < nb; ++ib) {
         const float d = GGML_FP16_TO_FP32(y[ib].d)*GGML_FP16_TO_FP32(x[ib].d);
         int sumi1 = 0, sumi2 = 0;
         for (int j = 0; j < QK4_NL/2; ++j) {
@@ -11329,7 +11794,6 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * restrict s, size_t bs, const void *
         sumf += d * (sumi1 + sumi2);
     }
     *s = sumf;
-#endif
 }
 
 void ggml_vec_dot_iq4_xs_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
@@ -11425,8 +11889,57 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * restrict s, size_t bs, const void *
 
     *s = hsum_float_8(accum);
 
+#elif defined __AVX__
+    const __m128i values128 = _mm_loadu_si128((const __m128i*)kvalues_iq4nl);
+    const __m128i m4b  = _mm_set1_epi8(0x0f);
+
+    __m256 accum = _mm256_setzero_ps();
+    for (int ibl = 0; ibl < nb; ++ibl) {
+        const uint8_t * qs = x[ibl].qs;
+        const int8_t  * q8 = y[ibl].qs;
+        uint16_t sh = x[ibl].scales_h;
+        __m128i sumi1_0 = _mm_setzero_si128();
+        __m128i sumi1_1 = _mm_setzero_si128();
+        __m128i sumi2_0 = _mm_setzero_si128();
+        __m128i sumi2_1 = _mm_setzero_si128();
+        for (int ib = 0; ib < QK_K/32; ib += 2) {
+            const __m128i q4bits_1 = _mm_loadu_si128((const __m128i *)qs); qs += 16;
+            const __m128i q4bits_2 = _mm_loadu_si128((const __m128i *)qs); qs += 16;
+            const __m128i q8b_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8b_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8b_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8b_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q4b_1_0 = _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_1, m4b));
+            const __m128i q4b_1_1 = _mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_1, 4), m4b));
+            const __m128i q4b_2_0 = _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_2, m4b));
+            const __m128i q4b_2_1 = _mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_2, 4), m4b));
+            const __m128i p16_1_0 = mul_add_epi8_sse(q4b_1_0, q8b_1_0);
+            const __m128i p16_1_1 = mul_add_epi8_sse(q4b_1_1, q8b_1_1);
+            const __m128i p16_2_0 = mul_add_epi8_sse(q4b_2_0, q8b_2_0);
+            const __m128i p16_2_1 = mul_add_epi8_sse(q4b_2_1, q8b_2_1);
+            const int16_t ls1 = ((x[ibl].scales_l[ib/2] & 0xf) | ((sh << 4) & 0x30)) - 32;
+            const int16_t ls2 = ((x[ibl].scales_l[ib/2] >>  4) | ((sh << 2) & 0x30)) - 32;
+            sh >>= 4;
+            const __m128i p_1_0 = _mm_madd_epi16(p16_1_0, _mm_set1_epi16(ls1));
+            const __m128i p_1_1 = _mm_madd_epi16(p16_1_1, _mm_set1_epi16(ls1));
+            const __m128i p_2_0 = _mm_madd_epi16(p16_2_0, _mm_set1_epi16(ls2));
+            const __m128i p_2_1 = _mm_madd_epi16(p16_2_1, _mm_set1_epi16(ls2));
+            sumi1_0 = _mm_add_epi32(p_1_0, sumi1_0);
+            sumi1_1 = _mm_add_epi32(p_1_1, sumi1_1);
+            sumi2_0 = _mm_add_epi32(p_2_0, sumi2_0);
+            sumi2_1 = _mm_add_epi32(p_2_1, sumi2_1);
+        }
+        __m128i sumi12_0 = _mm_add_epi32(sumi1_0, sumi2_0);
+        __m128i sumi12_1 = _mm_add_epi32(sumi1_1, sumi2_1);
+        accum = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(x[ibl].d)*y[ibl].d),
+                _mm256_cvtepi32_ps(MM256_SET_M128I(sumi12_1, sumi12_0))), accum);
+    }
+
+    *s = hsum_float_8(accum);
+
 #elif defined(__POWER9_VECTOR__)
     const vector signed char lowMask = vec_splats((signed char)0xF);
+    const vector int v0 = vec_splats((int32_t)0);
     const vector unsigned char v4 = vec_splats((unsigned char)0x4);
 
     vector float vsumf0 = vec_splats(0.0f);
@@ -11442,14 +11955,10 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * restrict s, size_t bs, const void *
         vector float vyd = vec_splats(y[ibl].d);
         vector float vd = vec_mul(vxd, vyd);
 
-        vector signed int vsumi0 = vec_splats((int32_t)0);
-        vector signed int vsumi1 = vec_splats((int32_t)0);
-        vector signed int vsumi2 = vec_splats((int32_t)0);
-        vector signed int vsumi3 = vec_splats((int32_t)0);
-        vector signed int vsumi4 = vec_splats((int32_t)0);
-        vector signed int vsumi5 = vec_splats((int32_t)0);
-        vector signed int vsumi6 = vec_splats((int32_t)0);
-        vector signed int vsumi7 = vec_splats((int32_t)0);
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
+        vector signed int vsumi2 = v0;
+        vector signed int vsumi3 = v0;
 
         uint16_t h = x[ibl].scales_h;
 
@@ -11494,21 +12003,12 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * restrict s, size_t bs, const void *
             vector signed short vscales01 = vec_splats((int16_t)ls0);
             vector signed short vscales23 = vec_splats((int16_t)ls1);
 
-            vsumi0 = vec_add(vec_mule(qv0, vscales01), vsumi0);
-            vsumi1 = vec_add(vec_mule(qv1, vscales01), vsumi1);
-            vsumi2 = vec_add(vec_mule(qv2, vscales23), vsumi2);
-            vsumi3 = vec_add(vec_mule(qv3, vscales23), vsumi3);
-            vsumi4 = vec_add(vec_mulo(qv0, vscales01), vsumi4);
-            vsumi5 = vec_add(vec_mulo(qv1, vscales01), vsumi5);
-            vsumi6 = vec_add(vec_mulo(qv2, vscales23), vsumi6);
-            vsumi7 = vec_add(vec_mulo(qv3, vscales23), vsumi7);
+            vsumi0 = vec_msum(qv0, vscales01, vsumi0);
+            vsumi1 = vec_msum(qv1, vscales01, vsumi1);
+            vsumi2 = vec_msum(qv2, vscales23, vsumi2);
+            vsumi3 = vec_msum(qv3, vscales23, vsumi3);
         }
 
-        vsumi0 = vec_add(vsumi0, vsumi4);
-        vsumi1 = vec_add(vsumi1, vsumi5);
-        vsumi2 = vec_add(vsumi2, vsumi6);
-        vsumi3 = vec_add(vsumi3, vsumi7);
-
         vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
         vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
         vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
@@ -12204,7 +12704,7 @@ static void quantize_row_iq2_xxs_impl(const float * restrict x, void * restrict
                     printf("Oops: found point %u not on grid:", u);
                     for (int i = 0; i < 8; ++i) printf(" %d", L[8*k+i]);
                     printf("\n");
-                    GGML_ASSERT(false);
+                    GGML_ABORT("fatal error");
                 }
                 q2[2*ib+0] |= ((uint32_t) grid_index << 8*k);
                 q2[2*ib+1] |= (block_signs[k] << 7*k);
@@ -12383,7 +12883,7 @@ static void quantize_row_iq2_xs_impl(const float * restrict x, void * restrict v
                     printf("Oops: found point %u not on grid:", u);
                     for (int i = 0; i < 8; ++i) printf(" %d", L[8*k+i]);
                     printf("\n");
-                    GGML_ASSERT(false);
+                    GGML_ABORT("fatal error");
                 }
                 q2[2*ib+k] = grid_index | (block_signs[k] << 9);
             }
@@ -12826,7 +13326,7 @@ static void quantize_row_iq3_xxs_impl(int grid_size, const float * restrict x, v
                     printf("Oops: found point %u not on grid:", u);
                     for (int i = 0; i < 4; ++i) printf(" %d", L[4*k+i]);
                     printf("\n");
-                    GGML_ASSERT(false);
+                    GGML_ABORT("fatal error");
                 }
                 if (grid_size == 256) {
                     q3[8*ib+k] = grid_index;
@@ -12880,10 +13380,10 @@ size_t quantize_iq3_xxs(const float * restrict src, void * restrict dst, int64_t
 void quantize_row_iq3_xxs(const float * restrict x, void * restrict vy, int64_t k) {
     assert(k % QK_K == 0);
     block_iq3_xxs * restrict y = vy;
-    quantize_row_iq3_xxs_reference(x, y, k);
+    quantize_row_iq3_xxs_ref(x, y, k);
 }
 
-void quantize_row_iq3_xxs_reference(const float * restrict x, block_iq3_xxs * restrict y, int64_t k) {
+void quantize_row_iq3_xxs_ref(const float * restrict x, block_iq3_xxs * restrict y, int64_t k) {
     assert(k % QK_K == 0);
     quantize_row_iq3_xxs_impl(256, x, y, k, NULL);
 }
@@ -13039,7 +13539,7 @@ static void quantize_row_iq3_s_impl(int block_size, const float * restrict x, vo
                     printf("Oops: found point %u not on grid:", u);
                     for (int i = 0; i < 4; ++i) printf(" %d", L[4*k+i]);
                     printf("\n");
-                    GGML_ASSERT(false);
+                    GGML_ABORT("fatal error");
                 }
                 qs[k] = grid_index & 255;
                 qh[(ib*bs4+k)/8] |= ((grid_index >> 8) << ((ib*bs4+k)%8));
@@ -13096,10 +13596,10 @@ size_t quantize_iq3_s(const float * restrict src, void * restrict dst, int64_t n
 void quantize_row_iq3_s(const float * restrict x, void * restrict vy, int64_t k) {
     assert(k % QK_K == 0);
     block_iq3_s * restrict y = vy;
-    quantize_row_iq3_s_reference(x, y, k);
+    quantize_row_iq3_s_ref(x, y, k);
 }
 
-void quantize_row_iq3_s_reference(const float * restrict x, block_iq3_s * restrict y, int64_t k) {
+void quantize_row_iq3_s_ref(const float * restrict x, block_iq3_s * restrict y, int64_t k) {
     assert(k % QK_K == 0);
     quantize_iq3_s(x, y, 1, k, NULL);
 }
@@ -13111,7 +13611,7 @@ static int iq1_find_best_neighbour(const uint16_t * restrict neighbours, const u
         const float * restrict xval, const float * restrict weight, float * scale, int8_t * restrict L, int ngrid) {
     int num_neighbors = neighbours[0];
     GGML_ASSERT(num_neighbors > 0);
-    float best_score = 0;
+    float best_score = -FLT_MAX;
     int grid_index = -1;
     for (int j = 1; j <= num_neighbors; ++j) {
         const int8_t * pg = (const int8_t *)(grid + neighbours[j]);
@@ -13309,7 +13809,7 @@ static void quantize_row_iq1_s_impl(const float * restrict x, void * restrict vy
                     sumw[j+1] = sumw[j] + weight[i];
                 }
             }
-            float best_score = 0, scale = max;
+            float best_score = -FLT_MIN, scale = max;
             int besti1 = -1, besti2 = -1, best_shift = 0;
             for (int i1 = 0; i1 <= block_size; ++i1) {
                 for (int i2 = i1; i2 <= block_size; ++i2) {
@@ -13485,7 +13985,7 @@ static void quantize_row_iq1_m_impl(const float * restrict x, void * restrict vy
                 idx[2*j] = j;
             }
             qsort(pairs, block_size, 2*sizeof(float), iq1_sort_helper);
-            float best_score = 0, scale = max;
+            float best_score = -FLT_MIN, scale = max;
             int besti1 = -1, besti2 = -1, best_k = -1;
             // 0: +, +
             // 1: +, -
@@ -13837,7 +14337,7 @@ void quantize_row_iq4_nl(const float * restrict x, void * restrict vy, int64_t k
     }
 }
 
-void quantize_row_iq4_nl_reference(const float * restrict x, block_iq4_nl * restrict y, int64_t k) {
+void quantize_row_iq4_nl_ref(const float * restrict x, block_iq4_nl * restrict y, int64_t k) {
     assert(k % QK4_NL == 0);
     quantize_row_iq4_nl(x, y, k);
 }
@@ -13865,10 +14365,10 @@ size_t quantize_iq4_xs(const float * restrict src, void * restrict dst, int64_t
 void quantize_row_iq4_xs(const float * restrict x, void * restrict vy, int64_t k) {
     assert(k % QK_K == 0);
     block_iq4_xs * restrict y = vy;
-    quantize_row_iq4_xs_reference(x, y, k);
+    quantize_row_iq4_xs_ref(x, y, k);
 }
 
-void quantize_row_iq4_xs_reference(const float * restrict x, block_iq4_xs * restrict y, int64_t k) {
+void quantize_row_iq4_xs_ref(const float * restrict x, block_iq4_xs * restrict y, int64_t k) {
     assert(k % QK_K == 0);
     quantize_iq4_xs(x, y, 1, k, NULL);
 }
@@ -14015,7 +14515,7 @@ static void quantize_row_iq2_s_impl(const float * restrict x, void * restrict vy
                     printf("Oops: found point %u not on grid:", u);
                     for (int i = 0; i < 8; ++i) printf(" %d", L[8*k+i]);
                     printf("\n");
-                    GGML_ASSERT(false);
+                    GGML_ABORT("fatal error");
                 }
                 const int i8 = 2*ib + k;
                 y[ibl].qs[i8] = grid_index & 255;
@@ -14055,7 +14555,7 @@ size_t quantize_iq2_s(const float * restrict src, void * restrict dst, int64_t n
     return nrow * nblock * sizeof(block_iq2_s);
 }
 
-void quantize_row_iq2_s_reference(const float * restrict x, block_iq2_s * restrict y, int64_t k) {
+void quantize_row_iq2_s_ref(const float * restrict x, block_iq2_s * restrict y, int64_t k) {
     assert(k % QK_K == 0);
     quantize_iq2_s(x, y, 1, k, NULL);
 }
@@ -14063,7 +14563,7 @@ void quantize_row_iq2_s_reference(const float * restrict x, block_iq2_s * restri
 void quantize_row_iq2_s(const float * restrict x, void * restrict vy, int64_t k) {
     assert(k % QK_K == 0);
     block_iq2_s * restrict y = vy;
-    quantize_row_iq2_s_reference(x, y, k);
+    quantize_row_iq2_s_ref(x, y, k);
 }
 
 static bool validate_float(float f, size_t i) {
@@ -14118,6 +14618,16 @@ static bool validate_fp16(ggml_fp16_t f, size_t i) {
         } \
     }
 
+#define VALIDATE_ROW_DATA_DVEC_F16_IMPL(type, data, nb, nr) \
+    const type * q = (const type *) (data); \
+    for (size_t i = 0; i < (nb); ++i) { \
+        for (size_t j = 0; j < (nr); ++j) { \
+            if (!validate_fp16(q[i].d[j], i)) { \
+                return false; \
+            } \
+        } \
+    }
+
 bool ggml_validate_row_data(enum ggml_type type, const void * data, size_t nbytes) {
     if (type < 0 || type >= GGML_TYPE_COUNT) {
         fprintf(stderr, "%s: invalid type %d\n", __func__, type);
@@ -14125,7 +14635,7 @@ bool ggml_validate_row_data(enum ggml_type type, const void * data, size_t nbyte
     }
 
     if (nbytes % ggml_type_size(type) != 0) {
-        fprintf(stderr, "%s: invalid size %zu for type %d\n", __func__, nbytes, type);
+        fprintf(stderr, "%s: invalid size %zu for type %s (type size = %zu)\n", __func__, nbytes, ggml_type_name(type), ggml_type_size(type));
         return false;
     }
 
@@ -14335,6 +14845,16 @@ bool ggml_validate_row_data(enum ggml_type type, const void * data, size_t nbyte
             {
                 VALIDATE_ROW_DATA_D_F16_IMPL(block_iq4_nl, data, nb);
             } break;
+        case GGML_TYPE_Q4_0_4_4:
+        case GGML_TYPE_Q4_0_4_8:
+            {
+                VALIDATE_ROW_DATA_DVEC_F16_IMPL(block_q4_0x4, data, nbytes / sizeof(block_q4_0x4), 4);
+            } break;
+        case GGML_TYPE_Q4_0_8_8:
+            {
+                VALIDATE_ROW_DATA_DVEC_F16_IMPL(block_q4_0x8, data, nbytes / sizeof(block_q4_0x8), 8);
+            } break;
+
         case GGML_TYPE_I8:
         case GGML_TYPE_I16:
         case GGML_TYPE_I32: