llama_cpp 0.12.4 → 0.12.6

data/vendor/tmp/llama.cpp/ggml-quants.c

@@ -49,6 +49,8 @@
 #define MIN(a, b) ((a) < (b) ? (a) : (b))
 #define MAX(a, b) ((a) > (b) ? (a) : (b))
 
+#define UNUSED GGML_UNUSED
+
 #define MM256_SET_M128I(a, b) _mm256_insertf128_si256(_mm256_castsi128_si256(b), (a), 1)
 
 #if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__)
@@ -268,6 +270,17 @@ static inline float hsum_float_4x4(const __m128 a, const __m128 b, const __m128
 #endif // defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__)
 
 #if defined(__ARM_NEON)
+
+#ifdef _MSC_VER
+
+#define ggml_vld1q_u32(w,x,y,z) { ((w) + ((uint64_t)(x) << 32)), ((y) + ((uint64_t)(z) << 32)) }
+
+#else
+
+#define ggml_vld1q_u32(w,x,y,z) { (w), (x), (y), (z) }
+
+#endif
+
 #if !defined(__aarch64__)
 
 // 64-bit compatibility
@@ -2381,19 +2394,20 @@ static void quantize_row_q4_K_impl(const float * restrict x, block_q4_K * restri
 
     uint8_t L[QK_K];
     uint8_t Laux[32];
+    uint8_t Ls[QK_K/32];
+    uint8_t Lm[QK_K/32];
     float   weights[32];
-    float mins[QK_K/32];
-    float scales[QK_K/32];
+    float   sw[QK_K/32];
+    float   mins[QK_K/32];
+    float   scales[QK_K/32];
 
     for (int i = 0; i < nb; i++) {
 
         float sum_x2 = 0;
         for (int l = 0; l < QK_K; ++l) sum_x2 += x[l] * x[l];
-        float sigma2 = sum_x2/QK_K;
+        float sigma2 = 2*sum_x2/QK_K;
         float av_x = sqrtf(sigma2);
 
-        float max_scale = 0; // as we are deducting the min, scales are always positive
-        float max_min = 0;
         for (int j = 0; j < QK_K/32; ++j) {
             if (quant_weights) {
                 const float * qw = quant_weights + QK_K*i + 32*j;
@@ -2401,25 +2415,17 @@ static void quantize_row_q4_K_impl(const float * restrict x, block_q4_K * restri
             } else {
                 for (int l = 0; l < 32; ++l) weights[l] = av_x + fabsf(x[32*j + l]);
             }
+            float sumw = 0;
+            for (int l = 0; l < 32; ++l) sumw += weights[l];
+            sw[j] = sumw;
             scales[j] = make_qkx3_quants(32, 15, x + 32*j, weights, L + 32*j, &mins[j], Laux, -0.9f, 0.05f, 36, false);
-          //scales[j] = make_qkx2_quants(32, 15, x + 32*j, weights, L + 32*j, &mins[j], Laux, -1.f, 0.1f, 20, false);
-            float scale = scales[j];
-            if (scale > max_scale) {
-                max_scale = scale;
-            }
-            float min = mins[j];
-            if (min > max_min) {
-                max_min = min;
-            }
         }
 
-        float inv_scale = max_scale > 0 ? 63.f/max_scale : 0.f;
-        float inv_min   = max_min   > 0 ? 63.f/max_min   : 0.f;
+        float d_block = make_qp_quants(QK_K/32, 63, scales, Ls, sw);
+        float m_block = make_qp_quants(QK_K/32, 63, mins,   Lm, sw);
         for (int j = 0; j < QK_K/32; ++j) {
-            uint8_t ls = nearest_int(inv_scale*scales[j]);
-            uint8_t lm = nearest_int(inv_min*mins[j]);
-            ls = MIN(63, ls);
-            lm = MIN(63, lm);
+            uint8_t ls = Ls[j];
+            uint8_t lm = Lm[j];
             if (j < 4) {
                 y[i].scales[j] = ls;
                 y[i].scales[j+4] = lm;
@@ -2429,8 +2435,8 @@ static void quantize_row_q4_K_impl(const float * restrict x, block_q4_K * restri
                 y[i].scales[j-0] |= ((lm >> 4) << 6);
             }
         }
-        y[i].d = GGML_FP32_TO_FP16(max_scale/63.f);
-        y[i].dmin = GGML_FP32_TO_FP16(max_min/63.f);
+        y[i].d = GGML_FP32_TO_FP16(d_block);
+        y[i].dmin = GGML_FP32_TO_FP16(m_block);
 
         uint8_t sc, m;
         for (int j = 0; j < QK_K/32; ++j) {
@@ -2688,20 +2694,21 @@ static void quantize_row_q5_K_impl(const float * restrict x, block_q5_K * restri
     const int nb = n_per_row / QK_K;
 
     uint8_t L[QK_K];
-    float mins[QK_K/32];
-    float scales[QK_K/32];
-    float weights[32];
     uint8_t Laux[32];
+    uint8_t Ls[QK_K/32];
+    uint8_t Lm[QK_K/32];
+    float   mins[QK_K/32];
+    float   scales[QK_K/32];
+    float   sw[QK_K/32];
+    float   weights[32];
 
     for (int i = 0; i < nb; i++) {
 
         float sum_x2 = 0;
         for (int l = 0; l < QK_K; ++l) sum_x2 += x[l] * x[l];
-        float sigma2 = sum_x2/QK_K;
+        float sigma2 = 2*sum_x2/QK_K;
         float av_x = sqrtf(sigma2);
 
-        float max_scale = 0; // as we are deducting the min, scales are always positive
-        float max_min = 0;
         for (int j = 0; j < QK_K/32; ++j) {
             if (quant_weights) {
                 const float * qw = quant_weights + QK_K*i + 32*j;
@@ -2709,22 +2716,19 @@ static void quantize_row_q5_K_impl(const float * restrict x, block_q5_K * restri
             } else {
                 for (int l = 0; l < 32; ++l) weights[l] = av_x + fabsf(x[32*j + l]);
             }
+            float sumw = 0;
+            for (int l = 0; l < 32; ++l) sumw += weights[l];
+            sw[j] = sumw;
+
             scales[j] = make_qkx3_quants(32, 31, x + 32*j, weights, L + 32*j, &mins[j], Laux, -0.9f, 0.05f, 36, false);
-            float scale = scales[j];
-            if (scale > max_scale) {
-                max_scale = scale;
-            }
-            float min = mins[j];
-            if (min > max_min) {
-                max_min = min;
-            }
         }
 
-        float inv_scale = max_scale > 0 ? 63.f/max_scale : 0.f;
-        float inv_min   = max_min   > 0 ? 63.f/max_min   : 0.f;
+        float d_block = make_qp_quants(QK_K/32, 63, scales, Ls, sw);
+        float m_block = make_qp_quants(QK_K/32, 63, mins,   Lm, sw);
+
         for (int j = 0; j < QK_K/32; ++j) {
-            uint8_t ls = nearest_int(inv_scale*scales[j]);
-            uint8_t lm = nearest_int(inv_min*mins[j]);
+            uint8_t ls = Ls[j];
+            uint8_t lm = Lm[j];
             ls = MIN(63, ls);
             lm = MIN(63, lm);
             if (j < 4) {
@@ -2736,8 +2740,8 @@ static void quantize_row_q5_K_impl(const float * restrict x, block_q5_K * restri
                 y[i].scales[j-0] |= ((lm >> 4) << 6);
             }
         }
-        y[i].d = GGML_FP32_TO_FP16(max_scale/63.f);
-        y[i].dmin = GGML_FP32_TO_FP16(max_min/63.f);
+        y[i].d = GGML_FP32_TO_FP16(d_block);
+        y[i].dmin = GGML_FP32_TO_FP16(m_block);
 
         uint8_t sc, m;
         for (int j = 0; j < QK_K/32; ++j) {
@@ -3675,15 +3679,92 @@ static inline __m128i get_scale_shuffle(int i) {
 }
 #endif
 
-void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
+void ggml_vec_dot_q4_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;
 
     assert(n % qk == 0);
+#if defined(__ARM_FEATURE_MATMUL_INT8)
+    assert((nrc == 2) || (nrc == 1));
+#else
+    assert(nrc == 1);
+#endif
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
 
     const block_q4_0 * restrict x = vx;
     const block_q8_0 * restrict y = vy;
 
+#if defined(__ARM_FEATURE_MATMUL_INT8)
+    if (nrc == 2) {
+        const block_q4_0 * restrict vx0 = vx;
+        const block_q4_0 * restrict vx1 = vx + bx;
+
+        const block_q8_0 * restrict vy0 = vy;
+        const block_q8_0 * restrict vy1 = vy + by;
+
+        float32x4_t sumv0 = vdupq_n_f32(0.0f);
+
+        for (int i = 0; i < nb; i++) {
+            const block_q4_0 * restrict b_x0 = &vx0[i];
+            const block_q4_0 * restrict b_x1 = &vx1[i];
+            const block_q8_0 * restrict b_y0 = &vy0[i];
+            const block_q8_0 * restrict b_y1 = &vy1[i];
+
+            const uint8x16_t m4b = vdupq_n_u8(0x0F);
+            const int8x16_t  s8b = vdupq_n_s8(0x8);
+
+            const uint8x16_t v0_0 = vld1q_u8(b_x0->qs);
+            const uint8x16_t v0_1 = vld1q_u8(b_x1->qs);
+
+            // 4-bit -> 8-bit
+            const int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8  (v0_0, m4b));
+            const int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4));
+            const int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8  (v0_1, m4b));
+            const int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4));
+
+            // sub 8
+            const int8x16_t x0_l = vsubq_s8(v0_0l, s8b);
+            const int8x16_t x0_h = vsubq_s8(v0_0h, s8b);
+            const int8x16_t x1_l = vsubq_s8(v0_1l, s8b);
+            const int8x16_t x1_h = vsubq_s8(v0_1h, s8b);
+
+            // load y
+            const int8x16_t y0_l = vld1q_s8(b_y0->qs);
+            const int8x16_t y0_h = vld1q_s8(b_y0->qs + 16);
+            const int8x16_t y1_l = vld1q_s8(b_y1->qs);
+            const int8x16_t y1_h = vld1q_s8(b_y1->qs + 16);
+
+            float32x4_t scale = {GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y0->d),
+                                 GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y1->d),
+                                 GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y0->d),
+                                 GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y1->d)};
+
+            int8x16_t l0 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(x0_l), vreinterpretq_s64_s8(x1_l)));
+            int8x16_t l1 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(x0_l), vreinterpretq_s64_s8(x1_l)));
+
+            int8x16_t l2 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(x0_h), vreinterpretq_s64_s8(x1_h)));
+            int8x16_t l3 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(x0_h), vreinterpretq_s64_s8(x1_h)));
+
+            int8x16_t r0 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(y0_l), vreinterpretq_s64_s8(y1_l)));
+            int8x16_t r1 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(y0_l), vreinterpretq_s64_s8(y1_l)));
+
+            int8x16_t r2 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(y0_h), vreinterpretq_s64_s8(y1_h)));
+            int8x16_t r3 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(y0_h), vreinterpretq_s64_s8(y1_h)));
+
+            sumv0 = vmlaq_f32(sumv0,(vcvtq_f32_s32(vmmlaq_s32((vmmlaq_s32((vmmlaq_s32((vmmlaq_s32(vdupq_n_s32(0), l0, r0)),
+                                                                                l1, r1)), l2, r2)), l3, r3))), scale);
+        }
+        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 + bs, vget_high_f32(sumv2));
+        return;
+    }
+#endif
 #if defined(__ARM_NEON)
     float32x4_t sumv0 = vdupq_n_f32(0.0f);
     float32x4_t sumv1 = vdupq_n_f32(0.0f);
@@ -3738,15 +3819,15 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, const void * restrict vx,
         /* 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 bx = bytes_from_nibbles_32(x[i].qs);
+        __m256i qx = bytes_from_nibbles_32(x[i].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 );
-        bx = _mm256_sub_epi8( bx, off );
+        qx = _mm256_sub_epi8( qx, off );
 
-        __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs);
+        __m256i qy = _mm256_loadu_si256((const __m256i *)y[i].qs);
 
-        const __m256 q = mul_sum_i8_pairs_float(bx, by);
+        const __m256 q = mul_sum_i8_pairs_float(qx, qy);
 
         /* Multiply q with scale and accumulate */
         acc = _mm256_fmadd_ps( d, q, acc );
@@ -3965,15 +4046,93 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, const void * restrict vx,
 #endif
 }
 
-void ggml_vec_dot_q4_1_q8_1(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
+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) {
     const int qk = QK8_1;
     const int nb = n / qk;
 
     assert(n % qk == 0);
+#if defined(__ARM_FEATURE_MATMUL_INT8)
+    assert((nrc == 2) || (nrc == 1));
+#else
+    assert(nrc == 1);
+#endif
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
 
     const block_q4_1 * restrict x = vx;
     const block_q8_1 * restrict y = vy;
 
+#if defined(__ARM_FEATURE_MATMUL_INT8)
+    if (nrc == 2) {
+        const block_q4_1 * restrict vx0 = vx;
+        const block_q4_1 * restrict vx1 = vx + bx;
+        const block_q8_1 * restrict vy0 = vy;
+        const block_q8_1 * restrict vy1 = vy + by;
+
+        float32x4_t sumv0 = vdupq_n_f32(0.0f);
+        float32x4_t summs0 = vdupq_n_f32(0.0f);
+
+        for (int i = 0; i < nb; i++) {
+            const block_q4_1 * restrict b_x0 = &vx0[i];
+            const block_q4_1 * restrict b_x1 = &vx1[i];
+            const block_q8_1 * restrict b_y0 = &vy0[i];
+            const block_q8_1 * restrict b_y1 = &vy1[i];
+
+            float32x4_t summs_t = {GGML_FP16_TO_FP32(b_x0->m) * b_y0->s,
+                                   GGML_FP16_TO_FP32(b_x1->m) * b_y0->s,
+                                   GGML_FP16_TO_FP32(b_x0->m) * b_y1->s,
+                                   GGML_FP16_TO_FP32(b_x1->m) * b_y1->s};
+            summs0 += summs_t;
+
+            const uint8x16_t m4b = vdupq_n_u8(0x0F);
+
+            const uint8x16_t v0_0 = vld1q_u8(b_x0->qs);
+            const uint8x16_t v0_1 = vld1q_u8(b_x1->qs);
+
+            // 4-bit -> 8-bit
+            const int8x16_t x0_l = vreinterpretq_s8_u8(vandq_u8  (v0_0, m4b));
+            const int8x16_t x0_h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4));
+            const int8x16_t x1_l = vreinterpretq_s8_u8(vandq_u8  (v0_1, m4b));
+            const int8x16_t x1_h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4));
+
+            // load y
+            const int8x16_t y0_l = vld1q_s8(b_y0->qs);
+            const int8x16_t y0_h = vld1q_s8(b_y0->qs + 16);
+            const int8x16_t y1_l = vld1q_s8(b_y1->qs);
+            const int8x16_t y1_h = vld1q_s8(b_y1->qs + 16);
+
+            // mmla into int32x4_t
+            float32x4_t scale = {GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y0->d),
+                                 GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y1->d),
+                                 GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y0->d),
+                                 GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y1->d)};
+
+            int8x16_t l0 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(x0_l), vreinterpretq_s64_s8(x1_l)));
+            int8x16_t l1 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(x0_l), vreinterpretq_s64_s8(x1_l)));
+
+            int8x16_t l2 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(x0_h), vreinterpretq_s64_s8(x1_h)));
+            int8x16_t l3 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(x0_h), vreinterpretq_s64_s8(x1_h)));
+
+            int8x16_t r0 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(y0_l), vreinterpretq_s64_s8(y1_l)));
+            int8x16_t r1 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(y0_l), vreinterpretq_s64_s8(y1_l)));
+
+            int8x16_t r2 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(y0_h), vreinterpretq_s64_s8(y1_h)));
+            int8x16_t r3 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(y0_h), vreinterpretq_s64_s8(y1_h)));
+            sumv0 = vmlaq_f32(sumv0,(vcvtq_f32_s32(vmmlaq_s32((vmmlaq_s32((vmmlaq_s32((vmmlaq_s32(vdupq_n_s32(0), l0, r0)),
+                                                                                l1, r1)), l2, r2)), l3, r3))), scale);
+        }
+
+        float32x4_t sumv1 = vextq_f32(sumv0, sumv0, 2);
+        float32x4_t sumv2 = vzip1q_f32(sumv0, sumv1);
+        sumv2 = sumv2 + summs0;
+
+        vst1_f32(s, vget_low_f32(sumv2));
+        vst1_f32(s + bs, vget_high_f32(sumv2));
+        return;
+    }
+#endif
     // TODO: add WASM SIMD
 #if defined(__ARM_NEON)
     float32x4_t sumv0 = vdupq_n_f32(0.0f);
@@ -4037,10 +4196,10 @@ void ggml_vec_dot_q4_1_q8_1(const int n, float * restrict s, const void * restri
         const __m256 d0d1 = _mm256_mul_ps( d0v, d1v );
 
         // Load 16 bytes, and unpack 4 bit fields into bytes, making 32 bytes
-        const __m256i bx = bytes_from_nibbles_32(x[i].qs);
-        const __m256i by = _mm256_loadu_si256( (const __m256i *)y[i].qs );
+        const __m256i qx = bytes_from_nibbles_32(x[i].qs);
+        const __m256i qy = _mm256_loadu_si256( (const __m256i *)y[i].qs );
 
-        const __m256 xy = mul_sum_us8_pairs_float(bx, by);
+        const __m256 xy = mul_sum_us8_pairs_float(qx, qy);
 
         // Accumulate d0*d1*x*y
 #if defined(__AVX2__)
@@ -4105,12 +4264,17 @@ void ggml_vec_dot_q4_1_q8_1(const int n, float * restrict s, const void * restri
 #endif
 }
 
-void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
+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;
 
     assert(n % qk == 0);
     assert(qk == QK5_0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
 
     const block_q5_0 * restrict x = vx;
     const block_q8_0 * restrict y = vy;
@@ -4254,14 +4418,14 @@ void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void * restri
         /* 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 bx = bytes_from_nibbles_32(x[i].qs);
+        __m256i qx = bytes_from_nibbles_32(x[i].qs);
         __m256i bxhi = bytes_from_bits_32(x[i].qh);
         bxhi = _mm256_andnot_si256(bxhi, _mm256_set1_epi8((char)0xF0));
-        bx = _mm256_or_si256(bx, bxhi);
+        qx = _mm256_or_si256(qx, bxhi);
 
-        __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs);
+        __m256i qy = _mm256_loadu_si256((const __m256i *)y[i].qs);
 
-        const __m256 q = mul_sum_i8_pairs_float(bx, by);
+        const __m256 q = mul_sum_i8_pairs_float(qx, qy);
 
         /* Multiply q with scale and accumulate */
         acc = _mm256_fmadd_ps(d, q, acc);
@@ -4391,12 +4555,17 @@ void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void * restri
 #endif
 }
 
-void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
+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;
 
     assert(n % qk == 0);
     assert(qk == QK5_1);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
 
     const block_q5_1 * restrict x = vx;
     const block_q8_1 * restrict y = vy;
@@ -4553,15 +4722,15 @@ void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void * restri
 
         summs += GGML_FP16_TO_FP32(x[i].m) * y[i].s;
 
-        __m256i bx = bytes_from_nibbles_32(x[i].qs);
+        __m256i qx = bytes_from_nibbles_32(x[i].qs);
         __m256i bxhi = bytes_from_bits_32(x[i].qh);
         bxhi = _mm256_and_si256(bxhi, _mm256_set1_epi8(0x10));
-        bx = _mm256_or_si256(bx, bxhi);
+        qx = _mm256_or_si256(qx, bxhi);
 
         const __m256 dy = _mm256_set1_ps(y[i].d);
-        const __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs);
+        const __m256i qy = _mm256_loadu_si256((const __m256i *)y[i].qs);
 
-        const __m256 q = mul_sum_us8_pairs_float(bx, by);
+        const __m256 q = mul_sum_us8_pairs_float(qx, qy);
 
         acc = _mm256_fmadd_ps(q, _mm256_mul_ps(dx, dy), acc);
     }
@@ -4690,15 +4859,79 @@ void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void * restri
 #endif
 }
 
-void ggml_vec_dot_q8_0_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
+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) {
     const int qk = QK8_0;
     const int nb = n / qk;
 
     assert(n % qk == 0);
+#if defined(__ARM_FEATURE_MATMUL_INT8)
+    assert((nrc == 2) || (nrc == 1));
+#else
+    assert(nrc == 1);
+#endif
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
 
     const block_q8_0 * restrict x = vx;
     const block_q8_0 * restrict y = vy;
 
+#if defined(__ARM_FEATURE_MATMUL_INT8)
+    if (nrc == 2) {
+        const block_q8_0 * restrict vx0 = vx;
+        const block_q8_0 * restrict vx1 = vx + bx;
+        const block_q8_0 * restrict vy0 = vy;
+        const block_q8_0 * restrict vy1 = vy + by;
+
+        float32x4_t sumv0 = vdupq_n_f32(0.0f);
+
+        for (int i = 0; i < nb; i++) {
+            const block_q8_0 * restrict b_x0 = &vx0[i];
+            const block_q8_0 * restrict b_y0 = &vy0[i];
+
+            const block_q8_0 * restrict b_x1 = &vx1[i];
+            const block_q8_0 * restrict b_y1 = &vy1[i];
+
+            const int8x16_t x0_l = vld1q_s8(b_x0->qs);
+            const int8x16_t x0_h = vld1q_s8(b_x0->qs + 16);
+            const int8x16_t x1_l = vld1q_s8(b_x1->qs);
+            const int8x16_t x1_h = vld1q_s8(b_x1->qs + 16);
+
+            // load y
+            const int8x16_t y0_l = vld1q_s8(b_y0->qs);
+            const int8x16_t y0_h = vld1q_s8(b_y0->qs + 16);
+            const int8x16_t y1_l = vld1q_s8(b_y1->qs);
+            const int8x16_t y1_h = vld1q_s8(b_y1->qs + 16);
+
+            float32x4_t scale = {GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y0->d),
+                             GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y1->d),
+                             GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y0->d),
+                             GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y1->d)};
+
+            int8x16_t l0 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(x0_l), vreinterpretq_s64_s8(x1_l)));
+            int8x16_t l1 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(x0_l), vreinterpretq_s64_s8(x1_l)));
+
+            int8x16_t l2 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(x0_h), vreinterpretq_s64_s8(x1_h)));
+            int8x16_t l3 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(x0_h), vreinterpretq_s64_s8(x1_h)));
+
+            int8x16_t r0 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(y0_l), vreinterpretq_s64_s8(y1_l)));
+            int8x16_t r1 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(y0_l), vreinterpretq_s64_s8(y1_l)));
+
+            int8x16_t r2 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(y0_h), vreinterpretq_s64_s8(y1_h)));
+            int8x16_t r3 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(y0_h), vreinterpretq_s64_s8(y1_h)));
+
+            sumv0 = vmlaq_f32(sumv0,(vcvtq_f32_s32(vmmlaq_s32((vmmlaq_s32((vmmlaq_s32((vmmlaq_s32(vdupq_n_s32(0), l0, r0)),
+                                                                                       l1, r1)), l2, r2)), l3, r3))), scale);
+        }
+        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 + bs, vget_high_f32(sumv2));
+        return;
+    }
+#endif
 #if defined(__ARM_NEON)
     float32x4_t sumv0 = vdupq_n_f32(0.0f);
     float32x4_t sumv1 = vdupq_n_f32(0.0f);
@@ -4740,10 +4973,10 @@ void ggml_vec_dot_q8_0_q8_0(const int n, float * restrict s, const void * restri
     for (int i = 0; i < nb; ++i) {
         // 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 bx = _mm256_loadu_si256((const __m256i *)x[i].qs);
-        __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs);
+        __m256i qx = _mm256_loadu_si256((const __m256i *)x[i].qs);
+        __m256i qy = _mm256_loadu_si256((const __m256i *)y[i].qs);
 
-        const __m256 q = mul_sum_i8_pairs_float(bx, by);
+        const __m256 q = mul_sum_i8_pairs_float(qx, qy);
 
         // Multiply q with scale and accumulate
 #if defined(__AVX2__)
@@ -4793,7 +5026,12 @@ void ggml_vec_dot_q8_0_q8_0(const int n, float * restrict s, const void * restri
 }
 
 #if QK_K == 256
-void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
+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) {
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
 
     const block_q2_K * restrict x = vx;
     const block_q8_K * restrict y = vy;
@@ -5169,7 +5407,12 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri
 
 #else
 
-void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
+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) {
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
 
     const block_q2_K * restrict x = vx;
     const block_q8_K * restrict y = vy;
@@ -5427,8 +5670,13 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri
 #endif
 
 #if QK_K == 256
-void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
+void ggml_vec_dot_q3_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) {
     assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
 
     const uint32_t kmask1 = 0x03030303;
     const uint32_t kmask2 = 0x0f0f0f0f;
@@ -5947,8 +6195,13 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri
 
 #else
 
-void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
+void ggml_vec_dot_q3_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) {
     assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
 
     const block_q3_K * restrict x = vx;
     const block_q8_K * restrict y = vy;
@@ -6290,8 +6543,13 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri
 #endif
 
 #if QK_K == 256
-void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
+void ggml_vec_dot_q4_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) {
     assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
 
     const block_q4_K * restrict x = vx;
     const block_q8_K * restrict y = vy;
@@ -6646,8 +6904,13 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri
 #endif
 }
 #else
-void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
+void ggml_vec_dot_q4_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) {
     assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
 
     const block_q4_K * restrict x = vx;
     const block_q8_K * restrict y = vy;
@@ -6889,8 +7152,13 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri
 #endif
 
 #if QK_K == 256
-void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
+void ggml_vec_dot_q5_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) {
     assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
 
     const block_q5_K * restrict x = vx;
     const block_q8_K * restrict y = vy;
@@ -7309,8 +7577,13 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri
 
 #else
 
-void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
+void ggml_vec_dot_q5_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) {
     assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
 
     const block_q5_K * restrict x = vx;
     const block_q8_K * restrict y = vy;
@@ -7575,8 +7848,13 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri
 
 
 #if QK_K == 256
-void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
+void ggml_vec_dot_q6_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) {
     assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
 
     const block_q6_K * restrict x = vx;
     const block_q8_K * restrict y = vy;
@@ -8007,8 +8285,13 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri
 
 #else
 
-void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
+void ggml_vec_dot_q6_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) {
     assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
 
     const block_q6_K * restrict x = vx;
     const block_q8_K * restrict y = vy;
@@ -8337,8 +8620,13 @@ 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,
 };
 
-void ggml_vec_dot_iq2_xxs_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
+void ggml_vec_dot_iq2_xxs_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) {
     assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
 
     const block_iq2_xxs * restrict x = vx;
     const block_q8_K    * restrict y = vy;
@@ -8460,8 +8748,13 @@ void ggml_vec_dot_iq2_xxs_q8_K(const int n, float * restrict s, const void * res
 #endif
 }
 
-void ggml_vec_dot_iq2_xs_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
+void ggml_vec_dot_iq2_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) {
     assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
 
     const block_iq2_xs * restrict x = vx;
     const block_q8_K   * restrict y = vy;
@@ -8680,8 +8973,13 @@ void ggml_vec_dot_iq2_xs_q8_K(const int n, float * restrict s, const void * rest
 }
 
 // TODO
-void ggml_vec_dot_iq3_xxs_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
+void ggml_vec_dot_iq3_xxs_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) {
     assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
 
     const block_iq3_xxs * restrict x = vx;
     const block_q8_K    * restrict y = vy;
@@ -8707,10 +9005,10 @@ void ggml_vec_dot_iq3_xxs_q8_K(const int n, float * restrict s, const void * res
         for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
             q8b = ggml_vld1q_s8_x4(q8); q8 += 64;
             memcpy(aux32, gas, 2*sizeof(uint32_t)); gas += 2*sizeof(uint32_t);
-            const uint32x4_t aux32x4_0 = {iq3xxs_grid[q3[ 0]], iq3xxs_grid[q3[ 1]], iq3xxs_grid[q3[ 2]], iq3xxs_grid[q3[ 3]]};
-            const uint32x4_t aux32x4_1 = {iq3xxs_grid[q3[ 4]], iq3xxs_grid[q3[ 5]], iq3xxs_grid[q3[ 6]], iq3xxs_grid[q3[ 7]]};
-            const uint32x4_t aux32x4_2 = {iq3xxs_grid[q3[ 8]], iq3xxs_grid[q3[ 9]], iq3xxs_grid[q3[10]], iq3xxs_grid[q3[11]]};
-            const uint32x4_t aux32x4_3 = {iq3xxs_grid[q3[12]], iq3xxs_grid[q3[13]], iq3xxs_grid[q3[14]], iq3xxs_grid[q3[15]]};
+            const uint32x4_t aux32x4_0 = ggml_vld1q_u32(iq3xxs_grid[q3[ 0]], iq3xxs_grid[q3[ 1]], iq3xxs_grid[q3[ 2]], iq3xxs_grid[q3[ 3]]);
+            const uint32x4_t aux32x4_1 = ggml_vld1q_u32(iq3xxs_grid[q3[ 4]], iq3xxs_grid[q3[ 5]], iq3xxs_grid[q3[ 6]], iq3xxs_grid[q3[ 7]]);
+            const uint32x4_t aux32x4_2 = ggml_vld1q_u32(iq3xxs_grid[q3[ 8]], iq3xxs_grid[q3[ 9]], iq3xxs_grid[q3[10]], iq3xxs_grid[q3[11]]);
+            const uint32x4_t aux32x4_3 = ggml_vld1q_u32(iq3xxs_grid[q3[12]], iq3xxs_grid[q3[13]], iq3xxs_grid[q3[14]], iq3xxs_grid[q3[15]]);
             q3 += 16;
             q3s.val[0] = vcombine_s8(vld1_s8((const void *)(signs64 + ((aux32[0] >>  0) & 127))), vld1_s8((const void *)(signs64 + ((aux32[0] >>  7) & 127))));
             q3s.val[1] = vcombine_s8(vld1_s8((const void *)(signs64 + ((aux32[0] >> 14) & 127))), vld1_s8((const void *)(signs64 + ((aux32[0] >> 21) & 127))));
@@ -9048,8 +9346,6 @@ static void quantize_row_iq2_xxs_impl(const float * restrict x, void * restrict
     int8_t L[32];
     int8_t Laux[32];
     float  waux[32];
-    bool   is_on_grid[4];
-    bool   is_on_grid_aux[4];
     uint8_t block_signs[4];
     uint32_t q2[2*(QK_K/32)];
 
@@ -9099,10 +9395,11 @@ static void quantize_row_iq2_xxs_impl(const float * restrict x, void * restrict
                 memset(L, 0, 32);
                 continue;
             }
+            float scale = make_qp_quants(32, kMaxQ+1, xval, (uint8_t*)L, weight);
+            float eff_max = scale*kMaxQ;
             float best = 0;
-            float scale = max/(2*kMaxQ-1);
-            for (int is = -9; is <= 9; ++is) {
-                float id = (2*kMaxQ-1+is*0.1f)/max;
+            for (int is = -6; is <= 6; ++is) {
+                float id = (2*kMaxQ-1+is*0.1f)/eff_max;
                 float this_scale = 1/id;
                 for (int k = 0; k < 4; ++k) {
                     for (int i = 0; i < 8; ++i) {
@@ -9112,9 +9409,7 @@ static void quantize_row_iq2_xxs_impl(const float * restrict x, void * restrict
                     uint16_t u = 0;
                     for (int i = 0; i < 8; ++i) u |= (Laux[8*k+i] << 2*i);
                     int grid_index = kmap_q2xs[u];
-                    is_on_grid_aux[k] = true;
                     if (grid_index < 0) {
-                        is_on_grid_aux[k] = false;
                         const uint16_t * neighbours = kneighbors_q2xs - kmap_q2xs[u] - 1;
                         grid_index = iq2_find_best_neighbour(neighbours, kgrid_q2xs, xval + 8*k, waux + 8*k, this_scale, Laux + 8*k);
                     }
@@ -9128,16 +9423,12 @@ static void quantize_row_iq2_xxs_impl(const float * restrict x, void * restrict
                 }
                 if (sumq2 > 0 && sumqx*sumqx > best*sumq2) {
                     scale = sumqx/sumq2; best = scale*sumqx;
-                    for (int i = 0; i < 32; ++i) L[i] = Laux[i];
-                    for (int k = 0; k <  4; ++k) is_on_grid[k] = is_on_grid_aux[k];
+                    memcpy(L, Laux, 32);
                 }
             }
-            int n_not_ongrid = 0;
-            for (int k = 0; k < 4; ++k) if (!is_on_grid[k]) ++n_not_ongrid;
-            if (n_not_ongrid > 0 && scale > 0) {
+            if (scale > 0) {
                 float id = 1/scale;
                 for (int k = 0; k < 4; ++k) {
-                    if (is_on_grid[k]) continue;
                     uint16_t u = 0;
                     for (int i = 0; i < 8; ++i) {
                         int l = nearest_int(0.5f*(id*xval[8*k+i]-1));
@@ -9193,49 +9484,10 @@ static void quantize_row_iq2_xxs_impl(const float * restrict x, void * restrict
         float d = max_scale/31;
         y[ibl].d = GGML_FP32_TO_FP16(d);
         float id = 1/d;
-        float sumqx = 0, sumq2 = 0;
         for (int ib = 0; ib < QK_K/32; ++ib) {
             int l = nearest_int(0.5f*(id*scales[ib]-1));
             l = MAX(0, MIN(15, l));
             q2[2*ib+1] |= ((uint32_t)l << 28);
-            const float * xb = xbl + 32*ib;
-            const float * qw = quant_weights + QK_K*ibl + 32*ib;
-            for (int i = 0; i < 32; ++i) weight[i] = qw[i] * sqrtf(sigma2 + xb[i]*xb[i]);
-            const uint8_t * aux8 = (const uint8_t *)(q2 + 2*ib);
-            const float db = d * (1 + 2*l);
-            uint32_t u = 0;
-            for (int k = 0; k < 4; ++k) {
-                const int8_t * signs = keven_signs_q2xs + 8*((q2[2*ib+1] >> 7*k) & 127);
-                const float * xk = xb + 8*k;
-                const float * wk = weight + 8*k;
-                const uint8_t * grid = (const uint8_t *)(kgrid_q2xs + aux8[k]);
-                float best_mse = 0; int best_index = aux8[k];
-                for (int j = 0; j < 8; ++j) {
-                    float diff = db * grid[j] * signs[j] - xk[j];
-                    best_mse += wk[j] * diff * diff;
-                }
-                for (int idx = 0; idx < 256; ++idx) {
-                    grid = (const uint8_t *)(kgrid_q2xs + idx);
-                    float mse = 0;
-                    for (int j = 0; j < 8; ++j) {
-                        float diff = db * grid[j] * signs[j] - xk[j];
-                        mse += wk[j] * diff * diff;
-                    }
-                    if (mse < best_mse) {
-                        best_mse = mse; best_index = idx;
-                    }
-                }
-                u |= (best_index << 8*k);
-                grid = (const uint8_t *)(kgrid_q2xs + best_index);
-                //grid = (const uint8_t *)(kgrid_q2xs + aux8[k]);
-                for (int j = 0; j < 8; ++j) {
-                    float q = db * grid[j] * signs[j];
-                    sumqx += wk[j] * q * xk[j];
-                    sumq2 += wk[j] * q * q;
-                }
-            }
-            q2[2*ib] = u;
-            if (sumq2 > 0) y[ibl].d = GGML_FP32_TO_FP16(d*sumqx/sumq2);
         }
         memcpy(y[ibl].qs, q2, QK_K/4);
     }