whispercpp 1.3.5 → 1.3.7

data/ext/sources/ggml/src/ggml-cpu/repack.cpp

@@ -48,6 +48,90 @@ static inline int nearest_int(float fval) {
 
 extern "C" {
 
+#if defined __riscv_zvfh
+void ggml_quantize_mat_q8_0_4x1_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
+    assert(QK8_0 == 32);
+    assert(k % QK8_0 == 0);
+    const int nb = k / QK8_0;
+
+    block_q8_0x4 * GGML_RESTRICT y = (block_q8_0x4 *) vy;
+
+    // scalar
+    const int blck_size_interleave = 1;
+    float srcv[4][QK8_0];
+    float id[4];
+
+    for (int i = 0; i < nb; i++) {
+        for (int row_iter = 0; row_iter < 4; row_iter++) {
+            float amax = 0.0f; // absolute max
+
+            for (int j = 0; j < QK8_0; j++) {
+                srcv[row_iter][j] = x[row_iter * k + i * QK8_0 + j];
+                amax = MAX(amax, fabsf(srcv[row_iter][j]));
+            }
+
+            const float d = amax / ((1 << 7) - 1);
+            id[row_iter] = d ? 1.0f / d : 0.0f;
+
+            y[i].d[row_iter] = GGML_CPU_FP32_TO_FP16(d);
+        }
+
+        for (int j = 0; j < QK8_0 * 4; j++) {
+            int src_offset = (j / (4 * blck_size_interleave)) * blck_size_interleave;
+            int src_id = (j % (4 * blck_size_interleave)) / blck_size_interleave;
+            src_offset += (j % blck_size_interleave);
+
+            float x0 = srcv[src_id][src_offset] * id[src_id];
+            y[i].qs[j] = roundf(x0);
+        }
+    }
+}
+
+void ggml_quantize_mat_q8_K_4x1_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
+    assert(QK_K == 256);
+    assert(k % QK_K == 0);
+    const int nb = k / QK_K;
+
+    block_q8_Kx4 * GGML_RESTRICT y = (block_q8_Kx4 *) vy;
+
+    const int blck_size_interleave = 1;
+    float srcv[4][QK_K];
+    float iscale[4];
+
+    for (int i = 0; i < nb; i++) {
+        for (int row_iter = 0; row_iter < 4; row_iter++) {
+            float amax = 0.0f; // absolute max
+            float max = 0;
+
+            for (int j = 0; j < QK_K; j++) {
+                srcv[row_iter][j] = x[row_iter * k + i * QK_K + j];
+                // Update the maximum value of the corresponding super block
+                if(amax < fabsf(srcv[row_iter][j])) {
+                    amax = fabsf(srcv[row_iter][j]);
+                    max = srcv[row_iter][j];
+                }
+            }
+
+            iscale[row_iter] = amax ? -127.f/max : 0;
+            y[i].d[row_iter] = amax ? 1/iscale[row_iter] : 0;
+        }
+
+        for (int j = 0; j < QK_K / 4; j++) {
+            y[i].bsums[j] = 0;
+        }
+        for (int j = 0; j < QK_K * 4; j++) {
+            int src_id = j % 4;
+            int src_offset = j / 4;
+            int index = ((j >> 6) << 2) + (j & 3);
+
+            float x0 = srcv[src_id][src_offset] * iscale[src_id];
+            y[i].qs[j] = nearest_int(x0);
+            y[i].bsums[index] += y[i].qs[j];
+        }
+    }
+}
+#endif
+
 void ggml_quantize_mat_q8_0_4x4_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
     assert(QK8_0 == 32);
     assert(k % QK8_0 == 0);
@@ -124,7 +208,6 @@ void ggml_quantize_mat_q8_0_4x8_generic(const float * GGML_RESTRICT x, void * GG
     }
 }
 
-
 void ggml_quantize_mat_q8_K_4x4_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
     assert(QK_K == 256);
     assert(k % QK_K == 0);
@@ -256,192 +339,289 @@ template <> void ggml_quantize_mat_t<8, GGML_TYPE_Q8_K>(const float * GGML_RESTR
     ggml_quantize_mat_q8_K_4x8(x, vy, n_per_row);
 }
 
-extern "C" {
+#if defined __riscv_zvfh
+template <> void ggml_quantize_mat_t<1, GGML_TYPE_Q8_0>(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t nrow, int64_t n_per_row) {
+    assert(nrow == 4);
+    UNUSED(nrow);
+    ggml_quantize_mat_q8_0_4x1(x, vy, n_per_row);
+}
 
-void ggml_gemv_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
-    const int qk = QK8_0;
-    const int nb = n / qk;
-    const int ncols_interleaved = 4;
-    const int blocklen = 4;
+template <> void ggml_quantize_mat_t<1, GGML_TYPE_Q8_K>(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t nrow, int64_t n_per_row) {
+    assert(nrow == 4);
+    UNUSED(nrow);
+    ggml_quantize_mat_q8_K_4x1(x, vy, n_per_row);
+}
+#endif
+
+template <int M, int N>
+static void ggml_gemv_q6_K_NxM_q8_K_generic_impl(int                        n,
+                                                 float * GGML_RESTRICT      s,
+                                                 size_t                     bs,
+                                                 const void * GGML_RESTRICT vx,
+                                                 const void * GGML_RESTRICT vy,
+                                                 int                        nr,
+                                                 int                        nc) {
+    constexpr int blocklen          = M;
+    constexpr int ncols_interleaved = N;
+    const int     qk                = QK_K;
+    const int     nb                = n / qk;
+    const int     blocks_per_half   = 64 / blocklen;
 
-    assert(nr == 1);
     assert(n % qk == 0);
     assert(nc % ncols_interleaved == 0);
 
-    UNUSED(s);
     UNUSED(bs);
-    UNUSED(vx);
-    UNUSED(vy);
     UNUSED(nr);
-    UNUSED(nc);
-    UNUSED(nb);
-    UNUSED(ncols_interleaved);
-    UNUSED(blocklen);
 
-    float sumf[4];
-    int sumi;
+    float sumf[8];
 
-    const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+    const block_q8_K * a_ptr = (const block_q8_K *) vy;
     for (int x = 0; x < nc / ncols_interleaved; x++) {
-        const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb);
+        const block_q6_Kx8 * b_ptr = (const block_q6_Kx8 *) vx + (x * nb);
+
+        for (int j = 0; j < ncols_interleaved; j++) {
+            sumf[j] = 0.0f;
+        }
 
-        for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
         for (int l = 0; l < nb; l++) {
             for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                const int base_l = (k / blocks_per_half) * 128 + (k % blocks_per_half) * blocklen;
+                const int base_h = base_l + 64;
+
+                const int scale_idx_l = base_l / 16;
+                const int scale_idx_h = base_h / 16;
+
+                const int qh_shift_l = ((base_l % 128) / 32) * 2;
+                const int qh_shift_h = ((base_h % 128) / 32) * 2;
+
+                const int qh_half_l = (base_l / 128) * 32;
+                const int qh_half_h = (base_h / 128) * 32;
+
                 for (int j = 0; j < ncols_interleaved; j++) {
-                    sumi = 0;
-                    for (int i = 0; i < blocklen; ++i) {
-                        const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
-                        const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
-                        sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
+                    const int8_t scale_l = b_ptr[l].scales[scale_idx_l * ncols_interleaved + j];
+                    const int8_t scale_h = b_ptr[l].scales[scale_idx_h * ncols_interleaved + j];
+
+                    int sumi_l = 0;
+                    int sumi_h = 0;
+
+                    for (int i = 0; i < blocklen; i++) {
+                        const int ql_pos = k * ncols_interleaved * blocklen + j * blocklen + i;
+                        const int l_4    = b_ptr[l].ql[ql_pos] & 0xF;
+                        const int hi_4   = (b_ptr[l].ql[ql_pos] >> 4) & 0xF;
+
+                        const int qh_idx_l    = qh_half_l + ((base_l + i) % 32);
+                        const int qh_chunk_l  = qh_idx_l / blocklen;
+                        const int qh_pos_l    = qh_idx_l % blocklen;
+                        const int qh_offset_l = qh_chunk_l * (blocklen * ncols_interleaved) + j * blocklen + qh_pos_l;
+                        const int hi_2_l      = (b_ptr[l].qh[qh_offset_l] >> qh_shift_l) & 0x3;
+
+                        const int qh_idx_h    = qh_half_h + ((base_h + i) % 32);
+                        const int qh_chunk_h  = qh_idx_h / blocklen;
+                        const int qh_pos_h    = qh_idx_h % blocklen;
+                        const int qh_offset_h = qh_chunk_h * (blocklen * ncols_interleaved) + j * blocklen + qh_pos_h;
+                        const int hi_2_h      = (b_ptr[l].qh[qh_offset_h] >> qh_shift_h) & 0x3;
+
+                        const int q_l = ((hi_2_l << 4) | l_4) - 32;
+                        const int q_h = ((hi_2_h << 4) | hi_4) - 32;
+
+                        const int8_t a_l = a_ptr[l].qs[base_l + i];
+                        const int8_t a_h = a_ptr[l].qs[base_h + i];
+
+                        sumi_l += q_l * a_l;
+                        sumi_h += q_h * a_h;
                     }
-                    sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
+
+                    sumf[j] +=
+                        (sumi_l * scale_l + sumi_h * scale_h) * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d;
                 }
             }
         }
-        for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
+
+        for (int j = 0; j < ncols_interleaved; j++) {
+            s[x * ncols_interleaved + j] = sumf[j];
+        }
     }
 }
 
-void ggml_gemv_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
-    const int qk = QK8_0;
-    const int nb = n / qk;
-    const int ncols_interleaved = 4;
-    const int blocklen = 8;
+template <int M, int N>
+static void ggml_gemm_q6_K_NxM_q8_K_generic_impl(int                        n,
+                                                 float * GGML_RESTRICT      s,
+                                                 size_t                     bs,
+                                                 const void * GGML_RESTRICT vx,
+                                                 const void * GGML_RESTRICT vy,
+                                                 int                        nr,
+                                                 int                        nc) {
+    constexpr int blocklen          = M;
+    constexpr int ncols_interleaved = N;
+    const int     qk                = QK_K;
+    const int     nb                = n / qk;
+    const int     blocks_per_half   = 64 / blocklen;
+    const int     q8_half_stride    = 512;
+    const int     q8_low_high_step  = 256;
 
-    assert (n % qk == 0);
-    assert (nc % ncols_interleaved == 0);
+    assert(n % qk == 0);
+    assert(nr % 4 == 0);
+    assert(nc % ncols_interleaved == 0);
 
-    UNUSED(s);
     UNUSED(bs);
-    UNUSED(vx);
-    UNUSED(vy);
-    UNUSED(nr);
-    UNUSED(nc);
-    UNUSED(nb);
-    UNUSED(ncols_interleaved);
-    UNUSED(blocklen);
 
-    float sumf[4];
-    int sumi;
+    float sumf[4][8];
 
-    const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
-    for (int x = 0; x < nc / ncols_interleaved; x++) {
-        const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb);
+    for (int y = 0; y < nr / 4; y++) {
+        const block_q8_Kx4 * a_ptr = (const block_q8_Kx4 *) vy + (y * nb);
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q6_Kx8 * b_ptr = (const block_q6_Kx8 *) vx + (x * nb);
 
-        for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
-        for (int l = 0; l < nb; l++) {
-            for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+            for (int m = 0; m < 4; m++) {
                 for (int j = 0; j < ncols_interleaved; j++) {
-                    sumi = 0;
-                    for (int i = 0; i < blocklen; ++i) {
-                        const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
-                        const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
-                        sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
-                    }
-                    sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
+                    sumf[m][j] = 0.0f;
                 }
             }
-        }
-        for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
-    }
-}
 
-void ggml_gemv_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
-    const int qk = QK8_0;
-    const int nb = n / qk;
-    const int ncols_interleaved = 8;
-    const int blocklen = 8;
+            for (int l = 0; l < nb; l++) {
+                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                    const int base_l = (k / blocks_per_half) * 128 + (k % blocks_per_half) * blocklen;
+                    const int base_h = base_l + 64;
 
-    assert (n % qk == 0);
-    assert (nc % ncols_interleaved == 0);
+                    const int scale_idx_l = base_l / 16;
+                    const int scale_idx_h = base_h / 16;
 
-    UNUSED(s);
-    UNUSED(bs);
-    UNUSED(vx);
-    UNUSED(vy);
-    UNUSED(nr);
-    UNUSED(nc);
-    UNUSED(nb);
-    UNUSED(ncols_interleaved);
-    UNUSED(blocklen);
+                    const int qh_shift_l = ((base_l % 128) / 32) * 2;
+                    const int qh_shift_h = ((base_h % 128) / 32) * 2;
 
-    float sumf[8];
-    int sumi;
+                    const int qh_half_l = (base_l / 128) * 32;
+                    const int qh_half_h = (base_h / 128) * 32;
 
-    const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
-    for (int x = 0; x < nc / ncols_interleaved; x++) {
-        const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
+                    const int q8_base = (k / blocks_per_half) * q8_half_stride + (k % blocks_per_half) * (blocklen * 4);
 
-        for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
-        for (int l = 0; l < nb; l++) {
-            for (int k = 0; k < (qk / (2 * blocklen)); k++) {
-                for (int j = 0; j < ncols_interleaved; j++) {
-                    sumi = 0;
-                    for (int i = 0; i < blocklen; ++i) {
-                        const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
-                        const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
-                        sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
+                    for (int m = 0; m < 4; m++) {
+                        for (int j = 0; j < ncols_interleaved; j++) {
+                            const int8_t scale_l = b_ptr[l].scales[scale_idx_l * ncols_interleaved + j];
+                            const int8_t scale_h = b_ptr[l].scales[scale_idx_h * ncols_interleaved + j];
+
+                            int sumi_l = 0;
+                            int sumi_h = 0;
+
+                            for (int i = 0; i < blocklen; i++) {
+                                const int ql_pos = k * ncols_interleaved * blocklen + j * blocklen + i;
+                                const int l_4    = b_ptr[l].ql[ql_pos] & 0xF;
+                                const int hi_4   = (b_ptr[l].ql[ql_pos] >> 4) & 0xF;
+
+                                const int qh_idx_l   = qh_half_l + ((base_l + i) % 32);
+                                const int qh_chunk_l = qh_idx_l / blocklen;
+                                const int qh_pos_l   = qh_idx_l % blocklen;
+                                const int qh_offset_l =
+                                    qh_chunk_l * (blocklen * ncols_interleaved) + j * blocklen + qh_pos_l;
+                                const int hi_2_l = (b_ptr[l].qh[qh_offset_l] >> qh_shift_l) & 0x3;
+
+                                const int qh_idx_h   = qh_half_h + ((base_h + i) % 32);
+                                const int qh_chunk_h = qh_idx_h / blocklen;
+                                const int qh_pos_h   = qh_idx_h % blocklen;
+                                const int qh_offset_h =
+                                    qh_chunk_h * (blocklen * ncols_interleaved) + j * blocklen + qh_pos_h;
+                                const int hi_2_h = (b_ptr[l].qh[qh_offset_h] >> qh_shift_h) & 0x3;
+
+                                const int q_l = ((hi_2_l << 4) | l_4) - 32;
+                                const int q_h = ((hi_2_h << 4) | hi_4) - 32;
+
+                                const int8_t q8_l = a_ptr[l].qs[q8_base + m * blocklen + i];
+                                const int8_t q8_h = a_ptr[l].qs[q8_base + m * blocklen + i + q8_low_high_step];
+
+                                sumi_l += q_l * q8_l;
+                                sumi_h += q_h * q8_h;
+                            }
+
+                            sumf[m][j] += (sumi_l * scale_l + sumi_h * scale_h) * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) *
+                                          a_ptr[l].d[m];
+                        }
                     }
-                    sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
+                }
+            }
+
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
                 }
             }
         }
-        for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
     }
 }
 
-void ggml_gemv_q4_K_8x4_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
-    const int qk = QK_K;
-    const int nb = n / qk;
-    const int ncols_interleaved = 8;
-    const int blocklen = 4;
-    static const uint32_t kmask1 = 0x3f3f3f3f;
-    static const uint32_t kmask2 = 0x0f0f0f0f;
-    static const uint32_t kmask3 = 0x03030303;
+template <int M, int N>
+static void ggml_gemv_q5_K_NxM_q8_K_generic_impl(int                        n,
+                                                 float * GGML_RESTRICT      s,
+                                                 size_t                     bs,
+                                                 const void * GGML_RESTRICT vx,
+                                                 const void * GGML_RESTRICT vy,
+                                                 int                        nr,
+                                                 int                        nc) {
+    constexpr int         blocklen          = M;
+    constexpr int         ncols_interleaved = N;
+    const int             qk                = QK_K;
+    const int             nb                = n / qk;
+    static const uint32_t kmask1            = 0x3f3f3f3f;
+    static const uint32_t kmask2            = 0x0f0f0f0f;
+    static const uint32_t kmask3            = 0x03030303;
 
-    assert (n % qk == 0);
-    assert (nc % ncols_interleaved == 0);
+    assert(n % qk == 0);
+    assert(nc % ncols_interleaved == 0);
 
     UNUSED(bs);
     UNUSED(nr);
 
-    float sumf[8];
-    float sum_minf[8];
+    float    sumf[ncols_interleaved];
+    float    sum_minf[ncols_interleaved];
     uint32_t utmp[32];
-    int sumi1;
-    int sumi2;
-    int sumi;
+    int      sumi1;
+    int      sumi2;
+    int      sumi;
 
     const block_q8_K * a_ptr = (const block_q8_K *) vy;
     for (int x = 0; x < nc / ncols_interleaved; x++) {
-        const block_q4_Kx8 * b_ptr = (const block_q4_Kx8 *) vx + (x * nb);
+        const block_q5_Kx8 * b_ptr = (const block_q5_Kx8 *) vx + (x * nb);
 
         for (int j = 0; j < ncols_interleaved; j++) {
-            sumf[j] = 0.0;
+            sumf[j]     = 0.0;
             sum_minf[j] = 0.0;
         }
         for (int l = 0; l < nb; l++) {
             for (int sb = 0; sb < 8; sb++) {
-                memcpy(utmp + sb * 4, b_ptr[l].scales + sb * 12, 12);
-                utmp[sb * 4 + 3] = ((utmp[sb * 4 + 2] >> 4) & kmask2) | (((utmp[sb * 4 + 1] >> 6) & kmask3) << 4);
+                memcpy(utmp + sb * 4, b_ptr[l].scales + sb * K_SCALE_SIZE, K_SCALE_SIZE);
+                utmp[sb * 4 + 3]      = ((utmp[sb * 4 + 2] >> 4) & kmask2) | (((utmp[sb * 4 + 1] >> 6) & kmask3) << 4);
                 const uint32_t uaux_0 = utmp[sb * 4 + 1] & kmask1;
-                utmp[sb * 4 + 1] = (utmp[sb * 4 + 2] & kmask2) | (((utmp[sb * 4 + 0] >> 6) & kmask3) << 4);
-                utmp[sb * 4 + 2] = uaux_0;
+                utmp[sb * 4 + 1]      = (utmp[sb * 4 + 2] & kmask2) | (((utmp[sb * 4 + 0] >> 6) & kmask3) << 4);
+                utmp[sb * 4 + 2]      = uaux_0;
                 utmp[sb * 4 + 0] &= kmask1;
             }
             for (int k = 0; k < (qk / (2 * blocklen)); k++) {
-                uint8_t * scales_0 = (uint8_t *) utmp + (k / 8) * 32;
-                uint8_t * scales_1 = (uint8_t *) utmp + (k / 8) * 32 + 16;
+                constexpr int scale_stride = 32;
+                uint8_t *     scales_0     = (uint8_t *) utmp + (k / (32 / blocklen)) * scale_stride;
+                uint8_t *     scales_1     = (uint8_t *) utmp + (k / (32 / blocklen)) * scale_stride + 16;
+
+                const int qh_shift = (k / (32 / blocklen)) * 2;
                 for (int j = 0; j < ncols_interleaved; j++) {
                     sumi1 = 0;
                     sumi2 = 0;
-                    sumi = 0;
+                    sumi  = 0;
                     for (int i = 0; i < blocklen; ++i) {
-                        const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF);
-                        const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4);
-                        sumi1 = (v0 * a_ptr[l].qs[(k / 8) * 64 + (k % 8) * blocklen + i]);
-                        sumi2 = (v1 * a_ptr[l].qs[(k / 8) * 64 + (k % 8) * blocklen + i + 32]);
+                        const int b_qs_offset = k * ncols_interleaved * blocklen + j * blocklen + i;
+
+                        const int qh_idx      = (k * blocklen + i) % 32;
+                        const int qh_chunk    = qh_idx / blocklen;
+                        const int qh_pos      = qh_idx % blocklen;
+                        const int b_qh_offset = qh_chunk * (blocklen * ncols_interleaved) + j * blocklen + qh_pos;
+
+                        const uint8_t qh_val = b_ptr[l].qh[b_qh_offset];
+                        const uint8_t h0     = (qh_val >> qh_shift) & 1;
+                        const uint8_t h1     = (qh_val >> (qh_shift + 1)) & 1;
+
+                        const int v0 = (int8_t) ((b_ptr[l].qs[b_qs_offset] & 0xF) | (h0 << 4));
+                        const int v1 = (int8_t) ((b_ptr[l].qs[b_qs_offset] >> 4) | (h1 << 4));
+
+                        const int q8_offset = (k / (32 / blocklen)) * 64 + (k % (32 / blocklen)) * blocklen + i;
+
+                        sumi1 = (v0 * a_ptr[l].qs[q8_offset]);
+                        sumi2 = (v1 * a_ptr[l].qs[q8_offset + 32]);
                         sumi1 = sumi1 * scales_0[j];
                         sumi2 = sumi2 * scales_1[j];
                         sumi += sumi1 + sumi2;
@@ -452,7 +632,8 @@ void ggml_gemv_q4_K_8x4_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
             for (int sb = 0; sb < 8; sb++) {
                 uint8_t * mins = (uint8_t *) utmp + 8 + sb * 16;
                 for (int j = 0; j < ncols_interleaved; j++) {
-                    sum_minf[j] += mins[j] * (a_ptr[l].bsums[sb * 2] + a_ptr[l].bsums[sb * 2 + 1]) * GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d;
+                    sum_minf[j] += mins[j] * (a_ptr[l].bsums[sb * 2] + a_ptr[l].bsums[sb * 2 + 1]) *
+                                   GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d;
                 }
             }
         }
@@ -462,17 +643,123 @@ void ggml_gemv_q4_K_8x4_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
     }
 }
 
-void ggml_gemv_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
-    const int qk = QK_K;
+template <int M, int N>
+static void ggml_gemm_q5_K_NxM_q8_K_generic_impl(int                        n,
+                                                 float * GGML_RESTRICT      s,
+                                                 size_t                     bs,
+                                                 const void * GGML_RESTRICT vx,
+                                                 const void * GGML_RESTRICT vy,
+                                                 int                        nr,
+                                                 int                        nc) {
+    constexpr int         blocklen          = M;
+    constexpr int         ncols_interleaved = N;
+    const int             qk                = QK_K;
+    const int             nb                = n / qk;
+    static const uint32_t kmask1            = 0x3f3f3f3f;
+    static const uint32_t kmask2            = 0x0f0f0f0f;
+    static const uint32_t kmask3            = 0x03030303;
+
+    assert(n % qk == 0);
+    assert(nr % 4 == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    float    sumf[4][ncols_interleaved];
+    float    sum_minf[4][ncols_interleaved];
+    uint32_t utmp[32];
+    int      sumi1;
+    int      sumi2;
+    int      sumi;
+
+    for (int y = 0; y < nr / 4; y++) {
+        const block_q8_Kx4 * a_ptr = (const block_q8_Kx4 *) vy + (y * nb);
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q5_Kx8 * b_ptr = (const block_q5_Kx8 *) vx + (x * nb);
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sumf[m][j]     = 0.0;
+                    sum_minf[m][j] = 0.0;
+                }
+            }
+            for (int l = 0; l < nb; l++) {
+                for (int sb = 0; sb < 8; sb++) {
+                    memcpy(utmp + sb * 4, b_ptr[l].scales + sb * K_SCALE_SIZE, K_SCALE_SIZE);
+                    utmp[sb * 4 + 3] = ((utmp[sb * 4 + 2] >> 4) & kmask2) | (((utmp[sb * 4 + 1] >> 6) & kmask3) << 4);
+                    const uint32_t uaux_0 = utmp[sb * 4 + 1] & kmask1;
+                    utmp[sb * 4 + 1]      = (utmp[sb * 4 + 2] & kmask2) | (((utmp[sb * 4 + 0] >> 6) & kmask3) << 4);
+                    utmp[sb * 4 + 2]      = uaux_0;
+                    utmp[sb * 4 + 0] &= kmask1;
+                }
+                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                    constexpr int scale_stride = 32;
+                    uint8_t *     scales_0     = (uint8_t *) utmp + (k / (32 / blocklen)) * scale_stride;
+                    uint8_t *     scales_1     = (uint8_t *) utmp + (k / (32 / blocklen)) * scale_stride + 16;
+
+                    const int qh_shift = (k / (32 / blocklen)) * 2;
+                    for (int m = 0; m < 4; m++) {
+                        for (int j = 0; j < ncols_interleaved; j++) {
+                            sumi1 = 0;
+                            sumi2 = 0;
+                            sumi  = 0;
+                            for (int i = 0; i < blocklen; ++i) {
+                                const int b_qs_offset = k * ncols_interleaved * blocklen + j * blocklen + i;
+
+                                const int qh_idx   = (k * blocklen + i) % 32;
+                                const int qh_chunk = qh_idx / blocklen;
+                                const int qh_pos   = qh_idx % blocklen;
+                                const int b_qh_offset =
+                                    qh_chunk * (blocklen * ncols_interleaved) + j * blocklen + qh_pos;
+
+                                const uint8_t qh_val = b_ptr[l].qh[b_qh_offset];
+                                const uint8_t h0     = (qh_val >> qh_shift) & 1;
+                                const uint8_t h1     = (qh_val >> (qh_shift + 1)) & 1;
+
+                                const int v0 = (int8_t) ((b_ptr[l].qs[b_qs_offset] & 0xF) | (h0 << 4));
+                                const int v1 = (int8_t) ((b_ptr[l].qs[b_qs_offset] >> 4) | (h1 << 4));
+
+                                const int q8_offset = (k / (32 / blocklen)) * 256 +
+                                                      (k % (32 / blocklen)) * 4 * blocklen + m * blocklen + i;
+
+                                sumi1 = (v0 * a_ptr[l].qs[q8_offset]);
+                                sumi2 = (v1 * a_ptr[l].qs[q8_offset + 128]);
+                                sumi1 = sumi1 * scales_0[j];
+                                sumi2 = sumi2 * scales_1[j];
+                                sumi += sumi1 + sumi2;
+                            }
+                            sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d[m];
+                        }
+                    }
+                }
+                for (int sb = 0; sb < 8; sb++) {
+                    uint8_t * mins = (uint8_t *) utmp + 8 + sb * 16;
+                    for (int m = 0; m < 4; m++) {
+                        const int16_t * bsums = a_ptr[l].bsums + (sb * 8) + (m * 4) - ((sb % 2) * 6);
+                        for (int j = 0; j < ncols_interleaved; j++) {
+                            sum_minf[m][j] += mins[j] * (bsums[0] + bsums[1]) *
+                                              GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d[m];
+                        }
+                    }
+                }
+            }
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j] - sum_minf[m][j];
+                }
+            }
+        }
+    }
+}
+
+extern "C" {
+
+void ggml_gemv_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    const int qk = QK8_0;
     const int nb = n / qk;
-    const int ncols_interleaved = 8;
-    const int blocklen = 8;
-    static const uint32_t kmask1 = 0x3f3f3f3f;
-    static const uint32_t kmask2 = 0x0f0f0f0f;
-    static const uint32_t kmask3 = 0x03030303;
+    const int ncols_interleaved = 4;
+    const int blocklen = 4;
 
-    assert (n % qk == 0);
-    assert (nc % ncols_interleaved == 0);
+    assert(nr == 1);
+    assert(n % qk == 0);
+    assert(nc % ncols_interleaved == 0);
 
     UNUSED(s);
     UNUSED(bs);
@@ -484,66 +771,35 @@ void ggml_gemv_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
     UNUSED(ncols_interleaved);
     UNUSED(blocklen);
 
-    float sumf[8];
-    float sum_minf[8];
-    uint32_t utmp[32];
-    int sumi1;
-    int sumi2;
+    float sumf[4];
     int sumi;
 
-    const block_q8_K * a_ptr = (const block_q8_K *) vy;
+    const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
     for (int x = 0; x < nc / ncols_interleaved; x++) {
-        const block_q4_Kx8 * b_ptr = (const block_q4_Kx8 *) vx + (x * nb);
+        const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb);
 
-        for (int j = 0; j < ncols_interleaved; j++) {
-            sumf[j] = 0.0;
-            sum_minf[j] = 0.0;
-        }
+        for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
         for (int l = 0; l < nb; l++) {
-            for (int sb = 0; sb < 8; sb++) {
-                memcpy(utmp + sb * 4, b_ptr[l].scales + sb * 12, 12);
-                utmp[sb * 4 + 3] = ((utmp[sb * 4 + 2] >> 4) & kmask2) | (((utmp[sb * 4 + 1] >> 6) & kmask3) << 4);
-                const uint32_t uaux_0 = utmp[sb * 4 + 1] & kmask1;
-                utmp[sb * 4 + 1] = (utmp[sb * 4 + 2] & kmask2) | (((utmp[sb * 4 + 0] >> 6) & kmask3) << 4);
-                utmp[sb * 4 + 2] = uaux_0;
-                utmp[sb * 4 + 0] &= kmask1;
-            }
             for (int k = 0; k < (qk / (2 * blocklen)); k++) {
-                uint8_t *scales_0 = (uint8_t*) utmp + (k / 4) * 32;
-                uint8_t *scales_1 = (uint8_t*) utmp + (k / 4) * 32 + 16;
                 for (int j = 0; j < ncols_interleaved; j++) {
-                    sumi1 = 0;
-                    sumi2 = 0;
                     sumi = 0;
                     for (int i = 0; i < blocklen; ++i) {
-                        const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF);
-                        const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4);
-                        sumi1 = (v0 * a_ptr[l].qs[(k >> 2) * 64 + (k % 4) * blocklen + i]);
-                        sumi2 = (v1 * a_ptr[l].qs[(k >> 2) * 64 + (k % 4) * blocklen + i + 32]);
-                        sumi1 = sumi1 * scales_0[j];
-                        sumi2 = sumi2 * scales_1[j];
-                        sumi += sumi1 + sumi2;
+                        const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
+                        const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                        sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
                     }
-                    sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d;
-                }
-            }
-            for (int sb = 0; sb < 8; sb++) {
-                uint8_t *mins = (uint8_t*) utmp + 8 + sb * 16;
-                for (int j = 0; j < ncols_interleaved; j++) {
-                    sum_minf[j] += mins[j] * (a_ptr[l].bsums[sb * 2] + a_ptr[l].bsums[sb * 2 + 1]) * GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d;
+                    sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
                 }
             }
         }
-        for (int j = 0; j < ncols_interleaved; j++) {
-            s[x * ncols_interleaved + j] = sumf[j] - sum_minf[j];
-        }
+        for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
     }
 }
 
-void ggml_gemv_q2_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
-    const int qk = QK_K;
+void ggml_gemv_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    const int qk = QK8_0;
     const int nb = n / qk;
-    const int ncols_interleaved = 8;
+    const int ncols_interleaved = 4;
     const int blocklen = 8;
 
     assert (n % qk == 0);
@@ -559,82 +815,56 @@ void ggml_gemv_q2_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
     UNUSED(ncols_interleaved);
     UNUSED(blocklen);
 
-    float sumf[8];
-    float sum_minf[8];
-    int sumi1,sumi2,sumi3,sumi4;
+    float sumf[4];
     int sumi;
 
-    const block_q8_K * a_ptr = (const block_q8_K *)vy;
-    for(int x = 0; x < nc / ncols_interleaved; x++) {
-        const block_q2_Kx8 * b_ptr = (const block_q2_Kx8 *) vx + (x * nb);
-        for (int j = 0; j < ncols_interleaved; j++) {
-            sumf[j] = 0.0;
-            sum_minf[j] = 0.0;
-        }
+    const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb);
+
+        for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
         for (int l = 0; l < nb; l++) {
-            for (int k = 0; k < (qk / (4 * blocklen)); k++) {
-                const uint8_t *scales_0 = b_ptr[l].scales + (k / 4) * 64 ;
-                const uint8_t *scales_1 = b_ptr[l].scales + (k / 4) * 64 + 16;
-                const uint8_t *scales_2 = b_ptr[l].scales + (k / 4) * 64 + 32;
-                const uint8_t *scales_3 = b_ptr[l].scales + (k / 4) * 64 + 48;
+            for (int k = 0; k < (qk / (2 * blocklen)); k++) {
                 for (int j = 0; j < ncols_interleaved; j++) {
-                    sumi1 = 0;
-                    sumi2 = 0;
-                    sumi3 = 0;
-                    sumi4 = 0;
                     sumi = 0;
-                    int offset = ((k / 2) % 2) + j * 2;
-                    for (int i = 0; i < blocklen; ++i){
-                        const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 3);
-                        const int v1 = (int8_t) ((b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 2 ) & 3);
-                        const int v2 = (int8_t) ((b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4 ) & 3);
-                        const int v3 = (int8_t) ((b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 6 ) & 3);
-                        sumi1 = (v0 * a_ptr[l].qs[(k >> 2) * 128 + (k % 4) * blocklen + i]);
-                        sumi2 = (v1 * a_ptr[l].qs[(k >> 2) * 128 + (k % 4) * blocklen + i + 32]);
-                        sumi3 = (v2 * a_ptr[l].qs[(k >> 2) * 128 + (k % 4) * blocklen + i + 64]);
-                        sumi4 = (v3 * a_ptr[l].qs[(k >> 2) * 128 + (k % 4) * blocklen + i + 96]);
-
-                        sumi1 = sumi1 * (scales_0[offset] & 0xF);
-                        sumi2 = sumi2 * (scales_1[offset] & 0xF);
-                        sumi3 = sumi3 * (scales_2[offset] & 0xF);
-                        sumi4 = sumi4 * (scales_3[offset] & 0xF);
-                        sumi += sumi1 + sumi2 + sumi3 + sumi4;
+                    for (int i = 0; i < blocklen; ++i) {
+                        const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
+                        const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                        sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
                     }
-                    sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d;
-                }
-            }
-            for(int sb = 0; sb < 8; sb++) {
-                const uint8_t *mins = b_ptr[l].scales + sb * 16;
-                for(int j = 0; j < ncols_interleaved; j++){
-                    sum_minf[j] += ((mins[j * 2] >> 4) * a_ptr[l].bsums[sb * 2] + (mins[(j * 2)+ 1] >> 4) * a_ptr[l].bsums[sb * 2 + 1]) * GGML_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d;
+                    sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
                 }
             }
         }
-        for (int j = 0; j < ncols_interleaved; j++) {
-            s[x * ncols_interleaved + j] = sumf[j] - sum_minf[j];
-        }
+        for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
     }
 }
 
-void ggml_gemv_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+void ggml_gemv_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
     const int qk = QK8_0;
     const int nb = n / qk;
-    const int ncols_interleaved = 4;
-    const int blocklen = 4;
+    const int ncols_interleaved = 8;
+    const int blocklen = 8;
 
-    assert(nr == 1);
-    assert(n % qk == 0);
-    assert(nc % ncols_interleaved == 0);
+    assert (n % qk == 0);
+    assert (nc % ncols_interleaved == 0);
 
+    UNUSED(s);
     UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
     UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
 
-    float sumf[4];
+    float sumf[8];
     int sumi;
 
     const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
     for (int x = 0; x < nc / ncols_interleaved; x++) {
-        const block_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) vx + (x * nb);
+        const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
 
         for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
         for (int l = 0; l < nb; l++) {
@@ -642,9 +872,9 @@ void ggml_gemv_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs
                 for (int j = 0; j < ncols_interleaved; j++) {
                     sumi = 0;
                     for (int i = 0; i < blocklen; ++i) {
-                        const int v0 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F];
-                        const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4];
-                        sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2]));
+                        const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
+                        const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                        sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
                     }
                     sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
                 }
@@ -654,139 +884,1212 @@ void ggml_gemv_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs
     }
 }
 
-void ggml_gemv_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
-    const int qk = QK8_0;
+void ggml_gemv_q4_K_8x4_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    const int qk = QK_K;
     const int nb = n / qk;
     const int ncols_interleaved = 8;
-    const int blocklen = 8;
+    const int blocklen = 4;
+    static const uint32_t kmask1 = 0x3f3f3f3f;
+    static const uint32_t kmask2 = 0x0f0f0f0f;
+    static const uint32_t kmask3 = 0x03030303;
 
-    assert(nr == 1);
-    assert(n % qk == 0);
-    assert(nc % ncols_interleaved == 0);
+    assert (n % qk == 0);
+    assert (nc % ncols_interleaved == 0);
 
     UNUSED(bs);
     UNUSED(nr);
 
     float sumf[8];
+    float sum_minf[8];
+    uint32_t utmp[32];
+    int sumi1;
+    int sumi2;
     int sumi;
 
-    const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+    const block_q8_K * a_ptr = (const block_q8_K *) vy;
     for (int x = 0; x < nc / ncols_interleaved; x++) {
-        const block_iq4_nlx8 * b_ptr = (const block_iq4_nlx8 *) vx + (x * nb);
+        const block_q4_Kx8 * b_ptr = (const block_q4_Kx8 *) vx + (x * nb);
 
-        for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
+        for (int j = 0; j < ncols_interleaved; j++) {
+            sumf[j] = 0.0;
+            sum_minf[j] = 0.0;
+        }
         for (int l = 0; l < nb; l++) {
+            for (int sb = 0; sb < 8; sb++) {
+                memcpy(utmp + sb * 4, b_ptr[l].scales + sb * 12, 12);
+                utmp[sb * 4 + 3] = ((utmp[sb * 4 + 2] >> 4) & kmask2) | (((utmp[sb * 4 + 1] >> 6) & kmask3) << 4);
+                const uint32_t uaux_0 = utmp[sb * 4 + 1] & kmask1;
+                utmp[sb * 4 + 1] = (utmp[sb * 4 + 2] & kmask2) | (((utmp[sb * 4 + 0] >> 6) & kmask3) << 4);
+                utmp[sb * 4 + 2] = uaux_0;
+                utmp[sb * 4 + 0] &= kmask1;
+            }
             for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                uint8_t * scales_0 = (uint8_t *) utmp + (k / 8) * 32;
+                uint8_t * scales_1 = (uint8_t *) utmp + (k / 8) * 32 + 16;
                 for (int j = 0; j < ncols_interleaved; j++) {
-                    sumi = 0;
+                    sumi1 = 0;
+                    sumi2 = 0;
+                    sumi = 0;
+                    for (int i = 0; i < blocklen; ++i) {
+                        const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF);
+                        const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4);
+                        sumi1 = (v0 * a_ptr[l].qs[(k / 8) * 64 + (k % 8) * blocklen + i]);
+                        sumi2 = (v1 * a_ptr[l].qs[(k / 8) * 64 + (k % 8) * blocklen + i + 32]);
+                        sumi1 = sumi1 * scales_0[j];
+                        sumi2 = sumi2 * scales_1[j];
+                        sumi += sumi1 + sumi2;
+                    }
+                    sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d;
+                }
+            }
+            for (int sb = 0; sb < 8; sb++) {
+                uint8_t * mins = (uint8_t *) utmp + 8 + sb * 16;
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sum_minf[j] += mins[j] * (a_ptr[l].bsums[sb * 2] + a_ptr[l].bsums[sb * 2 + 1]) * GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d;
+                }
+            }
+        }
+        for (int j = 0; j < ncols_interleaved; j++) {
+            s[x * ncols_interleaved + j] = sumf[j] - sum_minf[j];
+        }
+    }
+}
+
+void ggml_gemv_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    const int qk = QK_K;
+    const int nb = n / qk;
+    const int ncols_interleaved = 8;
+    const int blocklen = 8;
+    static const uint32_t kmask1 = 0x3f3f3f3f;
+    static const uint32_t kmask2 = 0x0f0f0f0f;
+    static const uint32_t kmask3 = 0x03030303;
+
+    assert (n % qk == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(bs);
+    UNUSED(nr);
+
+    float sumf[8];
+    float sum_minf[8];
+    uint32_t utmp[32];
+    int sumi1;
+    int sumi2;
+    int sumi;
+
+    const block_q8_K * a_ptr = (const block_q8_K *) vy;
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_q4_Kx8 * b_ptr = (const block_q4_Kx8 *) vx + (x * nb);
+
+        for (int j = 0; j < ncols_interleaved; j++) {
+            sumf[j] = 0.0;
+            sum_minf[j] = 0.0;
+        }
+        for (int l = 0; l < nb; l++) {
+            for (int sb = 0; sb < 8; sb++) {
+                memcpy(utmp + sb * 4, b_ptr[l].scales + sb * 12, 12);
+                utmp[sb * 4 + 3] = ((utmp[sb * 4 + 2] >> 4) & kmask2) | (((utmp[sb * 4 + 1] >> 6) & kmask3) << 4);
+                const uint32_t uaux_0 = utmp[sb * 4 + 1] & kmask1;
+                utmp[sb * 4 + 1] = (utmp[sb * 4 + 2] & kmask2) | (((utmp[sb * 4 + 0] >> 6) & kmask3) << 4);
+                utmp[sb * 4 + 2] = uaux_0;
+                utmp[sb * 4 + 0] &= kmask1;
+            }
+            for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                uint8_t *scales_0 = (uint8_t*) utmp + (k / 4) * 32;
+                uint8_t *scales_1 = (uint8_t*) utmp + (k / 4) * 32 + 16;
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sumi1 = 0;
+                    sumi2 = 0;
+                    sumi = 0;
+                    for (int i = 0; i < blocklen; ++i) {
+                        const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF);
+                        const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4);
+                        sumi1 = (v0 * a_ptr[l].qs[(k >> 2) * 64 + (k % 4) * blocklen + i]);
+                        sumi2 = (v1 * a_ptr[l].qs[(k >> 2) * 64 + (k % 4) * blocklen + i + 32]);
+                        sumi1 = sumi1 * scales_0[j];
+                        sumi2 = sumi2 * scales_1[j];
+                        sumi += sumi1 + sumi2;
+                    }
+                    sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d;
+                }
+            }
+            for (int sb = 0; sb < 8; sb++) {
+                uint8_t *mins = (uint8_t*) utmp + 8 + sb * 16;
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sum_minf[j] += mins[j] * (a_ptr[l].bsums[sb * 2] + a_ptr[l].bsums[sb * 2 + 1]) * GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d;
+                }
+            }
+        }
+        for (int j = 0; j < ncols_interleaved; j++) {
+            s[x * ncols_interleaved + j] = sumf[j] - sum_minf[j];
+        }
+    }
+}
+
+void ggml_gemv_q2_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    const int qk = QK_K;
+    const int nb = n / qk;
+    const int ncols_interleaved = 8;
+    const int blocklen = 8;
+
+    assert (n % qk == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(s);
+    UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
+    UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+    float sumf[8];
+    float sum_minf[8];
+    int sumi1,sumi2,sumi3,sumi4;
+    int sumi;
+
+    const block_q8_K * a_ptr = (const block_q8_K *)vy;
+    for(int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_q2_Kx8 * b_ptr = (const block_q2_Kx8 *) vx + (x * nb);
+        for (int j = 0; j < ncols_interleaved; j++) {
+            sumf[j] = 0.0;
+            sum_minf[j] = 0.0;
+        }
+        for (int l = 0; l < nb; l++) {
+            for (int k = 0; k < (qk / (4 * blocklen)); k++) {
+                const uint8_t *scales_0 = b_ptr[l].scales + (k / 4) * 64 ;
+                const uint8_t *scales_1 = b_ptr[l].scales + (k / 4) * 64 + 16;
+                const uint8_t *scales_2 = b_ptr[l].scales + (k / 4) * 64 + 32;
+                const uint8_t *scales_3 = b_ptr[l].scales + (k / 4) * 64 + 48;
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sumi1 = 0;
+                    sumi2 = 0;
+                    sumi3 = 0;
+                    sumi4 = 0;
+                    sumi = 0;
+                    int offset = ((k / 2) % 2) + j * 2;
+                    for (int i = 0; i < blocklen; ++i){
+                        const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 3);
+                        const int v1 = (int8_t) ((b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 2 ) & 3);
+                        const int v2 = (int8_t) ((b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4 ) & 3);
+                        const int v3 = (int8_t) ((b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 6 ) & 3);
+                        sumi1 = (v0 * a_ptr[l].qs[(k >> 2) * 128 + (k % 4) * blocklen + i]);
+                        sumi2 = (v1 * a_ptr[l].qs[(k >> 2) * 128 + (k % 4) * blocklen + i + 32]);
+                        sumi3 = (v2 * a_ptr[l].qs[(k >> 2) * 128 + (k % 4) * blocklen + i + 64]);
+                        sumi4 = (v3 * a_ptr[l].qs[(k >> 2) * 128 + (k % 4) * blocklen + i + 96]);
+
+                        sumi1 = sumi1 * (scales_0[offset] & 0xF);
+                        sumi2 = sumi2 * (scales_1[offset] & 0xF);
+                        sumi3 = sumi3 * (scales_2[offset] & 0xF);
+                        sumi4 = sumi4 * (scales_3[offset] & 0xF);
+                        sumi += sumi1 + sumi2 + sumi3 + sumi4;
+                    }
+                    sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d;
+                }
+            }
+            for(int sb = 0; sb < 8; sb++) {
+                const uint8_t *mins = b_ptr[l].scales + sb * 16;
+                for(int j = 0; j < ncols_interleaved; j++){
+                    sum_minf[j] += ((mins[j * 2] >> 4) * a_ptr[l].bsums[sb * 2] + (mins[(j * 2)+ 1] >> 4) * a_ptr[l].bsums[sb * 2 + 1]) * GGML_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d;
+                }
+            }
+        }
+        for (int j = 0; j < ncols_interleaved; j++) {
+            s[x * ncols_interleaved + j] = sumf[j] - sum_minf[j];
+        }
+    }
+}
+
+void ggml_gemv_q5_K_8x4_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    ggml_gemv_q5_K_NxM_q8_K_generic_impl<4, 8>(n, s, bs, vx, vy, nr, nc);
+}
+
+void ggml_gemv_q5_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    ggml_gemv_q5_K_NxM_q8_K_generic_impl<8, 8>(n, s, bs, vx, vy, nr, nc);
+}
+
+
+void ggml_gemv_q6_K_8x4_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    ggml_gemv_q6_K_NxM_q8_K_generic_impl<4, 8>(n, s, bs, vx, vy, nr, nc);
+}
+
+void ggml_gemv_q6_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    ggml_gemv_q6_K_NxM_q8_K_generic_impl<8, 8>(n, s, bs, vx, vy, nr, nc);
+}
+
+void ggml_gemv_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 4;
+    const int blocklen = 4;
+
+    assert(nr == 1);
+    assert(n % qk == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    UNUSED(bs);
+    UNUSED(nr);
+
+    float sumf[4];
+    int sumi;
+
+    const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) vx + (x * nb);
+
+        for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
+        for (int l = 0; l < nb; l++) {
+            for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sumi = 0;
+                    for (int i = 0; i < blocklen; ++i) {
+                        const int v0 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F];
+                        const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4];
+                        sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2]));
+                    }
+                    sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
+                }
+            }
+        }
+        for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
+    }
+}
+
+void ggml_gemv_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 8;
+    const int blocklen = 8;
+
+    assert(nr == 1);
+    assert(n % qk == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    UNUSED(bs);
+    UNUSED(nr);
+
+    float sumf[8];
+    int sumi;
+
+    const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_iq4_nlx8 * b_ptr = (const block_iq4_nlx8 *) vx + (x * nb);
+
+        for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
+        for (int l = 0; l < nb; l++) {
+            for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sumi = 0;
                     for (int i = 0; i < blocklen; ++i) {
                         const int v0 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F];
                         const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4];
                         sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2]));
                     }
-                    sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
+                    sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
+                }
+            }
+        }
+        for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
+    }
+}
+
+void ggml_gemv_mxfp4_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 4;
+    const int blocklen = 4;
+
+    assert(nr == 1);
+    assert(n % qk == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    UNUSED(bs);
+    UNUSED(nr);
+
+    float sumf[4];
+    int sumi;
+
+    const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_mxfp4x4 * b_ptr = (const block_mxfp4x4 *) vx + (x * nb);
+
+        for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
+        for (int l = 0; l < nb; l++) {
+            for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sumi = 0;
+                    for (int i = 0; i < blocklen; ++i) {
+                        const int v0 = kvalues_mxfp4[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F];
+                        const int v1 = kvalues_mxfp4[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4];
+                        sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2]));
+                    }
+                    sumf[j] += sumi * GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[l].e[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
+                }
+            }
+        }
+        for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
+    }
+}
+
+void ggml_gemv_mxfp4_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 8;
+    const int blocklen = 8;
+
+    assert(nr == 1);
+    assert(n % qk == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    UNUSED(bs);
+    UNUSED(nr);
+
+    float sumf[8];
+    int sumi;
+
+    const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_mxfp4x8 * b_ptr = (const block_mxfp4x8 *) vx + (x * nb);
+
+        for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
+        for (int l = 0; l < nb; l++) {
+            for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sumi = 0;
+                    for (int i = 0; i < blocklen; ++i) {
+                        const int v0 = kvalues_mxfp4[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F];
+                        const int v1 = kvalues_mxfp4[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4];
+                        sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2]));
+                    }
+                    sumf[j] += sumi * GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[l].e[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
+                }
+            }
+        }
+        for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
+    }
+}
+
+void ggml_gemv_q8_0_4x4_q8_0_generic(int                        n,
+                                     float * GGML_RESTRICT      s,
+                                     size_t                     bs,
+                                     const void * GGML_RESTRICT vx,
+                                     const void * GGML_RESTRICT vy,
+                                     int                        nr,
+                                     int                        nc) {
+    const int qk                = QK8_0;
+    const int nb                = n / qk;
+    const int ncols_interleaved = 4;
+    const int blocklen          = 4;
+
+    assert(nr == 1);
+    assert(n % qk == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    UNUSED(bs);
+    UNUSED(nr);
+
+    float sumf[4];
+    int   sumi;
+
+    const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_q8_0x4 * b_ptr = (const block_q8_0x4 *) vx + (x * nb);
+
+        for (int j = 0; j < ncols_interleaved; j++) {
+            sumf[j] = 0.0;
+        }
+        for (int l = 0; l < nb; l++) {
+            for (int k = 0; k < (qk / blocklen); k++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sumi = 0;
+                    for (int i = 0; i < blocklen; ++i) {
+                        const int v0 = b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i];
+                        sumi += v0 * a_ptr[l].qs[k * blocklen + i];
+                    }
+                    sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
+                }
+            }
+        }
+        for (int j = 0; j < ncols_interleaved; j++) {
+            s[x * ncols_interleaved + j] = sumf[j];
+        }
+    }
+}
+
+void ggml_gemv_q8_0_4x8_q8_0_generic(int                        n,
+                                     float * GGML_RESTRICT      s,
+                                     size_t                     bs,
+                                     const void * GGML_RESTRICT vx,
+                                     const void * GGML_RESTRICT vy,
+                                     int                        nr,
+                                     int                        nc) {
+    const int qk                = QK8_0;
+    const int nb                = n / qk;
+    const int ncols_interleaved = 4;
+    const int blocklen          = 8;
+
+    assert(nr == 1);
+    assert(n % qk == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    UNUSED(bs);
+    UNUSED(nr);
+
+    float sumf[4];
+    int   sumi;
+
+    const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_q8_0x4 * b_ptr = (const block_q8_0x4 *) vx + (x * nb);
+
+        for (int j = 0; j < ncols_interleaved; j++) {
+            sumf[j] = 0.0;
+        }
+        for (int l = 0; l < nb; l++) {
+            for (int k = 0; k < (qk / blocklen); k++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sumi = 0;
+                    for (int i = 0; i < blocklen; ++i) {
+                        const int v0 = b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i];
+                        sumi += v0 * a_ptr[l].qs[k * blocklen + i];
+                    }
+                    sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
+                }
+            }
+        }
+        for (int j = 0; j < ncols_interleaved; j++) {
+            s[x * ncols_interleaved + j] = sumf[j];
+        }
+    }
+}
+
+// Only enable these for RISC-V.
+#if defined __riscv_zvfh
+void ggml_gemv_q4_0_16x1_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 16;
+    const int blocklen = 1;
+
+    assert (n % qk == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(s);
+    UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
+    UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+    float sumf[16];
+    int sumi;
+
+    const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_q4_0x16 * b_ptr = (const block_q4_0x16 *) vx + (x * nb);
+
+        for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
+        for (int l = 0; l < nb; l++) {
+            for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sumi = 0;
+                    for (int i = 0; i < blocklen; ++i) {
+                        const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
+                        const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                        sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
+                    }
+                    sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
+                }
+            }
+        }
+        for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
+    }
+}
+
+void ggml_gemv_q4_K_16x1_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    const int qk = QK_K;
+    const int nb = n / qk;
+    const int ncols_interleaved = 16;
+    const int blocklen = 1;
+    assert (n % qk == 0);
+    assert (nc % ncols_interleaved == 0);
+    UNUSED(s);
+    UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
+    UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+    float sumf[16];
+    float sum_minf[16];
+    uint8_t scales[128];
+    uint8_t mins[128];
+    int sumi1;
+    int sumi2;
+    int sumi;
+    const block_q8_K * a_ptr = (const block_q8_K *) vy;
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_q4_Kx16 * b_ptr = (const block_q4_Kx16 *) vx + (x * nb);
+        for (int j = 0; j < ncols_interleaved; j++) {
+            sumf[j] = 0.0f;
+            sum_minf[j] = 0.0f;
+        }
+        for (int l = 0; l < nb; l++) {
+            for (int i = 0; i < 128; i++) {
+                scales[i] = b_ptr[l].scales[i] & 0x0F;
+                mins[i] = b_ptr[l].scales[i] >> 4;
+            }
+            for (int i = 0; i < 64; i++) {
+                scales[i] |= (b_ptr[l].scales[128 + i] & 0x03) << 4;
+                mins[i] |= (b_ptr[l].scales[128 + i] & 0x0C) << 2;
+                scales[i + 64] |= (b_ptr[l].scales[128 + i] & 0x30);
+                mins[i + 64] |= (b_ptr[l].scales[128 + i] & 0xC0) >> 2;
+            }
+            for (int sb = 0; sb < 8; sb++) {
+                uint8_t *min = &mins[sb * 16];
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sum_minf[j] += min[j] * (a_ptr[l].bsums[sb * 2] + a_ptr[l].bsums[sb * 2 + 1]) * GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d;
+                }
+            }
+            for (int sb = 0; sb < 8; sb += 2) {
+                uint8_t *scales_0 = &scales[sb * 16];
+                uint8_t *scales_1 = &scales[(sb + 1) * 16];
+                for (int i = 0; i < QK4_0; i++) {
+                    for (int j = 0; j < ncols_interleaved; j++) {
+                        sumi1 = 0;
+                        sumi2 = 0;
+                        sumi = 0;
+                        const int v0 = (int8_t) (b_ptr[l].qs[sb * 256 + i * 16 + j] & 0xF);
+                        const int v1 = (int8_t) (b_ptr[l].qs[sb * 256 + i * 16 + j] >> 4);
+                        sumi1 = (v0 * a_ptr[l].qs[sb * 32 + i]);
+                        sumi2 = (v1 * a_ptr[l].qs[sb * 32 + 32 + i]);
+                        sumi1 = sumi1 * scales_0[j];
+                        sumi2 = sumi2 * scales_1[j];
+                        sumi += sumi1 + sumi2;
+                        sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d;
+                    }
+                }
+            }
+        }
+        for (int j = 0; j < ncols_interleaved; j++) {
+            s[x * ncols_interleaved + j] = sumf[j] - sum_minf[j];
+        }
+    }
+}
+
+void ggml_gemv_iq4_nl_16x1_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 16;
+    const int blocklen = 1;
+
+    assert(nr == 1);
+    assert(n % qk == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    UNUSED(bs);
+    UNUSED(nr);
+
+    float sumf[16];
+    int sumi;
+
+    const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_iq4_nlx16 * b_ptr = (const block_iq4_nlx16 *) vx + (x * nb);
+
+        for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
+        for (int l = 0; l < nb; l++) {
+            for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sumi = 0;
+                    for (int i = 0; i < blocklen; ++i) {
+                        const int v0 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F];
+                        const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4];
+                        sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2]));
+                    }
+                    sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
+                }
+            }
+        }
+        for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
+    }
+}
+
+void ggml_gemv_q8_0_16x1_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    const int qk                = QK8_0;
+    const int nb                = n / qk;
+    const int ncols_interleaved = 16;
+    const int blocklen          = 1;
+
+    assert(nr == 1);
+    assert(n % qk == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    UNUSED(bs);
+    UNUSED(nr);
+
+    float sumf[16];
+    int   sumi;
+
+    const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_q8_0x16 * b_ptr = (const block_q8_0x16 *) vx + (x * nb);
+
+        for (int j = 0; j < ncols_interleaved; j++) {
+            sumf[j] = 0.0;
+        }
+        for (int l = 0; l < nb; l++) {
+            for (int k = 0; k < (qk / blocklen); k++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sumi = 0;
+                    for (int i = 0; i < blocklen; ++i) {
+                        const int v0 = b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i];
+                        sumi += v0 * a_ptr[l].qs[k * blocklen + i];
+                    }
+                    sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
+                }
+            }
+        }
+        for (int j = 0; j < ncols_interleaved; j++) {
+            s[x * ncols_interleaved + j] = sumf[j];
+        }
+    }
+}
+
+void ggml_gemv_q2_K_16x1_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    assert(n % QK_K == 0);
+    assert(nr == 1);
+    assert(nc % 16 == 0);
+
+    UNUSED(bs);
+    UNUSED(nr);
+
+    const int nb = n / QK_K;
+    const block_q2_Kx16 * x = (const block_q2_Kx16 *)vx;
+    const block_q8_K    * y = (const block_q8_K *)vy;
+
+    // Layout: Even-Low(0,2,4,6), Odd-Low(1,3,5,7), Even-High(8...), Odd-High(9...)
+    const int sb_perm[16] = {
+        0, 4, 1, 5, 2, 6, 3, 7,  // 0-7
+        8, 12, 9, 13, 10, 14, 11, 15 // 8-15
+    };
+
+    for (int col_tile = 0; col_tile < nc; col_tile += 16) {
+        const block_q2_Kx16 * x_ptr = x + (col_tile / 16) * nb;
+        const block_q8_K    * y_ptr = y;
+
+        float sumf[16] = {0};
+
+        // Loop over K-blocks
+        for (int k_block = 0; k_block < nb; ++k_block) {
+            int32_t isum[16]  = {0};
+            int32_t summs[16] = {0};
+
+            const uint8_t * qs_rhs = x_ptr[k_block].qs;
+            const uint8_t * sc_rhs = x_ptr[k_block].scales;
+            const int8_t  * qs_lhs = y_ptr[k_block].qs;
+            const int16_t * bs_lhs = y_ptr[k_block].bsums;
+
+            // Iterate over sub-blocks 0..15
+            for (int sb = 0; sb < 16; ++sb) {
+                // Correction Term
+                int16_t bsum = bs_lhs[sb];
+                int scale_offset = sb_perm[sb] * 16;
+
+                for (int col = 0; col < 16; ++col) {
+                    uint8_t sc_val = sc_rhs[scale_offset + col];
+                    summs[col] += bsum * (sc_val >> 4); // Min is high 4 bits
+                }
+
+                // Main Dot Product
+                // Calculate base offsets for Q2 unpacking based on SB
+                int byte_base;
+                if (sb < 8) byte_base = (sb % 2 == 0) ? 0 : 16;
+                else        byte_base = (sb % 2 == 0) ? 32 : 48;
+
+                int shift = ((sb / 2) % 4) * 2;
+
+                for (int col = 0; col < 16; ++col) {
+                    uint8_t sc_val = sc_rhs[scale_offset + col];
+                    int32_t d_sb = sc_val & 0xF; // Scale is low 4 bits
+
+                    // Process 16 elements (l=0..15)
+                    for (int l = 0; l < 16; ++l) {
+                        // Q2: Interleaved by column. Byte `l` contains 4 k-values.
+                        int qs_idx = (byte_base + l) * 16 + col;
+                        uint8_t q2_val = (qs_rhs[qs_idx] >> shift) & 3;
+
+                        // Q8: Linear access
+                        int k = sb * 16 + l;
+                        int8_t q8_val = qs_lhs[k];
+
+                        isum[col] += q8_val * q2_val * d_sb;
+                    }
+                }
+            }
+
+            // Finalize K-Block
+            for (int col = 0; col < 16; ++col) {
+                float d_lhs = y_ptr[k_block].d;
+                float d_rhs = GGML_FP16_TO_FP32(x_ptr[k_block].d[col]);
+                float dm_rhs = GGML_FP16_TO_FP32(x_ptr[k_block].dmin[col]);
+
+                float d_all = d_lhs * d_rhs;
+                float d_min = d_lhs * dm_rhs;
+
+                sumf[col] += (isum[col] * d_all) - (summs[col] * d_min);
+            }
+        }
+
+        for (int col = 0; col < 16; ++col) {
+            s[col_tile + col] = sumf[col];
+        }
+    }
+}
+#endif
+
+void ggml_gemm_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 4;
+    const int blocklen = 4;
+
+    assert (n % qk == 0);
+    assert (nr % 4 == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(s);
+    UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
+    UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+    {
+        float sumf[4][4];
+        int sumi;
+
+        for (int y = 0; y < nr / 4; y++) {
+            const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
+            for (int x = 0; x < nc / ncols_interleaved; x++) {
+                const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb);
+                for (int m = 0; m < 4; m++) {
+                    for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
+                }
+                for (int l = 0; l < nb; l++) {
+                    for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                        for (int m = 0; m < 4; m++) {
+                            for (int j = 0; j < ncols_interleaved; j++) {
+                                sumi = 0;
+                                for (int i = 0; i < blocklen; ++i) {
+                                    const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
+                                    const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                                    sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
+                                            (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
+                                }
+                                sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
+                            }
+                        }
+                    }
+                }
+                for (int m = 0; m < 4; m++) {
+                    for (int j = 0; j < ncols_interleaved; j++)
+                        s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
+                }
+            }
+        }
+    }
+}
+
+void ggml_gemm_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 4;
+    const int blocklen = 8;
+
+    assert (n % qk == 0);
+    assert (nr % 4 == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(s);
+    UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
+    UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+    float sumf[4][4];
+    int sumi;
+
+    for (int y = 0; y < nr / 4; y++) {
+        const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb);
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
+            }
+            for (int l = 0; l < nb; l++) {
+                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                    for (int m = 0; m < 4; m++) {
+                        for (int j = 0; j < ncols_interleaved; j++) {
+                            sumi = 0;
+                            for (int i = 0; i < blocklen; ++i) {
+                                const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
+                                const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                                sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
+                                        (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
+                            }
+                            sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
+                        }
+                    }
+                }
+            }
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++)
+                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
+            }
+        }
+    }
+}
+
+void ggml_gemm_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 8;
+    const int blocklen = 8;
+
+    assert (n % qk == 0);
+    assert (nr % 4 == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(s);
+    UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
+    UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+    float sumf[4][8];
+    int sumi;
+
+    for (int y = 0; y < nr / 4; y++) {
+        const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
+            }
+            for (int l = 0; l < nb; l++) {
+                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                    for (int m = 0; m < 4; m++) {
+                        for (int j = 0; j < ncols_interleaved; j++) {
+                            sumi = 0;
+                            for (int i = 0; i < blocklen; ++i) {
+                                const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
+                                const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                                sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
+                                         (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
+                            }
+                            sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
+                        }
+                    }
+                }
+            }
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++)
+                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
+            }
+        }
+    }
+}
+
+void ggml_gemm_q4_K_8x4_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    const int qk = QK_K;
+    const int nb = n / qk;
+    const int ncols_interleaved = 8;
+    const int blocklen = 4;
+    static const uint32_t kmask1 = 0x3f3f3f3f;
+    static const uint32_t kmask2 = 0x0f0f0f0f;
+    static const uint32_t kmask3 = 0x03030303;
+
+    assert (n % qk == 0);
+    assert (nr % 4 == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+    float sumf[4][8];
+    float sum_minf[4][8];
+    uint32_t utmp[32];
+    int sumi1;
+    int sumi2;
+    int sumi;
+
+    for (int y = 0; y < nr / 4; y++) {
+        const block_q8_Kx4 * a_ptr = (const block_q8_Kx4 *) vy + (y * nb);
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q4_Kx8 * b_ptr = (const block_q4_Kx8 *) vx + (x * nb);
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sumf[m][j] = 0.0;
+                    sum_minf[m][j] = 0.0;
+                }
+            }
+            for (int l = 0; l < nb; l++) {
+                for (int sb = 0; sb < 8; sb++) {
+                    memcpy(utmp + sb * 4, b_ptr[l].scales + sb * 12, 12);
+                    utmp[sb * 4 + 3] = ((utmp[sb * 4 + 2] >> 4) & kmask2) | (((utmp[sb * 4 + 1] >> 6) & kmask3) << 4);
+                    const uint32_t uaux_0 = utmp[sb * 4 + 1] & kmask1;
+                    utmp[sb * 4 + 1] = (utmp[sb * 4 + 2] & kmask2) | (((utmp[sb * 4 + 0] >> 6) & kmask3) << 4);
+                    utmp[sb * 4 + 2] = uaux_0;
+                    utmp[sb * 4 + 0] &= kmask1;
+                }
+                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                    uint8_t * scales_0 = (uint8_t *) utmp + (k / 8) * 32;
+                    uint8_t * scales_1 = (uint8_t *) utmp + (k / 8) * 32 + 16;
+                    for (int m = 0; m < 4; m++) {
+                        for (int j = 0; j < ncols_interleaved; j++) {
+                            sumi1 = 0;
+                            sumi2 = 0;
+                            sumi = 0;
+                            for (int i = 0; i < blocklen; ++i) {
+                                const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF);
+                                const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4);
+                                sumi1 = (v0 * a_ptr[l].qs[(k / 8) * 256 + (k % 8) * 4 * blocklen + m * blocklen + i]);
+                                sumi2 = (v1 * a_ptr[l].qs[(k / 8) * 256 + (k % 8) * 4 * blocklen + m * blocklen + i + 128]);
+                                sumi1 = sumi1 * scales_0[j];
+                                sumi2 = sumi2 * scales_1[j];
+                                sumi += sumi1 + sumi2;
+                            }
+                            sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d[m];
+                        }
+                    }
+                }
+                for (int sb = 0; sb < 8; sb++) {
+                    uint8_t * mins = (uint8_t *) utmp + 8 + sb * 16;
+                    for(int m = 0; m < 4; m++) {
+                        const int16_t * bsums = a_ptr[l].bsums + (sb * 8) + (m * 4) - ((sb % 2) * 6);
+                        for(int j = 0; j < ncols_interleaved; j++) {
+                            sum_minf[m][j] += mins[j] * (bsums[0] + bsums[1]) * GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d[m];
+                        }
+                    }
+                }
+            }
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j] - sum_minf[m][j];
+                }
+            }
+        }
+    }
+}
+
+void ggml_gemm_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    const int qk = QK_K;
+    const int nb = n / qk;
+    const int ncols_interleaved = 8;
+    const int blocklen = 8;
+    static const uint32_t kmask1 = 0x3f3f3f3f;
+    static const uint32_t kmask2 = 0x0f0f0f0f;
+    static const uint32_t kmask3 = 0x03030303;
+
+    assert (n % qk == 0);
+    assert (nr % 4 == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(bs);
+
+    float sumf[4][8];
+    float sum_minf[4][8];
+    uint32_t utmp[32];
+    int sumi1;
+    int sumi2;
+    int sumi;
+
+    for (int y = 0; y < nr / 4; y++) {
+        const block_q8_Kx4 * a_ptr = (const block_q8_Kx4 *) vy + (y * nb);
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q4_Kx8 * b_ptr = (const block_q4_Kx8 *) vx + (x * nb);
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sumf[m][j] = 0.0;
+                    sum_minf[m][j] = 0.0;
+                }
+            }
+            for (int l = 0; l < nb; l++) {
+                for (int sb = 0; sb < 8; sb++) {
+                    memcpy(utmp + sb * 4, b_ptr[l].scales + sb * 12, 12);
+                    utmp[sb * 4 + 3] = ((utmp[sb * 4 + 2] >> 4) & kmask2) | (((utmp[sb * 4 + 1] >> 6) & kmask3) << 4);
+                    const uint32_t uaux_0 = utmp[sb * 4 + 1] & kmask1;
+                    utmp[sb * 4 + 1] = (utmp[sb * 4 + 2] & kmask2) | (((utmp[sb * 4 + 0] >> 6) & kmask3) << 4);
+                    utmp[sb * 4 + 2] = uaux_0;
+                    utmp[sb * 4 + 0] &= kmask1;
+                }
+                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                    uint8_t *scales_0 = (uint8_t*) utmp + (k / 4) * 32;
+                    uint8_t *scales_1 = (uint8_t*) utmp + (k / 4) * 32 + 16;
+                    for (int m = 0; m < 4; m++) {
+                        for (int j = 0; j < ncols_interleaved; j++) {
+                            sumi1 = 0;
+                            sumi2 = 0;
+                            sumi = 0;
+                            for (int i = 0; i < blocklen; ++i) {
+                                const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF);
+                                const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4);
+                                sumi1 = (v0 * a_ptr[l].qs[(k >> 2) * 256 + (k % 4) * 4 * blocklen + m * blocklen + i]);
+                                sumi2 = (v1 * a_ptr[l].qs[(k >> 2) * 256 + (k % 4) * 4 * blocklen + m * blocklen + i + 128]);
+                                sumi1 = sumi1 * scales_0[j];
+                                sumi2 = sumi2 * scales_1[j];
+                                sumi += sumi1 + sumi2;
+                            }
+                            sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d[m];
+                        }
+                    }
+                }
+                for (int sb = 0; sb < 8; sb++) {
+                    uint8_t *mins = (uint8_t*) utmp + 8 + sb * 16;
+                    for(int m = 0; m < 4; m++) {
+                        const int16_t *bsums = a_ptr[l].bsums + (sb * 8) + (m * 4) - ((sb % 2) * 6);
+                        for(int j = 0; j < ncols_interleaved; j++) {
+                            sum_minf[m][j] += mins[j] * (bsums[0] + bsums[1]) * GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d[m];
+                        }
+                    }
+                }
+            }
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j] - sum_minf[m][j];
                 }
             }
         }
-        for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
     }
 }
 
-void ggml_gemv_q8_0_4x4_q8_0_generic(int                        n,
-                                     float * GGML_RESTRICT      s,
-                                     size_t                     bs,
-                                     const void * GGML_RESTRICT vx,
-                                     const void * GGML_RESTRICT vy,
-                                     int                        nr,
-                                     int                        nc) {
-    const int qk                = QK8_0;
-    const int nb                = n / qk;
-    const int ncols_interleaved = 4;
-    const int blocklen          = 4;
+void ggml_gemm_q2_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    const int qk = QK_K;
+    const int nb = n / qk;
+    const int ncols_interleaved = 8;
+    const int blocklen = 8;
 
-    assert(nr == 1);
-    assert(n % qk == 0);
-    assert(nc % ncols_interleaved == 0);
+    assert (n % qk == 0);
+    assert (nr % 4 == 0);
+    assert (nc % ncols_interleaved == 0);
 
+    UNUSED(s);
     UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
     UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
 
-    float sumf[4];
-    int   sumi;
-
-    const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
-    for (int x = 0; x < nc / ncols_interleaved; x++) {
-        const block_q8_0x4 * b_ptr = (const block_q8_0x4 *) vx + (x * nb);
+    float sumf[4][8];
+    float sum_minf[4][8];
+    int sumi1, sumi2, sumi3, sumi4;
+    int sumi;
 
-        for (int j = 0; j < ncols_interleaved; j++) {
-            sumf[j] = 0.0;
-        }
-        for (int l = 0; l < nb; l++) {
-            for (int k = 0; k < (qk / blocklen); k++) {
+    for (int y = 0; y < nr / 4; y++) {
+        const block_q8_Kx4 * a_ptr = (const block_q8_Kx4 *) vy + (y * nb);
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q2_Kx8 * b_ptr = (const block_q2_Kx8 *) vx + (x * nb);
+            for (int m = 0; m < 4; m++) {
                 for (int j = 0; j < ncols_interleaved; j++) {
-                    sumi = 0;
-                    for (int i = 0; i < blocklen; ++i) {
-                        const int v0 = b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i];
-                        sumi += v0 * a_ptr[l].qs[k * blocklen + i];
+                    sumf[m][j] = 0.0;
+                    sum_minf[m][j] = 0.0;
+                }
+            }
+            for (int l = 0; l < nb; l++) {
+                for (int k = 0; k < (qk / (4 * blocklen)); k++) {
+
+                    const uint8_t *scales_0 = b_ptr[l].scales + (k / 4) * 64 ;
+                    const uint8_t *scales_1 = b_ptr[l].scales + (k / 4) * 64 + 16;
+                    const uint8_t *scales_2 = b_ptr[l].scales + (k / 4) * 64 + 32;
+                    const uint8_t *scales_3 = b_ptr[l].scales + (k / 4) * 64 + 48;
+                    for (int m = 0; m < 4; m++) {
+                        for (int j = 0; j < ncols_interleaved; j++) {
+                            sumi1 = 0;
+                            sumi2 = 0;
+                            sumi3 = 0;
+                            sumi4 = 0;
+                            sumi = 0;
+                            int offset = ((k / 2) % 2) + j * 2;
+                            for (int i = 0; i < blocklen; ++i){
+                                const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 3);
+                                const int v1 = (int8_t) ((b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 2 ) & 3);
+                                const int v2 = (int8_t) ((b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4 ) & 3);
+                                const int v3 = (int8_t) ((b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 6 ) & 3);
+                                sumi1 = (v0 * a_ptr[l].qs[(k >> 2) * 512 + (k % 4) * 4 * blocklen + m * blocklen + i]);
+                                sumi2 = (v1 * a_ptr[l].qs[(k >> 2) * 512  + (k % 4) * 4 * blocklen + m * blocklen + i + 128]);
+                                sumi3 = (v2 * a_ptr[l].qs[(k >> 2) * 512  + (k % 4) * 4 * blocklen + m * blocklen + i + 256]);
+                                sumi4 = (v3 * a_ptr[l].qs[(k >> 2) * 512  + (k % 4) * 4 * blocklen + m * blocklen + i + 384]);
+                                sumi1 = sumi1 * (scales_0[offset] & 0xF);
+                                sumi2 = sumi2 * (scales_1[offset] & 0xF);
+                                sumi3 = sumi3 * (scales_2[offset] & 0xF);
+                                sumi4 = sumi4 * (scales_3[offset] & 0xF);
+                                sumi += sumi1 + sumi2 + sumi3 + sumi4;
+                            }
+                            sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d[m];
+                        }
+                    }
+                }
+                for(int sb = 0; sb < 8; sb++) {
+                    const uint8_t *mins = b_ptr[l].scales + sb * 16;
+                    for(int m = 0; m < 4; m++) {
+                        const int16_t *bsums = a_ptr[l].bsums + (sb * 8) + (m * 4) - ((sb % 2) *  6);
+                        for(int j = 0; j < ncols_interleaved; j++) {
+                            int mins_prod = ((mins[j * 2] >> 4) * bsums[0] + (mins[(j * 2)+ 1] >> 4) * bsums[1]);
+                            sum_minf[m][j] += (mins_prod) * GGML_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d[m];
+                        }
                     }
-                    sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
                 }
             }
-        }
-        for (int j = 0; j < ncols_interleaved; j++) {
-            s[x * ncols_interleaved + j] = sumf[j];
+
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j] - sum_minf[m][j];
+                }
+            }
         }
     }
 }
 
-void ggml_gemv_q8_0_4x8_q8_0_generic(int                        n,
-                                     float * GGML_RESTRICT      s,
-                                     size_t                     bs,
-                                     const void * GGML_RESTRICT vx,
-                                     const void * GGML_RESTRICT vy,
-                                     int                        nr,
-                                     int                        nc) {
-    const int qk                = QK8_0;
-    const int nb                = n / qk;
-    const int ncols_interleaved = 4;
-    const int blocklen          = 8;
-
-    assert(nr == 1);
-    assert(n % qk == 0);
-    assert(nc % ncols_interleaved == 0);
-
-    UNUSED(bs);
-    UNUSED(nr);
+void ggml_gemm_q5_K_8x4_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    ggml_gemm_q5_K_NxM_q8_K_generic_impl<4, 8>(n, s, bs, vx, vy, nr, nc);
+}
 
-    float sumf[4];
-    int   sumi;
+void ggml_gemm_q5_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    ggml_gemm_q5_K_NxM_q8_K_generic_impl<8, 8>(n, s, bs, vx, vy, nr, nc);
+}
 
-    const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
-    for (int x = 0; x < nc / ncols_interleaved; x++) {
-        const block_q8_0x4 * b_ptr = (const block_q8_0x4 *) vx + (x * nb);
+void ggml_gemm_q6_K_8x4_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    ggml_gemm_q6_K_NxM_q8_K_generic_impl<4, 8>(n, s, bs, vx, vy, nr, nc);
+}
 
-        for (int j = 0; j < ncols_interleaved; j++) {
-            sumf[j] = 0.0;
-        }
-        for (int l = 0; l < nb; l++) {
-            for (int k = 0; k < (qk / blocklen); k++) {
-                for (int j = 0; j < ncols_interleaved; j++) {
-                    sumi = 0;
-                    for (int i = 0; i < blocklen; ++i) {
-                        const int v0 = b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i];
-                        sumi += v0 * a_ptr[l].qs[k * blocklen + i];
-                    }
-                    sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
-                }
-            }
-        }
-        for (int j = 0; j < ncols_interleaved; j++) {
-            s[x * ncols_interleaved + j] = sumf[j];
-        }
-    }
+void ggml_gemm_q6_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+   ggml_gemm_q6_K_NxM_q8_K_generic_impl<8, 8>(n, s, bs, vx, vy, nr, nc);
 }
 
-void ggml_gemm_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+void ggml_gemm_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
     const int qk = QK8_0;
     const int nb = n / qk;
     const int ncols_interleaved = 4;
@@ -813,7 +2116,7 @@ void ggml_gemm_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
         for (int y = 0; y < nr / 4; y++) {
             const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
             for (int x = 0; x < nc / ncols_interleaved; x++) {
-                const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb);
+                const block_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) vx + (x * nb);
                 for (int m = 0; m < 4; m++) {
                     for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
                 }
@@ -823,10 +2126,10 @@ void ggml_gemm_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
                             for (int j = 0; j < ncols_interleaved; j++) {
                                 sumi = 0;
                                 for (int i = 0; i < blocklen; ++i) {
-                                    const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
-                                    const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                                    const int v0 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F];
+                                    const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4];
                                     sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
-                                            (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
+                                            (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4]));
                                 }
                                 sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
                             }
@@ -842,33 +2145,23 @@ void ggml_gemm_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
     }
 }
 
-void ggml_gemm_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+void ggml_gemm_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
     const int qk = QK8_0;
     const int nb = n / qk;
-    const int ncols_interleaved = 4;
+    const int ncols_interleaved = 8;
     const int blocklen = 8;
 
-    assert (n % qk == 0);
-    assert (nr % 4 == 0);
-    assert (nc % ncols_interleaved == 0);
-
-    UNUSED(s);
-    UNUSED(bs);
-    UNUSED(vx);
-    UNUSED(vy);
-    UNUSED(nr);
-    UNUSED(nc);
-    UNUSED(nb);
-    UNUSED(ncols_interleaved);
-    UNUSED(blocklen);
+    assert(n % qk == 0);
+    assert(nr % 4 == 0);
+    assert(nc % ncols_interleaved == 0);
 
-    float sumf[4][4];
+    float sumf[4][8];
     int sumi;
 
     for (int y = 0; y < nr / 4; y++) {
         const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
         for (int x = 0; x < nc / ncols_interleaved; x++) {
-            const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb);
+            const block_iq4_nlx8 * b_ptr = (const block_iq4_nlx8 *) vx + (x * nb);
             for (int m = 0; m < 4; m++) {
                 for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
             }
@@ -878,10 +2171,10 @@ void ggml_gemm_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
                         for (int j = 0; j < ncols_interleaved; j++) {
                             sumi = 0;
                             for (int i = 0; i < blocklen; ++i) {
-                                const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
-                                const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                                const int v0 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F];
+                                const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4];
                                 sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
-                                        (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
+                                         (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4]));
                             }
                             sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
                         }
@@ -896,25 +2189,59 @@ void ggml_gemm_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
     }
 }
 
-void ggml_gemm_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+void ggml_gemm_mxfp4_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
     const int qk = QK8_0;
     const int nb = n / qk;
-    const int ncols_interleaved = 8;
-    const int blocklen = 8;
+    const int ncols_interleaved = 4;
+    const int blocklen = 4;
 
-    assert (n % qk == 0);
-    assert (nr % 4 == 0);
-    assert (nc % ncols_interleaved == 0);
+    assert(n % qk == 0);
+    assert(nr % 4 == 0);
+    assert(nc % ncols_interleaved == 0);
 
-    UNUSED(s);
-    UNUSED(bs);
-    UNUSED(vx);
-    UNUSED(vy);
-    UNUSED(nr);
-    UNUSED(nc);
-    UNUSED(nb);
-    UNUSED(ncols_interleaved);
-    UNUSED(blocklen);
+    float sumf[4][4];
+    int sumi;
+
+    for (int y = 0; y < nr / 4; y++) {
+        const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_mxfp4x4 * b_ptr = (const block_mxfp4x4 *) vx + (x * nb);
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
+            }
+            for (int l = 0; l < nb; l++) {
+                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                    for (int m = 0; m < 4; m++) {
+                        for (int j = 0; j < ncols_interleaved; j++) {
+                            sumi = 0;
+                            for (int i = 0; i < blocklen; ++i) {
+                                const int v0 = kvalues_mxfp4[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F];
+                                const int v1 = kvalues_mxfp4[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4];
+                                sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
+                                         (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4]));
+                            }
+                            sumf[m][j] += sumi * GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[l].e[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
+                        }
+                    }
+                }
+            }
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++)
+                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
+            }
+        }
+    }
+}
+
+void ggml_gemm_mxfp4_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 8;
+    const int blocklen = 8;
+
+    assert(n % qk == 0);
+    assert(nr % 4 == 0);
+    assert(nc % ncols_interleaved == 0);
 
     float sumf[4][8];
     int sumi;
@@ -922,7 +2249,7 @@ void ggml_gemm_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
     for (int y = 0; y < nr / 4; y++) {
         const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
         for (int x = 0; x < nc / ncols_interleaved; x++) {
-            const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
+            const block_mxfp4x8 * b_ptr = (const block_mxfp4x8 *) vx + (x * nb);
             for (int m = 0; m < 4; m++) {
                 for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
             }
@@ -932,12 +2259,12 @@ void ggml_gemm_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
                         for (int j = 0; j < ncols_interleaved; j++) {
                             sumi = 0;
                             for (int i = 0; i < blocklen; ++i) {
-                                const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
-                                const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                                const int v0 = kvalues_mxfp4[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F];
+                                const int v1 = kvalues_mxfp4[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4];
                                 sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
-                                         (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
+                                         (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4]));
                             }
-                            sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
+                            sumf[m][j] += sumi * GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[l].e[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
                         }
                     }
                 }
@@ -950,183 +2277,119 @@ void ggml_gemm_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
     }
 }
 
-void ggml_gemm_q4_K_8x4_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
-    const int qk = QK_K;
-    const int nb = n / qk;
-    const int ncols_interleaved = 8;
-    const int blocklen = 4;
-    static const uint32_t kmask1 = 0x3f3f3f3f;
-    static const uint32_t kmask2 = 0x0f0f0f0f;
-    static const uint32_t kmask3 = 0x03030303;
-
-    assert (n % qk == 0);
-    assert (nr % 4 == 0);
-    assert (nc % ncols_interleaved == 0);
+void ggml_gemm_q8_0_4x4_q8_0_generic(int                        n,
+                                     float * GGML_RESTRICT      s,
+                                     size_t                     bs,
+                                     const void * GGML_RESTRICT vx,
+                                     const void * GGML_RESTRICT vy,
+                                     int                        nr,
+                                     int                        nc) {
+    const int qk                = QK8_0;
+    const int nb                = n / qk;
+    const int ncols_interleaved = 4;
+    const int blocklen          = 4;
 
-    UNUSED(nb);
-    UNUSED(ncols_interleaved);
-    UNUSED(blocklen);
+    assert(n % qk == 0);
+    assert(nr % 4 == 0);
+    assert(nc % ncols_interleaved == 0);
 
-    float sumf[4][8];
-    float sum_minf[4][8];
-    uint32_t utmp[32];
-    int sumi1;
-    int sumi2;
-    int sumi;
+    float sumf[4][4];
+    int   sumi;
 
     for (int y = 0; y < nr / 4; y++) {
-        const block_q8_Kx4 * a_ptr = (const block_q8_Kx4 *) vy + (y * nb);
+        const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
         for (int x = 0; x < nc / ncols_interleaved; x++) {
-            const block_q4_Kx8 * b_ptr = (const block_q4_Kx8 *) vx + (x * nb);
+            const block_q8_0x4 * b_ptr = (const block_q8_0x4 *) vx + (x * nb);
             for (int m = 0; m < 4; m++) {
                 for (int j = 0; j < ncols_interleaved; j++) {
                     sumf[m][j] = 0.0;
-                    sum_minf[m][j] = 0.0;
                 }
             }
             for (int l = 0; l < nb; l++) {
-                for (int sb = 0; sb < 8; sb++) {
-                    memcpy(utmp + sb * 4, b_ptr[l].scales + sb * 12, 12);
-                    utmp[sb * 4 + 3] = ((utmp[sb * 4 + 2] >> 4) & kmask2) | (((utmp[sb * 4 + 1] >> 6) & kmask3) << 4);
-                    const uint32_t uaux_0 = utmp[sb * 4 + 1] & kmask1;
-                    utmp[sb * 4 + 1] = (utmp[sb * 4 + 2] & kmask2) | (((utmp[sb * 4 + 0] >> 6) & kmask3) << 4);
-                    utmp[sb * 4 + 2] = uaux_0;
-                    utmp[sb * 4 + 0] &= kmask1;
-                }
-                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
-                    uint8_t * scales_0 = (uint8_t *) utmp + (k / 8) * 32;
-                    uint8_t * scales_1 = (uint8_t *) utmp + (k / 8) * 32 + 16;
+                for (int k = 0; k < (qk / blocklen); k++) {
                     for (int m = 0; m < 4; m++) {
                         for (int j = 0; j < ncols_interleaved; j++) {
-                            sumi1 = 0;
-                            sumi2 = 0;
                             sumi = 0;
                             for (int i = 0; i < blocklen; ++i) {
-                                const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF);
-                                const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4);
-                                sumi1 = (v0 * a_ptr[l].qs[(k / 8) * 256 + (k % 8) * 4 * blocklen + m * blocklen + i]);
-                                sumi2 = (v1 * a_ptr[l].qs[(k / 8) * 256 + (k % 8) * 4 * blocklen + m * blocklen + i + 128]);
-                                sumi1 = sumi1 * scales_0[j];
-                                sumi2 = sumi2 * scales_1[j];
-                                sumi += sumi1 + sumi2;
+                                const int v0 = b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i];
+                                sumi += v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i];
                             }
-                            sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d[m];
-                        }
-                    }
-                }
-                for (int sb = 0; sb < 8; sb++) {
-                    uint8_t * mins = (uint8_t *) utmp + 8 + sb * 16;
-                    for(int m = 0; m < 4; m++) {
-                        const int16_t * bsums = a_ptr[l].bsums + (sb * 8) + (m * 4) - ((sb % 2) * 6);
-                        for(int j = 0; j < ncols_interleaved; j++) {
-                            sum_minf[m][j] += mins[j] * (bsums[0] + bsums[1]) * GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d[m];
+                            sumf[m][j] +=
+                                sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
                         }
                     }
                 }
             }
             for (int m = 0; m < 4; m++) {
                 for (int j = 0; j < ncols_interleaved; j++) {
-                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j] - sum_minf[m][j];
+                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
                 }
             }
         }
     }
 }
 
-void ggml_gemm_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
-    const int qk = QK_K;
-    const int nb = n / qk;
-    const int ncols_interleaved = 8;
-    const int blocklen = 8;
-    static const uint32_t kmask1 = 0x3f3f3f3f;
-    static const uint32_t kmask2 = 0x0f0f0f0f;
-    static const uint32_t kmask3 = 0x03030303;
 
-    assert (n % qk == 0);
-    assert (nr % 4 == 0);
-    assert (nc % ncols_interleaved == 0);
 
-    UNUSED(s);
-    UNUSED(bs);
-    UNUSED(vx);
-    UNUSED(vy);
-    UNUSED(nr);
-    UNUSED(nc);
-    UNUSED(nb);
-    UNUSED(ncols_interleaved);
-    UNUSED(blocklen);
+void ggml_gemm_q8_0_4x8_q8_0_generic(int                        n,
+                                     float * GGML_RESTRICT      s,
+                                     size_t                     bs,
+                                     const void * GGML_RESTRICT vx,
+                                     const void * GGML_RESTRICT vy,
+                                     int                        nr,
+                                     int                        nc) {
+    const int qk                = QK8_0;
+    const int nb                = n / qk;
+    const int ncols_interleaved = 4;
+    const int blocklen          = 8;
 
-    float sumf[4][8];
-    float sum_minf[4][8];
-    uint32_t utmp[32];
-    int sumi1;
-    int sumi2;
-    int sumi;
+    assert(n % qk == 0);
+    assert(nr % 4 == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    float sumf[4][4];
+    int   sumi;
 
     for (int y = 0; y < nr / 4; y++) {
-        const block_q8_Kx4 * a_ptr = (const block_q8_Kx4 *) vy + (y * nb);
+        const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
         for (int x = 0; x < nc / ncols_interleaved; x++) {
-            const block_q4_Kx8 * b_ptr = (const block_q4_Kx8 *) vx + (x * nb);
+            const block_q8_0x4 * b_ptr = (const block_q8_0x4 *) vx + (x * nb);
             for (int m = 0; m < 4; m++) {
                 for (int j = 0; j < ncols_interleaved; j++) {
                     sumf[m][j] = 0.0;
-                    sum_minf[m][j] = 0.0;
                 }
             }
             for (int l = 0; l < nb; l++) {
-                for (int sb = 0; sb < 8; sb++) {
-                    memcpy(utmp + sb * 4, b_ptr[l].scales + sb * 12, 12);
-                    utmp[sb * 4 + 3] = ((utmp[sb * 4 + 2] >> 4) & kmask2) | (((utmp[sb * 4 + 1] >> 6) & kmask3) << 4);
-                    const uint32_t uaux_0 = utmp[sb * 4 + 1] & kmask1;
-                    utmp[sb * 4 + 1] = (utmp[sb * 4 + 2] & kmask2) | (((utmp[sb * 4 + 0] >> 6) & kmask3) << 4);
-                    utmp[sb * 4 + 2] = uaux_0;
-                    utmp[sb * 4 + 0] &= kmask1;
-                }
-                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
-                    uint8_t *scales_0 = (uint8_t*) utmp + (k / 4) * 32;
-                    uint8_t *scales_1 = (uint8_t*) utmp + (k / 4) * 32 + 16;
+                for (int k = 0; k < (qk / blocklen); k++) {
                     for (int m = 0; m < 4; m++) {
                         for (int j = 0; j < ncols_interleaved; j++) {
-                            sumi1 = 0;
-                            sumi2 = 0;
                             sumi = 0;
                             for (int i = 0; i < blocklen; ++i) {
-                                const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF);
-                                const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4);
-                                sumi1 = (v0 * a_ptr[l].qs[(k >> 2) * 256 + (k % 4) * 4 * blocklen + m * blocklen + i]);
-                                sumi2 = (v1 * a_ptr[l].qs[(k >> 2) * 256 + (k % 4) * 4 * blocklen + m * blocklen + i + 128]);
-                                sumi1 = sumi1 * scales_0[j];
-                                sumi2 = sumi2 * scales_1[j];
-                                sumi += sumi1 + sumi2;
+                                const int v0 = b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i];
+                                sumi += v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i];
                             }
-                            sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d[m];
-                        }
-                    }
-                }
-                for (int sb = 0; sb < 8; sb++) {
-                    uint8_t *mins = (uint8_t*) utmp + 8 + sb * 16;
-                    for(int m = 0; m < 4; m++) {
-                        const int16_t *bsums = a_ptr[l].bsums + (sb * 8) + (m * 4) - ((sb % 2) * 6);
-                        for(int j = 0; j < ncols_interleaved; j++) {
-                            sum_minf[m][j] += mins[j] * (bsums[0] + bsums[1]) * GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d[m];
+                            sumf[m][j] +=
+                                sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
                         }
                     }
                 }
             }
             for (int m = 0; m < 4; m++) {
                 for (int j = 0; j < ncols_interleaved; j++) {
-                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j] - sum_minf[m][j];
+                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
                 }
             }
         }
     }
 }
 
-void ggml_gemm_q2_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
-    const int qk = QK_K;
+// Only enable these for RISC-V.
+#if defined __riscv_zvfh
+void ggml_gemm_q4_0_16x1_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    const int qk = QK8_0;
     const int nb = n / qk;
-    const int ncols_interleaved = 8;
-    const int blocklen = 8;
+    const int ncols_interleaved = 16;
+    const int blocklen = 1;
 
     assert (n % qk == 0);
     assert (nr % 4 == 0);
@@ -1142,82 +2405,45 @@ void ggml_gemm_q2_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
     UNUSED(ncols_interleaved);
     UNUSED(blocklen);
 
-    float sumf[4][8];
-    float sum_minf[4][8];
-    int sumi1, sumi2, sumi3, sumi4;
+    float sumf[4][16];
     int sumi;
 
     for (int y = 0; y < nr / 4; y++) {
-        const block_q8_Kx4 * a_ptr = (const block_q8_Kx4 *) vy + (y * nb);
+        const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
         for (int x = 0; x < nc / ncols_interleaved; x++) {
-            const block_q2_Kx8 * b_ptr = (const block_q2_Kx8 *) vx + (x * nb);
+            const block_q4_0x16 * b_ptr = (const block_q4_0x16 *) vx + (x * nb);
             for (int m = 0; m < 4; m++) {
-                for (int j = 0; j < ncols_interleaved; j++) {
-                    sumf[m][j] = 0.0;
-                    sum_minf[m][j] = 0.0;
-                }
+                for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
             }
             for (int l = 0; l < nb; l++) {
-                for (int k = 0; k < (qk / (4 * blocklen)); k++) {
-
-                    const uint8_t *scales_0 = b_ptr[l].scales + (k / 4) * 64 ;
-                    const uint8_t *scales_1 = b_ptr[l].scales + (k / 4) * 64 + 16;
-                    const uint8_t *scales_2 = b_ptr[l].scales + (k / 4) * 64 + 32;
-                    const uint8_t *scales_3 = b_ptr[l].scales + (k / 4) * 64 + 48;
+                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
                     for (int m = 0; m < 4; m++) {
                         for (int j = 0; j < ncols_interleaved; j++) {
-                            sumi1 = 0;
-                            sumi2 = 0;
-                            sumi3 = 0;
-                            sumi4 = 0;
                             sumi = 0;
-                            int offset = ((k / 2) % 2) + j * 2;
-                            for (int i = 0; i < blocklen; ++i){
-                                const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 3);
-                                const int v1 = (int8_t) ((b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 2 ) & 3);
-                                const int v2 = (int8_t) ((b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4 ) & 3);
-                                const int v3 = (int8_t) ((b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 6 ) & 3);
-                                sumi1 = (v0 * a_ptr[l].qs[(k >> 2) * 512 + (k % 4) * 4 * blocklen + m * blocklen + i]);
-                                sumi2 = (v1 * a_ptr[l].qs[(k >> 2) * 512  + (k % 4) * 4 * blocklen + m * blocklen + i + 128]);
-                                sumi3 = (v2 * a_ptr[l].qs[(k >> 2) * 512  + (k % 4) * 4 * blocklen + m * blocklen + i + 256]);
-                                sumi4 = (v3 * a_ptr[l].qs[(k >> 2) * 512  + (k % 4) * 4 * blocklen + m * blocklen + i + 384]);
-                                sumi1 = sumi1 * (scales_0[offset] & 0xF);
-                                sumi2 = sumi2 * (scales_1[offset] & 0xF);
-                                sumi3 = sumi3 * (scales_2[offset] & 0xF);
-                                sumi4 = sumi4 * (scales_3[offset] & 0xF);
-                                sumi += sumi1 + sumi2 + sumi3 + sumi4;
+                            for (int i = 0; i < blocklen; ++i) {
+                                const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
+                                const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                                sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
+                                         (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
                             }
-                            sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d[m];
-                        }
-                    }
-                }
-                for(int sb = 0; sb < 8; sb++) {
-                    const uint8_t *mins = b_ptr[l].scales + sb * 16;
-                    for(int m = 0; m < 4; m++) {
-                        const int16_t *bsums = a_ptr[l].bsums + (sb * 8) + (m * 4) - ((sb % 2) *  6);
-                        for(int j = 0; j < ncols_interleaved; j++) {
-                            int mins_prod = ((mins[j * 2] >> 4) * bsums[0] + (mins[(j * 2)+ 1] >> 4) * bsums[1]);
-                            sum_minf[m][j] += (mins_prod) * GGML_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d[m];
+                            sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
                         }
                     }
                 }
             }
-
             for (int m = 0; m < 4; m++) {
-                for (int j = 0; j < ncols_interleaved; j++) {
-                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j] - sum_minf[m][j];
-                }
+                for (int j = 0; j < ncols_interleaved; j++)
+                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
             }
         }
     }
 }
 
-
-void ggml_gemm_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
-    const int qk = QK8_0;
+void ggml_gemm_q4_K_16x1_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    const int qk = QK_K;
     const int nb = n / qk;
-    const int ncols_interleaved = 4;
-    const int blocklen = 4;
+    const int ncols_interleaved = 16;
+    const int blocklen = 1;
 
     assert (n % qk == 0);
     assert (nr % 4 == 0);
@@ -1233,59 +2459,97 @@ void ggml_gemm_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs
     UNUSED(ncols_interleaved);
     UNUSED(blocklen);
 
-    {
-        float sumf[4][4];
-        int sumi;
+    float sumf[4][16];
+    float sum_minf[4][16];
+    uint8_t scales[128];
+    uint8_t mins[128];
+    int sumi1;
+    int sumi2;
+    int sumi;
+
+    for (int y = 0; y < nr / 4; y++) {
+        const block_q8_Kx4 * a_ptr = (const block_q8_Kx4 *) vy + (y * nb);
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q4_Kx16 * b_ptr = (const block_q4_Kx16 *) vx + (x * nb);
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sumf[m][j] = 0.0;
+                    sum_minf[m][j] = 0.0;
+                }
+            }
+            for (int l = 0; l < nb; l++) {
+                for (int i = 0; i < 128; i++) {
+                    scales[i] = b_ptr[l].scales[i] & 0x0F;
+                    mins[i] = b_ptr[l].scales[i] >> 4;
+                }
+                for (int i = 0; i < 64; i++) {
+                    scales[i] |= (b_ptr[l].scales[128 + i] & 0x03) << 4;
+                    mins[i] |= (b_ptr[l].scales[128 + i] & 0x0C) << 2;
+                    scales[i + 64] |= (b_ptr[l].scales[128 + i] & 0x30);
+                    mins[i + 64] |= (b_ptr[l].scales[128 + i] & 0xC0) >> 2;
+                }
 
-        for (int y = 0; y < nr / 4; y++) {
-            const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
-            for (int x = 0; x < nc / ncols_interleaved; x++) {
-                const block_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) vx + (x * nb);
-                for (int m = 0; m < 4; m++) {
-                    for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
+                for (int sb = 0; sb < 8; sb++) {
+                    uint8_t *min = &mins[sb * 16];
+                    for(int m = 0; m < 4; m++) {
+                        const int16_t bsums = a_ptr[l].bsums[sb * 8 + m] + a_ptr[l].bsums[sb * 8 + m + 4];
+                        for(int j = 0; j < ncols_interleaved; j++) {
+                            sum_minf[m][j] += min[j] * bsums * GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d[m];
+                        }
+                    }
                 }
-                for (int l = 0; l < nb; l++) {
-                    for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+
+                for (int sb = 0; sb < 8; sb += 2) {
+                    uint8_t *scales_0 = &scales[sb * 16];
+                    uint8_t *scales_1 = &scales[(sb + 1) * 16];
+
+                    for (int i = 0; i < QK4_0; i++) {
                         for (int m = 0; m < 4; m++) {
                             for (int j = 0; j < ncols_interleaved; j++) {
+                                sumi1 = 0;
+                                sumi2 = 0;
                                 sumi = 0;
-                                for (int i = 0; i < blocklen; ++i) {
-                                    const int v0 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F];
-                                    const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4];
-                                    sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
-                                            (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4]));
-                                }
-                                sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
+
+                                const int v0 = (int8_t) (b_ptr[l].qs[sb * 256 + i * 16 + j] & 0xF);
+                                const int v1 = (int8_t) (b_ptr[l].qs[sb * 256 + i * 16 + j] >> 4);
+                                sumi1 = (v0 * a_ptr[l].qs[sb * 4 * 32 + i * 4 + m]);
+                                sumi2 = (v1 * a_ptr[l].qs[sb * 4 * 32 + 32 * 4 + i * 4 + m]);
+                                sumi1 = sumi1 * scales_0[j];
+                                sumi2 = sumi2 * scales_1[j];
+                                sumi += sumi1 + sumi2;
+
+                                sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d[m];
                             }
                         }
                     }
                 }
-                for (int m = 0; m < 4; m++) {
-                    for (int j = 0; j < ncols_interleaved; j++)
-                        s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
+            }
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j] - sum_minf[m][j];
                 }
             }
         }
     }
 }
 
-void ggml_gemm_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+void ggml_gemm_iq4_nl_16x1_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
     const int qk = QK8_0;
     const int nb = n / qk;
-    const int ncols_interleaved = 8;
-    const int blocklen = 8;
+    const int ncols_interleaved = 16;
+    const int blocklen = 1;
 
     assert(n % qk == 0);
     assert(nr % 4 == 0);
     assert(nc % ncols_interleaved == 0);
 
-    float sumf[4][8];
+    float sumf[4][16];
     int sumi;
 
     for (int y = 0; y < nr / 4; y++) {
         const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
         for (int x = 0; x < nc / ncols_interleaved; x++) {
-            const block_iq4_nlx8 * b_ptr = (const block_iq4_nlx8 *) vx + (x * nb);
+            const block_iq4_nlx16 * b_ptr = (const block_iq4_nlx16 *) vx + (x * nb);
             for (int m = 0; m < 4; m++) {
                 for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
             }
@@ -1298,7 +2562,7 @@ void ggml_gemm_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs
                                 const int v0 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F];
                                 const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4];
                                 sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
-                                         (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4]));
+                                         (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + (qk / 2) * 4]));
                             }
                             sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
                         }
@@ -1313,29 +2577,23 @@ void ggml_gemm_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs
     }
 }
 
-void ggml_gemm_q8_0_4x4_q8_0_generic(int                        n,
-                                     float * GGML_RESTRICT      s,
-                                     size_t                     bs,
-                                     const void * GGML_RESTRICT vx,
-                                     const void * GGML_RESTRICT vy,
-                                     int                        nr,
-                                     int                        nc) {
+void ggml_gemm_q8_0_16x1_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
     const int qk                = QK8_0;
     const int nb                = n / qk;
-    const int ncols_interleaved = 4;
-    const int blocklen          = 4;
+    const int ncols_interleaved = 16;
+    const int blocklen          = 1;
 
     assert(n % qk == 0);
     assert(nr % 4 == 0);
     assert(nc % ncols_interleaved == 0);
 
-    float sumf[4][4];
+    float sumf[4][16];
     int   sumi;
 
     for (int y = 0; y < nr / 4; y++) {
         const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
         for (int x = 0; x < nc / ncols_interleaved; x++) {
-            const block_q8_0x4 * b_ptr = (const block_q8_0x4 *) vx + (x * nb);
+            const block_q8_0x16 * b_ptr = (const block_q8_0x16 *) vx + (x * nb);
             for (int m = 0; m < 4; m++) {
                 for (int j = 0; j < ncols_interleaved; j++) {
                     sumf[m][j] = 0.0;
@@ -1365,57 +2623,102 @@ void ggml_gemm_q8_0_4x4_q8_0_generic(int                        n,
     }
 }
 
-void ggml_gemm_q8_0_4x8_q8_0_generic(int                        n,
-                                     float * GGML_RESTRICT      s,
-                                     size_t                     bs,
-                                     const void * GGML_RESTRICT vx,
-                                     const void * GGML_RESTRICT vy,
-                                     int                        nr,
-                                     int                        nc) {
-    const int qk                = QK8_0;
-    const int nb                = n / qk;
-    const int ncols_interleaved = 4;
-    const int blocklen          = 8;
 
-    assert(n % qk == 0);
+void ggml_gemm_q2_K_16x1_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    assert(n % QK_K == 0);
     assert(nr % 4 == 0);
-    assert(nc % ncols_interleaved == 0);
+    assert(nc % 16 == 0);
+    const int nb = n / QK_K;
+    const block_q2_Kx16 * x = (const block_q2_Kx16 *)vx;
+    const block_q8_Kx4  * y = (const block_q8_Kx4 *)vy;
+
+    const int sb_perm[16] = {
+        0, 4, 1, 5, 2, 6, 3, 7,
+        8, 12, 9, 13, 10, 14, 11, 15
+    };
 
-    float sumf[4][4];
-    int   sumi;
+    // Iterate Rows in tiles of 4
+    for (int row_tile = 0; row_tile < nr; row_tile += 4) {
+        // Iterate Columns in tiles of 16
+        for (int col_tile = 0; col_tile < nc; col_tile += 16) {
+
+            const block_q2_Kx16 * x_ptr = x + (col_tile / 16) * nb;
+            const block_q8_Kx4  * y_ptr = y + (row_tile / 4) * nb;
+
+            float sumf[4][16];
+            memset(sumf, 0, sizeof(sumf));
+
+            for (int k_block = 0; k_block < nb; ++k_block) {
+                int32_t isum[4][16];
+                int32_t summs[4][16];
+                memset(isum, 0, sizeof(isum));
+                memset(summs, 0, sizeof(summs));
+
+                const uint8_t * qs_rhs = x_ptr[k_block].qs;
+                const uint8_t * sc_rhs = x_ptr[k_block].scales;
+                const int8_t  * qs_lhs = y_ptr[k_block].qs;
+                const int16_t * bs_lhs = y_ptr[k_block].bsums;
+
+                for (int sb = 0; sb < 16; ++sb) {
+                    int scale_offset = sb_perm[sb] * 16;
+
+                    int byte_base;
+                    if (sb < 8) byte_base = (sb % 2 == 0) ? 0 : 16;
+                    else        byte_base = (sb % 2 == 0) ? 32 : 48;
+                    int shift = ((sb / 2) % 4) * 2;
+
+                    for (int col = 0; col < 16; ++col) {
+                        uint8_t sc_val = sc_rhs[scale_offset + col];
+                        int32_t d_sb = sc_val & 0xF;
+                        int32_t m_sb = sc_val >> 4;
+
+                        // Correction Term
+                        for (int r = 0; r < 4; ++r) {
+                            int bsum_idx = (sb / 4) * 16 + r * 4 + (sb % 4);
+                            summs[r][col] += bs_lhs[bsum_idx] * m_sb;
+                        }
 
-    for (int y = 0; y < nr / 4; y++) {
-        const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
-        for (int x = 0; x < nc / ncols_interleaved; x++) {
-            const block_q8_0x4 * b_ptr = (const block_q8_0x4 *) vx + (x * nb);
-            for (int m = 0; m < 4; m++) {
-                for (int j = 0; j < ncols_interleaved; j++) {
-                    sumf[m][j] = 0.0;
-                }
-            }
-            for (int l = 0; l < nb; l++) {
-                for (int k = 0; k < (qk / blocklen); k++) {
-                    for (int m = 0; m < 4; m++) {
-                        for (int j = 0; j < ncols_interleaved; j++) {
-                            sumi = 0;
-                            for (int i = 0; i < blocklen; ++i) {
-                                const int v0 = b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i];
-                                sumi += v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i];
+                        // Main Dot Product
+                        for (int l = 0; l < 16; ++l) {
+                            int qs_idx = (byte_base + l) * 16 + col;
+                            uint8_t q2_val = (qs_rhs[qs_idx] >> shift) & 3;
+
+                            // Calculate Q8 index for this specific k and row
+                            int k = sb * 16 + l;
+                            int q8_idx = (k / 4) * 16 + (k % 4);
+
+                            for (int r = 0; r < 4; ++r) {
+                                // Add r*4 to jump to the correct row within the 4x4 chunk
+                                int8_t q8_val = qs_lhs[q8_idx + r * 4];
+                                isum[r][col] += q8_val * q2_val * d_sb;
                             }
-                            sumf[m][j] +=
-                                sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
                         }
                     }
                 }
+
+                // Finalize K-Block
+                for (int col = 0; col < 16; ++col) {
+                    float d_rhs = GGML_FP16_TO_FP32(x_ptr[k_block].d[col]);
+                    float dm_rhs = GGML_FP16_TO_FP32(x_ptr[k_block].dmin[col]);
+
+                    for (int r = 0; r < 4; ++r) {
+                        float d_lhs = y_ptr[k_block].d[r];
+                        float d_all = d_lhs * d_rhs;
+                        float d_min = d_lhs * dm_rhs;
+                        sumf[r][col] += (isum[r][col] * d_all) - (summs[r][col] * d_min);
+                    }
+                }
             }
-            for (int m = 0; m < 4; m++) {
-                for (int j = 0; j < ncols_interleaved; j++) {
-                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
+
+            for (int r = 0; r < 4; ++r) {
+                for (int col = 0; col < 16; ++col) {
+                    s[(row_tile + r) * bs + (col_tile + col)] = sumf[r][col];
                 }
             }
         }
     }
 }
+#endif
 
 } // extern "C"
 
@@ -1498,16 +2801,212 @@ static block_q4_0x8 make_block_q4_0x8(block_q4_0 * in, unsigned int blck_size_in
 
         uint64_t elems;
         memcpy(&elems, &in[src_id].qs[src_offset], sizeof(uint64_t));
-        elems ^= xor_mask;
+        elems ^= xor_mask;
+        memcpy(&out.qs[dst_offset], &elems, sizeof(uint64_t));
+    }
+
+    return out;
+}
+
+static block_q4_0x16 make_block_q4_0x16(block_q4_0 * in, unsigned int blck_size_interleave) {
+    block_q4_0x16 out;
+
+    for (int i = 0; i < 16; i++) {
+        out.d[i] = in[i].d;
+    }
+
+    const int end = QK4_0 * 8 / blck_size_interleave;
+
+    if (blck_size_interleave == 1) {
+        const uint8_t xor_mask = 0x88;
+        for (int i = 0; i < end; ++i) {
+            int src_id = i % 16;
+            int src_offset = i / 16;
+            int dst_offset = i;
+
+            out.qs[dst_offset] = in[src_id].qs[src_offset] ^ xor_mask;
+        }
+    } else {
+        GGML_ASSERT(false);
+    }
+
+    return out;
+}
+
+static block_q4_Kx8 make_block_q4_Kx8(block_q4_K * in, unsigned int blck_size_interleave) {
+    block_q4_Kx8 out;
+    //Delta(scale) and dmin values of the eight Q4_K structures are copied onto the output interleaved structure
+    for (int i = 0; i < 8; i++) {
+        out.d[i] = in[i].GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.d;
+    }
+
+    for (int i = 0; i < 8; i++) {
+        out.dmin[i] = in[i].GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.dmin;
+    }
+
+    const int end = QK_K * 4 / blck_size_interleave;
+
+    // Interleave Q4_K quants by taking 8 bytes at a time
+    for (int i = 0; i < end; ++i) {
+        int src_id = i % 8;
+        int src_offset = (i / 8) * blck_size_interleave;
+        int dst_offset = i * blck_size_interleave;
+
+        // buffer large enough for the max interleave block size (8 bytes)
+        uint64_t elems;
+        memcpy(&elems, &in[src_id].qs[src_offset], blck_size_interleave);
+        memcpy(&out.qs[dst_offset], &elems, blck_size_interleave);
+    }
+
+    // The below logic is designed so as to unpack and rearrange scales and mins values in Q4_K
+    // Currently the Q4_K structure has 8 scales and 8 mins packed in 12 bytes ( 6 bits for each value)
+    // The output Q4_Kx8 structure has 96 bytes
+    // Every 12 byte is packed such that it contains scales and mins for corresponding sub blocks from Q4_K structure
+    // For eg - First 12 bytes contains 8 scales and 8 mins - each of first sub block from different Q4_K structures
+    uint8_t s[8], m[8];
+
+    for (int i = 0; i < 4; i++) {
+        for (int j = 0; j < 8; j++) {
+            s[j] = in[j].scales[i] & 63;
+            m[j] = in[j].scales[i + 4] & 63;
+        }
+
+        out.scales[i * 12]      = (s[0] & 63) + ((s[4] & 48) << 2);
+        out.scales[i * 12 + 1]  = (s[1] & 63) + ((s[5] & 48) << 2);
+        out.scales[i * 12 + 2]  = (s[2] & 63) + ((s[6] & 48) << 2);
+        out.scales[i * 12 + 3]  = (s[3] & 63) + ((s[7] & 48) << 2);
+        out.scales[i * 12 + 4]  = (m[0] & 63) + ((m[4] & 48) << 2);
+        out.scales[i * 12 + 5]  = (m[1] & 63) + ((m[5] & 48) << 2);
+        out.scales[i * 12 + 6]  = (m[2] & 63) + ((m[6] & 48) << 2);
+        out.scales[i * 12 + 7]  = (m[3] & 63) + ((m[7] & 48) << 2);
+        out.scales[i * 12 + 8]  = (s[4] & 15) + ((m[4] & 15) << 4);
+        out.scales[i * 12 + 9]  = (s[5] & 15) + ((m[5] & 15) << 4);
+        out.scales[i * 12 + 10] = (s[6] & 15) + ((m[6] & 15) << 4);
+        out.scales[i * 12 + 11] = (s[7] & 15) + ((m[7] & 15) << 4);
+
+    }
+
+    for (int i = 0; i < 4; i++) {
+        for (int j = 0; j < 8; j++) {
+            s[j] = ((in[j].scales[i] & 192) >> 2) | (in[j].scales[i+8] & 15);
+            m[j] = ((in[j].scales[i + 4] & 192) >> 2) | ((in[j].scales[i+8] & 240) >> 4);
+        }
+
+        out.scales[i * 12 + 48] = (s[0] & 63) + ((s[4] & 48) << 2);
+        out.scales[i * 12 + 49] = (s[1] & 63) + ((s[5] & 48) << 2);
+        out.scales[i * 12 + 50] = (s[2] & 63) + ((s[6] & 48) << 2);
+        out.scales[i * 12 + 51] = (s[3] & 63) + ((s[7] & 48) << 2);
+        out.scales[i * 12 + 52] = (m[0] & 63) + ((m[4] & 48) << 2);
+        out.scales[i * 12 + 53] = (m[1] & 63) + ((m[5] & 48) << 2);
+        out.scales[i * 12 + 54] = (m[2] & 63) + ((m[6] & 48) << 2);
+        out.scales[i * 12 + 55] = (m[3] & 63) + ((m[7] & 48) << 2);
+        out.scales[i * 12 + 56] = (s[4] & 15) + ((m[4] & 15) << 4);
+        out.scales[i * 12 + 57] = (s[5] & 15) + ((m[5] & 15) << 4);
+        out.scales[i * 12 + 58] = (s[6] & 15) + ((m[6] & 15) << 4);
+        out.scales[i * 12 + 59] = (s[7] & 15) + ((m[7] & 15) << 4);
+
+    }
+
+    return out;
+}
+
+static block_q4_Kx16 make_block_q4_Kx16(block_q4_K * in, unsigned int blck_size_interleave) {
+    block_q4_Kx16 out;
+    //Delta(scale) and dmin values of the 16 Q4_K structures are copied onto the output interleaved structure
+    for (int i = 0; i < 16; i++) {
+        out.d[i] = in[i].GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.d;
+    }
+
+    for (int i = 0; i < 16; i++) {
+        out.dmin[i] = in[i].GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.dmin;
+    }
+
+    const int end = QK_K * 8 / blck_size_interleave;
+
+    if (blck_size_interleave == 1) {
+        for (int i = 0; i < end; ++i) {
+            int src_id = i % 16;
+            int src_offset = i / 16;
+            int dst_offset = i;
+
+            out.qs[dst_offset] = in[src_id].qs[src_offset];
+        }
+
+        // RVV repacking.
+        //
+        // Extract sums and mins for all 8 sub-blocks for each block of Q4_K.
+        uint8_t s[128], m[128];
+        for (int i = 0; i < 4; i++) {
+            for (int j = 0; j < 16; j++) {
+                s[i * 16 + j] = in[j].scales[i] & 63;
+                m[i * 16 + j] = in[j].scales[i + 4] & 63;
+            }
+        }
+        for (int i = 0; i < 4; i++) {
+            for (int j = 0; j < 16; j++) {
+                s[64 + i * 16 + j] = ((in[j].scales[i] & 192) >> 2) | (in[j].scales[i+8] & 15);
+                m[64 + i * 16 + j] = ((in[j].scales[i + 4] & 192) >> 2) | ((in[j].scales[i+8] & 240) >> 4);
+            }
+        }
+
+        for (int i = 0; i < 128; i++) {
+            out.scales[i] = (s[i] & 15) | ((m[i] & 15) << 4);
+        }
+        for (int i = 0; i < 64; i++) {
+            out.scales[128 + i] = ((s[i] & 48) >> 4) | ((m[i] & 48) >> 2) | (s[64 + i] & 48) | ((m[64 + i] & 48) << 2);
+        }
+    } else {
+        GGML_ASSERT(false);
+    }
+
+    return out;
+}
+
+static block_q2_Kx8 make_block_q2_Kx8(block_q2_K * in, unsigned int blck_size_interleave) {
+    block_q2_Kx8 out;
+
+    // Delta(scale) and dmin values of the eight Q2_K structures are copied onto the output interleaved structure
+    for (int i = 0; i < 8; i++) {
+        out.d[i] = in[i].GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.d;
+    }
+
+    for (int i = 0; i < 8; i++) {
+        out.dmin[i] = in[i].GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.dmin;
+    }
+
+    const int end = QK_K * 2 / blck_size_interleave;
+
+    // Interleave Q2_K quants by taking 8 bytes at a time
+    for (int i = 0; i < end; ++i) {
+        int src_id = i % 8;
+        int src_offset = (i / 8) * blck_size_interleave;
+        int dst_offset = i * blck_size_interleave;
+
+        uint64_t elems;
+        memcpy(&elems, &in[src_id].qs[src_offset], sizeof(uint64_t));
         memcpy(&out.qs[dst_offset], &elems, sizeof(uint64_t));
     }
 
+    // The below logic is designed so as to unpack and rearrange scales and mins values in Q2_K
+    // Currently the Q2_K structure has 16 scales and 16 mins packed in 16 bytes ( 4 bits for each value)
+    // The output Q2_Kx8 structure has 128 bytes for storing scales and mins
+    // Every 16 byte is packed such that it contains scales and mins for corresponding sub blocks from Q2_K structure
+    // For eg - First 16 bytes contains 16 scales and 16 mins - each of first and second sub blocks from different Q2_K structures
+
+    for (int i = 0; i < 128; i++) {
+        // Index for selecting which q2k super block
+        int src1 = (i % 16) / 2;
+        // Index for selecting scale
+        int src2 = ((i / 16) * 2) + (i % 2);
+
+        out.scales[i] = in[src1].scales[src2];
+    }
     return out;
 }
 
-static block_q4_Kx8 make_block_q4_Kx8(block_q4_K * in, unsigned int blck_size_interleave) {
-    block_q4_Kx8 out;
-    //Delta(scale) and dmin values of the eight Q4_K structures are copied onto the output interleaved structure
+static block_q5_Kx8 make_block_q5_Kx8(block_q5_K * in, unsigned int blck_size_interleave) {
+    block_q5_Kx8 out;
+    //Delta(scale) and dmin values of the eight Q5_K structures are copied onto the output interleaved structure
     for (int i = 0; i < 8; i++) {
         out.d[i] = in[i].GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.d;
     }
@@ -1518,22 +3017,33 @@ static block_q4_Kx8 make_block_q4_Kx8(block_q4_K * in, unsigned int blck_size_in
 
     const int end = QK_K * 4 / blck_size_interleave;
 
-    // Interleave Q4_K quants by taking 8 bytes at a time
+    // Interleave Q5_K quants by taking blck_size_interleave bytes at a time
     for (int i = 0; i < end; ++i) {
-        int src_id = i % 8;
+        int src_id     = i % 8;
         int src_offset = (i / 8) * blck_size_interleave;
         int dst_offset = i * blck_size_interleave;
 
-        uint64_t elems;
-        memcpy(&elems, &in[src_id].qs[src_offset], sizeof(uint64_t));
-        memcpy(&out.qs[dst_offset], &elems, sizeof(uint64_t));
+        memcpy(&out.qs[dst_offset], &in[src_id].qs[src_offset], blck_size_interleave);
     }
 
-    // The below logic is designed so as to unpack and rearrange scales and mins values in Q4_K
-    // Currently the Q4_K structure has 8 scales and 8 mins packed in 12 bytes ( 6 bits for each value)
-    // The output Q4_Kx8 structure has 96 bytes
-    // Every 12 byte is packed such that it contains scales and mins for corresponding sub blocks from Q4_K structure
-    // For eg - First 12 bytes contains 8 scales and 8 mins - each of first sub block from different Q4_K structures
+    // Repeat for high bits with the same chunk size, since
+    // the high bits are interleaved in Q5_K and the index is
+    // qh_idx = (qs_idx % 32);
+    // qh_val = qh[qh_idx] >> (qs_idx / 32);
+    for (int i = 0; i < end / 4; ++i) {
+        int src_id     = i % 8;
+        int src_offset = (i / 8) * blck_size_interleave;
+        int dst_offset = i * blck_size_interleave;
+
+        memcpy(&out.qh[dst_offset], &in[src_id].qh[src_offset], blck_size_interleave);
+    }
+
+    // The below logic is copied over from Q4_K
+    // The point is to unpack all the scales and mins for each sub block every time we load 12 bytes.
+    // Currently the Q5_K structure has 8 scales and 8 mins packed in 12 bytes ( 6 bits for each value)
+    // The output Q5_Kx8 structure has 96 bytes
+    // Every 12 byte is packed such that it contains scales and mins for corresponding sub blocks from Q5_K structure
+    // For eg - First 12 bytes contains 8 scales and 8 mins - each of first sub block from different Q5_K structures
     uint8_t s[8], m[8];
 
     for (int i = 0; i < 4; i++) {
@@ -1554,13 +3064,12 @@ static block_q4_Kx8 make_block_q4_Kx8(block_q4_K * in, unsigned int blck_size_in
         out.scales[i * 12 + 9]  = (s[5] & 15) + ((m[5] & 15) << 4);
         out.scales[i * 12 + 10] = (s[6] & 15) + ((m[6] & 15) << 4);
         out.scales[i * 12 + 11] = (s[7] & 15) + ((m[7] & 15) << 4);
-
     }
 
     for (int i = 0; i < 4; i++) {
         for (int j = 0; j < 8; j++) {
-            s[j] = ((in[j].scales[i] & 192) >> 2) | (in[j].scales[i+8] & 15);
-            m[j] = ((in[j].scales[i + 4] & 192) >> 2) | ((in[j].scales[i+8] & 240) >> 4);
+            s[j] = ((in[j].scales[i] & 192) >> 2) | (in[j].scales[i + 8] & 15);
+            m[j] = ((in[j].scales[i + 4] & 192) >> 2) | ((in[j].scales[i + 8] & 240) >> 4);
         }
 
         out.scales[i * 12 + 48] = (s[0] & 63) + ((s[4] & 48) << 2);
@@ -1575,54 +3084,117 @@ static block_q4_Kx8 make_block_q4_Kx8(block_q4_K * in, unsigned int blck_size_in
         out.scales[i * 12 + 57] = (s[5] & 15) + ((m[5] & 15) << 4);
         out.scales[i * 12 + 58] = (s[6] & 15) + ((m[6] & 15) << 4);
         out.scales[i * 12 + 59] = (s[7] & 15) + ((m[7] & 15) << 4);
-
     }
 
     return out;
 }
 
-static block_q2_Kx8 make_block_q2_Kx8(block_q2_K * in, unsigned int blck_size_interleave) {
-    block_q2_Kx8 out;
+static block_q6_Kx8 make_block_q6_Kx8(block_q6_K * in, unsigned int blck_size_interleave) {
+    block_q6_Kx8  out;
+    constexpr int n_blocks = 8;  // Kx8
+    for (int i = 0; i < n_blocks; i++) {
+        out.d[i] = in[i].d;
+    }
 
-    // Delta(scale) and dmin values of the eight Q2_K structures are copied onto the output interleaved structure
-    for (int i = 0; i < 8; i++) {
-        out.d[i] = in[i].GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.d;
+    const int end_ls = QK_K * 4 / blck_size_interleave;
+    // Interleave Q6_K quants by taking blck_size_interleave bytes at a time
+    for (int i = 0; i < end_ls; ++i) {
+        int src_id     = i % n_blocks;
+        int src_offset = (i / n_blocks) * blck_size_interleave;
+        int dst_offset = i * blck_size_interleave;
+
+        uint64_t elem_ls;
+        memcpy(&elem_ls, &in[src_id].ql[src_offset], blck_size_interleave);
+        memcpy(&out.ql[dst_offset], &elem_ls, blck_size_interleave);
     }
 
-    for (int i = 0; i < 8; i++) {
+    // Interleave high bits using same chunk size as low bits
+    const int end_hs = end_ls / 2;
+    for (int i = 0; i < end_hs; ++i) {
+        int src_id     = i % n_blocks;
+        int src_offset = (i / n_blocks) * blck_size_interleave;
+        int dst_offset = i * blck_size_interleave;
+
+        uint64_t elem_hs;
+        memcpy(&elem_hs, &in[src_id].qh[src_offset], blck_size_interleave);
+        memcpy(&out.qh[dst_offset], &elem_hs, blck_size_interleave);
+    }
+
+    // The below logic is designed so as to unpack and rearrange scales in Q6_K
+    // The output Q6_Kx8 structure interleaves the 8 bit scales in the same fashion as the quants
+    // Q6_K structure has an 8-bit scale per 16 elements -> 16 scales
+    // scales: [0 bl0 0 bl1 ... 0 bl7][1 bl0 ... 1 bl7] ... [15 bl0 ... 15 bl7]  (bl = block)
+    constexpr int n_scales = QK_K / 16;
+
+    for (int i = 0; i < n_blocks; i++) {
+        for (int j = 0; j < n_scales; j++) {
+            out.scales[j * n_blocks + i] = in[i].scales[j];
+        }
+    }
+
+    return out;
+}
+
+static block_q2_Kx16 make_block_q2_Kx16(const block_q2_K * in, unsigned int blck_size_interleave) {
+    block_q2_Kx16 out;
+    constexpr int N_COLS = 16;
+
+    // 1. Copy Super-Scales (d) and Super-Mins (dmin)
+    for (int i = 0; i < N_COLS; i++) {
+        out.d[i]    = in[i].GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.d;
         out.dmin[i] = in[i].GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.dmin;
     }
 
-    const int end = QK_K * 2 / blck_size_interleave;
+    // 2. Interleave Q2_K Data
+    const int bytes_per_col = 64;
+    const int total_bytes = N_COLS * bytes_per_col;
+    const int end = total_bytes / blck_size_interleave;
 
-    // Interleave Q2_K quants by taking 8 bytes at a time
     for (int i = 0; i < end; ++i) {
-        int src_id = i % 8;
-        int src_offset = (i / 8) * blck_size_interleave;
+        int src_col_id = i % N_COLS;
+        int src_offset = (i / N_COLS) * blck_size_interleave;
         int dst_offset = i * blck_size_interleave;
-
-        uint64_t elems;
-        memcpy(&elems, &in[src_id].qs[src_offset], sizeof(uint64_t));
-        memcpy(&out.qs[dst_offset], &elems, sizeof(uint64_t));
+        memcpy(&out.qs[dst_offset], &in[src_col_id].qs[src_offset], blck_size_interleave);
     }
 
-    // The below logic is designed so as to unpack and rearrange scales and mins values in Q2_K
-    // Currently the Q2_K structure has 16 scales and 16 mins packed in 16 bytes ( 4 bits for each value)
-    // The output Q2_Kx8 structure has 128 bytes for storing scales and mins
-    // Every 16 byte is packed such that it contains scales and mins for corresponding sub blocks from Q2_K structure
-    // For eg - First 16 bytes contains 16 scales and 16 mins - each of first and second sub blocks from different Q2_K structures
+    // 3. Repack Scales into the Optimized "Sequential-Parallel" Layout
+    int out_idx = 0;
 
-    for(int i = 0; i < 128; i++){
+    // Arrays define the sub-block order for each group
+    const int even_low_sbs[]  = {0, 2, 4, 6};
+    const int odd_low_sbs[]   = {1, 3, 5, 7};
+    const int even_high_sbs[] = {8, 10, 12, 14};
+    const int odd_high_sbs[]  = {9, 11, 13, 15};
 
-        // Index for selecting which q2k super block
-        int src1 = (i % 16) / 2;
-        // Index for selecting scale
-        int src2 = ((i / 16) * 2) + (i % 2);
+    // Pack Group 1: Even-Low
+    for (int sb : even_low_sbs) {
+        for (int col = 0; col < N_COLS; col++) {
+            out.scales[out_idx++] = in[col].scales[sb];
+        }
+    }
 
-        out.scales[i] = in[src1].scales[src2];
+    // Pack Group 2: Odd-Low
+    for (int sb : odd_low_sbs) {
+        for (int col = 0; col < N_COLS; col++) {
+            out.scales[out_idx++] = in[col].scales[sb];
+        }
+    }
+
+    // Pack Group 3: Even-High
+    for (int sb : even_high_sbs) {
+        for (int col = 0; col < N_COLS; col++) {
+            out.scales[out_idx++] = in[col].scales[sb];
+        }
     }
-    return out;
 
+    // Pack Group 4: Odd-High
+    for (int sb : odd_high_sbs) {
+        for (int col = 0; col < N_COLS; col++) {
+            out.scales[out_idx++] = in[col].scales[sb];
+        }
+    }
+
+    return out;
 }
 
 static int repack_q4_0_to_q4_0_4_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) {
@@ -1687,6 +3259,36 @@ static int repack_q4_K_to_q4_K_8_bl(struct ggml_tensor * t, int interleave_block
     GGML_UNUSED(data_size);
 }
 
+static int repack_q4_K_to_q4_K_16_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) {
+    GGML_ASSERT(t->type == GGML_TYPE_Q4_K);
+    constexpr int nrows_interleaved = 16;
+
+    block_q4_Kx16 * dst = (block_q4_Kx16*)t->data;
+    const block_q4_K * src = (const block_q4_K*) data;
+    block_q4_K dst_tmp[16];
+    int nrow = ggml_nrows(t);
+    int nblocks = t->ne[0] / QK_K;
+
+    GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_q4_K));
+
+    if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % 8 != 0) {
+        return -1;
+    }
+
+    for (int b = 0; b < nrow; b += nrows_interleaved) {
+        for (int64_t x = 0; x < nblocks; x++) {
+            for (int i  = 0; i < nrows_interleaved; i++ ) {
+                dst_tmp[i] = src[x + i * nblocks];
+            }
+            *dst++ = make_block_q4_Kx16(dst_tmp, interleave_block);
+        }
+        src += nrows_interleaved * nblocks;
+    }
+    return 0;
+
+    GGML_UNUSED(data_size);
+}
+
 static int repack_q2_K_to_q2_K_8_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) {
     GGML_ASSERT(t->type == GGML_TYPE_Q2_K);
     GGML_ASSERT(interleave_block == 8);
@@ -1706,7 +3308,7 @@ static int repack_q2_K_to_q2_K_8_bl(struct ggml_tensor * t, int interleave_block
 
     for (int b = 0; b < nrow; b += nrows_interleaved) {
         for (int64_t x = 0; x < nblocks; x++) {
-            for (int i  = 0; i < nrows_interleaved; i++ ) {
+            for (int i = 0; i < nrows_interleaved; i++) {
                 dst_tmp[i] = src[x + i * nblocks];
             }
             *dst++ = make_block_q2_Kx8(dst_tmp, interleave_block);
@@ -1718,6 +3320,132 @@ static int repack_q2_K_to_q2_K_8_bl(struct ggml_tensor * t, int interleave_block
     GGML_UNUSED(data_size);
 }
 
+static int repack_q2_K_to_q2_K_16_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) {
+    GGML_ASSERT(t->type == GGML_TYPE_Q2_K);
+    constexpr int nrows_interleaved = 16;
+
+    block_q2_Kx16 * dst = (block_q2_Kx16*)t->data;
+    const block_q2_K * src = (const block_q2_K*) data;
+
+    block_q2_K dst_tmp[nrows_interleaved];
+
+    int nrow = ggml_nrows(t);
+    int nblocks = t->ne[0] / QK_K;
+
+    GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_q2_K));
+
+    if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % 8 != 0) {
+        return -1;
+    }
+
+    for (int b = 0; b < nrow; b += nrows_interleaved) {
+        for (int64_t x = 0; x < nblocks; x++) {
+            // This loop gathers 16 separate blocks (one from each column)
+            // that correspond to the same K-dimension chunk.
+            for (int i  = 0; i < nrows_interleaved; i++ ) {
+                dst_tmp[i] = src[x + i * nblocks];
+            }
+
+            *dst++ = make_block_q2_Kx16(dst_tmp, interleave_block);
+        }
+        src += nrows_interleaved * nblocks;
+    }
+    return 0;
+
+    GGML_UNUSED(data_size);
+}
+
+static int repack_q4_0_to_q4_0_16_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) {
+    GGML_ASSERT(t->type == GGML_TYPE_Q4_0);
+    constexpr int nrows_interleaved = 16;
+
+    block_q4_0x16 * dst = (block_q4_0x16*)t->data;
+    const block_q4_0 * src = (const block_q4_0*) data;
+    block_q4_0 dst_tmp[16];
+    int nrow = ggml_nrows(t);
+    int nblocks = t->ne[0] / QK4_0;
+
+    GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_q4_0));
+
+    if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % 8 != 0) {
+        return -1;
+    }
+
+    for (int b = 0; b < nrow; b += nrows_interleaved) {
+        for (int64_t x = 0; x < nblocks; x++) {
+            for (int i  = 0; i < nrows_interleaved; i++ ) {
+                dst_tmp[i] = src[x + i * nblocks];
+            }
+            *dst++ = make_block_q4_0x16(dst_tmp, interleave_block);
+        }
+        src += nrows_interleaved * nblocks;
+    }
+    return 0;
+
+    GGML_UNUSED(data_size);
+}
+
+static int repack_q5_K_to_q5_K_8_bl(struct ggml_tensor *       t,
+                                    int                        interleave_block,
+                                    const void * GGML_RESTRICT data,
+                                    size_t                     data_size) {
+    GGML_ASSERT(t->type == GGML_TYPE_Q5_K);
+    GGML_ASSERT(interleave_block == 4 || interleave_block == 8);
+    constexpr int nrows_interleaved = 8;
+
+    block_q5_Kx8 *     dst = (block_q5_Kx8 *) t->data;
+    const block_q5_K * src = (const block_q5_K *) data;
+    block_q5_K         dst_tmp[8];
+    int                nrow    = ggml_nrows(t);
+    int                nblocks = t->ne[0] / QK_K;
+
+    GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_q5_K));
+
+    if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % 8 != 0) {
+        return -1;
+    }
+
+    for (int b = 0; b < nrow; b += nrows_interleaved) {
+        for (int64_t x = 0; x < nblocks; x++) {
+            for (int i = 0; i < nrows_interleaved; i++) {
+                dst_tmp[i] = src[x + i * nblocks];
+            }
+            *dst++ = make_block_q5_Kx8(dst_tmp, interleave_block);
+        }
+        src += nrows_interleaved * nblocks;
+    }
+    return 0;
+}
+
+static int repack_q6_K_to_q6_K_8_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) {
+    GGML_ASSERT(t->type == GGML_TYPE_Q6_K);
+    GGML_ASSERT(interleave_block == 4 || interleave_block == 8);
+    constexpr int nrows_interleaved = 8;
+
+    block_q6_Kx8 * dst = (block_q6_Kx8 *)t->data;
+    const block_q6_K * src = (const block_q6_K *) data;
+    block_q6_K dst_tmp[8];
+    int nrow = ggml_nrows(t);
+    int nblocks = t->ne[0] / QK_K;
+
+    GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_q6_K));
+
+    if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % 8 != 0) {
+        return -1;
+    }
+
+    for (int b = 0; b < nrow; b += nrows_interleaved) {
+        for (int64_t x = 0; x < nblocks; x++) {
+            for (int i = 0; i < nrows_interleaved; i++) {
+                dst_tmp[i] = src[x + i * nblocks];
+            }
+            *dst++ = make_block_q6_Kx8(dst_tmp, interleave_block);
+        }
+        src += nrows_interleaved * nblocks;
+    }
+    return 0;
+}
+
 static int repack_q4_0_to_q4_0_8_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) {
     GGML_ASSERT(t->type == GGML_TYPE_Q4_0);
     GGML_ASSERT(interleave_block == 8);
@@ -1757,9 +3485,63 @@ static int repack_q8_0_to_q8_0_4_bl(struct ggml_tensor *       t,
     GGML_ASSERT(interleave_block == 4 || interleave_block == 8);
     constexpr int nrows_interleaved = 4;
 
-    block_q8_0x4 *     dst = (block_q8_0x4 *) t->data;
+    block_q8_0x4 *     dst = (block_q8_0x4 *) t->data;
+    const block_q8_0 * src = (const block_q8_0 *) data;
+    block_q8_0         dst_tmp[4];
+    int                nrow    = ggml_nrows(t);
+    int                nblocks = t->ne[0] / QK8_0;
+
+    GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_q8_0));
+
+    if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % 8 != 0) {
+        return -1;
+    }
+
+    for (int b = 0; b < nrow; b += nrows_interleaved) {
+        for (int64_t x = 0; x < nblocks; x++) {
+            for (int i = 0; i < nrows_interleaved; i++) {
+                dst_tmp[i] = src[x + i * nblocks];
+            }
+            *dst++ = make_block_q8_0x4(dst_tmp, interleave_block);
+        }
+        src += nrows_interleaved * nblocks;
+    }
+    return 0;
+}
+
+static block_q8_0x16 make_block_q8_0x16(block_q8_0 * in, unsigned int blck_size_interleave) {
+    block_q8_0x16 out;
+
+    for (int i = 0; i < 16; i++) {
+        out.d[i] = in[i].d;
+    }
+
+    const int end = QK8_0 * 16 / blck_size_interleave;
+
+    if (blck_size_interleave == 1) {
+        for (int i = 0; i < end; ++i) {
+            int src_id     = i % 16;
+            int src_offset = i / 16;
+            int dst_offset = i;
+            out.qs[dst_offset] = in[src_id].qs[src_offset];
+        }
+    } else {
+        GGML_ASSERT(false);
+    }
+
+    return out;
+}
+
+static int repack_q8_0_to_q8_0_16_bl(struct ggml_tensor *       t,
+                                    int                        interleave_block,
+                                    const void * GGML_RESTRICT data,
+                                    size_t                     data_size) {
+    GGML_ASSERT(t->type == GGML_TYPE_Q8_0);
+    constexpr int nrows_interleaved = 16;
+
+    block_q8_0x16 *     dst = (block_q8_0x16 *) t->data;
     const block_q8_0 * src = (const block_q8_0 *) data;
-    block_q8_0         dst_tmp[4];
+    block_q8_0         dst_tmp[16];
     int                nrow    = ggml_nrows(t);
     int                nblocks = t->ne[0] / QK8_0;
 
@@ -1774,7 +3556,7 @@ static int repack_q8_0_to_q8_0_4_bl(struct ggml_tensor *       t,
             for (int i = 0; i < nrows_interleaved; i++) {
                 dst_tmp[i] = src[x + i * nblocks];
             }
-            *dst++ = make_block_q8_0x4(dst_tmp, interleave_block);
+            *dst++ = make_block_q8_0x16(dst_tmp, interleave_block);
         }
         src += nrows_interleaved * nblocks;
     }
@@ -1906,6 +3688,177 @@ static int repack_iq4_nl_to_iq4_nl_8_bl(struct ggml_tensor * t, int interleave_b
     GGML_UNUSED(data_size);
 }
 
+static block_iq4_nlx16 make_block_iq4_nlx16(block_iq4_nl * in, unsigned int blck_size_interleave) {
+    block_iq4_nlx16 out;
+
+    for (int i = 0; i < 16; i++) {
+        out.d[i] = in[i].d;
+    }
+
+    const int end = QK4_NL * 8 / blck_size_interleave;
+
+    if (blck_size_interleave == 1) {
+        for (int i = 0; i < end; ++i) {
+            int src_id = i % 16;
+            int src_offset = i / 16;
+            int dst_offset = i;
+
+            out.qs[dst_offset] = in[src_id].qs[src_offset];
+        }
+    } else {
+        GGML_ASSERT(false);
+    }
+
+    return out;
+}
+
+static int repack_iq4_nl_to_iq4_nl_16_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) {
+    GGML_ASSERT(t->type == GGML_TYPE_IQ4_NL);
+    GGML_ASSERT(interleave_block == 1);
+
+    const block_iq4_nl    * src = (const block_iq4_nl   *)data;
+          block_iq4_nlx16 * dst = (      block_iq4_nlx16 *)t->data;
+
+    block_iq4_nl dst_tmp[16];
+
+    int nrow = ggml_nrows(t);
+    int nrows_interleaved = 16;
+    int nblocks = t->ne[0] / QK4_NL;
+
+    GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_iq4_nl));
+
+    if (t->ne[1] % nrows_interleaved != 0) {
+        return -1;
+    }
+
+    for (int b = 0; b < nrow; b += nrows_interleaved) {
+        for (int64_t x = 0; x < nblocks; x++) {
+            for (int i = 0; i < nrows_interleaved; i++) {
+                dst_tmp[i] = src[x + i * nblocks];
+            }
+            *dst++ = make_block_iq4_nlx16(dst_tmp, interleave_block);
+        }
+        src += nrows_interleaved * nblocks;
+    }
+    return 0;
+
+    GGML_UNUSED(data_size);
+}
+
+static block_mxfp4x4 make_block_mxfp4x4(block_mxfp4 * in, unsigned int blck_size_interleave) {
+    block_mxfp4x4 out;
+
+    for (int i = 0; i < 4; i++) {
+        out.e[i] = in[i].e;
+    }
+
+    const int end = QK_MXFP4 * 2 / blck_size_interleave;
+
+    if (blck_size_interleave == 4) {
+        for (int i = 0; i < end; ++i) {
+            int src_id = i % 4;
+            int src_offset = (i / 4) * blck_size_interleave;
+            int dst_offset = i * blck_size_interleave;
+
+            memcpy(&out.qs[dst_offset], &in[src_id].qs[src_offset], sizeof(uint32_t));
+        }
+    } else {
+        GGML_ASSERT(false);
+    }
+
+    return out;
+}
+
+static int repack_mxfp4_to_mxfp4_4_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) {
+    GGML_ASSERT(t->type == GGML_TYPE_MXFP4);
+    GGML_ASSERT(interleave_block == 4);
+
+    const block_mxfp4   * src = (const block_mxfp4   *)data;
+          block_mxfp4x4 * dst = (      block_mxfp4x4 *)t->data;
+
+    block_mxfp4 dst_tmp[4];
+
+    int nrow = ggml_nrows(t);
+    int nrows_interleaved = 4;
+    int nblocks = t->ne[0] / QK_MXFP4;
+
+    GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_mxfp4));
+
+    if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % 8 != 0) {
+        return -1;
+    }
+
+    for (int b = 0; b < nrow; b += nrows_interleaved) {
+        for (int64_t x = 0; x < nblocks; x++) {
+            for (int i = 0; i < nrows_interleaved; i++) {
+                dst_tmp[i] = src[x + i * nblocks];
+            }
+            *dst++ = make_block_mxfp4x4(dst_tmp, interleave_block);
+        }
+        src += nrows_interleaved * nblocks;
+    }
+    return 0;
+
+    GGML_UNUSED(data_size);
+}
+
+static block_mxfp4x8 make_block_mxfp4x8(block_mxfp4 * in, unsigned int blck_size_interleave) {
+    block_mxfp4x8 out;
+
+    for (int i = 0; i < 8; i++) {
+        out.e[i] = in[i].e;
+    }
+
+    const int end = QK_MXFP4 * 4 / blck_size_interleave;
+
+    if (blck_size_interleave == 8) {
+        for (int i = 0; i < end; ++i) {
+            int src_id = i % 8;
+            int src_offset = (i / 8) * blck_size_interleave;
+            int dst_offset = i * blck_size_interleave;
+
+            memcpy(&out.qs[dst_offset], &in[src_id].qs[src_offset], sizeof(uint64_t));
+        }
+    } else {
+        GGML_ASSERT(false);
+    }
+
+    return out;
+}
+
+static int repack_mxfp4_to_mxfp4_8_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) {
+    GGML_ASSERT(t->type == GGML_TYPE_MXFP4);
+    GGML_ASSERT(interleave_block == 8);
+
+    const block_mxfp4   * src = (const block_mxfp4   *)data;
+          block_mxfp4x8 * dst = (      block_mxfp4x8 *)t->data;
+
+    block_mxfp4 dst_tmp[8];
+
+    int nrow = ggml_nrows(t);
+    int nrows_interleaved = 8;
+    int nblocks = t->ne[0] / QK_MXFP4;
+
+    GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_mxfp4));
+
+    if (t->ne[1] % nrows_interleaved != 0) {
+        return -1;
+    }
+
+    for (int b = 0; b < nrow; b += nrows_interleaved) {
+        for (int64_t x = 0; x < nblocks; x++) {
+            for (int i = 0; i < nrows_interleaved; i++) {
+                dst_tmp[i] = src[x + i * nblocks];
+            }
+            *dst++ = make_block_mxfp4x8(dst_tmp, interleave_block);
+        }
+        src += nrows_interleaved * nblocks;
+    }
+    return 0;
+
+    GGML_UNUSED(data_size);
+}
+
 namespace ggml::cpu::repack {
 // repack
 template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS>
@@ -1936,6 +3889,22 @@ template <> int repack<block_q2_K, 8, 8>(struct ggml_tensor * t, const void * da
     return repack_q2_K_to_q2_K_8_bl(t, 8, data, data_size);
 }
 
+template <> int repack<block_q5_K, 4, 8>(struct ggml_tensor * t, const void * data, size_t data_size) {
+    return repack_q5_K_to_q5_K_8_bl(t, 4, data, data_size);
+}
+
+template <> int repack<block_q5_K, 8, 8>(struct ggml_tensor * t, const void * data, size_t data_size) {
+    return repack_q5_K_to_q5_K_8_bl(t, 8, data, data_size);
+}
+
+template <> int repack<block_q6_K, 4, 8>(struct ggml_tensor * t, const void * data, size_t data_size) {
+    return repack_q6_K_to_q6_K_8_bl(t, 4, data, data_size);
+}
+
+template <> int repack<block_q6_K, 8, 8>(struct ggml_tensor * t, const void * data, size_t data_size) {
+    return repack_q6_K_to_q6_K_8_bl(t, 8, data, data_size);
+}
+
 template <> int repack<block_iq4_nl, 4, 4>(struct ggml_tensor * t, const void * data, size_t data_size) {
     return repack_iq4_nl_to_iq4_nl_4_bl(t, 4, data, data_size);
 }
@@ -1949,6 +3918,14 @@ template <> int repack<block_iq4_nl, 8, 8>(struct ggml_tensor * t, const void *
     return repack_iq4_nl_to_iq4_nl_8_bl(t, 8, data, data_size);
 }
 
+template <> int repack<block_mxfp4, 4, 4>(struct ggml_tensor * t, const void * data, size_t data_size) {
+    return repack_mxfp4_to_mxfp4_4_bl(t, 4, data, data_size);
+}
+
+template <> int repack<block_mxfp4, 8, 8>(struct ggml_tensor * t, const void * data, size_t data_size) {
+    return repack_mxfp4_to_mxfp4_8_bl(t, 8, data, data_size);
+}
+
 template <> int repack<block_q8_0, 4, 4>(struct ggml_tensor * t, const void * data, size_t data_size) {
     return repack_q8_0_to_q8_0_4_bl(t, 4, data, data_size);
 }
@@ -1957,6 +3934,28 @@ template <> int repack<block_q8_0, 8, 4>(struct ggml_tensor * t, const void * da
     return repack_q8_0_to_q8_0_4_bl(t, 8, data, data_size);
 }
 
+#if defined __riscv_zvfh
+template <> int repack<block_q4_0, 1, 16>(struct ggml_tensor * t, const void * data, size_t data_size) {
+    return repack_q4_0_to_q4_0_16_bl(t, 1, data, data_size);
+}
+
+template <> int repack<block_q4_K, 1, 16>(struct ggml_tensor * t, const void * data, size_t data_size) {
+    return repack_q4_K_to_q4_K_16_bl(t, 1, data, data_size);
+}
+
+template <> int repack<block_iq4_nl, 1, 16>(struct ggml_tensor * t, const void * data, size_t data_size) {
+    return repack_iq4_nl_to_iq4_nl_16_bl(t, 1, data, data_size);
+}
+
+template <> int repack<block_q8_0, 1, 16>(struct ggml_tensor * t, const void * data, size_t data_size) {
+    return repack_q8_0_to_q8_0_16_bl(t, 1, data, data_size);
+}
+
+template <> int repack<block_q2_K, 1, 16>(struct ggml_tensor * t, const void * data, size_t data_size) {
+    return repack_q2_K_to_q2_K_16_bl(t, 1, data, data_size);
+}
+#endif
+
 // gemv
 template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PARAM_TYPE>
 void gemv(int, float *, size_t, const void *, const void *, int, int);
@@ -1973,6 +3972,17 @@ template <> void gemv<block_q4_0, 8, 8, GGML_TYPE_Q8_0>(int n, float * s, size_t
     ggml_gemv_q4_0_8x8_q8_0(n, s, bs, vx, vy, nr, nc);
 }
 
+template <>
+void gemv<block_q2_K, 8, 8, GGML_TYPE_Q8_K>(int          n,
+                                            float *      s,
+                                            size_t       bs,
+                                            const void * vx,
+                                            const void * vy,
+                                            int          nr,
+                                            int          nc) {
+    ggml_gemv_q2_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
+}
+
 template <> void gemv<block_q4_K, 4, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
     ggml_gemv_q4_K_8x4_q8_K(n, s, bs, vx, vy, nr, nc);
 }
@@ -1981,8 +3991,20 @@ template <> void gemv<block_q4_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t
     ggml_gemv_q4_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
 }
 
-template <> void gemv<block_q2_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
-    ggml_gemv_q2_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
+template <> void gemv<block_q5_K, 4, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemv_q5_K_8x4_q8_K(n, s, bs, vx, vy, nr, nc);
+}
+
+template <> void gemv<block_q5_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemv_q5_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
+}
+
+template <> void gemv<block_q6_K, 4, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemv_q6_K_8x4_q8_K(n, s, bs, vx, vy, nr, nc);
+}
+
+template <> void gemv<block_q6_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemv_q6_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
 }
 
 template <> void gemv<block_iq4_nl, 4, 4, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
@@ -1993,6 +4015,14 @@ template <> void gemv<block_iq4_nl, 8, 8, GGML_TYPE_Q8_0>(int n, float * s, size
     ggml_gemv_iq4_nl_8x8_q8_0(n, s, bs, vx, vy, nr, nc);
 }
 
+template <> void gemv<block_mxfp4, 4, 4, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemv_mxfp4_4x4_q8_0(n, s, bs, vx, vy, nr, nc);
+}
+
+template <> void gemv<block_mxfp4, 8, 8, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemv_mxfp4_8x8_q8_0(n, s, bs, vx, vy, nr, nc);
+}
+
 template <> void gemv<block_q8_0, 4, 4, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
     ggml_gemv_q8_0_4x4_q8_0(n, s, bs, vx, vy, nr, nc);
 }
@@ -2001,6 +4031,28 @@ template <> void gemv<block_q8_0, 8, 4, GGML_TYPE_Q8_0>(int n, float * s, size_t
     ggml_gemv_q8_0_4x8_q8_0(n, s, bs, vx, vy, nr, nc);
 }
 
+#if defined __riscv_zvfh
+template <> void gemv<block_q4_0, 1, 16, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemv_q4_0_16x1_q8_0(n, s, bs, vx, vy, nr, nc);
+}
+
+template <> void gemv<block_q4_K, 1, 16, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemv_q4_K_16x1_q8_K(n, s, bs, vx, vy, nr, nc);
+}
+
+template <> void gemv<block_iq4_nl, 1, 16, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemv_iq4_nl_16x1_q8_0(n, s, bs, vx, vy, nr, nc);
+}
+
+template <> void gemv<block_q8_0, 1, 16, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemv_q8_0_16x1_q8_0(n, s, bs, vx, vy, nr, nc);
+}
+
+template <> void gemv<block_q2_K, 1, 16, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemv_q2_K_16x1_q8_K(n, s, bs, vx, vy, nr, nc);
+}
+#endif
+
 // gemm
 template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PARAM_TYPE>
 void gemm(int, float *, size_t, const void *, const void *, int, int);
@@ -2013,20 +4065,43 @@ template <> void gemm<block_q4_0, 8, 4, GGML_TYPE_Q8_0>(int n, float * s, size_t
     ggml_gemm_q4_0_4x8_q8_0(n, s, bs, vx, vy, nr, nc);
 }
 
-template <> void gemm<block_q4_K, 4, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
-    ggml_gemm_q4_K_8x4_q8_K(n, s, bs, vx, vy, nr, nc);
+template <>
+void gemm<block_q4_0, 8, 8, GGML_TYPE_Q8_0>(int          n,
+                                            float *      s,
+                                            size_t       bs,
+                                            const void * vx,
+                                            const void * vy,
+                                            int          nr,
+                                            int          nc) {
+    ggml_gemm_q4_0_8x8_q8_0(n, s, bs, vx, vy, nr, nc);
 }
 
-template <> void gemm<block_q4_0, 8, 8, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
-    ggml_gemm_q4_0_8x8_q8_0(n, s, bs, vx, vy, nr, nc);
+template <> void gemm<block_q2_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemm_q2_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
+}
+
+template <> void gemm<block_q4_K, 4, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemm_q4_K_8x4_q8_K(n, s, bs, vx, vy, nr, nc);
 }
 
 template <> void gemm<block_q4_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
     ggml_gemm_q4_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
 }
 
-template <> void gemm<block_q2_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
-    ggml_gemm_q2_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
+template <> void gemm<block_q5_K, 4, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemm_q5_K_8x4_q8_K(n, s, bs, vx, vy, nr, nc);
+}
+
+template <> void gemm<block_q5_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemm_q5_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
+}
+
+template <> void gemm<block_q6_K, 4, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemm_q6_K_8x4_q8_K(n, s, bs, vx, vy, nr, nc);
+}
+
+template <> void gemm<block_q6_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemm_q6_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
 }
 
 template <> void gemm<block_iq4_nl, 4, 4, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
@@ -2037,6 +4112,14 @@ template <> void gemm<block_iq4_nl, 8, 8, GGML_TYPE_Q8_0>(int n, float * s, size
     ggml_gemm_iq4_nl_8x8_q8_0(n, s, bs, vx, vy, nr, nc);
 }
 
+template <> void gemm<block_mxfp4, 4, 4, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemm_mxfp4_4x4_q8_0(n, s, bs, vx, vy, nr, nc);
+}
+
+template <> void gemm<block_mxfp4, 8, 8, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemm_mxfp4_8x8_q8_0(n, s, bs, vx, vy, nr, nc);
+}
+
 template <> void gemm<block_q8_0, 4, 4, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
     ggml_gemm_q8_0_4x4_q8_0(n, s, bs, vx, vy, nr, nc);
 }
@@ -2045,6 +4128,28 @@ template <> void gemm<block_q8_0, 8, 4, GGML_TYPE_Q8_0>(int n, float * s, size_t
     ggml_gemm_q8_0_4x8_q8_0(n, s, bs, vx, vy, nr, nc);
 }
 
+#if defined __riscv_zvfh
+template <> void gemm<block_q4_0, 1, 16, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemm_q4_0_16x1_q8_0(n, s, bs, vx, vy, nr, nc);
+}
+
+template <> void gemm<block_q4_K, 1, 16, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemm_q4_K_16x1_q8_K(n, s, bs, vx, vy, nr, nc);
+}
+
+template <> void gemm<block_iq4_nl, 1, 16, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemm_iq4_nl_16x1_q8_0(n, s, bs, vx, vy, nr, nc);
+}
+
+template <> void gemm<block_q8_0, 1, 16, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemm_q8_0_16x1_q8_0(n, s, bs, vx, vy, nr, nc);
+}
+
+template <> void gemm<block_q2_K, 1, 16, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemm_q2_K_16x1_q8_K(n, s, bs, vx, vy, nr, nc);
+}
+#endif
+
 class tensor_traits_base : public ggml::cpu::tensor_traits {
   public:
     virtual int repack(struct ggml_tensor * t, const void * data, size_t data_size) = 0;
@@ -2063,7 +4168,7 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PAR
             case GGML_OP_MUL_MAT_ID:
                 {
                     size = ggml_row_size(PARAM_TYPE, ggml_nelements(op->src[1]));
-                    size = GGML_PAD(size, sizeof(int64_t)); // + padding for next bloc.
+                    size = GGML_PAD(size, sizeof(int64_t)); // + padding for next block.
 
                     const int64_t ne02 = op->src[0]->ne[2]; // n_as, n_expert
                     const int64_t ne12 = op->src[1]->ne[2]; // n_tokens
@@ -2328,7 +4433,7 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PAR
         auto * wdata          = (char *)params->wdata;
         auto * wdata_src1_end = (char *)wdata + GGML_PAD(nbw3, sizeof(int64_t));
 
-        // total of [n_as][ne12 + 1] elemets of type mmid_row_mapping (2*int32_t = int64_t)
+        // total of [n_as][ne12 + 1] elements of type mmid_row_mapping (2*int32_t = int64_t)
         auto * matrix_row_counts = (int64_t *) (wdata_src1_end);                                        // [n_as]
         struct mmid_row_mapping * matrix_rows = (struct mmid_row_mapping *) (matrix_row_counts + n_as); // [n_as][ne12]
 
@@ -2393,20 +4498,19 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PAR
             for (int ir1 = 0; ir1 < nr1; ir1++) {
                 struct mmid_row_mapping row_mapping = MMID_MATRIX_ROW(cur_a, ir1);
 
-                const int id = row_mapping.i1; // selected expert index
+                const int id = row_mapping.i1;  // selected expert index
 
                 const int64_t i11 = id % ne11;
-                const int64_t i12 = row_mapping.i2; // row index in src1
+                const int64_t i12 = row_mapping.i2;  // row index in src1
 
-                const int64_t i1 = id;  // selected expert index
-                const int64_t i2 = i12; // row
+                const int64_t i1 = id;               // selected expert index
+                const int64_t i2 = i12;              // row
 
                 const auto * src1_col = (const char *) wdata + (i11 * nbw1 + i12 * nbw2);
 
-                gemv<BLOC_TYPE, INTER_SIZE, NB_COLS, PARAM_TYPE>(ne00,
-                        (float *)((char *) dst->data + (i1 * nb1 + i2 * nb2)) + src0_cur_start, ne01,
-                        src0_cur + src0_cur_start * nb01,
-                        src1_col, 1, src0_cur_end - src0_cur_start);
+                gemv<BLOC_TYPE, INTER_SIZE, NB_COLS, PARAM_TYPE>(
+                    ne00, (float *) ((char *) dst->data + (i1 * nb1 + i2 * nb2)) + src0_cur_start, ne01,
+                    src0_cur + src0_cur_start * nb01, src1_col, 1, src0_cur_end - src0_cur_start);
             }
         }
 #undef MMID_MATRIX_ROW
@@ -2422,7 +4526,6 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PAR
 }  // namespace ggml::cpu::repack
 
 static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(const struct ggml_tensor * cur) {
-
     // instance for Q4
     static const ggml::cpu::repack::tensor_traits<block_q4_0, 4, 4, GGML_TYPE_Q8_0> q4_0_4x4_q8_0;
     static const ggml::cpu::repack::tensor_traits<block_q4_0, 8, 4, GGML_TYPE_Q8_0> q4_0_4x8_q8_0;
@@ -2432,6 +4535,14 @@ static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(cons
     static const ggml::cpu::repack::tensor_traits<block_q4_K, 4, 8, GGML_TYPE_Q8_K> q4_K_8x4_q8_K;
     static const ggml::cpu::repack::tensor_traits<block_q4_K, 8, 8, GGML_TYPE_Q8_K> q4_K_8x8_q8_K;
 
+    // instance for Q5_K
+    static const ggml::cpu::repack::tensor_traits<block_q5_K, 4, 8, GGML_TYPE_Q8_K> q5_K_8x4_q8_K;
+    static const ggml::cpu::repack::tensor_traits<block_q5_K, 8, 8, GGML_TYPE_Q8_K> q5_K_8x8_q8_K;
+
+    // instance for Q6_K
+    static const ggml::cpu::repack::tensor_traits<block_q6_K, 4, 8, GGML_TYPE_Q8_K> q6_K_8x4_q8_K;
+    static const ggml::cpu::repack::tensor_traits<block_q6_K, 8, 8, GGML_TYPE_Q8_K> q6_K_8x8_q8_K;
+
     // instance for Q2
     static const ggml::cpu::repack::tensor_traits<block_q2_K, 8, 8, GGML_TYPE_Q8_K> q2_K_8x8_q8_K;
 
@@ -2439,13 +4550,28 @@ static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(cons
     static const ggml::cpu::repack::tensor_traits<block_iq4_nl, 4, 4, GGML_TYPE_Q8_0> iq4_nl_4x4_q8_0;
     static const ggml::cpu::repack::tensor_traits<block_iq4_nl, 8, 8, GGML_TYPE_Q8_0> iq4_nl_8x8_q8_0;
 
+    // instance for MXFP4
+    static const ggml::cpu::repack::tensor_traits<block_mxfp4, 4, 4, GGML_TYPE_Q8_0> mxfp4_4x4_q8_0;
+    static const ggml::cpu::repack::tensor_traits<block_mxfp4, 8, 8, GGML_TYPE_Q8_0> mxfp4_8x8_q8_0;
+
     // instance for Q8_0
     static const ggml::cpu::repack::tensor_traits<block_q8_0, 4, 4, GGML_TYPE_Q8_0> q8_0_4x4_q8_0;
     static const ggml::cpu::repack::tensor_traits<block_q8_0, 8, 4, GGML_TYPE_Q8_0> q8_0_4x8_q8_0;
 
+    // instances for RISC-V
+    //
+    // These implement outer-product style matrix multiplication kernels with
+    // an interleave of 1.
+#if defined __riscv_zvfh
+    static const ggml::cpu::repack::tensor_traits<block_q4_0, 1, 16, GGML_TYPE_Q8_0> q4_0_16x1_q8_0;
+    static const ggml::cpu::repack::tensor_traits<block_q4_K, 1, 16, GGML_TYPE_Q8_K> q4_K_16x1_q8_K;
+    static const ggml::cpu::repack::tensor_traits<block_iq4_nl, 1, 16, GGML_TYPE_Q8_0> iq4_nl_16x1_q8_0;
+    static const ggml::cpu::repack::tensor_traits<block_q8_0, 1, 16, GGML_TYPE_Q8_0> q8_0_16x1_q8_0;
+    static const ggml::cpu::repack::tensor_traits<block_q2_K, 1, 16, GGML_TYPE_Q8_K> q2_K_16x1_q8_K;
+#endif
+
     if (cur->type == GGML_TYPE_Q4_0) {
-        if (ggml_cpu_has_avx2() || (ggml_cpu_has_sve() && ggml_cpu_has_matmul_int8() && ggml_cpu_get_sve_cnt() == QK8_0)
-            || (ggml_cpu_has_riscv_v() && (ggml_cpu_get_rvv_vlen() >= QK4_0))) {
+        if (ggml_cpu_has_avx2() || (ggml_cpu_has_sve() && ggml_cpu_has_matmul_int8() && ggml_cpu_get_sve_cnt() == QK8_0)) {
             if (cur->ne[1] % 8 == 0) {
                 return &q4_0_8x8_q8_0;
             }
@@ -2460,6 +4586,17 @@ static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(cons
                 return &q4_0_4x4_q8_0;
             }
         }
+        if (ggml_cpu_has_riscv_v()) {
+            #if defined __riscv_zvfh
+            switch (__riscv_vlenb() * 8) {
+                case 128:  { break; } // TODO
+                case 256:  { if (cur->ne[1] % 16 == 0) { return &q4_0_16x1_q8_0; } break; }
+                case 512:  { break; } // TODO
+                case 1024: { break; } // TODO
+                default:   { return nullptr; }
+            }
+            #endif
+        }
     } else if (cur->type == GGML_TYPE_Q4_K) {
         if (ggml_cpu_has_avx2()) {
             if (cur->ne[1] % 8 == 0) {
@@ -2476,12 +4613,56 @@ static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(cons
                 return &q4_K_8x4_q8_K;
             }
         }
+        if (ggml_cpu_has_riscv_v()) {
+            #if defined __riscv_zvfh
+            switch (__riscv_vlenb() * 8) {
+                case 128:  { break; } // TODO
+                case 256:  { if (cur->ne[1] % 16 == 0) { return &q4_K_16x1_q8_K; } break; }
+                case 512:  { break; } // TODO
+                case 1024: { break; } // TODO
+                default:   { return nullptr; }
+            }
+            #endif
+        }
     } else if (cur->type == GGML_TYPE_Q2_K) {
         if (ggml_cpu_has_avx512()) {
             if (cur->ne[1] % 8 == 0) {
                 return &q2_K_8x8_q8_K;
             }
         }
+        if (ggml_cpu_has_riscv_v()) {
+            #if defined __riscv_zvfh
+            switch (__riscv_vlenb() * 8) {
+                case 128:  { break; } // TODO
+                case 256:  { if (cur->ne[1] % 16 == 0) { return &q2_K_16x1_q8_K; } break; }
+                case 512:  { break; } // TODO
+                case 1024: { break; } // TODO
+                default:   { return nullptr; }
+            }
+            #endif
+        }
+    } else if (cur->type == GGML_TYPE_Q5_K) {
+        if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) {
+            if (cur->ne[1] % 8 == 0) {
+                return &q5_K_8x8_q8_K;
+            }
+        }
+        if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
+            if (cur->ne[1] % 8 == 0) {
+                return &q5_K_8x4_q8_K;
+            }
+        }
+    } else if (cur->type == GGML_TYPE_Q6_K) {
+        if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) {
+            if (cur->ne[1] % 8 == 0) {
+                return &q6_K_8x8_q8_K;
+            }
+        }
+        if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
+            if (cur->ne[1] % 8 == 0) {
+                return &q6_K_8x4_q8_K;
+            }
+        }
     } else if (cur->type == GGML_TYPE_IQ4_NL) {
         if (ggml_cpu_has_avx2()) {
             if (cur->ne[1] % 8 == 0) {
@@ -2493,6 +4674,28 @@ static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(cons
                 return &iq4_nl_4x4_q8_0;
             }
         }
+        if (ggml_cpu_has_riscv_v()) {
+            #if defined __riscv_zvfh
+            switch (__riscv_vlenb() * 8) {
+                case 128:  { break; } // TODO
+                case 256:  { if (cur->ne[1] % 16 == 0) { return &iq4_nl_16x1_q8_0; } break; }
+                case 512:  { break; } // TODO
+                case 1024: { break; } // TODO
+                default:   { return nullptr; }
+            }
+            #endif
+        }
+    } else if (cur->type == GGML_TYPE_MXFP4) {
+        if (ggml_cpu_has_avx2()) {
+            if (cur->ne[1] % 8 == 0) {
+                return &mxfp4_8x8_q8_0;
+            }
+        }
+        if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
+            if (cur->ne[1] % 4 == 0) {
+                return &mxfp4_4x4_q8_0;
+            }
+        }
     } else if (cur->type == GGML_TYPE_Q8_0) {
         if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) {
             if (cur->ne[1] % 4 == 0) {
@@ -2504,6 +4707,17 @@ static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(cons
                 return &q8_0_4x4_q8_0;
             }
         }
+        if (ggml_cpu_has_riscv_v()) {
+            #if defined __riscv_zvfh
+            switch (__riscv_vlenb() * 8) {
+                case 128:  { break; } // TODO
+                case 256:  { if (cur->ne[1] % 16 == 0) { return &q8_0_16x1_q8_0; } break; }
+                case 512:  { break; } // TODO
+                case 1024: { break; } // TODO
+                default:   { return nullptr; }
+            }
+            #endif
+        }
     }
 
     return nullptr;