whispercpp 1.3.4 → 1.3.5

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

@@ -124,6 +124,58 @@ 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);
+    const int nb = k / QK_K;
+
+    block_q8_Kx4 * GGML_RESTRICT y = (block_q8_Kx4 *) vy;
+
+    // scalar
+    const int blck_size_interleave = 4;
+    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;
+        }
+
+        // Quants values are interleaved in sequence of four bytes from corresponding super blocks
+        // Bsums values are interleaved in sequence of four bsums from each super block taken for interleaving
+        // i.e first four bsums from the first super block, followed by first four bsums from second super block and so on
+        for (int j = 0; j < QK_K * 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);
+            int index = (((j & 15) >> 2) << 2) + ((j >> 8) << 4) + ((j >> 6) & 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];
+        }
+    }
+}
+
 void ggml_quantize_mat_q8_K_4x8_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
     assert(QK_K == 256);
     assert(k % QK_K == 0);
@@ -192,6 +244,12 @@ template <> void ggml_quantize_mat_t<8, GGML_TYPE_Q8_0>(const float * GGML_RESTR
     ggml_quantize_mat_q8_0_4x8(x, vy, n_per_row);
 }
 
+template <> void ggml_quantize_mat_t<4, 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_4x4(x, vy, n_per_row);
+}
+
 template <> void ggml_quantize_mat_t<8, 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);
@@ -333,6 +391,77 @@ void ggml_gemv_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
     }
 }
 
+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;
+
+    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 / 8) * 32;
+                uint8_t * scales_1 = (uint8_t *) utmp + (k / 8) * 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 / 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;
@@ -563,6 +692,100 @@ void ggml_gemv_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs
     }
 }
 
+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];
+        }
+    }
+}
+
 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;
@@ -727,6 +950,89 @@ 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);
+
+    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;
@@ -1007,8 +1313,129 @@ 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) {
+    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);
+
+    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_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];
+                            }
+                            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_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);
+    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_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];
+                            }
+                            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];
+                }
+            }
+        }
+    }
+}
+
 } // extern "C"
 
+static block_q8_0x4 make_block_q8_0x4(block_q8_0 * in, unsigned int blck_size_interleave) {
+    block_q8_0x4 out;
+
+    for (int i = 0; i < 4; i++) {
+        out.d[i] = in[i].d;
+    }
+
+    const int end = QK8_0 * 4 / blck_size_interleave;
+    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], blck_size_interleave);
+    }
+    return out;
+}
+
 static block_q4_0x4 make_block_q4_0x4(block_q4_0 * in, unsigned int blck_size_interleave) {
     block_q4_0x4 out;
 
@@ -1228,9 +1655,10 @@ static int repack_q4_0_to_q4_0_4_bl(struct ggml_tensor * t, int interleave_block
 
     GGML_UNUSED(data_size);
 }
+
 static int repack_q4_K_to_q4_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_Q4_K);
-    GGML_ASSERT(interleave_block == 8);
+    GGML_ASSERT(interleave_block == 8 || interleave_block == 4);
     constexpr int nrows_interleaved = 8;
 
     block_q4_Kx8 * dst = (block_q4_Kx8*)t->data;
@@ -1321,6 +1749,38 @@ static int repack_q4_0_to_q4_0_8_bl(struct ggml_tensor * t, int interleave_block
     GGML_UNUSED(data_size);
 }
 
+static int repack_q8_0_to_q8_0_4_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);
+    GGML_ASSERT(interleave_block == 4 || interleave_block == 8);
+    constexpr int nrows_interleaved = 4;
+
+    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_iq4_nlx4 make_block_iq4_nlx4(block_iq4_nl * in, unsigned int blck_size_interleave) {
     block_iq4_nlx4 out;
 
@@ -1468,6 +1928,10 @@ template <> int repack<block_q4_K, 8, 8>(struct ggml_tensor * t, const void * da
     return repack_q4_K_to_q4_K_8_bl(t, 8, data, data_size);
 }
 
+template <> int repack<block_q4_K, 4, 8>(struct ggml_tensor * t, const void * data, size_t data_size) {
+    return repack_q4_K_to_q4_K_8_bl(t, 4, data, data_size);
+}
+
 template <> int repack<block_q2_K, 8, 8>(struct ggml_tensor * t, const void * data, size_t data_size) {
     return repack_q2_K_to_q2_K_8_bl(t, 8, data, data_size);
 }
@@ -1485,6 +1949,14 @@ 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_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);
+}
+
+template <> int repack<block_q8_0, 8, 4>(struct ggml_tensor * t, const void * data, size_t data_size) {
+    return repack_q8_0_to_q8_0_4_bl(t, 8, data, data_size);
+}
+
 // 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);
@@ -1501,6 +1973,10 @@ 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_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);
+}
+
 template <> void gemv<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_gemv_q4_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
 }
@@ -1517,6 +1993,14 @@ 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_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);
+}
+
+template <> void gemv<block_q8_0, 8, 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_4x8_q8_0(n, s, bs, vx, vy, nr, nc);
+}
+
 // 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);
@@ -1529,6 +2013,10 @@ 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);
 }
@@ -1549,6 +2037,14 @@ 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_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);
+}
+
+template <> void gemm<block_q8_0, 8, 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_4x8_q8_0(n, s, bs, vx, vy, nr, nc);
+}
+
 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;
@@ -1600,6 +2096,55 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PAR
         return false;
     }
 
+    void forward_mul_mat_one_chunk(ggml_compute_params * params,
+                                   ggml_tensor *         op,
+                                   int64_t               src0_start,
+                                   int64_t               src0_end,
+                                   int64_t               src1_start,
+                                   int64_t               src1_end) {
+        const ggml_tensor * src0 = op->src[0];
+        const ggml_tensor * src1 = op->src[1];
+        ggml_tensor *       dst  = op;
+
+        GGML_TENSOR_BINARY_OP_LOCALS
+
+        const size_t src1_col_stride = ggml_row_size(PARAM_TYPE, ne10);
+
+        GGML_ASSERT(ne03 == 1 && ne13 == 1);
+        GGML_ASSERT(ne12 % ne02 == 0);
+        const int64_t r2 = ne12 / ne02;
+
+        const int64_t i12 = src1_start / ne1;
+        const int64_t i11 = src1_start - i12 * ne1;
+
+        // Determine batch index
+        const int64_t i02 = i12 / r2;
+
+        const int64_t i1 = i11;
+        const int64_t i2 = i12;
+
+        const char * src0_ptr = (const char *) src0->data + i02 * nb02;
+        const char * src1_ptr = (const char *) params->wdata + (i11 + i12 * ne11) * src1_col_stride;
+        char *       dst_ptr  = ((char *) dst->data + (i1 * nb1 + i2 * nb2));
+
+        const int64_t nrows = src1_end - src1_start;
+        const int64_t ncols = src0_end - src0_start;
+
+        GGML_ASSERT(src1_ptr + src1_col_stride * nrows <= (const char *) params->wdata + params->wsize);
+
+        // If there are more than three rows in src1, use gemm; otherwise, use gemv.
+        if (nrows > 3) {
+            gemm<BLOC_TYPE, INTER_SIZE, NB_COLS, PARAM_TYPE>(ne00, (float *) (dst_ptr) + src0_start, nb1 / nb0,
+                                                             src0_ptr + src0_start * nb01, src1_ptr,
+                                                             nrows - (nrows % 4), ncols);
+        }
+        for (int iter = nrows - (nrows % 4); iter < nrows; iter++) {
+            gemv<BLOC_TYPE, INTER_SIZE, NB_COLS, PARAM_TYPE>(ne00, (float *) (dst_ptr + (iter * nb1)) + src0_start,
+                                                             ne01, src0_ptr + src0_start * nb01,
+                                                             src1_ptr + (src1_col_stride * iter), 1 /* nrows */, ncols);
+        }
+    }
+
     void forward_mul_mat(ggml_compute_params * params, ggml_tensor * op) {
         const ggml_tensor * src0 = op->src[0];
         const ggml_tensor * src1 = op->src[1];
@@ -1621,6 +2166,12 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PAR
         GGML_ASSERT(nb1 <= nb2);
         GGML_ASSERT(nb2 <= nb3);
 
+        // TODO: General batched mul mat for 4D tensors
+        // Currently only supports 3D tensors
+        GGML_ASSERT(ne03 == 1);
+        GGML_ASSERT(ne13 == 1);
+        GGML_ASSERT(ne3 == 1);
+
         GGML_ASSERT(src1->type == GGML_TYPE_F32);
 
         GGML_ASSERT(ggml_n_dims(op->src[0]) == 2);
@@ -1628,46 +2179,102 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PAR
 
         char *       wdata = static_cast<char *>(params->wdata);
         const size_t nbw1  = ggml_row_size(PARAM_TYPE, ne10);
+        const size_t nbw2  = nbw1 * ne11;
 
-        assert(params->wsize >= nbw1 * ne11);
+        assert(params->wsize >= nbw2 * ne12);
 
         const ggml_from_float_t from_float = ggml_get_type_traits_cpu(PARAM_TYPE)->from_float;
 
-        int64_t i11_processed = 0;
-        for (int64_t i11 = ith * 4; i11 < ne11 - ne11 % 4; i11 += nth * 4) {
-            ggml_quantize_mat_t<INTER_SIZE, PARAM_TYPE>((float *) ((char *) src1->data + i11 * nb11), (void *) (wdata + i11 * nbw1), 4, ne10);
-        }
+        // INFO: Quantization is done in planes to avoid extra complexity in chunking.
+        // Flattening dimensions not multiple of INTER_SIZE would require extra handling depending on how
+        // the planes are broadcast.
+        for (int64_t i12 = 0; i12 < ne12; i12++) {
+            char * data_ptr  = (char *) src1->data + i12 * nb12;
+            char * wdata_ptr = wdata + i12 * nbw2;
 
-        i11_processed = ne11 - ne11 % 4;
-        for (int64_t i11 = i11_processed + ith; i11 < ne11; i11 += nth) {
-            from_float((float *) ((char *) src1->data + i11 * nb11), (void *) (wdata + i11 * nbw1), ne10);
+            for (int64_t i11 = ith * 4; i11 < ne11 - ne11 % 4; i11 += nth * 4) {
+                ggml_quantize_mat_t<INTER_SIZE, PARAM_TYPE>((float *) (data_ptr + i11 * nb11),
+                                                            (void *) (wdata_ptr + i11 * nbw1), 4, ne10);
+            }
+
+            const int64_t i11_processed = ne11 - ne11 % 4;
+            for (int64_t i11 = i11_processed + ith; i11 < ne11; i11 += nth) {
+                from_float((float *) (data_ptr + i11 * nb11), (void *) (wdata_ptr + i11 * nbw1), ne10);
+            }
         }
 
-        ggml_barrier(params->threadpool);
+        // disable for NUMA
+        const bool disable_chunking = ggml_is_numa();
 
-        const void * src1_wdata      = params->wdata;
-        const size_t src1_col_stride = ggml_row_size(PARAM_TYPE, ne10);
-        int64_t      src0_start      = (ith * ne01) / nth;
-        int64_t      src0_end        = ((ith + 1) * ne01) / nth;
-        src0_start = (src0_start % NB_COLS) ? src0_start + NB_COLS - (src0_start % NB_COLS) : src0_start;
-        src0_end   = (src0_end   % NB_COLS) ? src0_end   + NB_COLS - (src0_end   % NB_COLS) : src0_end;
-        if (src0_start >= src0_end) {
-            return;
+        // 4x chunks per thread
+        const int64_t nr0 = ggml_nrows(op->src[0]);
+
+        int     nth_scaled  = nth * 4;
+        int64_t chunk_size0 = (nr0 + nth_scaled - 1) / nth_scaled;
+        int64_t nchunk0     = (nr0 + chunk_size0 - 1) / chunk_size0;
+
+        // src1 is chunked only by full planes.
+        // When we flatten we need to address dimensions not multiple of the q8 INTER_SIZE
+        // to route them thorugh GEMV.
+        // nchunk1 = ne12 also avoids messing the chunking for models with no 3d tensors
+        // to avoid affecting their performance
+        int64_t nchunk1 = ne12;
+
+        // Ensure minimum chunk size to avoid alignment issues with high thread counts
+        // Minimum chunk size should be at least NB_COLS to prevent overlapping chunks after alignment
+        const int64_t min_chunk_size = NB_COLS;
+        if (nchunk0 > 0 && (nr0 / nchunk0) < min_chunk_size && nr0 >= min_chunk_size) {
+            nchunk0 = (nr0 + min_chunk_size - 1) / min_chunk_size;
         }
 
-        // If there are more than three rows in src1, use gemm; otherwise, use gemv.
-        if (ne11 > 3) {
-            gemm<BLOC_TYPE, INTER_SIZE, NB_COLS, PARAM_TYPE>(ne00,
-                    (float *) ((char *) dst->data) + src0_start, ne01,
-                    (const char *) src0->data + src0_start * nb01,
-                    (const char *) src1_wdata, ne11 - ne11 % 4, src0_end - src0_start);
+        int64_t dr0 = (nr0 + nchunk0 - 1) / nchunk0;
+        // Only increase nchunk0 to nth if it won't make chunks too small
+        if (nth == 1 || ((nchunk0 < nth || disable_chunking) && (nr0 + nth - 1) / nth >= min_chunk_size)) {
+            nchunk0 = nth;
+            dr0 = (nr0 + nchunk0 - 1) / nchunk0;
         }
-        for (int iter = ne11 - ne11 % 4; iter < ne11; iter++) {
-            gemv<BLOC_TYPE, INTER_SIZE, NB_COLS, PARAM_TYPE>(ne00,
-                    (float *) ((char *) dst->data + (iter * nb1)) + src0_start, ne01,
-                    (const char *) src0->data + src0_start * nb01,
-                    (const char *) src1_wdata + (src1_col_stride * iter), 1,
-                    src0_end - src0_start);
+
+        // Ensure nchunk doesn't exceed the number of rows divided by minimum chunk size
+        // This prevents creating too many tiny chunks that could overlap after alignment
+        const int64_t max_nchunk = (nr0 + min_chunk_size - 1) / min_chunk_size;
+        nchunk0                  = MIN(nchunk0, max_nchunk);
+
+        if (ith == 0) {
+            // Every thread starts at ith, so the first unprocessed chunk is nth.  This save a bit of coordination right at the start.
+            ggml_threadpool_chunk_set(params->threadpool, nth);
+        }
+
+        ggml_barrier(params->threadpool);
+
+        // The first chunk comes from our thread_id, the rest will get auto-assigned.
+        int current_chunk = ith;
+
+        while (current_chunk < nchunk0 * nchunk1) {
+            const int64_t ith0 = current_chunk % nchunk0;
+            const int64_t ith1 = current_chunk / nchunk0;
+
+            int64_t src0_start = dr0 * ith0;
+            int64_t src0_end   = MIN(src0_start + dr0, nr0);
+
+            // full-plane range for src1
+            int64_t src1_start = ith1 * ne11;
+            int64_t src1_end = (ith1 + 1) * ne11;
+
+            // Align boundaries to NB_COLS - round up to ensure all data is included
+            // The chunk size limiting above ensures chunks are large enough to prevent overlaps
+            src0_start = (src0_start % NB_COLS) ? src0_start + NB_COLS - (src0_start % NB_COLS) : src0_start;
+            src0_end   = (src0_end % NB_COLS) ? src0_end + NB_COLS - (src0_end % NB_COLS) : src0_end;
+            src0_end   = MIN(src0_end, ne01);
+
+            // Make sure current plane is the last one before exiting
+            if (src0_start >= src0_end) {
+                current_chunk = ggml_threadpool_chunk_add(params->threadpool, 1);
+                continue;
+            }
+
+            forward_mul_mat_one_chunk(params, dst, src0_start, src0_end, src1_start, src1_end);
+
+            current_chunk = ggml_threadpool_chunk_add(params->threadpool, 1);
         }
     }
 
@@ -1772,8 +2379,12 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PAR
             int64_t src0_cur_start = (ith * ne01) / nth;
             int64_t src0_cur_end   = ((ith + 1) * ne01) / nth;
 
+            // Align boundaries to NB_COLS - round up to ensure all data is included
             src0_cur_start = (src0_cur_start % NB_COLS) ? src0_cur_start + NB_COLS - (src0_cur_start % NB_COLS) : src0_cur_start;
             src0_cur_end   = (src0_cur_end   % NB_COLS) ? src0_cur_end   + NB_COLS - (src0_cur_end   % NB_COLS) : src0_cur_end;
+            if (src0_cur_end > ne01) {
+                src0_cur_end = ne01;
+            }
 
             if (src0_cur_start >= src0_cur_end) {
                 return;
@@ -1816,6 +2427,9 @@ static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(cons
     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;
     static const ggml::cpu::repack::tensor_traits<block_q4_0, 8, 8, GGML_TYPE_Q8_0> q4_0_8x8_q8_0;
+
+    // instance for Q4_K
+    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 Q2
@@ -1825,8 +2439,13 @@ 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 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;
+
     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)) {
+        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 (cur->ne[1] % 8 == 0) {
                 return &q4_0_8x8_q8_0;
             }
@@ -1847,6 +2466,16 @@ static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(cons
                 return &q4_K_8x8_q8_K;
             }
         }
+        if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) {
+            if (cur->ne[1] % 8 == 0) {
+                return &q4_K_8x8_q8_K;
+            }
+        }
+        if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
+            if (cur->ne[1] % 8 == 0) {
+                return &q4_K_8x4_q8_K;
+            }
+        }
     } else if (cur->type == GGML_TYPE_Q2_K) {
         if (ggml_cpu_has_avx512()) {
             if (cur->ne[1] % 8 == 0) {
@@ -1864,6 +2493,17 @@ static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(cons
                 return &iq4_nl_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) {
+                return &q8_0_4x8_q8_0;
+            }
+        }
+        if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
+            if (cur->ne[1] % 4 == 0) {
+                return &q8_0_4x4_q8_0;
+            }
+        }
     }
 
     return nullptr;