llama_cpp 0.12.6 → 0.13.0

data/vendor/tmp/llama.cpp/ggml-metal.metal

@@ -351,12 +351,17 @@ kernel void kernel_sum_rows(
 kernel void kernel_soft_max(
         device const float * src0,
         device const float * src1,
+        device const float * src2,
         device       float * dst,
         constant   int64_t & ne00,
         constant   int64_t & ne01,
         constant   int64_t & ne02,
         constant     float & scale,
-        threadgroup float  * buf [[threadgroup(0)]],
+        constant     float & max_bias,
+        constant     float & m0,
+        constant     float & m1,
+        constant  uint32_t & n_head_log2,
+        threadgroup  float * buf [[threadgroup(0)]],
         uint  tgpig[[threadgroup_position_in_grid]],
         uint  tpitg[[thread_position_in_threadgroup]],
         uint  sgitg[[simdgroup_index_in_threadgroup]],
@@ -368,13 +373,26 @@ kernel void kernel_soft_max(
 
     device const float * psrc0 =         src0 + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
     device const float * pmask = src1 != src0 ? src1                               + i01*ne00 : nullptr;
+    device const float * ppos  = src2 != src0 ? src2                                          : nullptr;
     device       float * pdst  =         dst  + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
 
+    float slope = 0.0f;
+
+    // ALiBi
+    if (max_bias > 0.0f) {
+        const int64_t h = i02;
+
+        const float base = h < n_head_log2 ? m0 : m1;
+        const int   exp  = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
+
+        slope = pow(base, exp);
+    }
+
     // parallel max
     float lmax = -INFINITY;
 
     for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
-        lmax = MAX(lmax, psrc0[i00]*scale + (pmask ? pmask[i00] : 0.0f));
+        lmax = MAX(lmax, psrc0[i00]*scale + (pmask ? pmask[i00] : 0.0f) + (ppos ? slope*ppos[i00] : 0.0f));
     }
 
     // find the max value in the block
@@ -399,7 +417,7 @@ kernel void kernel_soft_max(
     // parallel sum
     float lsum = 0.0f;
     for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
-        const float exp_psrc0 = exp((psrc0[i00]*scale + (pmask ? pmask[i00] : 0.0f)) - max_val);
+        const float exp_psrc0 = exp((psrc0[i00]*scale + (pmask ? pmask[i00] : 0.0f) + (ppos ? slope*ppos[i00] : 0.0f)) - max_val);
         lsum += exp_psrc0;
         pdst[i00] = exp_psrc0;
     }
@@ -437,12 +455,17 @@ kernel void kernel_soft_max(
 kernel void kernel_soft_max_4(
         device const float * src0,
         device const float * src1,
+        device const float * src2,
         device       float * dst,
         constant   int64_t & ne00,
         constant   int64_t & ne01,
         constant   int64_t & ne02,
         constant     float & scale,
-        threadgroup float  * buf [[threadgroup(0)]],
+        constant     float & max_bias,
+        constant     float & m0,
+        constant     float & m1,
+        constant  uint32_t & n_head_log2,
+        threadgroup  float * buf [[threadgroup(0)]],
         uint  tgpig[[threadgroup_position_in_grid]],
         uint  tpitg[[thread_position_in_threadgroup]],
         uint  sgitg[[simdgroup_index_in_threadgroup]],
@@ -454,13 +477,25 @@ kernel void kernel_soft_max_4(
 
     device const float4 * psrc4 =                (device const float4 *)(src0 + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00);
     device const float4 * pmask = src1 != src0 ? (device const float4 *)(src1 +                                      i01*ne00) : nullptr;
+    device const float4 * ppos  = src2 != src0 ? (device const float4 *)(src2)                                                 : nullptr;
     device       float4 * pdst4 =                (device       float4 *)(dst  + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00);
 
+    float slope = 0.0f;
+
+    if (max_bias > 0.0f) {
+        const int64_t h = i02;
+
+        const float base = h < n_head_log2 ? m0 : m1;
+        const int   exp  = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
+
+        slope = pow(base, exp);
+    }
+
     // parallel max
     float4 lmax4 = -INFINITY;
 
     for (int i00 = tpitg; i00 < ne00/4; i00 += ntg) {
-        lmax4 = fmax(lmax4, psrc4[i00]*scale + (pmask ? pmask[i00] : 0.0f));
+        lmax4 = fmax(lmax4, psrc4[i00]*scale + (pmask ? pmask[i00] : 0.0f) + (ppos ? slope*ppos[i00] : 0.0f));
     }
 
     const float lmax = MAX(MAX(lmax4[0], lmax4[1]), MAX(lmax4[2], lmax4[3]));
@@ -486,7 +521,7 @@ kernel void kernel_soft_max_4(
     // parallel sum
     float4 lsum4 = 0.0f;
     for (int i00 = tpitg; i00 < ne00/4; i00 += ntg) {
-        const float4 exp_psrc4 = exp((psrc4[i00]*scale + (pmask ? pmask[i00] : 0.0f)) - max_val);
+        const float4 exp_psrc4 = exp((psrc4[i00]*scale + (pmask ? pmask[i00] : 0.0f) + (ppos ? slope*ppos[i00] : 0.0f)) - max_val);
         lsum4 += exp_psrc4;
         pdst4[i00] = exp_psrc4;
     }
@@ -2484,12 +2519,58 @@ typedef struct {
 } block_iq2_xs;
 // 74 bytes / block for QK_K = 256, so 2.3125 bpw
 
+// 2.5625 bpw quants
+typedef struct {
+    half d;
+    uint8_t qs[QK_K/4];
+    uint8_t qh[QK_K/32];
+    uint8_t scales[QK_K/32];
+} block_iq2_s;
+
 typedef struct {
     half d;
     uint8_t qs[3*QK_K/8];
 } block_iq3_xxs;
 // 98 bytes / block for QK_K = 256, so 3.0625 bpw
 
+// 3.4375 bpw
+#if QK_K == 64
+#define IQ3S_N_SCALE 2
+#else
+#define IQ3S_N_SCALE QK_K/64
+#endif
+typedef struct {
+    half d;
+    uint8_t qs[QK_K/4];
+    uint8_t qh[QK_K/32];
+    uint8_t signs[QK_K/8];
+    uint8_t scales[IQ3S_N_SCALE];
+} block_iq3_s;
+
+typedef struct {
+    half d;
+    uint8_t qs[QK_K/8];
+    uint8_t scales[QK_K/16];
+} block_iq1_s;
+
+// Non-linear quants
+#define QK4_NL 32
+typedef struct {
+    half    d;
+    uint8_t qs[QK4_NL/2];
+} block_iq4_nl;
+
+#if QK_K == 64
+#define block_iq4_xs block_iq4_nl
+#else
+typedef struct {
+    half d;
+    uint16_t scales_h;
+    uint8_t  scales_l[QK_K/64];
+    uint8_t  qs[QK_K/2];
+} block_iq4_xs;
+#endif
+
 //====================================== dot products =========================
 
 void kernel_mul_mv_q2_K_f32_impl(
@@ -3712,6 +3793,265 @@ constexpr constant static uint64_t iq2xs_grid[512] = {
     0x2b2b2b2b082b2b08, 0x2b2b2b2b082b2b2b, 0x2b2b2b2b2b190819, 0x2b2b2b2b2b2b2b2b,
 };
 
+constexpr constant static uint64_t iq2s_grid[1024] = {
+    0x0808080808080808, 0x080808080808082b, 0x0808080808081919, 0x0808080808082b08,
+    0x0808080808082b2b, 0x0808080808190819, 0x0808080808191908, 0x080808080819192b,
+    0x0808080808192b19, 0x08080808082b0808, 0x08080808082b082b, 0x08080808082b1919,
+    0x08080808082b2b08, 0x0808080819080819, 0x0808080819081908, 0x080808081908192b,
+    0x0808080819082b19, 0x0808080819190808, 0x080808081919082b, 0x0808080819191919,
+    0x0808080819192b08, 0x08080808192b0819, 0x08080808192b1908, 0x08080808192b192b,
+    0x08080808192b2b19, 0x080808082b080808, 0x080808082b08082b, 0x080808082b081919,
+    0x080808082b082b08, 0x080808082b190819, 0x080808082b191908, 0x080808082b2b0808,
+    0x080808082b2b1919, 0x080808082b2b2b2b, 0x0808081908080819, 0x0808081908081908,
+    0x080808190808192b, 0x0808081908082b19, 0x0808081908190808, 0x080808190819082b,
+    0x0808081908191919, 0x0808081908192b08, 0x08080819082b0819, 0x08080819082b1908,
+    0x0808081919080808, 0x080808191908082b, 0x0808081919081919, 0x0808081919082b08,
+    0x0808081919190819, 0x0808081919191908, 0x080808191919192b, 0x0808081919192b19,
+    0x08080819192b0808, 0x08080819192b1919, 0x08080819192b2b08, 0x080808192b080819,
+    0x080808192b081908, 0x080808192b190808, 0x080808192b19082b, 0x080808192b191919,
+    0x080808192b2b0819, 0x080808192b2b1908, 0x0808082b08080808, 0x0808082b0808082b,
+    0x0808082b08081919, 0x0808082b08082b08, 0x0808082b08190819, 0x0808082b08191908,
+    0x0808082b082b0808, 0x0808082b082b2b2b, 0x0808082b19080819, 0x0808082b19081908,
+    0x0808082b1908192b, 0x0808082b19082b19, 0x0808082b19190808, 0x0808082b19191919,
+    0x0808082b2b080808, 0x0808082b2b081919, 0x0808082b2b082b2b, 0x0808082b2b191908,
+    0x0808082b2b2b082b, 0x0808190808080819, 0x0808190808081908, 0x080819080808192b,
+    0x0808190808082b19, 0x0808190808190808, 0x080819080819082b, 0x0808190808191919,
+    0x0808190808192b08, 0x08081908082b0819, 0x08081908082b1908, 0x08081908082b192b,
+    0x08081908082b2b19, 0x0808190819080808, 0x080819081908082b, 0x0808190819081919,
+    0x0808190819082b08, 0x0808190819082b2b, 0x0808190819190819, 0x0808190819191908,
+    0x080819081919192b, 0x0808190819192b19, 0x08081908192b0808, 0x08081908192b082b,
+    0x08081908192b1919, 0x080819082b080819, 0x080819082b081908, 0x080819082b08192b,
+    0x080819082b082b19, 0x080819082b190808, 0x080819082b191919, 0x080819082b192b08,
+    0x080819082b2b0819, 0x080819082b2b1908, 0x0808191908080808, 0x080819190808082b,
+    0x0808191908081919, 0x0808191908082b08, 0x0808191908082b2b, 0x0808191908190819,
+    0x0808191908191908, 0x080819190819192b, 0x0808191908192b19, 0x08081919082b0808,
+    0x08081919082b1919, 0x08081919082b2b08, 0x0808191919080819, 0x0808191919081908,
+    0x080819191908192b, 0x0808191919082b19, 0x0808191919190808, 0x080819191919082b,
+    0x0808191919191919, 0x0808191919192b08, 0x08081919192b0819, 0x08081919192b1908,
+    0x080819192b080808, 0x080819192b08082b, 0x080819192b081919, 0x080819192b082b08,
+    0x080819192b190819, 0x080819192b191908, 0x080819192b2b0808, 0x0808192b08080819,
+    0x0808192b08081908, 0x0808192b0808192b, 0x0808192b08082b19, 0x0808192b08190808,
+    0x0808192b08191919, 0x0808192b19080808, 0x0808192b19081919, 0x0808192b19082b08,
+    0x0808192b19190819, 0x0808192b19191908, 0x0808192b192b0808, 0x0808192b2b080819,
+    0x0808192b2b081908, 0x0808192b2b190808, 0x08082b0808080808, 0x08082b080808082b,
+    0x08082b0808081919, 0x08082b0808082b08, 0x08082b0808190819, 0x08082b0808191908,
+    0x08082b080819192b, 0x08082b0808192b19, 0x08082b08082b0808, 0x08082b08082b1919,
+    0x08082b08082b2b2b, 0x08082b0819080819, 0x08082b0819081908, 0x08082b081908192b,
+    0x08082b0819082b19, 0x08082b0819190808, 0x08082b081919082b, 0x08082b0819191919,
+    0x08082b0819192b08, 0x08082b08192b0819, 0x08082b08192b1908, 0x08082b082b080808,
+    0x08082b082b081919, 0x08082b082b191908, 0x08082b082b2b2b2b, 0x08082b1908080819,
+    0x08082b1908081908, 0x08082b1908190808, 0x08082b190819082b, 0x08082b1908191919,
+    0x08082b1908192b08, 0x08082b19082b0819, 0x08082b1919080808, 0x08082b1919081919,
+    0x08082b1919082b08, 0x08082b1919190819, 0x08082b1919191908, 0x08082b19192b0808,
+    0x08082b192b080819, 0x08082b192b190808, 0x08082b2b08080808, 0x08082b2b08190819,
+    0x08082b2b08191908, 0x08082b2b082b082b, 0x08082b2b082b2b08, 0x08082b2b082b2b2b,
+    0x08082b2b19190808, 0x08082b2b2b192b19, 0x0819080808080819, 0x0819080808081908,
+    0x081908080808192b, 0x0819080808082b19, 0x0819080808190808, 0x081908080819082b,
+    0x0819080808191919, 0x0819080808192b08, 0x08190808082b0819, 0x08190808082b1908,
+    0x08190808082b192b, 0x0819080819080808, 0x081908081908082b, 0x0819080819081919,
+    0x0819080819082b08, 0x0819080819190819, 0x0819080819191908, 0x081908081919192b,
+    0x0819080819192b19, 0x08190808192b0808, 0x08190808192b082b, 0x08190808192b1919,
+    0x08190808192b2b08, 0x081908082b080819, 0x081908082b081908, 0x081908082b08192b,
+    0x081908082b190808, 0x081908082b191919, 0x081908082b192b08, 0x081908082b2b0819,
+    0x081908082b2b1908, 0x0819081908080808, 0x081908190808082b, 0x0819081908081919,
+    0x0819081908082b08, 0x0819081908082b2b, 0x0819081908190819, 0x0819081908191908,
+    0x081908190819192b, 0x0819081908192b19, 0x08190819082b0808, 0x08190819082b082b,
+    0x08190819082b1919, 0x08190819082b2b08, 0x0819081919080819, 0x0819081919081908,
+    0x081908191908192b, 0x0819081919082b19, 0x0819081919190808, 0x081908191919082b,
+    0x0819081919191919, 0x0819081919192b08, 0x08190819192b0819, 0x08190819192b1908,
+    0x081908192b080808, 0x081908192b08082b, 0x081908192b081919, 0x081908192b082b08,
+    0x081908192b190819, 0x081908192b191908, 0x0819082b08080819, 0x0819082b08081908,
+    0x0819082b08082b19, 0x0819082b08190808, 0x0819082b08191919, 0x0819082b082b0819,
+    0x0819082b082b1908, 0x0819082b19080808, 0x0819082b19081919, 0x0819082b19190819,
+    0x0819082b19191908, 0x0819082b2b080819, 0x0819082b2b081908, 0x0819082b2b190808,
+    0x0819190808080808, 0x081919080808082b, 0x0819190808081919, 0x0819190808082b08,
+    0x0819190808190819, 0x0819190808191908, 0x081919080819192b, 0x0819190808192b19,
+    0x08191908082b0808, 0x08191908082b1919, 0x08191908082b2b08, 0x0819190819080819,
+    0x0819190819081908, 0x081919081908192b, 0x0819190819082b19, 0x0819190819190808,
+    0x081919081919082b, 0x0819190819191919, 0x0819190819192b08, 0x08191908192b0819,
+    0x08191908192b1908, 0x081919082b080808, 0x081919082b08082b, 0x081919082b081919,
+    0x081919082b082b08, 0x081919082b190819, 0x081919082b191908, 0x081919082b2b0808,
+    0x0819191908080819, 0x0819191908081908, 0x081919190808192b, 0x0819191908082b19,
+    0x0819191908190808, 0x081919190819082b, 0x0819191908191919, 0x0819191908192b08,
+    0x08191919082b0819, 0x08191919082b1908, 0x0819191919080808, 0x081919191908082b,
+    0x0819191919081919, 0x0819191919082b08, 0x0819191919190819, 0x0819191919191908,
+    0x08191919192b0808, 0x081919192b080819, 0x081919192b081908, 0x081919192b190808,
+    0x0819192b08080808, 0x0819192b08081919, 0x0819192b08082b08, 0x0819192b08190819,
+    0x0819192b08191908, 0x0819192b082b0808, 0x0819192b19080819, 0x0819192b19081908,
+    0x0819192b19190808, 0x0819192b2b080808, 0x0819192b2b2b2b2b, 0x08192b0808080819,
+    0x08192b0808081908, 0x08192b080808192b, 0x08192b0808082b19, 0x08192b0808190808,
+    0x08192b0808191919, 0x08192b0808192b08, 0x08192b08082b0819, 0x08192b0819080808,
+    0x08192b081908082b, 0x08192b0819081919, 0x08192b0819082b08, 0x08192b0819190819,
+    0x08192b0819191908, 0x08192b08192b0808, 0x08192b082b080819, 0x08192b082b081908,
+    0x08192b1908080808, 0x08192b190808082b, 0x08192b1908081919, 0x08192b1908082b08,
+    0x08192b1908190819, 0x08192b1908191908, 0x08192b19082b0808, 0x08192b1919080819,
+    0x08192b1919081908, 0x08192b1919190808, 0x08192b19192b2b19, 0x08192b192b2b082b,
+    0x08192b2b08081908, 0x08192b2b08190808, 0x08192b2b19080808, 0x08192b2b1919192b,
+    0x082b080808080808, 0x082b08080808082b, 0x082b080808081919, 0x082b080808082b08,
+    0x082b080808190819, 0x082b080808191908, 0x082b08080819192b, 0x082b080808192b19,
+    0x082b0808082b0808, 0x082b0808082b1919, 0x082b0808082b2b2b, 0x082b080819080819,
+    0x082b080819081908, 0x082b080819190808, 0x082b08081919082b, 0x082b080819191919,
+    0x082b0808192b1908, 0x082b08082b080808, 0x082b08082b082b2b, 0x082b08082b191908,
+    0x082b08082b2b2b2b, 0x082b081908080819, 0x082b081908081908, 0x082b081908190808,
+    0x082b08190819082b, 0x082b081908191919, 0x082b0819082b0819, 0x082b081919080808,
+    0x082b08191908082b, 0x082b081919081919, 0x082b081919190819, 0x082b081919191908,
+    0x082b0819192b0808, 0x082b08192b080819, 0x082b08192b081908, 0x082b08192b190808,
+    0x082b082b08080808, 0x082b082b08082b2b, 0x082b082b082b082b, 0x082b082b082b2b08,
+    0x082b082b082b2b2b, 0x082b082b19081908, 0x082b082b19190808, 0x082b082b2b082b08,
+    0x082b082b2b082b2b, 0x082b082b2b2b2b08, 0x082b190808080819, 0x082b190808081908,
+    0x082b19080808192b, 0x082b190808082b19, 0x082b190808190808, 0x082b190808191919,
+    0x082b190808192b08, 0x082b1908082b0819, 0x082b1908082b1908, 0x082b190819080808,
+    0x082b19081908082b, 0x082b190819081919, 0x082b190819082b08, 0x082b190819190819,
+    0x082b190819191908, 0x082b1908192b0808, 0x082b19082b080819, 0x082b19082b081908,
+    0x082b19082b190808, 0x082b191908080808, 0x082b191908081919, 0x082b191908082b08,
+    0x082b191908190819, 0x082b191908191908, 0x082b1919082b0808, 0x082b191919080819,
+    0x082b191919081908, 0x082b191919190808, 0x082b1919192b192b, 0x082b19192b080808,
+    0x082b192b08080819, 0x082b192b08081908, 0x082b192b08190808, 0x082b192b19080808,
+    0x082b192b19192b19, 0x082b2b0808080808, 0x082b2b0808081919, 0x082b2b0808190819,
+    0x082b2b0808191908, 0x082b2b0819080819, 0x082b2b0819081908, 0x082b2b0819190808,
+    0x082b2b082b082b2b, 0x082b2b082b2b2b2b, 0x082b2b1908080819, 0x082b2b1908081908,
+    0x082b2b1908190808, 0x082b2b192b191919, 0x082b2b2b08082b2b, 0x082b2b2b082b082b,
+    0x082b2b2b192b1908, 0x082b2b2b2b082b08, 0x082b2b2b2b082b2b, 0x1908080808080819,
+    0x1908080808081908, 0x190808080808192b, 0x1908080808082b19, 0x1908080808190808,
+    0x190808080819082b, 0x1908080808191919, 0x1908080808192b08, 0x1908080808192b2b,
+    0x19080808082b0819, 0x19080808082b1908, 0x19080808082b192b, 0x1908080819080808,
+    0x190808081908082b, 0x1908080819081919, 0x1908080819082b08, 0x1908080819082b2b,
+    0x1908080819190819, 0x1908080819191908, 0x190808081919192b, 0x1908080819192b19,
+    0x19080808192b0808, 0x19080808192b082b, 0x19080808192b1919, 0x190808082b080819,
+    0x190808082b081908, 0x190808082b190808, 0x190808082b191919, 0x190808082b192b08,
+    0x190808082b2b0819, 0x190808082b2b1908, 0x1908081908080808, 0x190808190808082b,
+    0x1908081908081919, 0x1908081908082b08, 0x1908081908190819, 0x1908081908191908,
+    0x190808190819192b, 0x1908081908192b19, 0x19080819082b0808, 0x19080819082b082b,
+    0x19080819082b1919, 0x1908081919080819, 0x1908081919081908, 0x190808191908192b,
+    0x1908081919082b19, 0x1908081919190808, 0x190808191919082b, 0x1908081919191919,
+    0x1908081919192b08, 0x19080819192b0819, 0x19080819192b1908, 0x190808192b080808,
+    0x190808192b08082b, 0x190808192b081919, 0x190808192b082b08, 0x190808192b190819,
+    0x190808192b191908, 0x190808192b2b0808, 0x1908082b08080819, 0x1908082b08081908,
+    0x1908082b08190808, 0x1908082b0819082b, 0x1908082b08191919, 0x1908082b08192b08,
+    0x1908082b082b1908, 0x1908082b19080808, 0x1908082b19081919, 0x1908082b19082b08,
+    0x1908082b19190819, 0x1908082b19191908, 0x1908082b192b0808, 0x1908082b2b080819,
+    0x1908082b2b081908, 0x1908190808080808, 0x190819080808082b, 0x1908190808081919,
+    0x1908190808082b08, 0x1908190808082b2b, 0x1908190808190819, 0x1908190808191908,
+    0x190819080819192b, 0x1908190808192b19, 0x19081908082b0808, 0x19081908082b082b,
+    0x19081908082b1919, 0x19081908082b2b08, 0x1908190819080819, 0x1908190819081908,
+    0x190819081908192b, 0x1908190819082b19, 0x1908190819190808, 0x190819081919082b,
+    0x1908190819191919, 0x1908190819192b08, 0x19081908192b0819, 0x19081908192b1908,
+    0x190819082b080808, 0x190819082b08082b, 0x190819082b081919, 0x190819082b082b08,
+    0x190819082b190819, 0x190819082b191908, 0x190819082b2b0808, 0x1908191908080819,
+    0x1908191908081908, 0x190819190808192b, 0x1908191908082b19, 0x1908191908190808,
+    0x190819190819082b, 0x1908191908191919, 0x1908191908192b08, 0x19081919082b0819,
+    0x19081919082b1908, 0x1908191919080808, 0x190819191908082b, 0x1908191919081919,
+    0x1908191919082b08, 0x1908191919190819, 0x1908191919191908, 0x19081919192b0808,
+    0x19081919192b2b2b, 0x190819192b080819, 0x190819192b081908, 0x190819192b190808,
+    0x1908192b08080808, 0x1908192b0808082b, 0x1908192b08081919, 0x1908192b08082b08,
+    0x1908192b08190819, 0x1908192b08191908, 0x1908192b082b0808, 0x1908192b19080819,
+    0x1908192b19081908, 0x1908192b19190808, 0x1908192b2b080808, 0x1908192b2b2b1919,
+    0x19082b0808080819, 0x19082b0808081908, 0x19082b0808082b19, 0x19082b0808190808,
+    0x19082b080819082b, 0x19082b0808191919, 0x19082b0808192b08, 0x19082b08082b0819,
+    0x19082b08082b1908, 0x19082b0819080808, 0x19082b081908082b, 0x19082b0819081919,
+    0x19082b0819082b08, 0x19082b0819190819, 0x19082b0819191908, 0x19082b08192b0808,
+    0x19082b082b081908, 0x19082b082b190808, 0x19082b1908080808, 0x19082b190808082b,
+    0x19082b1908081919, 0x19082b1908082b08, 0x19082b1908190819, 0x19082b1908191908,
+    0x19082b19082b0808, 0x19082b1919080819, 0x19082b1919081908, 0x19082b1919190808,
+    0x19082b192b080808, 0x19082b192b19192b, 0x19082b2b08080819, 0x19082b2b08081908,
+    0x19082b2b08190808, 0x19082b2b19080808, 0x1919080808080808, 0x191908080808082b,
+    0x1919080808081919, 0x1919080808082b08, 0x1919080808190819, 0x1919080808191908,
+    0x191908080819192b, 0x1919080808192b19, 0x19190808082b0808, 0x19190808082b082b,
+    0x19190808082b1919, 0x19190808082b2b08, 0x1919080819080819, 0x1919080819081908,
+    0x191908081908192b, 0x1919080819082b19, 0x1919080819190808, 0x191908081919082b,
+    0x1919080819191919, 0x1919080819192b08, 0x19190808192b0819, 0x19190808192b1908,
+    0x191908082b080808, 0x191908082b08082b, 0x191908082b081919, 0x191908082b082b08,
+    0x191908082b190819, 0x191908082b191908, 0x1919081908080819, 0x1919081908081908,
+    0x191908190808192b, 0x1919081908082b19, 0x1919081908190808, 0x191908190819082b,
+    0x1919081908191919, 0x1919081908192b08, 0x19190819082b0819, 0x19190819082b1908,
+    0x1919081919080808, 0x191908191908082b, 0x1919081919081919, 0x1919081919082b08,
+    0x1919081919190819, 0x1919081919191908, 0x19190819192b0808, 0x191908192b080819,
+    0x191908192b081908, 0x191908192b190808, 0x1919082b08080808, 0x1919082b08081919,
+    0x1919082b08082b08, 0x1919082b08190819, 0x1919082b08191908, 0x1919082b082b0808,
+    0x1919082b19080819, 0x1919082b19081908, 0x1919082b19190808, 0x1919082b192b2b19,
+    0x1919082b2b080808, 0x1919190808080819, 0x1919190808081908, 0x191919080808192b,
+    0x1919190808082b19, 0x1919190808190808, 0x191919080819082b, 0x1919190808191919,
+    0x1919190808192b08, 0x19191908082b0819, 0x19191908082b1908, 0x1919190819080808,
+    0x191919081908082b, 0x1919190819081919, 0x1919190819082b08, 0x1919190819190819,
+    0x1919190819191908, 0x19191908192b0808, 0x191919082b080819, 0x191919082b081908,
+    0x191919082b190808, 0x1919191908080808, 0x191919190808082b, 0x1919191908081919,
+    0x1919191908082b08, 0x1919191908190819, 0x1919191908191908, 0x19191919082b0808,
+    0x1919191919080819, 0x1919191919081908, 0x1919191919190808, 0x191919192b080808,
+    0x1919192b08080819, 0x1919192b08081908, 0x1919192b08190808, 0x1919192b082b192b,
+    0x1919192b19080808, 0x19192b0808080808, 0x19192b080808082b, 0x19192b0808081919,
+    0x19192b0808082b08, 0x19192b0808190819, 0x19192b0808191908, 0x19192b08082b0808,
+    0x19192b0819080819, 0x19192b0819081908, 0x19192b0819190808, 0x19192b0819192b2b,
+    0x19192b082b080808, 0x19192b1908080819, 0x19192b1908081908, 0x19192b1908190808,
+    0x19192b1919080808, 0x19192b2b08080808, 0x19192b2b08192b19, 0x19192b2b2b081919,
+    0x19192b2b2b2b2b08, 0x192b080808080819, 0x192b080808081908, 0x192b08080808192b,
+    0x192b080808190808, 0x192b08080819082b, 0x192b080808191919, 0x192b080808192b08,
+    0x192b0808082b0819, 0x192b0808082b1908, 0x192b080819080808, 0x192b080819081919,
+    0x192b080819082b08, 0x192b080819190819, 0x192b080819191908, 0x192b0808192b0808,
+    0x192b08082b081908, 0x192b08082b190808, 0x192b081908080808, 0x192b08190808082b,
+    0x192b081908081919, 0x192b081908082b08, 0x192b081908190819, 0x192b081908191908,
+    0x192b0819082b0808, 0x192b081919080819, 0x192b081919081908, 0x192b081919190808,
+    0x192b08192b080808, 0x192b08192b192b19, 0x192b082b08081908, 0x192b082b08190808,
+    0x192b082b19080808, 0x192b082b1919192b, 0x192b082b2b2b0819, 0x192b190808080808,
+    0x192b190808081919, 0x192b190808082b08, 0x192b190808190819, 0x192b190808191908,
+    0x192b1908082b0808, 0x192b190819080819, 0x192b190819081908, 0x192b190819190808,
+    0x192b19082b080808, 0x192b191908080819, 0x192b191908081908, 0x192b191908190808,
+    0x192b191919080808, 0x192b191919082b2b, 0x192b1919192b2b08, 0x192b19192b19082b,
+    0x192b192b08080808, 0x192b192b2b191908, 0x192b2b0808080819, 0x192b2b0808081908,
+    0x192b2b0808190808, 0x192b2b08192b1919, 0x192b2b082b192b08, 0x192b2b1908080808,
+    0x192b2b19082b2b2b, 0x192b2b2b1908082b, 0x192b2b2b2b2b0819, 0x2b08080808080808,
+    0x2b0808080808082b, 0x2b08080808081919, 0x2b08080808082b08, 0x2b08080808190819,
+    0x2b08080808191908, 0x2b08080808192b19, 0x2b080808082b0808, 0x2b080808082b1919,
+    0x2b08080819080819, 0x2b08080819081908, 0x2b08080819190808, 0x2b0808081919082b,
+    0x2b08080819191919, 0x2b08080819192b08, 0x2b080808192b0819, 0x2b0808082b080808,
+    0x2b0808082b081919, 0x2b0808082b190819, 0x2b0808082b191908, 0x2b08081908080819,
+    0x2b08081908081908, 0x2b08081908082b19, 0x2b08081908190808, 0x2b0808190819082b,
+    0x2b08081908191919, 0x2b08081908192b08, 0x2b080819082b0819, 0x2b080819082b1908,
+    0x2b08081919080808, 0x2b0808191908082b, 0x2b08081919081919, 0x2b08081919082b08,
+    0x2b08081919190819, 0x2b08081919191908, 0x2b0808192b080819, 0x2b0808192b081908,
+    0x2b0808192b190808, 0x2b0808192b2b2b19, 0x2b08082b08080808, 0x2b08082b08081919,
+    0x2b08082b08082b2b, 0x2b08082b08190819, 0x2b08082b08191908, 0x2b08082b19080819,
+    0x2b08082b19081908, 0x2b08082b19190808, 0x2b08190808080819, 0x2b08190808081908,
+    0x2b0819080808192b, 0x2b08190808082b19, 0x2b08190808190808, 0x2b0819080819082b,
+    0x2b08190808191919, 0x2b08190808192b08, 0x2b081908082b0819, 0x2b08190819080808,
+    0x2b0819081908082b, 0x2b08190819081919, 0x2b08190819082b08, 0x2b08190819190819,
+    0x2b08190819191908, 0x2b081908192b0808, 0x2b0819082b080819, 0x2b0819082b081908,
+    0x2b0819082b190808, 0x2b08191908080808, 0x2b0819190808082b, 0x2b08191908081919,
+    0x2b08191908082b08, 0x2b08191908190819, 0x2b08191908191908, 0x2b081919082b0808,
+    0x2b08191919080819, 0x2b08191919081908, 0x2b08191919190808, 0x2b0819192b080808,
+    0x2b0819192b082b2b, 0x2b08192b08080819, 0x2b08192b08081908, 0x2b08192b08190808,
+    0x2b08192b082b2b19, 0x2b08192b19080808, 0x2b082b0808080808, 0x2b082b0808081919,
+    0x2b082b0808190819, 0x2b082b0808191908, 0x2b082b0819080819, 0x2b082b0819081908,
+    0x2b082b0819190808, 0x2b082b082b2b082b, 0x2b082b1908080819, 0x2b082b1908081908,
+    0x2b082b1919080808, 0x2b082b19192b1919, 0x2b082b2b082b082b, 0x2b082b2b19192b08,
+    0x2b082b2b19192b2b, 0x2b082b2b2b08082b, 0x2b082b2b2b2b082b, 0x2b19080808080819,
+    0x2b19080808081908, 0x2b19080808082b19, 0x2b19080808190808, 0x2b1908080819082b,
+    0x2b19080808191919, 0x2b19080808192b08, 0x2b190808082b1908, 0x2b19080819080808,
+    0x2b1908081908082b, 0x2b19080819081919, 0x2b19080819082b08, 0x2b19080819190819,
+    0x2b19080819191908, 0x2b190808192b0808, 0x2b1908082b080819, 0x2b1908082b081908,
+    0x2b1908082b190808, 0x2b19081908080808, 0x2b19081908081919, 0x2b19081908190819,
+    0x2b19081908191908, 0x2b19081919080819, 0x2b19081919081908, 0x2b19081919190808,
+    0x2b19081919192b2b, 0x2b19082b08080819, 0x2b19082b08081908, 0x2b19082b08190808,
+    0x2b19082b19080808, 0x2b19082b2b2b192b, 0x2b19190808080808, 0x2b1919080808082b,
+    0x2b19190808081919, 0x2b19190808082b08, 0x2b19190808190819, 0x2b19190808191908,
+    0x2b191908082b0808, 0x2b19190819080819, 0x2b19190819081908, 0x2b19190819190808,
+    0x2b1919082b080808, 0x2b1919082b19192b, 0x2b19191908080819, 0x2b19191908081908,
+    0x2b19191908190808, 0x2b19191919080808, 0x2b1919192b192b08, 0x2b1919192b2b0819,
+    0x2b19192b08080808, 0x2b19192b1908192b, 0x2b19192b192b1908, 0x2b192b0808080819,
+    0x2b192b0808081908, 0x2b192b0808190808, 0x2b192b08082b192b, 0x2b192b0819080808,
+    0x2b192b082b2b2b19, 0x2b192b1908080808, 0x2b192b1919082b19, 0x2b192b191919082b,
+    0x2b192b2b2b190808, 0x2b2b080808080808, 0x2b2b080808081919, 0x2b2b080808082b2b,
+    0x2b2b080808191908, 0x2b2b0808082b082b, 0x2b2b0808082b2b2b, 0x2b2b080819080819,
+    0x2b2b080819081908, 0x2b2b080819190808, 0x2b2b08082b2b082b, 0x2b2b08082b2b2b2b,
+    0x2b2b081919080808, 0x2b2b0819192b1919, 0x2b2b082b0808082b, 0x2b2b082b08082b2b,
+    0x2b2b082b082b082b, 0x2b2b082b082b2b08, 0x2b2b082b082b2b2b, 0x2b2b082b2b08082b,
+    0x2b2b082b2b082b08, 0x2b2b082b2b082b2b, 0x2b2b082b2b2b2b08, 0x2b2b190808080819,
+    0x2b2b190808081908, 0x2b2b190808190808, 0x2b2b190819080808, 0x2b2b19082b082b19,
+    0x2b2b19082b2b1908, 0x2b2b191908080808, 0x2b2b191908192b19, 0x2b2b192b19190819,
+    0x2b2b2b0808082b2b, 0x2b2b2b08082b2b08, 0x2b2b2b082b2b082b, 0x2b2b2b1919191908,
+    0x2b2b2b192b08192b, 0x2b2b2b2b08082b08, 0x2b2b2b2b08082b2b, 0x2b2b2b2b082b0808,
+    0x2b2b2b2b082b082b, 0x2b2b2b2b082b2b08, 0x2b2b2b2b2b082b08, 0x2b2b2b2b2b2b2b2b,
+};
+
 constexpr constant static uint32_t iq3xxs_grid[256] = {
     0x04040404, 0x04040414, 0x04040424, 0x04040c0c, 0x04040c1c, 0x04040c3e, 0x04041404, 0x04041414,
     0x04041c0c, 0x04042414, 0x04043e1c, 0x04043e2c, 0x040c040c, 0x040c041c, 0x040c0c04, 0x040c0c14,
@@ -3747,6 +4087,204 @@ constexpr constant static uint32_t iq3xxs_grid[256] = {
     0x3e1c1c1c, 0x3e1c3404, 0x3e24140c, 0x3e24240c, 0x3e2c0404, 0x3e2c0414, 0x3e2c1424, 0x3e341c04,
 };
 
+constexpr constant static uint32_t iq3xs_grid[512] = {
+    0x04040404, 0x0404040c, 0x04040414, 0x0404042c, 0x0404043e, 0x04040c04, 0x04040c0c, 0x04040c14,
+    0x04040c24, 0x04040c34, 0x04041404, 0x0404140c, 0x0404142c, 0x04041c1c, 0x04042404, 0x04042414,
+    0x0404242c, 0x0404243e, 0x04042c0c, 0x04042c1c, 0x04043404, 0x04043414, 0x04043e0c, 0x04043e24,
+    0x04043e3e, 0x040c0404, 0x040c040c, 0x040c0414, 0x040c0424, 0x040c0c04, 0x040c0c0c, 0x040c0c2c,
+    0x040c1404, 0x040c141c, 0x040c143e, 0x040c1c0c, 0x040c1c2c, 0x040c2424, 0x040c340c, 0x040c342c,
+    0x040c3e14, 0x04140404, 0x0414040c, 0x0414042c, 0x0414043e, 0x04140c04, 0x04140c1c, 0x04140c34,
+    0x0414140c, 0x0414142c, 0x04141c04, 0x04141c24, 0x04142414, 0x0414242c, 0x0414243e, 0x04142c0c,
+    0x04142c1c, 0x04143e04, 0x04143e1c, 0x041c041c, 0x041c0c0c, 0x041c0c2c, 0x041c1404, 0x041c1414,
+    0x041c1c0c, 0x041c1c1c, 0x041c1c34, 0x041c2424, 0x041c2c04, 0x041c2c14, 0x041c343e, 0x041c3e0c,
+    0x041c3e2c, 0x04240404, 0x04240c1c, 0x04240c3e, 0x0424140c, 0x04241424, 0x04241c14, 0x04242404,
+    0x0424241c, 0x04242c0c, 0x04243e04, 0x042c0414, 0x042c0424, 0x042c1404, 0x042c1414, 0x042c1434,
+    0x042c1c1c, 0x042c240c, 0x042c242c, 0x042c243e, 0x042c3434, 0x042c3e1c, 0x04340434, 0x04340c0c,
+    0x04340c1c, 0x04341c0c, 0x04342c14, 0x04343e0c, 0x043e0404, 0x043e0414, 0x043e0424, 0x043e1404,
+    0x043e1414, 0x043e1434, 0x043e1c1c, 0x043e2c04, 0x043e2c24, 0x0c040404, 0x0c04040c, 0x0c040414,
+    0x0c040424, 0x0c040c04, 0x0c040c0c, 0x0c040c1c, 0x0c040c2c, 0x0c040c3e, 0x0c041404, 0x0c041414,
+    0x0c041c0c, 0x0c041c24, 0x0c041c34, 0x0c042c24, 0x0c042c34, 0x0c04340c, 0x0c043e14, 0x0c0c0404,
+    0x0c0c040c, 0x0c0c041c, 0x0c0c0434, 0x0c0c0c04, 0x0c0c0c24, 0x0c0c140c, 0x0c0c1c04, 0x0c0c1c1c,
+    0x0c0c240c, 0x0c0c2c04, 0x0c0c2c14, 0x0c0c3e04, 0x0c0c3e34, 0x0c140404, 0x0c140c14, 0x0c140c2c,
+    0x0c140c3e, 0x0c141404, 0x0c141424, 0x0c141c14, 0x0c142404, 0x0c14241c, 0x0c142c2c, 0x0c143404,
+    0x0c143e14, 0x0c1c040c, 0x0c1c0424, 0x0c1c043e, 0x0c1c0c04, 0x0c1c0c1c, 0x0c1c140c, 0x0c1c143e,
+    0x0c1c1c04, 0x0c1c1c24, 0x0c1c240c, 0x0c1c3414, 0x0c1c3e04, 0x0c24041c, 0x0c24042c, 0x0c240c14,
+    0x0c240c24, 0x0c241c0c, 0x0c241c1c, 0x0c242414, 0x0c242434, 0x0c242c04, 0x0c242c24, 0x0c2c040c,
+    0x0c2c0c04, 0x0c2c0c1c, 0x0c2c140c, 0x0c2c1c04, 0x0c2c1c14, 0x0c2c2c0c, 0x0c341404, 0x0c341424,
+    0x0c34143e, 0x0c342424, 0x0c342434, 0x0c3e040c, 0x0c3e041c, 0x0c3e0c04, 0x0c3e0c14, 0x0c3e140c,
+    0x0c3e1c2c, 0x0c3e240c, 0x0c3e3414, 0x0c3e3e04, 0x14040404, 0x1404040c, 0x1404041c, 0x1404042c,
+    0x1404043e, 0x14040c04, 0x14040c14, 0x14040c24, 0x14040c34, 0x1404140c, 0x1404141c, 0x1404143e,
+    0x14041c04, 0x14041c14, 0x1404240c, 0x1404241c, 0x1404242c, 0x14042c04, 0x14042c14, 0x1404343e,
+    0x14043e04, 0x14043e1c, 0x14043e2c, 0x140c0404, 0x140c0414, 0x140c0c04, 0x140c0c1c, 0x140c0c3e,
+    0x140c1414, 0x140c142c, 0x140c1c0c, 0x140c1c24, 0x140c2414, 0x140c2c0c, 0x1414040c, 0x14140424,
+    0x1414043e, 0x1414140c, 0x1414141c, 0x14141c04, 0x14141c3e, 0x1414240c, 0x14142c1c, 0x14142c3e,
+    0x14143e0c, 0x14143e24, 0x141c0404, 0x141c0414, 0x141c042c, 0x141c0c0c, 0x141c1414, 0x141c1424,
+    0x141c1c0c, 0x141c1c1c, 0x141c2414, 0x141c2c04, 0x141c3434, 0x1424040c, 0x1424043e, 0x14241404,
+    0x1424141c, 0x14241c14, 0x14241c2c, 0x1424240c, 0x14243e14, 0x14243e2c, 0x142c0424, 0x142c0c0c,
+    0x142c1414, 0x142c1c3e, 0x142c2404, 0x142c2c1c, 0x142c3e04, 0x14340404, 0x14340414, 0x1434043e,
+    0x1434140c, 0x14342c2c, 0x1434340c, 0x143e042c, 0x143e0c0c, 0x143e1434, 0x143e1c04, 0x143e241c,
+    0x143e2c04, 0x1c040414, 0x1c040c0c, 0x1c040c1c, 0x1c040c2c, 0x1c040c3e, 0x1c041414, 0x1c041c0c,
+    0x1c041c1c, 0x1c041c2c, 0x1c042414, 0x1c042424, 0x1c04243e, 0x1c042c0c, 0x1c04341c, 0x1c043e0c,
+    0x1c0c040c, 0x1c0c041c, 0x1c0c042c, 0x1c0c0c24, 0x1c0c140c, 0x1c0c141c, 0x1c0c2404, 0x1c0c3404,
+    0x1c0c3e14, 0x1c0c3e34, 0x1c140404, 0x1c140c14, 0x1c141404, 0x1c141c14, 0x1c141c24, 0x1c142c04,
+    0x1c1c040c, 0x1c1c0c04, 0x1c1c0c24, 0x1c1c140c, 0x1c1c141c, 0x1c1c143e, 0x1c1c1c04, 0x1c1c240c,
+    0x1c1c241c, 0x1c1c243e, 0x1c1c2c2c, 0x1c1c3e1c, 0x1c24041c, 0x1c240c0c, 0x1c240c34, 0x1c241414,
+    0x1c241c0c, 0x1c242c14, 0x1c243404, 0x1c243424, 0x1c2c040c, 0x1c2c0c04, 0x1c2c0c14, 0x1c2c142c,
+    0x1c2c1c14, 0x1c2c2424, 0x1c2c2c34, 0x1c2c3e1c, 0x1c340c34, 0x1c34240c, 0x1c3e040c, 0x1c3e041c,
+    0x1c3e1404, 0x1c3e1414, 0x1c3e1c2c, 0x24040404, 0x24040424, 0x24040c14, 0x24041404, 0x24041424,
+    0x2404143e, 0x24041c14, 0x2404240c, 0x24042c04, 0x24043e04, 0x240c0414, 0x240c043e, 0x240c0c0c,
+    0x240c0c1c, 0x240c1414, 0x240c1c04, 0x240c1c2c, 0x240c241c, 0x240c2c0c, 0x240c2c2c, 0x2414040c,
+    0x2414041c, 0x24140c04, 0x24140c2c, 0x2414140c, 0x24141c1c, 0x24142404, 0x24142c3e, 0x24143414,
+    0x24143e04, 0x241c0424, 0x241c0c0c, 0x241c0c1c, 0x241c1404, 0x241c1414, 0x241c1c0c, 0x241c1c2c,
+    0x24240404, 0x24240414, 0x24241424, 0x24241c3e, 0x24242404, 0x24243e0c, 0x242c042c, 0x242c043e,
+    0x242c140c, 0x242c3414, 0x24340c1c, 0x24341c24, 0x24343404, 0x243e0c04, 0x243e0c2c, 0x243e1c04,
+    0x243e241c, 0x243e2c0c, 0x2c040414, 0x2c040c04, 0x2c040c24, 0x2c041414, 0x2c042404, 0x2c042424,
+    0x2c04243e, 0x2c042c14, 0x2c043434, 0x2c043e24, 0x2c0c040c, 0x2c0c041c, 0x2c0c042c, 0x2c0c0c14,
+    0x2c0c140c, 0x2c0c1c14, 0x2c0c3e14, 0x2c140404, 0x2c140c0c, 0x2c14141c, 0x2c141c04, 0x2c141c34,
+    0x2c142c1c, 0x2c1c0414, 0x2c1c043e, 0x2c1c0c04, 0x2c1c143e, 0x2c1c2424, 0x2c1c2c0c, 0x2c1c342c,
+    0x2c1c3e1c, 0x2c24040c, 0x2c240424, 0x2c241404, 0x2c241c14, 0x2c242434, 0x2c2c0c14, 0x2c2c1434,
+    0x2c2c2c0c, 0x2c2c2c1c, 0x2c342414, 0x2c3e0414, 0x2c3e0424, 0x2c3e1414, 0x34040c0c, 0x34040c1c,
+    0x34040c2c, 0x34041c0c, 0x34041c1c, 0x34043404, 0x340c0404, 0x340c1404, 0x340c143e, 0x340c3424,
+    0x34140c14, 0x34141c24, 0x34142414, 0x34142c2c, 0x34143414, 0x34143e04, 0x341c0404, 0x341c0c24,
+    0x341c140c, 0x341c2404, 0x3424142c, 0x3424241c, 0x34243414, 0x342c0404, 0x342c041c, 0x342c1c24,
+    0x342c3404, 0x3434042c, 0x34342404, 0x343e0c0c, 0x343e0c1c, 0x3e040404, 0x3e040424, 0x3e04043e,
+    0x3e041404, 0x3e041414, 0x3e041c34, 0x3e042404, 0x3e042c24, 0x3e043414, 0x3e0c0414, 0x3e0c0c0c,
+    0x3e0c1424, 0x3e0c241c, 0x3e0c242c, 0x3e14040c, 0x3e140424, 0x3e140c04, 0x3e140c34, 0x3e14140c,
+    0x3e141c04, 0x3e142c0c, 0x3e1c0414, 0x3e1c1c14, 0x3e1c1c2c, 0x3e1c2c1c, 0x3e24040c, 0x3e24042c,
+    0x3e240c1c, 0x3e241404, 0x3e242c04, 0x3e2c1414, 0x3e2c2414, 0x3e340414, 0x3e341c0c, 0x3e3e0404,
+};
+
+#define NGRID_IQ1S 512
+constexpr constant static uint64_t iq1s_grid[NGRID_IQ1S] = {
+    0xffffffffffff0101, 0xffffffffff01ff00, 0xffffffffff010100, 0xffffffff00000000,
+    0xffffffff01ff00ff, 0xffffffff01ff0001, 0xffffffff0101ffff, 0xffffffff0101ff01,
+    0xffffff00ff000000, 0xffffff000000ff00, 0xffffff00000000ff, 0xffffff0000000100,
+    0xffffff0000010000, 0xffffff0001000000, 0xffffff01ffff00ff, 0xffffff01ff01ff00,
+    0xffffff01ff010100, 0xffffff0100000001, 0xffffff0101ffff00, 0xffffff0101ff0101,
+    0xffffff0101010100, 0xffff00ffff00ff01, 0xffff00ffff0000ff, 0xffff00ff00ff0100,
+    0xffff00ff0100ff00, 0xffff00ff010001ff, 0xffff0000ff0101ff, 0xffff000000ffff00,
+    0xffff000000000000, 0xffff00000001ff01, 0xffff000001000101, 0xffff0000010100ff,
+    0xffff0001ffff0100, 0xffff00010000ff00, 0xffff000100010101, 0xffff000101000000,
+    0xffff01ffffff0000, 0xffff01ffff01ffff, 0xffff01ffff010100, 0xffff01ff00000000,
+    0xffff01ff01ffffff, 0xffff01ff01ff0001, 0xffff01ff0101ffff, 0xffff01ff01010001,
+    0xffff0100ffffff01, 0xffff01000000ffff, 0xffff010000000100, 0xffff010001ff01ff,
+    0xffff010001000000, 0xffff0101ff000000, 0xffff0101000101ff, 0xffff010101ffff01,
+    0xffff01010101ff00, 0xff00ffffff000000, 0xff00ffff00ffff00, 0xff00ffff00000001,
+    0xff00ffff000001ff, 0xff00ffff01010000, 0xff00ff00ffff0000, 0xff00ff00ff00ff00,
+    0xff00ff00ff0000ff, 0xff00ff00ff000100, 0xff00ff00ff010001, 0xff00ff0000ff0001,
+    0xff00ff000000ffff, 0xff00ff0000000000, 0xff00ff000001ff00, 0xff00ff0000010100,
+    0xff00ff0001ff0000, 0xff00ff000100ff00, 0xff00ff0001000100, 0xff00ff01ff000000,
+    0xff00ff0100ff0000, 0xff00ff01000001ff, 0xff00ff0101010001, 0xff0000ff00000000,
+    0xff0000ff0001ff00, 0xff0000ff00010100, 0xff000000ffff0101, 0xff000000ff000000,
+    0xff000000ff01ff00, 0xff00000000ff0000, 0xff0000000000ff00, 0xff000000000000ff,
+    0xff00000000000000, 0xff00000000000001, 0xff00000000000100, 0xff0000000001ffff,
+    0xff00000000010000, 0xff00000001000000, 0xff00000001010100, 0xff000001ff00ff01,
+    0xff000001ff0100ff, 0xff00000100000000, 0xff0000010001ff00, 0xff00000101ff0100,
+    0xff0000010100ff00, 0xff0001ff00ff00ff, 0xff0001ff00000101, 0xff0001ff000100ff,
+    0xff0001ff01000000, 0xff000100ff0001ff, 0xff0001000000ff01, 0xff00010000000000,
+    0xff00010000010001, 0xff00010000010100, 0xff00010001ffff00, 0xff00010001ff0101,
+    0xff00010001010000, 0xff000101ffffffff, 0xff000101ff000101, 0xff00010101ff00ff,
+    0xff00010101000001, 0xff000101010100ff, 0xff01ffffff000101, 0xff01ffffff01ffff,
+    0xff01ffffff01ff01, 0xff01ffffff0101ff, 0xff01ffff00000000, 0xff01ffff01ff0001,
+    0xff01ffff0101ff01, 0xff01ff00ff000000, 0xff01ff0000ff0100, 0xff01ff000000ff01,
+    0xff01ff0000010000, 0xff01ff00010000ff, 0xff01ff01ff01ff00, 0xff01ff0100000101,
+    0xff0100ffffff0000, 0xff0100ffff010000, 0xff0100ff01ff00ff, 0xff0100ff01000100,
+    0xff0100ff010100ff, 0xff010000ffffff01, 0xff01000000000000, 0xff0100000101ff00,
+    0xff010001ffff00ff, 0xff010001ff000100, 0xff01000100ffff00, 0xff01000100010001,
+    0xff01000101ff0001, 0xff010001010001ff, 0xff0101ffffffffff, 0xff0101ffff01ffff,
+    0xff0101ffff010101, 0xff0101ff0000ff00, 0xff0101ff01010001, 0xff010100ff000000,
+    0xff010100ff01ff01, 0xff01010000ff0001, 0xff01010000000100, 0xff01010001000000,
+    0xff0101010100ffff, 0x00ffffff0000ff01, 0x00ffffff000000ff, 0x00ffffff00000100,
+    0x00ffffff00010000, 0x00ffff00ffff0001, 0x00ffff00ff0000ff, 0x00ffff00ff000100,
+    0x00ffff0000000000, 0x00ffff0001000100, 0x00ffff0001010001, 0x00ffff01ff00ff01,
+    0x00ffff0100ff0100, 0x00ffff010000ff00, 0x00ffff01000100ff, 0x00ffff0101ff00ff,
+    0x00ffff010101ff00, 0x00ff00ffffffffff, 0x00ff00ffffff01ff, 0x00ff00ffff000101,
+    0x00ff00ff00000000, 0x00ff00ff000101ff, 0x00ff00ff01010101, 0x00ff0000ff000000,
+    0x00ff0000ff01ffff, 0x00ff000000ff0000, 0x00ff00000000ff00, 0x00ff0000000000ff,
+    0x00ff000000000000, 0x00ff000000000001, 0x00ff000000000100, 0x00ff000000010000,
+    0x00ff000001ffff01, 0x00ff000001000000, 0x00ff0001ff000101, 0x00ff000100ffffff,
+    0x00ff000100000000, 0x00ff0001010001ff, 0x00ff01ffff000000, 0x00ff01ff0001ff00,
+    0x00ff01ff01ff0100, 0x00ff0100ff01ff01, 0x00ff010000ff00ff, 0x00ff010000ff0101,
+    0x00ff010000000000, 0x00ff010000010101, 0x00ff01000100ff00, 0x00ff010001010000,
+    0x00ff0101ffffff00, 0x00ff01010000ff01, 0x00ff010100000100, 0x00ff010101ff0000,
+    0x0000ffffffff0100, 0x0000ffffff00ff00, 0x0000ffffff0000ff, 0x0000ffffff010000,
+    0x0000ffff00000000, 0x0000ffff00010101, 0x0000ffff01ffff01, 0x0000ffff01000100,
+    0x0000ff00ff000000, 0x0000ff00ff01ff00, 0x0000ff00ff0101ff, 0x0000ff0000ff0000,
+    0x0000ff000000ff00, 0x0000ff00000000ff, 0x0000ff0000000000, 0x0000ff0000000001,
+    0x0000ff0000000100, 0x0000ff0000010000, 0x0000ff0001ffffff, 0x0000ff0001ff01ff,
+    0x0000ff0001000000, 0x0000ff000101ffff, 0x0000ff01ffff0101, 0x0000ff01ff010000,
+    0x0000ff0100000000, 0x0000ff0101000101, 0x000000ffffff0001, 0x000000ffff000000,
+    0x000000ff00ff0000, 0x000000ff0000ff00, 0x000000ff000000ff, 0x000000ff00000000,
+    0x000000ff00000001, 0x000000ff00000100, 0x000000ff00010000, 0x000000ff01000000,
+    0x000000ff0101ff00, 0x00000000ffff0000, 0x00000000ff00ff00, 0x00000000ff0000ff,
+    0x00000000ff000000, 0x00000000ff000001, 0x00000000ff000100, 0x00000000ff010000,
+    0x0000000000ffff00, 0x0000000000ff00ff, 0x0000000000ff0000, 0x0000000000ff0001,
+    0x0000000000ff0100, 0x000000000000ffff, 0x000000000000ff00, 0x000000000000ff01,
+    0x00000000000000ff, 0x0000000000000001, 0x00000000000001ff, 0x0000000000000100,
+    0x0000000000000101, 0x000000000001ff00, 0x00000000000100ff, 0x0000000000010000,
+    0x0000000000010001, 0x0000000000010100, 0x0000000001ff0000, 0x000000000100ff00,
+    0x00000000010000ff, 0x0000000001000000, 0x0000000001000001, 0x0000000001000100,
+    0x0000000001010000, 0x00000001ffff01ff, 0x00000001ff000000, 0x0000000100ff0000,
+    0x000000010000ff00, 0x00000001000000ff, 0x0000000100000000, 0x0000000100000001,
+    0x0000000100000100, 0x0000000100010000, 0x0000000101000000, 0x000001ffff00ff00,
+    0x000001ffff010001, 0x000001ffff0101ff, 0x000001ff00ffff01, 0x000001ff0000ffff,
+    0x000001ff00000000, 0x000001ff010000ff, 0x000001ff01010100, 0x00000100ffff0100,
+    0x00000100ff000000, 0x0000010000ff0000, 0x000001000000ff00, 0x00000100000000ff,
+    0x0000010000000000, 0x0000010000000001, 0x0000010000000100, 0x0000010000010000,
+    0x0000010001000000, 0x000001000101ff01, 0x00000101ffff0001, 0x00000101ff01ffff,
+    0x0000010100000000, 0x0000010101010100, 0x0001ffffff000000, 0x0001ffff00ffffff,
+    0x0001ffff00000100, 0x0001ffff0001ff00, 0x0001ffff01000000, 0x0001ff00ffffff00,
+    0x0001ff00ffff01ff, 0x0001ff00ff010000, 0x0001ff0000000000, 0x0001ff0000010001,
+    0x0001ff0001ff0000, 0x0001ff0001010100, 0x0001ff01ff0000ff, 0x0001ff01ff000001,
+    0x0001ff0100ffffff, 0x0001ff010001ffff, 0x0001ff01000101ff, 0x0001ff010100ff01,
+    0x000100ffff00ffff, 0x000100ffff00ff01, 0x000100ffff000100, 0x000100ff00000000,
+    0x000100ff000101ff, 0x000100ff01ff0101, 0x000100ff0100ffff, 0x000100ff01010101,
+    0x00010000ff000000, 0x00010000ff010100, 0x0001000000ff0000, 0x000100000000ff00,
+    0x00010000000000ff, 0x0001000000000000, 0x0001000000000001, 0x0001000000000100,
+    0x0001000000010000, 0x0001000001ffff01, 0x0001000001000000, 0x0001000100ff0101,
+    0x0001000100000000, 0x00010001010100ff, 0x000101ffffff01ff, 0x000101ffffff0101,
+    0x000101ff00010000, 0x000101ff01ff0000, 0x000101ff0100ff01, 0x00010100ffff0000,
+    0x0001010000000000, 0x000101000001ffff, 0x0001010000010101, 0x00010100010001ff,
+    0x00010101ff00ff00, 0x00010101ff010001, 0x0001010100ffffff, 0x0001010100ff01ff,
+    0x00010101000101ff, 0x0001010101ff0000, 0x000101010100ff01, 0x0001010101000101,
+    0x01ffffffffff0101, 0x01ffffffff01ffff, 0x01ffffffff01ff01, 0x01ffffffff0101ff,
+    0x01ffffffff010101, 0x01ffffff00000000, 0x01ffffff01ff01ff, 0x01ffffff01000101,
+    0x01ffffff0101ff01, 0x01ffffff010100ff, 0x01ffff000000ff00, 0x01ffff0000000001,
+    0x01ffff00000001ff, 0x01ffff0000010000, 0x01ffff0001ff0000, 0x01ffff01ffffffff,
+    0x01ffff01ffff01ff, 0x01ffff01ff000000, 0x01ffff01ff01ffff, 0x01ffff01ff0101ff,
+    0x01ffff010100ffff, 0x01ff00ffffff0000, 0x01ff00ffff010000, 0x01ff00ff00ffff01,
+    0x01ff0000ff0000ff, 0x01ff000000000000, 0x01ff00000001ff01, 0x01ff000001ffffff,
+    0x01ff000001010100, 0x01ff0001ffffff01, 0x01ff0001ff010001, 0x01ff000101ff0100,
+    0x01ff000101000001, 0x01ff0001010100ff, 0x01ff01ffff00ffff, 0x01ff01ff00010001,
+    0x01ff01ff01000000, 0x01ff01ff010101ff, 0x01ff0100ff000001, 0x01ff010000ffff00,
+    0x01ff010000000100, 0x01ff010001ff01ff, 0x01ff01000101ffff, 0x01ff0101ffff00ff,
+    0x01ff0101ffff0101, 0x01ff0101ff0101ff, 0x01ff010100010000, 0x0100ffff00ff00ff,
+    0x0100ffff00ff0001, 0x0100ffff00000100, 0x0100ffff0100ff00, 0x0100ff00ffff0000,
+    0x0100ff00ff00ffff, 0x0100ff00ff00ff01, 0x0100ff00ff000100, 0x0100ff00ff010000,
+    0x0100ff0000000000, 0x0100ff00000100ff, 0x0100ff0001ff0101, 0x0100ff0001010101,
+    0x0100ff0100ff00ff, 0x0100ff0100ff0001, 0x0100ff0100000100, 0x0100ff0100010001,
+    0x0100ff0101000000, 0x010000ffff00ff00, 0x010000ff0000ffff, 0x010000ff00000000,
+    0x010000ff010001ff, 0x010000ff01010001, 0x01000000ffffff00, 0x01000000ffff0101,
+    0x01000000ff000000, 0x01000000ff0100ff, 0x01000000ff010101, 0x0100000000ff0000,
+    0x010000000000ff00, 0x01000000000000ff, 0x0100000000000000, 0x0100000000000001,
+    0x0100000000000100, 0x0100000000010000, 0x0100000001000000, 0x0100000100000000,
+    0x01000001000101ff, 0x0100000101ffff01, 0x010001ffff000101, 0x010001ff00ff0100,
+    0x010001ff0000ff00, 0x010001ff000100ff, 0x010001ff01ffffff, 0x01000100ffff0000,
+    0x01000100ff0001ff, 0x0100010000000000, 0x010001000001ff00, 0x0100010001ff0000,
+    0x01000100010000ff, 0x0100010001000101, 0x01000101ff00ff01, 0x0100010100ff0100,
+    0x010001010000ffff, 0x0100010101010001, 0x0101ffffffff0101, 0x0101ffffff0001ff,
+    0x0101ffffff01ffff, 0x0101ffffff010101, 0x0101ffff00000000, 0x0101ffff0101ffff,
+    0x0101ffff010101ff, 0x0101ff00ff000000, 0x0101ff0000ff0100, 0x0101ff000000ff00,
+    0x0101ff0000010000, 0x0101ff00010000ff, 0x0101ff0001000001, 0x0101ff01ff010101,
+    0x0101ff0100000000, 0x0101ff010101ff00, 0x010100ffffff0000, 0x010100ffff010000,
+    0x010100ff00ff01ff, 0x010100ff000000ff, 0x010100ff00000101, 0x010100ff01ffff00,
+    0x01010000ffffff01, 0x01010000ff000100, 0x01010000ff01ff01, 0x0101000000000000,
+    0x01010000000100ff, 0x010100000101ff01, 0x01010001ffff0000, 0x01010001ff00ffff,
+    0x01010001ff010000, 0x0101000101ffffff, 0x0101000101ff01ff, 0x0101000101010101,
+    0x010101ffff01ffff, 0x010101ff00000000, 0x010101ff0001ff01, 0x010101ff0101ffff,
+    0x010101ff010101ff, 0x01010100ffffffff, 0x01010100ff000001, 0x010101000000ff00,
+    0x0101010001010000, 0x0101010100ff0001, 0x010101010001ff01, 0x010101010101ffff,
+};
 
 constexpr constant static uint8_t ksigns_iq2xs[128] = {
       0, 129, 130,   3, 132,   5,   6, 135, 136,   9,  10, 139,  12, 141, 142,  15,
@@ -3812,7 +4350,6 @@ void kernel_mul_mv_iq2_xxs_f32_impl(
         threadgroup_barrier(mem_flags::mem_threadgroup);
     }
 
-#if QK_K == 256
     const int ix = tiisg;
 
     device const float * y4 = y + 32 * ix;
@@ -3853,9 +4390,6 @@ void kernel_mul_mv_iq2_xxs_f32_impl(
 
         y4 += 32 * 32;
     }
-#else
-    // TODO
-#endif
 
     for (int row = 0; row < N_DST; ++row) {
         all_sum = simd_sum(sumf[row]);
@@ -3945,7 +4479,6 @@ void kernel_mul_mv_iq2_xs_f32_impl(
         threadgroup_barrier(mem_flags::mem_threadgroup);
     }
 
-#if QK_K == 256
     const int ix = tiisg;
 
     device const float * y4 = y + 32 * ix;
@@ -3996,9 +4529,6 @@ void kernel_mul_mv_iq2_xs_f32_impl(
 
         y4 += 32 * 32;
     }
-#else
-    // TODO
-#endif
 
     for (int row = 0; row < N_DST; ++row) {
         all_sum = simd_sum(sumf[row]);
@@ -4088,7 +4618,6 @@ void kernel_mul_mv_iq3_xxs_f32_impl(
         threadgroup_barrier(mem_flags::mem_threadgroup);
     }
 
-#if QK_K == 256
     const int ix = tiisg;
 
     device const float * y4 = y + 32 * ix;
@@ -4132,9 +4661,6 @@ void kernel_mul_mv_iq3_xxs_f32_impl(
 
         y4 += 32 * 32;
     }
-#else
-    // TODO
-#endif
 
     for (int row = 0; row < N_DST; ++row) {
         all_sum = simd_sum(sumf[row]);
@@ -4173,767 +4699,1802 @@ kernel void kernel_mul_mv_iq3_xxs_f32(
     kernel_mul_mv_iq3_xxs_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
 }
 
+void kernel_mul_mv_iq3_s_f32_impl(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   uint    & r2,
+        constant   uint    & r3,
+        threadgroup int8_t * shared_values [[threadgroup(0)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
 
-//============================= templates and their specializations =============================
-
-// NOTE: this is not dequantizing - we are simply fitting the template
-template <typename type4x4>
-void dequantize_f32(device const float4x4 * src, short il, thread type4x4 & reg) {
-    float4x4 temp = *(((device float4x4 *)src));
-    for (int i = 0; i < 16; i++){
-        reg[i/4][i%4] = temp[i/4][i%4];
-    }
-}
-
-template <typename type4x4>
-void dequantize_f16(device const half4x4 * src, short il, thread type4x4 & reg) {
-    half4x4 temp = *(((device half4x4 *)src));
-    for (int i = 0; i < 16; i++){
-        reg[i/4][i%4] = temp[i/4][i%4];
-    }
-}
+    const int nb = ne00/QK_K;
+    const int r0 = tgpig.x;
+    const int r1 = tgpig.y;
+    const int im = tgpig.z;
 
-template <typename type4x4>
-void dequantize_q4_0(device const block_q4_0 *xb, short il, thread type4x4 & reg) {
-    device const uint16_t * qs = ((device const uint16_t *)xb + 1);
-    const float d1 = il ? (xb->d / 16.h) : xb->d;
-    const float d2 = d1 / 256.f;
-    const float md = -8.h * xb->d;
-    const ushort mask0 = il ? 0x00F0 : 0x000F;
-    const ushort mask1 = mask0 << 8;
+    const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
+    const int ib_row = first_row * nb;
 
-    for (int i=0;i<8;i++) {
-        reg[i/2][2*(i%2)+0] = d1 * (qs[i] & mask0) + md;
-        reg[i/2][2*(i%2)+1] = d2 * (qs[i] & mask1) + md;
-    }
-}
+    const uint i12 = im%ne12;
+    const uint i13 = im/ne12;
 
-template <typename type4x4>
-void dequantize_q4_1(device const block_q4_1 *xb, short il, thread type4x4 & reg) {
-    device const uint16_t * qs = ((device const uint16_t *)xb + 2);
-    const float d1 = il ? (xb->d / 16.h) : xb->d;
-    const float d2 = d1 / 256.f;
-    const float  m = xb->m;
-    const ushort mask0 = il ? 0x00F0 : 0x000F;
-    const ushort mask1 = mask0 << 8;
+    const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
 
-    for (int i=0;i<8;i++) {
-        reg[i/2][2*(i%2)+0] = ((qs[i] & mask0) * d1) + m;
-        reg[i/2][2*(i%2)+1] = ((qs[i] & mask1) * d2) + m;
-    }
-}
+    device const block_iq3_s * x = (device const block_iq3_s *) src0 + ib_row + offset0;
+    device const float       * y = (device const float       *) src1 + r1*ne10 + im*ne00*ne1;
 
-template <typename type4x4>
-void dequantize_q5_0(device const block_q5_0 *xb, short il, thread type4x4 & reg) {
-    device const uint16_t * qs = ((device const uint16_t *)xb + 3);
-    const float d = xb->d;
-    const float md = -16.h * xb->d;
-    const ushort mask = il ? 0x00F0 : 0x000F;
+    float yl[32];
+    float sumf[N_DST]={0.f}, all_sum;
 
-    const uint32_t qh = *((device const uint32_t *)xb->qh);
+    const int nb32 = nb * (QK_K / 32);
 
-    const int x_mv = il ? 4 : 0;
+    threadgroup uint32_t * values = (threadgroup uint32_t *)shared_values;
+    {
+        int nval = 8;
+        int pos  = (32*sgitg + tiisg)*nval;
+        for (int i = 0; i < nval; ++i) values[pos + i] = iq3xs_grid[pos + i];
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+    }
 
-    const int gh_mv = il ? 12 : 0;
-    const int gh_bk = il ?  0 : 4;
+    const int ix = tiisg;
 
-    for (int i = 0; i < 8; i++) {
-        // extract the 5-th bits for x0 and x1
-        const uint8_t xh_0 = ((qh >> (gh_mv + 2*i  )) << gh_bk) & 0x10;
-        const uint8_t xh_1 = ((qh >> (gh_mv + 2*i+1)) << gh_bk) & 0x10;
+    device const float * y4 = y + 32 * ix;
 
-        // combine the 4-bits from qs with the 5th bit
-        const int32_t x0 = ((((qs[i]     ) & mask) >> x_mv) | xh_0);
-        const int32_t x1 = ((((qs[i] >> 8) & mask) >> x_mv) | xh_1);
+    for (int ib32 = ix; ib32 < nb32; ib32 += 32) {
 
-        reg[i/2][2*(i%2)+0] = d * x0 + md;
-        reg[i/2][2*(i%2)+1] = d * x1 + md;
-    }
-}
+        for (int i = 0; i < 32; ++i) {
+            yl[i] = y4[i];
+        }
 
-template <typename type4x4>
-void dequantize_q5_1(device const block_q5_1 *xb, short il, thread type4x4 & reg) {
-    device const uint16_t * qs = ((device const uint16_t *)xb + 4);
-    const float d = xb->d;
-    const float m = xb->m;
-    const ushort mask = il ? 0x00F0 : 0x000F;
+        const int ibl = ib32 / (QK_K / 32);
+        const int ib  = ib32 % (QK_K / 32);
 
-    const uint32_t qh = *((device const uint32_t *)xb->qh);
+        device const block_iq3_s * xr = x + ibl;
+        device const uint8_t * qs = xr->qs + 8 * ib;
+        device const uint8_t * qh = xr->qh + ib;
+        device const uint8_t * sc = xr->scales + (ib/2);
+        device const uint8_t * signs = xr->signs + 4 * ib;
+        device const half * dh = &xr->d;
 
-    const int x_mv = il ? 4 : 0;
+        for (int row = 0; row < N_DST; row++) {
 
-    const int gh_mv = il ? 12 : 0;
-    const int gh_bk = il ?  0 : 4;
+            const float db = dh[0];
+            const float d = db * (0.5f + ((sc[0] >> 4*(ib%2)) & 0xf));
 
-    for (int i = 0; i < 8; i++) {
-        // extract the 5-th bits for x0 and x1
-        const uint8_t xh_0 = ((qh >> (gh_mv + 2*i  )) << gh_bk) & 0x10;
-        const uint8_t xh_1 = ((qh >> (gh_mv + 2*i+1)) << gh_bk) & 0x10;
+            float2 sum = {0};
+            for (int l = 0; l < 4; ++l) {
+                const threadgroup uint8_t * grid1 = (const threadgroup uint8_t *)(values + (qs[2*l+0] | ((qh[0] << (8-2*l)) & 256)));
+                const threadgroup uint8_t * grid2 = (const threadgroup uint8_t *)(values + (qs[2*l+1] | ((qh[0] << (7-2*l)) & 256)));
+                for (int j = 0; j < 4; ++j) {
+                    sum[0] += yl[8*l + j + 0] * grid1[j] * select(1, -1, signs[l] & kmask_iq2xs[j+0]);
+                    sum[1] += yl[8*l + j + 4] * grid2[j] * select(1, -1, signs[l] & kmask_iq2xs[j+4]);
+                }
+            }
+            sumf[row] += d * (sum[0] + sum[1]);
 
-        // combine the 4-bits from qs with the 5th bit
-        const int32_t x0 = ((((qs[i]     ) & mask) >> x_mv) | xh_0);
-        const int32_t x1 = ((((qs[i] >> 8) & mask) >> x_mv) | xh_1);
+            dh  += nb*sizeof(block_iq3_s)/2;
+            qs  += nb*sizeof(block_iq3_s);
+            qh  += nb*sizeof(block_iq3_s);
+            sc  += nb*sizeof(block_iq3_s);
+            signs += nb*sizeof(block_iq3_s);
+        }
 
-        reg[i/2][2*(i%2)+0] = d * x0 + m;
-        reg[i/2][2*(i%2)+1] = d * x1 + m;
+        y4 += 32 * 32;
     }
-}
-
-template <typename type4x4>
-void dequantize_q8_0(device const block_q8_0 *xb, short il, thread type4x4 & reg) {
-    device const int8_t * qs = ((device const int8_t *)xb->qs);
-    const half d = xb->d;
 
-    for (int i = 0; i < 16; i++) {
-        reg[i/4][i%4] = (qs[i + 16*il] * d);
+    for (int row = 0; row < N_DST; ++row) {
+        all_sum = simd_sum(sumf[row]);
+        if (tiisg == 0) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum * 0.5f;
+        }
     }
 }
 
-template <typename type4x4>
-void dequantize_q2_K(device const block_q2_K *xb, short il, thread type4x4 & reg) {
-    const float d = xb->d;
-    const float min = xb->dmin;
-    device const uint8_t * q = (device const uint8_t *)xb->qs;
-    float dl, ml;
-    uint8_t sc = xb->scales[il];
+[[host_name("kernel_mul_mv_iq3_s_f32")]]
+kernel void kernel_mul_mv_iq3_s_f32(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne11,
+        constant   int64_t & ne12,
+        constant  uint64_t & nb10,
+        constant  uint64_t & nb11,
+        constant  uint64_t & nb12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   uint    & r2,
+        constant   uint    & r3,
+        threadgroup int8_t * shared_values [[threadgroup(0)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
 
-#if QK_K == 256
-    q = q + 32*(il/8) + 16*(il&1);
-    il = (il/2)%4;
-#endif
-    half  coef = il>1 ? (il>2 ? 1/64.h : 1/16.h) : (il>0 ? 1/4.h : 1.h);
-    uchar mask = il>1 ? (il>2 ? 192    : 48)     : (il>0 ? 12    : 3);
-    dl = d * (sc & 0xF) * coef, ml = min * (sc >> 4);
-    for (int i = 0; i < 16; ++i) {
-        reg[i/4][i%4] = dl * (q[i] & mask) - ml;
-    }
+    kernel_mul_mv_iq3_s_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
 }
 
-template <typename type4x4>
-void dequantize_q3_K(device const block_q3_K *xb, short il, thread type4x4 & reg) {
-    const half d_all = xb->d;
-    device const uint8_t * q = (device const uint8_t *)xb->qs;
-    device const uint8_t * h = (device const uint8_t *)xb->hmask;
-    device const int8_t * scales = (device const int8_t *)xb->scales;
-
-#if QK_K == 256
-    q = q + 32 * (il/8) + 16 * (il&1);
-    h = h + 16 * (il&1);
-    uint8_t m = 1 << (il/2);
-    uint16_t kmask1 = (il/4)>1 ? ((il/4)>2 ? 192 : 48) : \
-                                 ((il/4)>0 ? 12  : 3);
-    uint16_t kmask2 = il/8 ? 0xF0 : 0x0F;
-    uint16_t scale_2 = scales[il%8], scale_1 = scales[8 + il%4];
-    int16_t  dl_int = (il/4)&1 ? (scale_2&kmask2) | ((scale_1&kmask1) << 2)
-                               : (scale_2&kmask2) | ((scale_1&kmask1) << 4);
-    float dl = il<8 ? d_all * (dl_int - 32.f) : d_all * (dl_int / 16.f - 32.f);
-    const float ml = 4.f * dl;
+void kernel_mul_mv_iq2_s_f32_impl(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   uint    & r2,
+        constant   uint    & r3,
+        threadgroup int8_t * shared_values [[threadgroup(0)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
 
-    il = (il/2) & 3;
-    const half    coef = il>1 ? (il>2 ? 1/64.h : 1/16.h) : (il>0 ? 1/4.h : 1.h);
-    const uint8_t mask = il>1 ? (il>2 ? 192    : 48)     : (il>0 ? 12    : 3);
-    dl *= coef;
+    const int nb = ne00/QK_K;
+    const int r0 = tgpig.x;
+    const int r1 = tgpig.y;
+    const int im = tgpig.z;
 
-    for (int i = 0; i < 16; ++i) {
-        reg[i/4][i%4] = dl * (q[i] & mask) - (h[i] & m ? 0 : ml);
-    }
-#else
-    float    kcoef = il&1 ? 1.f/16.f : 1.f;
-    uint16_t kmask = il&1 ? 0xF0     : 0x0F;
-    float    dl = d_all * ((scales[il/2] & kmask) * kcoef - 8);
-    float    coef = il>1 ? (il>2 ? 1/64.h : 1/16.h) : (il>0 ? 1/4.h : 1.h);
-    uint8_t  mask = il>1 ? (il>2 ? 192    : 48)     : (il>0 ? 12    : 3);
-    uint8_t  m = 1<<(il*2);
-    for (int i = 0; i < 16; ++i) {
-        reg[i/4][i%4] = coef * dl * ((q[i] & mask) - ((h[i%8] & (m * (1 + i/8))) ? 0 : 4.f/coef));
-    }
-#endif
-}
+    const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
+    const int ib_row = first_row * nb;
 
-static inline uchar2 get_scale_min_k4_just2(int j, int k, device const uchar * q) {
-    return j < 4 ? uchar2{uchar(q[j+0+k] & 63), uchar(q[j+4+k] & 63)}
-                 : uchar2{uchar((q[j+4+k] & 0xF) | ((q[j-4+k] & 0xc0) >> 2)), uchar((q[j+4+k] >> 4) | ((q[j-0+k] & 0xc0) >> 2))};
-}
+    const uint i12 = im%ne12;
+    const uint i13 = im/ne12;
 
-template <typename type4x4>
-void dequantize_q4_K(device const block_q4_K *xb, short il, thread type4x4 & reg) {
-    device const uchar * q = xb->qs;
+    const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
 
-#if QK_K == 256
-    short is = (il/4) * 2;
-    q = q + (il/4) * 32 + 16 * (il&1);
-    il = il & 3;
-    const uchar2 sc = get_scale_min_k4_just2(is, il/2, xb->scales);
-    const float d   = il < 2 ? xb->d : xb->d / 16.h;
-    const float min = xb->dmin;
-    const float dl = d * sc[0];
-    const float ml = min * sc[1];
-#else
-    q = q + 16 * (il&1);
-    device const uint8_t * s = xb->scales;
-    device const half2 * dh = (device const half2 *)xb->d;
-    const float2 d = (float2)dh[0];
-    const float dl = il<2 ? d[0] * (s[0]&0xF) : d[0] * (s[1]&0xF)/16.h;
-    const float ml = il<2 ? d[1] * (s[0]>>4)  : d[1] * (s[1]>>4);
-#endif
-    const ushort mask = il<2 ? 0x0F : 0xF0;
-    for (int i = 0; i < 16; ++i) {
-        reg[i/4][i%4] = dl * (q[i] & mask) - ml;
-    }
-}
+    device const block_iq2_s * x = (device const block_iq2_s *) src0 + ib_row + offset0;
+    device const float       * y = (device const float       *) src1 + r1*ne10 + im*ne00*ne1;
 
-template <typename type4x4>
-void dequantize_q5_K(device const block_q5_K *xb, short il, thread type4x4 & reg) {
-    device const uint8_t * q  = xb->qs;
-    device const uint8_t * qh = xb->qh;
+    float yl[32];
+    float sumf[N_DST]={0.f}, all_sum;
 
-#if QK_K == 256
-    short is = (il/4) * 2;
-    q  = q + 32 * (il/4) + 16 * (il&1);
-    qh = qh + 16 * (il&1);
-    uint8_t ul = 1 << (il/2);
-    il = il & 3;
-    const uchar2 sc = get_scale_min_k4_just2(is, il/2, xb->scales);
-    const float d = il < 2 ? xb->d : xb->d / 16.f;
-    const float min = xb->dmin;
-    const float dl = d * sc[0];
-    const float ml = min * sc[1];
+    const int nb32 = nb * (QK_K / 32);
 
-    const ushort mask  = il<2 ? 0x0F : 0xF0;
-    const float qh_val = il<2 ? 16.f : 256.f;
-    for (int i = 0; i < 16; ++i) {
-        reg[i/4][i%4] = dl * ((q[i] & mask) + (qh[i] & ul ? qh_val : 0)) - ml;
-    }
-#else
-    q = q + 16 * (il&1);
-    device const int8_t * s = xb->scales;
-    const float dl = xb->d * s[il];
-    uint8_t m = 1<<(il*2);
-    const float  coef = il<2 ? 1.f  : 1.f/16.f;
-    const ushort mask = il<2 ? 0x0F : 0xF0;
-    for (int i = 0; i < 16; ++i) {
-        reg[i/4][i%4] = coef * dl * ((q[i] & mask) - (qh[i%8] & (m*(1+i/8)) ? 0.f : 16.f/coef));
-    }
-#endif
-}
+    //threadgroup uint64_t * values = (threadgroup uint64_t *)shared_values;
+    //{
+    //    int nval = 32;
+    //    int pos  = (32*sgitg + tiisg)*nval;
+    //    for (int i = 0; i < nval; ++i) values[pos + i] = iq2s_grid[pos + i];
+    //    threadgroup_barrier(mem_flags::mem_threadgroup);
+    //}
 
-template <typename type4x4>
-void dequantize_q6_K(device const block_q6_K *xb, short il, thread type4x4 & reg) {
-    const half d_all = xb->d;
-    device const uint8_t * ql = (device const uint8_t *)xb->ql;
-    device const uint8_t * qh = (device const uint8_t *)xb->qh;
-    device const int8_t * scales = (device const int8_t *)xb->scales;
+    const int ix = tiisg;
 
-#if QK_K == 256
-    ql = ql + 64*(il/8) + 32*((il/2)&1) + 16*(il&1);
-    qh = qh + 32*(il/8) + 16*(il&1);
-    float sc = scales[(il%2) + 2 * ((il/2))];
-    il = (il/2) & 3;
-#else
-    ql = ql + 16 * (il&1);
-    float sc = scales[il];
-#endif
-    const uint16_t  kmask1 = il>1 ? (il>2 ? 192 : 48) : (il>0 ? 12 : 3);
-    const uint16_t  kmask2 = il>1 ? 0xF0              : 0x0F;
-    const float       coef = il>1 ? 1.f/16.f          : 1.f;
-    const float ml = d_all * sc * 32.f;
-    const float dl = d_all * sc * coef;
-    for (int i = 0; i < 16; ++i) {
-        const half q = il&1 ? ((ql[i] & kmask2) | ((qh[i] & kmask1) << 2))
-                            : ((ql[i] & kmask2) | ((qh[i] & kmask1) << 4));
-        reg[i/4][i%4] = dl * q - ml;
-    }
-}
+    device const float * y4 = y + 32 * ix;
 
-template <typename type4x4>
-void dequantize_iq2_xxs(device const block_iq2_xxs * xb, short il, thread type4x4 & reg) {
-    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
-    const float d = xb->d;
-    const int ib32 = il/2;
-    il = il%2;
-    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
-    // each block of 32 needs 2 uint32_t's for the quants & scale, so 4 uint16_t's.
-    device const uint16_t * q2 = xb->qs + 4*ib32;
-    const uint32_t aux32_g = q2[0] | (q2[1] << 16);
-    const uint32_t aux32_s = q2[2] | (q2[3] << 16);
-    thread const uint8_t * aux8 = (thread const uint8_t *)&aux32_g;
-    const float dl = d * (0.5f + (aux32_s >> 28)) * 0.25f;
-    constant uint8_t * grid = (constant uint8_t *)(iq2xxs_grid + aux8[2*il+0]);
-    uint8_t signs = ksigns_iq2xs[(aux32_s >> 14*il) & 127];
-    for (int i = 0; i < 8; ++i) {
-        reg[i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
-    }
-    grid = (constant uint8_t *)(iq2xxs_grid + aux8[2*il+1]);
-    signs = ksigns_iq2xs[(aux32_s >> (14*il+7)) & 127];
-    for (int i = 0; i < 8; ++i) {
-        reg[2+i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
-    }
-}
+    for (int ib32 = ix; ib32 < nb32; ib32 += 32) {
 
-template <typename type4x4>
-void dequantize_iq2_xs(device const block_iq2_xs * xb, short il, thread type4x4 & reg) {
-    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
-    const float d = xb->d;
-    const int ib32 = il/2;
-    il = il%2;
-    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
-    device const uint16_t * q2 = xb->qs + 4*ib32;
-    const float dl = d * (0.5f + ((xb->scales[ib32] >> 4*il) & 0xf)) * 0.25f;
-    constant uint8_t * grid = (constant uint8_t *)(iq2xs_grid + (q2[2*il+0] & 511));
-    uint8_t signs = ksigns_iq2xs[q2[2*il+0] >> 9];
-    for (int i = 0; i < 8; ++i) {
-        reg[i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
-    }
-    grid = (constant uint8_t *)(iq2xs_grid + (q2[2*il+1] & 511));
-    signs = ksigns_iq2xs[q2[2*il+1] >> 9];
-    for (int i = 0; i < 8; ++i) {
-        reg[2+i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
-    }
-}
+        for (int i = 0; i < 32; ++i) {
+            yl[i] = y4[i];
+        }
 
-template <typename type4x4>
-void dequantize_iq3_xxs(device const block_iq3_xxs * xb, short il, thread type4x4 & reg) {
-    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
-    const float d = xb->d;
-    const int ib32 = il/2;
-    il = il%2;
-    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
-    device const uint8_t * q3 = xb->qs + 8*ib32;
-    device const uint16_t * gas = (device const uint16_t *)(xb->qs + QK_K/4) + 2*ib32;
-    const uint32_t aux32 = gas[0] | (gas[1] << 16);
-    const float dl = d * (0.5f + (aux32 >> 28)) * 0.5f;
-    constant uint8_t * grid1 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+0]);
-    constant uint8_t * grid2 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+1]);
-    uint8_t signs = ksigns_iq2xs[(aux32 >> 14*il) & 127];
-    for (int i = 0; i < 4; ++i) {
-        reg[0][i] = dl * grid1[i] * (signs & kmask_iq2xs[i+0] ? -1.f : 1.f);
-        reg[1][i] = dl * grid2[i] * (signs & kmask_iq2xs[i+4] ? -1.f : 1.f);
+        const int ibl = ib32 / (QK_K / 32);
+        const int ib  = ib32 % (QK_K / 32);
+
+        device const block_iq2_s * xr = x + ibl;
+        device const uint8_t * qs = xr->qs + 4 * ib;
+        device const uint8_t * qh = xr->qh + ib;
+        device const uint8_t * sc = xr->scales + ib;
+        device const uint8_t * signs = qs + QK_K/8;
+        device const half * dh = &xr->d;
+
+        for (int row = 0; row < N_DST; row++) {
+
+            const float db = dh[0];
+            const float d1 = db * (0.5f + (sc[0] & 0xf));
+            const float d2 = db * (0.5f + (sc[0] >>  4));
+
+            float2 sum = {0};
+            for (int l = 0; l < 2; ++l) {
+                //const threadgroup uint8_t * grid1 = (const threadgroup uint8_t *)(values + (qs[l+0] | ((qh[0] << (8-2*l)) & 0x300)));
+                //const threadgroup uint8_t * grid2 = (const threadgroup uint8_t *)(values + (qs[l+2] | ((qh[0] << (4-2*l)) & 0x300)));
+                constant uint8_t * grid1 = (constant uint8_t *)(iq2s_grid + (qs[l+0] | ((qh[0] << (8-2*l)) & 0x300)));
+                constant uint8_t * grid2 = (constant uint8_t *)(iq2s_grid + (qs[l+2] | ((qh[0] << (4-2*l)) & 0x300)));
+                for (int j = 0; j < 8; ++j) {
+                    sum[0] += yl[8*l + j +  0] * grid1[j] * select(1, -1, signs[l+0] & kmask_iq2xs[j]);
+                    sum[1] += yl[8*l + j + 16] * grid2[j] * select(1, -1, signs[l+2] & kmask_iq2xs[j]);
+                }
+            }
+            sumf[row] += d1 * sum[0] + d2 * sum[1];
+
+            dh  += nb*sizeof(block_iq2_s)/2;
+            qs  += nb*sizeof(block_iq2_s);
+            qh  += nb*sizeof(block_iq2_s);
+            sc  += nb*sizeof(block_iq2_s);
+            signs += nb*sizeof(block_iq2_s);
+        }
+
+        y4 += 32 * 32;
     }
-    grid1 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+2]);
-    grid2 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+3]);
-    signs = ksigns_iq2xs[(aux32 >> (14*il+7)) & 127];
-    for (int i = 0; i < 4; ++i) {
-        reg[2][i] = dl * grid1[i] * (signs & kmask_iq2xs[i+0] ? -1.f : 1.f);
-        reg[3][i] = dl * grid2[i] * (signs & kmask_iq2xs[i+4] ? -1.f : 1.f);
+
+    for (int row = 0; row < N_DST; ++row) {
+        all_sum = simd_sum(sumf[row]);
+        if (tiisg == 0) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum * 0.25f;
+        }
     }
 }
 
-template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread float4x4 &)>
-kernel void kernel_get_rows(
+[[host_name("kernel_mul_mv_iq2_s_f32")]]
+kernel void kernel_mul_mv_iq2_s_f32(
         device const  void * src0,
-        device const  char * src1,
+        device const float * src1,
         device       float * dst,
         constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant  uint64_t & nb00,
         constant  uint64_t & nb01,
         constant  uint64_t & nb02,
         constant   int64_t & ne10,
+        constant   int64_t & ne11,
+        constant   int64_t & ne12,
         constant  uint64_t & nb10,
         constant  uint64_t & nb11,
-        constant  uint64_t & nb1,
-        constant  uint64_t & nb2,
-        uint3                tgpig[[threadgroup_position_in_grid]],
-        uint                 tiitg[[thread_index_in_threadgroup]],
-        uint3                tptg [[threads_per_threadgroup]]) {
-    //const int64_t i = tgpig;
-    //const int64_t r = ((device int32_t *) src1)[i];
-
-    const int64_t i10 = tgpig.x;
-    const int64_t i11 = tgpig.y;
-
-    const int64_t r = ((device int32_t *) ((device char *) src1 + i11*nb11 + i10*nb10))[0];
-
-    const int64_t i02 = i11;
+        constant  uint64_t & nb12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   uint    & r2,
+        constant   uint    & r3,
+        threadgroup int8_t * shared_values [[threadgroup(0)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
 
-    for (int64_t ind = tiitg; ind < ne00/16; ind += tptg.x) {
-        float4x4 temp;
-        dequantize_func(
-            ((device const block_q *) ((device char *) src0 + r*nb01 + i02*nb02)) + ind/nl, ind%nl, temp);
-        *(((device float4x4 *) ((device char *) dst + i11*nb2 + i10*nb1)) + ind) = temp;
-    }
+    kernel_mul_mv_iq2_s_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
 }
 
-kernel void kernel_get_rows_f32(
+void kernel_mul_mv_iq1_s_f32_impl(
         device const  void * src0,
-        device const  char * src1,
+        device const float * src1,
         device       float * dst,
         constant   int64_t & ne00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
         constant   int64_t & ne10,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb1,
-        constant  uint64_t & nb2,
-        uint3                tgpig[[threadgroup_position_in_grid]],
-        uint                 tiitg[[thread_index_in_threadgroup]],
-        uint3                tptg [[threads_per_threadgroup]]) {
-    const int64_t i10 = tgpig.x;
-    const int64_t i11 = tgpig.y;
-
-    const int64_t r = ((device int32_t *) ((device char *) src1 + i11*nb11 + i10*nb10))[0];
+        constant   int64_t & ne12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   uint    & r2,
+        constant   uint    & r3,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
 
-    const int64_t i02 = i11;
+    const int nb = ne00/QK_K;
+    const int r0 = tgpig.x;
+    const int r1 = tgpig.y;
+    const int im = tgpig.z;
 
-    for (int ind = tiitg; ind < ne00; ind += tptg.x) {
-        ((device float *) ((device char *) dst + i11*nb2 + i10*nb1))[ind] =
-            ((device float *) ((device char *) src0 + r*nb01 + i02*nb02))[ind];
-    }
-}
+    const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
+    const int ib_row = first_row * nb;
 
-kernel void kernel_get_rows_f16(
-        device const  void * src0,
-        device const  char * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant   int64_t & ne10,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb1,
-        constant  uint64_t & nb2,
-        uint3                tgpig[[threadgroup_position_in_grid]],
-        uint                 tiitg[[thread_index_in_threadgroup]],
-        uint3                tptg [[threads_per_threadgroup]]) {
-    const int64_t i10 = tgpig.x;
-    const int64_t i11 = tgpig.y;
+    const uint i12 = im%ne12;
+    const uint i13 = im/ne12;
 
-    const int64_t r = ((device int32_t *) ((device char *) src1 + i11*nb11 + i10*nb10))[0];
+    const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+    device const block_iq1_s * x = (device const block_iq1_s *) src0 + ib_row + offset0;
+    device const float       * y = (device const float       *) src1 + r1*ne10 + im*ne00*ne1;
 
-    const int64_t i02 = i11;
+    float yl[16];
+    float sumf[N_DST]={0.f}, all_sum;
 
-    for (int ind = tiitg; ind < ne00; ind += tptg.x) {
-        ((device float *) ((device char *) dst + i11*nb2 + i10*nb1))[ind] =
-            ((device half *) ((device char *) src0 + r*nb01 + i02*nb02))[ind];
-    }
-}
+    const int nb32 = nb * (QK_K / 32);
 
-kernel void kernel_get_rows_i32(
-        device const  void * src0,
-        device const  char * src1,
-        device     int32_t * dst,
-        constant   int64_t & ne00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant   int64_t & ne10,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb1,
-        constant  uint64_t & nb2,
-        uint3                tgpig[[threadgroup_position_in_grid]],
-        uint                 tiitg[[thread_index_in_threadgroup]],
-        uint3                tptg [[threads_per_threadgroup]]) {
-    const int64_t i10 = tgpig.x;
-    const int64_t i11 = tgpig.y;
+    const int ix = tiisg/2;
+    const int il = tiisg%2;
 
-    const int64_t r = ((device int32_t *) ((device char *) src1 + i11*nb11 + i10*nb10))[0];
+    device const float * y4 = y + 32 * ix + 16 * il;
 
-    const int64_t i02 = i11;
+    for (int ib32 = ix; ib32 < nb32; ib32 += 16) {
 
-    for (int ind = tiitg; ind < ne00; ind += tptg.x) {
-        ((device int32_t *) ((device char *) dst + i11*nb2 + i10*nb1))[ind] =
-            ((device int32_t *) ((device char *) src0 + r*nb01 + i02*nb02))[ind];
-    }
-}
+        for (int i = 0; i < 16; ++i) {
+            yl[i] = y4[i];
+        }
 
+        const int ibl = ib32 / (QK_K / 32);
+        const int ib  = ib32 % (QK_K / 32);
 
-#define BLOCK_SIZE_M 64 // 8 simdgroup matrices from matrix A
-#define BLOCK_SIZE_N 32 // 4 simdgroup matrices from matrix B
-#define BLOCK_SIZE_K 32
-#define THREAD_MAT_M 4 // each thread take 4 simdgroup matrices from matrix A
-#define THREAD_MAT_N 2 // each thread take 2 simdgroup matrices from matrix B
-#define THREAD_PER_BLOCK 128
-#define THREAD_PER_ROW 2 // 2 thread for each row in matrix A to load numbers
-#define THREAD_PER_COL 4 // 4 thread for each row in matrix B to load numbers
-#define SG_MAT_SIZE 64 // simdgroup matrix is of shape 8x8
-#define SG_MAT_ROW 8
+        device const block_iq1_s * xr = x + ibl;
+        device const uint8_t * qs = xr->qs + 4 * ib + 2 * il;
+        device const uint8_t * sc = xr->scales + 2 * ib + il;
+        device const half    * dh = &xr->d;
 
-// each block_q contains 16*nl weights
-template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread half4x4 &)>
-void kernel_mul_mm_impl(device const  uchar * src0,
-                        device const  uchar * src1,
-                        device        float * dst,
-                        constant    int64_t & ne00,
-                        constant    int64_t & ne02,
-                        constant   uint64_t & nb01,
-                        constant   uint64_t & nb02,
-                        constant    int64_t & ne12,
-                        constant   uint64_t & nb10,
-                        constant   uint64_t & nb11,
-                        constant   uint64_t & nb12,
-                        constant    int64_t & ne0,
-                        constant    int64_t & ne1,
-                        constant       uint & r2,
-                        constant       uint & r3,
-                        threadgroup   uchar * shared_memory [[threadgroup(0)]],
-                        uint3                 tgpig[[threadgroup_position_in_grid]],
-                        uint                  tiitg[[thread_index_in_threadgroup]],
-                        uint                  sgitg[[simdgroup_index_in_threadgroup]]) {
+        for (int row = 0; row < N_DST; row++) {
 
-    threadgroup half  * sa = (threadgroup half  *)(shared_memory);
-    threadgroup float * sb = (threadgroup float *)(shared_memory + 4096);
+            constant int8_t * grid1 = (constant int8_t *)(iq1s_grid + (qs[0] | ((sc[0] & 0x08) << 5)));
+            constant int8_t * grid2 = (constant int8_t *)(iq1s_grid + (qs[1] | ((sc[0] & 0x80) << 1)));
 
-    const uint r0 = tgpig.y;
-    const uint r1 = tgpig.x;
-    const uint im = tgpig.z;
+            float2 sum = {0};
+            for (int j = 0; j < 8; ++j) {
+                sum[0] += yl[j+ 0] * grid1[j];
+                sum[1] += yl[j+ 8] * grid2[j];
+            }
+            sumf[row] += (float)dh[0] * (sum[0] * (2*(sc[0] & 7) + 1) + sum[1] * (2*((sc[0] >> 4) & 7) + 1));
 
-    // if this block is of 64x32 shape or smaller
-    short n_rows = (ne0 - r0 * BLOCK_SIZE_M < BLOCK_SIZE_M) ? (ne0 - r0 * BLOCK_SIZE_M) : BLOCK_SIZE_M;
-    short n_cols = (ne1 - r1 * BLOCK_SIZE_N < BLOCK_SIZE_N) ? (ne1 - r1 * BLOCK_SIZE_N) : BLOCK_SIZE_N;
+            dh += nb*sizeof(block_iq1_s)/2;
+            qs += nb*sizeof(block_iq1_s);
+            sc += nb*sizeof(block_iq1_s);
+        }
 
-    // a thread shouldn't load data outside of the matrix
-    short thread_row = ((short)tiitg/THREAD_PER_ROW) < n_rows ? ((short)tiitg/THREAD_PER_ROW) : n_rows - 1;
-    short thread_col = ((short)tiitg/THREAD_PER_COL) < n_cols ? ((short)tiitg/THREAD_PER_COL) : n_cols - 1;
+        y4 += 16 * 32;
+    }
 
-    simdgroup_half8x8  ma[4];
-    simdgroup_float8x8 mb[2];
-    simdgroup_float8x8 c_res[8];
-    for (int i = 0; i < 8; i++){
-        c_res[i] = make_filled_simdgroup_matrix<float, 8>(0.f);
+    for (int row = 0; row < N_DST; ++row) {
+        all_sum = simd_sum(sumf[row]);
+        if (tiisg == 0) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum;
+        }
     }
+}
 
-    short il = (tiitg % THREAD_PER_ROW);
+constexpr constant static float kvalues_iq4nl_f[16] = {
+    -127.f, -104.f, -83.f, -65.f, -49.f, -35.f, -22.f, -10.f, 1.f, 13.f, 25.f, 38.f, 53.f, 69.f, 89.f, 113.f
+};
+
+void kernel_mul_mv_iq4_nl_f32_impl(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   uint    & r2,
+        constant   uint    & r3,
+        threadgroup float  * shared_values [[threadgroup(0)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    const int nb = ne00/QK4_NL;
+    const int r0 = tgpig.x;
+    const int r1 = tgpig.y;
+    const int im = tgpig.z;
+    const int first_row = (r0 * 2 + sgitg) * 2;
+    const int ib_row = first_row * nb;
 
     const uint i12 = im%ne12;
     const uint i13 = im/ne12;
 
-    uint   offset0 = (i12/r2)*nb02 + (i13/r3)*(nb02*ne02);
-    ushort offset1 = il/nl;
+    const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+    device const block_iq4_nl * x = (device const block_iq4_nl *) src0 + ib_row + offset0;
+    device const float        * y = (device const float        *) src1 + r1*ne10 + im*ne00*ne1;
 
-    device const block_q * x = (device const block_q *)(src0 + (r0 * BLOCK_SIZE_M + thread_row) * nb01 + offset0) + offset1;
-    device const float   * y = (device const float   *)(src1
-        + nb12 * im
-        + nb11 * (r1 * BLOCK_SIZE_N + thread_col)
-        + nb10 * (BLOCK_SIZE_K / THREAD_PER_COL * (tiitg % THREAD_PER_COL)));
+    const int ix = tiisg/2;  // 0...15
+    const int it = tiisg%2;  // 0 or 1
 
-    for (int loop_k = 0; loop_k < ne00; loop_k += BLOCK_SIZE_K) {
-        // load data and store to threadgroup memory
-        half4x4 temp_a;
-        dequantize_func(x, il, temp_a);
-        threadgroup_barrier(mem_flags::mem_threadgroup);
+    shared_values[tiisg] = kvalues_iq4nl_f[tiisg%16];
+    threadgroup_barrier(mem_flags::mem_threadgroup);
 
-        #pragma unroll(16)
-        for (int i = 0; i < 16; i++) {
-            *(sa + SG_MAT_SIZE * ((tiitg / THREAD_PER_ROW / 8) \
-            +                     (tiitg % THREAD_PER_ROW) * 16 + (i / 8) * 8) \
-            +                     (tiitg / THREAD_PER_ROW) % 8  + (i & 7) * 8) = temp_a[i/4][i%4];
-        }
+    float4 yl[4];
+    float sumf[2]={0.f}, all_sum;
 
-        *(threadgroup float2x4 *)(sb + (tiitg % THREAD_PER_COL) * 8 * 32 + 8 * (tiitg / THREAD_PER_COL)) = *((device float2x4 *)y);
+    device const float * yb = y + ix * QK4_NL + it * 8;
 
-        il = (il + 2 < nl) ? il + 2 : il % 2;
-        x  = (il < 2) ? x + (2+nl-1)/nl : x;
-        y += BLOCK_SIZE_K;
+    uint32_t aux32[2];
+    thread const uint8_t * q8 = (thread const uint8_t *)aux32;
 
-        threadgroup_barrier(mem_flags::mem_threadgroup);
+    float4 qf1, qf2;
 
-        // load matrices from threadgroup memory and conduct outer products
-        threadgroup half  * lsma = (sa + THREAD_MAT_M * SG_MAT_SIZE * (sgitg % 2));
-        threadgroup float * lsmb = (sb + THREAD_MAT_N * SG_MAT_SIZE * (sgitg / 2));
+    for (int ib = ix; ib < nb; ib += 16) {
 
-        #pragma unroll(4)
-        for (int ik = 0; ik < BLOCK_SIZE_K / 8; ik++) {
-            #pragma unroll(4)
-            for (int i = 0; i < 4; i++) {
-                simdgroup_load(ma[i],lsma + SG_MAT_SIZE * i);
-            }
-            simdgroup_barrier(mem_flags::mem_none);
-            #pragma unroll(2)
-            for (int i = 0; i < 2; i++) {
-                simdgroup_load(mb[i],lsmb + SG_MAT_SIZE * i);
-            }
+        device const float4 * y4 = (device const float4 *)yb;
+        yl[0] = y4[0]; yl[1] = y4[4]; yl[2] = y4[1]; yl[3] = y4[5];
 
-            lsma += BLOCK_SIZE_M / SG_MAT_ROW * SG_MAT_SIZE;
-            lsmb += BLOCK_SIZE_N / SG_MAT_ROW * SG_MAT_SIZE;
+        for (int row = 0; row < 2; ++row) {
+
+            device const block_iq4_nl & xb = x[row*nb + ib];
+            device const uint16_t * q4 = (device const uint16_t *)(xb.qs + 8*it);
+
+            float4 acc1 = {0.f}, acc2 = {0.f};
+
+            aux32[0] = q4[0] | (q4[1] << 16);
+            aux32[1] = (aux32[0] >> 4) & 0x0f0f0f0f;
+            aux32[0] &= 0x0f0f0f0f;
+            qf1 = {shared_values[q8[0]], shared_values[q8[1]], shared_values[q8[2]], shared_values[q8[3]]};
+            qf2 = {shared_values[q8[4]], shared_values[q8[5]], shared_values[q8[6]], shared_values[q8[7]]};
+            acc1 += yl[0] * qf1;
+            acc2 += yl[1] * qf2;
+
+            aux32[0] = q4[2] | (q4[3] << 16);
+            aux32[1] = (aux32[0] >> 4) & 0x0f0f0f0f;
+            aux32[0] &= 0x0f0f0f0f;
+            qf1 = {shared_values[q8[0]], shared_values[q8[1]], shared_values[q8[2]], shared_values[q8[3]]};
+            qf2 = {shared_values[q8[4]], shared_values[q8[5]], shared_values[q8[6]], shared_values[q8[7]]};
+            acc1 += yl[2] * qf1;
+            acc2 += yl[3] * qf2;
+
+            acc1 += acc2;
+
+            sumf[row] += (float)xb.d * (acc1[0] + acc1[1] + acc1[2] + acc1[3]);
 
-            #pragma unroll(8)
-            for (int i = 0; i < 8; i++){
-                simdgroup_multiply_accumulate(c_res[i], mb[i/4], ma[i%4], c_res[i]);
-            }
         }
+
+        yb += 16 * QK4_NL;
     }
 
-    if ((r0 + 1) * BLOCK_SIZE_M <= ne0 && (r1 + 1) * BLOCK_SIZE_N <= ne1) {
-        device float * C = dst + (BLOCK_SIZE_M * r0 + 32 * (sgitg &  1)) \
-                               + (BLOCK_SIZE_N * r1 + 16 * (sgitg >> 1)) * ne0 + im*ne1*ne0;
-        for (int i = 0; i < 8; i++) {
-            simdgroup_store(c_res[i], C + 8 * (i%4) + 8 * ne0 * (i/4), ne0);
-        }
-    } else {
-        // block is smaller than 64x32, we should avoid writing data outside of the matrix
-        threadgroup_barrier(mem_flags::mem_threadgroup);
-        threadgroup float * temp_str = ((threadgroup float *)shared_memory) \
-                                      + 32 * (sgitg&1) + (16 * (sgitg>>1)) * BLOCK_SIZE_M;
-        for (int i = 0; i < 8; i++) {
-            simdgroup_store(c_res[i], temp_str + 8 * (i%4) + 8 * BLOCK_SIZE_M * (i/4), BLOCK_SIZE_M);
-        }
-
-        threadgroup_barrier(mem_flags::mem_threadgroup);
-
-        device float * C = dst + (BLOCK_SIZE_M * r0) + (BLOCK_SIZE_N * r1) * ne0 + im*ne1*ne0;
-        if (sgitg == 0) {
-            for (int i = 0; i < n_rows; i++) {
-                for (int j = tiitg; j < n_cols; j += BLOCK_SIZE_N) {
-                    *(C + i + j * ne0) = *(temp_str + i + j * BLOCK_SIZE_M);
-                }
-            }
+    for (int row = 0; row < 2; ++row) {
+        all_sum = simd_sum(sumf[row]);
+        if (tiisg == 0) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum;
         }
     }
 }
 
-// same as kernel_mul_mm_impl, but src1 and dst are accessed via indices stored in src1ids
-template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread half4x4 &)>
-void kernel_mul_mm_id_impl(
-        device const  uchar * src0,
-        device const  uchar * src1,
-        thread        short * src1ids,
-        device        float * dst,
-        constant    int64_t & ne00,
-        constant    int64_t & ne02,
-        constant   uint64_t & nb01,
-        constant   uint64_t & nb02,
-        constant    int64_t & ne12,
-        constant   uint64_t & nb10,
-        constant   uint64_t & nb11,
-        constant   uint64_t & nb12,
-        constant    int64_t & ne0,
-                    int64_t   ne1,
-        constant       uint & r2,
-        constant       uint & r3,
-        threadgroup   uchar * shared_memory,
-        uint3                 tgpig[[threadgroup_position_in_grid]],
-        uint                  tiitg[[thread_index_in_threadgroup]],
-        uint                  sgitg[[simdgroup_index_in_threadgroup]]) {
+#if QK_K != 64
+void kernel_mul_mv_iq4_xs_f32_impl(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   uint    & r2,
+        constant   uint    & r3,
+        threadgroup float  * shared_values [[threadgroup(0)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
 
-    threadgroup half  * sa = (threadgroup half  *)(shared_memory);
-    threadgroup float * sb = (threadgroup float *)(shared_memory + 4096);
+    const int nb = ne00/QK_K;
+    const int r0 = tgpig.x;
+    const int r1 = tgpig.y;
+    const int im = tgpig.z;
+    const int first_row = (r0 * 2 + sgitg) * 2;
+    const int ib_row = first_row * nb;
 
-    const uint r0 = tgpig.y;
-    const uint r1 = tgpig.x;
-    const uint im = tgpig.z;
+    const uint i12 = im%ne12;
+    const uint i13 = im/ne12;
 
-    if (r1 * BLOCK_SIZE_N >= ne1) return;
+    const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+    device const block_iq4_xs * x = (device const block_iq4_xs *) src0 + ib_row + offset0;
+    device const float        * y = (device const float        *) src1 + r1*ne10 + im*ne00*ne1;
 
-    // if this block is of 64x32 shape or smaller
-    short n_rows = (ne0 - r0 * BLOCK_SIZE_M < BLOCK_SIZE_M) ? (ne0 - r0 * BLOCK_SIZE_M) : BLOCK_SIZE_M;
-    short n_cols = (ne1 - r1 * BLOCK_SIZE_N < BLOCK_SIZE_N) ? (ne1 - r1 * BLOCK_SIZE_N) : BLOCK_SIZE_N;
+    const int ix = tiisg/16;  // 0 or 1
+    const int it = tiisg%16;  // 0...15
+    const int ib = it/2;
+    const int il = it%2;
 
-    // a thread shouldn't load data outside of the matrix
-    short thread_row = ((short)tiitg/THREAD_PER_ROW) < n_rows ? ((short)tiitg/THREAD_PER_ROW) : n_rows - 1;
-    short thread_col = ((short)tiitg/THREAD_PER_COL) < n_cols ? ((short)tiitg/THREAD_PER_COL) : n_cols - 1;
+    shared_values[tiisg] = kvalues_iq4nl_f[tiisg%16];
+    threadgroup_barrier(mem_flags::mem_threadgroup);
 
-    simdgroup_half8x8  ma[4];
-    simdgroup_float8x8 mb[2];
-    simdgroup_float8x8 c_res[8];
-    for (int i = 0; i < 8; i++){
-        c_res[i] = make_filled_simdgroup_matrix<float, 8>(0.f);
-    }
+    float4 yl[4];
+    float sumf[2]={0.f}, all_sum;
 
-    short il = (tiitg % THREAD_PER_ROW);
+    device const float * yb = y + ix * QK_K + ib * 32 + il * 8;
 
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
+    uint32_t aux32[2];
+    thread const uint8_t * q8 = (thread const uint8_t *)aux32;
 
-    uint   offset0 = (i12/r2)*nb02 + (i13/r3)*(nb02*ne02);
-    ushort offset1 = il/nl;
+    float4 qf1, qf2;
 
-    device const block_q * x = (device const block_q *)(src0 + (r0 * BLOCK_SIZE_M + thread_row) * nb01 + offset0) + offset1;
-    device const float   * y = (device const float   *)(src1
-        + nb12 * im
-        + nb11 * src1ids[r1 * BLOCK_SIZE_N + thread_col]
-        + nb10 * (BLOCK_SIZE_K / THREAD_PER_COL * (tiitg % THREAD_PER_COL)));
+    for (int ibl = ix; ibl < nb; ibl += 2) {
 
-    for (int loop_k = 0; loop_k < ne00; loop_k += BLOCK_SIZE_K) {
-        // load data and store to threadgroup memory
-        half4x4 temp_a;
-        dequantize_func(x, il, temp_a);
-        threadgroup_barrier(mem_flags::mem_threadgroup);
+        device const float4 * y4 = (device const float4 *)yb;
+        yl[0] = y4[0]; yl[1] = y4[4]; yl[2] = y4[1]; yl[3] = y4[5];
 
-        for (int i = 0; i < 16; i++) {
-            *(sa + SG_MAT_SIZE * ((tiitg / THREAD_PER_ROW / 8) \
-            +                     (tiitg % THREAD_PER_ROW) * 16 + (i / 8) * 8) \
-            +                     (tiitg / THREAD_PER_ROW) % 8  + (i & 7) * 8) = temp_a[i/4][i%4];
-        }
+        for (int row = 0; row < 2; ++row) {
 
-        *(threadgroup float2x4 *)(sb + (tiitg % THREAD_PER_COL) * 8 * 32 + 8 * (tiitg / THREAD_PER_COL)) = *((device float2x4 *)y);
+            device const block_iq4_xs & xb = x[row*nb + ibl];
+            device const uint32_t * q4 = (device const uint32_t *)(xb.qs + 16*ib + 8*il);
 
-        il = (il + 2 < nl) ? il + 2 : il % 2;
-        x  = (il < 2) ? x + (2+nl-1)/nl : x;
-        y += BLOCK_SIZE_K;
+            float4 acc1 = {0.f}, acc2 = {0.f};
 
-        threadgroup_barrier(mem_flags::mem_threadgroup);
+            aux32[0] = q4[0] & 0x0f0f0f0f;
+            aux32[1] = (q4[0] >> 4) & 0x0f0f0f0f;
+            qf1 = {shared_values[q8[0]], shared_values[q8[1]], shared_values[q8[2]], shared_values[q8[3]]};
+            qf2 = {shared_values[q8[4]], shared_values[q8[5]], shared_values[q8[6]], shared_values[q8[7]]};
+            acc1 += yl[0] * qf1;
+            acc2 += yl[1] * qf2;
 
-        // load matrices from threadgroup memory and conduct outer products
-        threadgroup half  * lsma = (sa + THREAD_MAT_M * SG_MAT_SIZE * (sgitg % 2));
-        threadgroup float * lsmb = (sb + THREAD_MAT_N * SG_MAT_SIZE * (sgitg / 2));
+            aux32[0] = q4[1] & 0x0f0f0f0f;
+            aux32[1] = (q4[1] >> 4) & 0x0f0f0f0f;
+            qf1 = {shared_values[q8[0]], shared_values[q8[1]], shared_values[q8[2]], shared_values[q8[3]]};
+            qf2 = {shared_values[q8[4]], shared_values[q8[5]], shared_values[q8[6]], shared_values[q8[7]]};
+            acc1 += yl[2] * qf1;
+            acc2 += yl[3] * qf2;
 
-        for (int ik = 0; ik < BLOCK_SIZE_K / 8; ik++) {
-            for (int i = 0; i < 4; i++) {
-                simdgroup_load(ma[i],lsma + SG_MAT_SIZE * i);
-            }
-            simdgroup_barrier(mem_flags::mem_none);
-            for (int i = 0; i < 2; i++) {
-                simdgroup_load(mb[i],lsmb + SG_MAT_SIZE * i);
-            }
+            acc1 += acc2;
 
-            lsma += BLOCK_SIZE_M / SG_MAT_ROW * SG_MAT_SIZE;
-            lsmb += BLOCK_SIZE_N / SG_MAT_ROW * SG_MAT_SIZE;
+            const int ls = (((xb.scales_l[ib/2] >> 4*(ib%2)) & 0xf) | (((xb.scales_h >> 2*ib) & 3) << 4)) - 32;
+            sumf[row] += (float)xb.d * ls * (acc1[0] + acc1[1] + acc1[2] + acc1[3]);
 
-            for (int i = 0; i < 8; i++){
-                simdgroup_multiply_accumulate(c_res[i], mb[i/4], ma[i%4], c_res[i]);
-            }
         }
+
+        yb += 2 * QK_K;
+    }
+
+    for (int row = 0; row < 2; ++row) {
+        all_sum = simd_sum(sumf[row]);
+        if (tiisg == 0) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum;
+        }
+    }
+}
+#endif
+
+[[host_name("kernel_mul_mv_iq1_s_f32")]]
+kernel void kernel_mul_mv_iq1_s_f32(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne11,
+        constant   int64_t & ne12,
+        constant  uint64_t & nb10,
+        constant  uint64_t & nb11,
+        constant  uint64_t & nb12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   uint    & r2,
+        constant   uint    & r3,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    kernel_mul_mv_iq1_s_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, tgpig, tiisg, sgitg);
+}
+
+[[host_name("kernel_mul_mv_iq4_nl_f32")]]
+kernel void kernel_mul_mv_iq4_nl_f32(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne11,
+        constant   int64_t & ne12,
+        constant  uint64_t & nb10,
+        constant  uint64_t & nb11,
+        constant  uint64_t & nb12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   uint    & r2,
+        constant   uint    & r3,
+        threadgroup float * shared_values [[threadgroup(0)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint tiisg[[thread_index_in_simdgroup]],
+        uint sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    kernel_mul_mv_iq4_nl_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
+}
+
+[[host_name("kernel_mul_mv_iq4_xs_f32")]]
+kernel void kernel_mul_mv_iq4_xs_f32(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne11,
+        constant   int64_t & ne12,
+        constant  uint64_t & nb10,
+        constant  uint64_t & nb11,
+        constant  uint64_t & nb12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   uint    & r2,
+        constant   uint    & r3,
+        threadgroup float * shared_values [[threadgroup(0)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint tiisg[[thread_index_in_simdgroup]],
+        uint sgitg[[simdgroup_index_in_threadgroup]]) {
+
+#if QK_K == 64
+    kernel_mul_mv_iq4_nl_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
+#else
+    kernel_mul_mv_iq4_xs_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
+#endif
+}
+
+//============================= templates and their specializations =============================
+
+// NOTE: this is not dequantizing - we are simply fitting the template
+template <typename type4x4>
+void dequantize_f32(device const float4x4 * src, short il, thread type4x4 & reg) {
+    float4x4 temp = *(((device float4x4 *)src));
+    for (int i = 0; i < 16; i++){
+        reg[i/4][i%4] = temp[i/4][i%4];
+    }
+}
+
+template <typename type4x4>
+void dequantize_f16(device const half4x4 * src, short il, thread type4x4 & reg) {
+    half4x4 temp = *(((device half4x4 *)src));
+    for (int i = 0; i < 16; i++){
+        reg[i/4][i%4] = temp[i/4][i%4];
     }
+}
+
+template <typename type4x4>
+void dequantize_q4_0(device const block_q4_0 *xb, short il, thread type4x4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 1);
+    const float d1 = il ? (xb->d / 16.h) : xb->d;
+    const float d2 = d1 / 256.f;
+    const float md = -8.h * xb->d;
+    const ushort mask0 = il ? 0x00F0 : 0x000F;
+    const ushort mask1 = mask0 << 8;
+
+    for (int i=0;i<8;i++) {
+        reg[i/2][2*(i%2)+0] = d1 * (qs[i] & mask0) + md;
+        reg[i/2][2*(i%2)+1] = d2 * (qs[i] & mask1) + md;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q4_1(device const block_q4_1 *xb, short il, thread type4x4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 2);
+    const float d1 = il ? (xb->d / 16.h) : xb->d;
+    const float d2 = d1 / 256.f;
+    const float  m = xb->m;
+    const ushort mask0 = il ? 0x00F0 : 0x000F;
+    const ushort mask1 = mask0 << 8;
+
+    for (int i=0;i<8;i++) {
+        reg[i/2][2*(i%2)+0] = ((qs[i] & mask0) * d1) + m;
+        reg[i/2][2*(i%2)+1] = ((qs[i] & mask1) * d2) + m;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q5_0(device const block_q5_0 *xb, short il, thread type4x4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 3);
+    const float d = xb->d;
+    const float md = -16.h * xb->d;
+    const ushort mask = il ? 0x00F0 : 0x000F;
+
+    const uint32_t qh = *((device const uint32_t *)xb->qh);
+
+    const int x_mv = il ? 4 : 0;
+
+    const int gh_mv = il ? 12 : 0;
+    const int gh_bk = il ?  0 : 4;
+
+    for (int i = 0; i < 8; i++) {
+        // extract the 5-th bits for x0 and x1
+        const uint8_t xh_0 = ((qh >> (gh_mv + 2*i  )) << gh_bk) & 0x10;
+        const uint8_t xh_1 = ((qh >> (gh_mv + 2*i+1)) << gh_bk) & 0x10;
+
+        // combine the 4-bits from qs with the 5th bit
+        const int32_t x0 = ((((qs[i]     ) & mask) >> x_mv) | xh_0);
+        const int32_t x1 = ((((qs[i] >> 8) & mask) >> x_mv) | xh_1);
+
+        reg[i/2][2*(i%2)+0] = d * x0 + md;
+        reg[i/2][2*(i%2)+1] = d * x1 + md;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q5_1(device const block_q5_1 *xb, short il, thread type4x4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 4);
+    const float d = xb->d;
+    const float m = xb->m;
+    const ushort mask = il ? 0x00F0 : 0x000F;
+
+    const uint32_t qh = *((device const uint32_t *)xb->qh);
+
+    const int x_mv = il ? 4 : 0;
+
+    const int gh_mv = il ? 12 : 0;
+    const int gh_bk = il ?  0 : 4;
+
+    for (int i = 0; i < 8; i++) {
+        // extract the 5-th bits for x0 and x1
+        const uint8_t xh_0 = ((qh >> (gh_mv + 2*i  )) << gh_bk) & 0x10;
+        const uint8_t xh_1 = ((qh >> (gh_mv + 2*i+1)) << gh_bk) & 0x10;
+
+        // combine the 4-bits from qs with the 5th bit
+        const int32_t x0 = ((((qs[i]     ) & mask) >> x_mv) | xh_0);
+        const int32_t x1 = ((((qs[i] >> 8) & mask) >> x_mv) | xh_1);
+
+        reg[i/2][2*(i%2)+0] = d * x0 + m;
+        reg[i/2][2*(i%2)+1] = d * x1 + m;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q8_0(device const block_q8_0 *xb, short il, thread type4x4 & reg) {
+    device const int8_t * qs = ((device const int8_t *)xb->qs);
+    const half d = xb->d;
+
+    for (int i = 0; i < 16; i++) {
+        reg[i/4][i%4] = (qs[i + 16*il] * d);
+    }
+}
+
+template <typename type4x4>
+void dequantize_q2_K(device const block_q2_K *xb, short il, thread type4x4 & reg) {
+    const float d = xb->d;
+    const float min = xb->dmin;
+    device const uint8_t * q = (device const uint8_t *)xb->qs;
+    float dl, ml;
+    uint8_t sc = xb->scales[il];
+
+#if QK_K == 256
+    q = q + 32*(il/8) + 16*(il&1);
+    il = (il/2)%4;
+#endif
+    half  coef = il>1 ? (il>2 ? 1/64.h : 1/16.h) : (il>0 ? 1/4.h : 1.h);
+    uchar mask = il>1 ? (il>2 ? 192    : 48)     : (il>0 ? 12    : 3);
+    dl = d * (sc & 0xF) * coef, ml = min * (sc >> 4);
+    for (int i = 0; i < 16; ++i) {
+        reg[i/4][i%4] = dl * (q[i] & mask) - ml;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q3_K(device const block_q3_K *xb, short il, thread type4x4 & reg) {
+    const half d_all = xb->d;
+    device const uint8_t * q = (device const uint8_t *)xb->qs;
+    device const uint8_t * h = (device const uint8_t *)xb->hmask;
+    device const int8_t * scales = (device const int8_t *)xb->scales;
+
+#if QK_K == 256
+    q = q + 32 * (il/8) + 16 * (il&1);
+    h = h + 16 * (il&1);
+    uint8_t m = 1 << (il/2);
+    uint16_t kmask1 = (il/4)>1 ? ((il/4)>2 ? 192 : 48) : \
+                                 ((il/4)>0 ? 12  : 3);
+    uint16_t kmask2 = il/8 ? 0xF0 : 0x0F;
+    uint16_t scale_2 = scales[il%8], scale_1 = scales[8 + il%4];
+    int16_t  dl_int = (il/4)&1 ? (scale_2&kmask2) | ((scale_1&kmask1) << 2)
+                               : (scale_2&kmask2) | ((scale_1&kmask1) << 4);
+    float dl = il<8 ? d_all * (dl_int - 32.f) : d_all * (dl_int / 16.f - 32.f);
+    const float ml = 4.f * dl;
+
+    il = (il/2) & 3;
+    const half    coef = il>1 ? (il>2 ? 1/64.h : 1/16.h) : (il>0 ? 1/4.h : 1.h);
+    const uint8_t mask = il>1 ? (il>2 ? 192    : 48)     : (il>0 ? 12    : 3);
+    dl *= coef;
+
+    for (int i = 0; i < 16; ++i) {
+        reg[i/4][i%4] = dl * (q[i] & mask) - (h[i] & m ? 0 : ml);
+    }
+#else
+    float    kcoef = il&1 ? 1.f/16.f : 1.f;
+    uint16_t kmask = il&1 ? 0xF0     : 0x0F;
+    float    dl = d_all * ((scales[il/2] & kmask) * kcoef - 8);
+    float    coef = il>1 ? (il>2 ? 1/64.h : 1/16.h) : (il>0 ? 1/4.h : 1.h);
+    uint8_t  mask = il>1 ? (il>2 ? 192    : 48)     : (il>0 ? 12    : 3);
+    uint8_t  m = 1<<(il*2);
+    for (int i = 0; i < 16; ++i) {
+        reg[i/4][i%4] = coef * dl * ((q[i] & mask) - ((h[i%8] & (m * (1 + i/8))) ? 0 : 4.f/coef));
+    }
+#endif
+}
+
+static inline uchar2 get_scale_min_k4_just2(int j, int k, device const uchar * q) {
+    return j < 4 ? uchar2{uchar(q[j+0+k] & 63), uchar(q[j+4+k] & 63)}
+                 : uchar2{uchar((q[j+4+k] & 0xF) | ((q[j-4+k] & 0xc0) >> 2)), uchar((q[j+4+k] >> 4) | ((q[j-0+k] & 0xc0) >> 2))};
+}
+
+template <typename type4x4>
+void dequantize_q4_K(device const block_q4_K *xb, short il, thread type4x4 & reg) {
+    device const uchar * q = xb->qs;
+
+#if QK_K == 256
+    short is = (il/4) * 2;
+    q = q + (il/4) * 32 + 16 * (il&1);
+    il = il & 3;
+    const uchar2 sc = get_scale_min_k4_just2(is, il/2, xb->scales);
+    const float d   = il < 2 ? xb->d : xb->d / 16.h;
+    const float min = xb->dmin;
+    const float dl = d * sc[0];
+    const float ml = min * sc[1];
+#else
+    (void) get_scale_min_k4_just2;
+
+    q = q + 16 * (il&1);
+    device const uint8_t * s = xb->scales;
+    device const half2 * dh = (device const half2 *)xb->d;
+    const float2 d = (float2)dh[0];
+    const float dl = il<2 ? d[0] * (s[0]&0xF) : d[0] * (s[1]&0xF)/16.h;
+    const float ml = il<2 ? d[1] * (s[0]>>4)  : d[1] * (s[1]>>4);
+#endif
+    const ushort mask = il<2 ? 0x0F : 0xF0;
+    for (int i = 0; i < 16; ++i) {
+        reg[i/4][i%4] = dl * (q[i] & mask) - ml;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q5_K(device const block_q5_K *xb, short il, thread type4x4 & reg) {
+    device const uint8_t * q  = xb->qs;
+    device const uint8_t * qh = xb->qh;
+
+#if QK_K == 256
+    short is = (il/4) * 2;
+    q  = q + 32 * (il/4) + 16 * (il&1);
+    qh = qh + 16 * (il&1);
+    uint8_t ul = 1 << (il/2);
+    il = il & 3;
+    const uchar2 sc = get_scale_min_k4_just2(is, il/2, xb->scales);
+    const float d = il < 2 ? xb->d : xb->d / 16.f;
+    const float min = xb->dmin;
+    const float dl = d * sc[0];
+    const float ml = min * sc[1];
+
+    const ushort mask  = il<2 ? 0x0F : 0xF0;
+    const float qh_val = il<2 ? 16.f : 256.f;
+    for (int i = 0; i < 16; ++i) {
+        reg[i/4][i%4] = dl * ((q[i] & mask) + (qh[i] & ul ? qh_val : 0)) - ml;
+    }
+#else
+    q = q + 16 * (il&1);
+    device const int8_t * s = xb->scales;
+    const float dl = xb->d * s[il];
+    uint8_t m = 1<<(il*2);
+    const float  coef = il<2 ? 1.f  : 1.f/16.f;
+    const ushort mask = il<2 ? 0x0F : 0xF0;
+    for (int i = 0; i < 16; ++i) {
+        reg[i/4][i%4] = coef * dl * ((q[i] & mask) - (qh[i%8] & (m*(1+i/8)) ? 0.f : 16.f/coef));
+    }
+#endif
+}
+
+template <typename type4x4>
+void dequantize_q6_K(device const block_q6_K *xb, short il, thread type4x4 & reg) {
+    const half d_all = xb->d;
+    device const uint8_t * ql = (device const uint8_t *)xb->ql;
+    device const uint8_t * qh = (device const uint8_t *)xb->qh;
+    device const int8_t * scales = (device const int8_t *)xb->scales;
+
+#if QK_K == 256
+    ql = ql + 64*(il/8) + 32*((il/2)&1) + 16*(il&1);
+    qh = qh + 32*(il/8) + 16*(il&1);
+    float sc = scales[(il%2) + 2 * ((il/2))];
+    il = (il/2) & 3;
+#else
+    ql = ql + 16 * (il&1);
+    float sc = scales[il];
+#endif
+    const uint16_t  kmask1 = il>1 ? (il>2 ? 192 : 48) : (il>0 ? 12 : 3);
+    const uint16_t  kmask2 = il>1 ? 0xF0              : 0x0F;
+    const float       coef = il>1 ? 1.f/16.f          : 1.f;
+    const float ml = d_all * sc * 32.f;
+    const float dl = d_all * sc * coef;
+    for (int i = 0; i < 16; ++i) {
+        const half q = il&1 ? ((ql[i] & kmask2) | ((qh[i] & kmask1) << 2))
+                            : ((ql[i] & kmask2) | ((qh[i] & kmask1) << 4));
+        reg[i/4][i%4] = dl * q - ml;
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq2_xxs(device const block_iq2_xxs * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const float d = xb->d;
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    // each block of 32 needs 2 uint32_t's for the quants & scale, so 4 uint16_t's.
+    device const uint16_t * q2 = xb->qs + 4*ib32;
+    const uint32_t aux32_g = q2[0] | (q2[1] << 16);
+    const uint32_t aux32_s = q2[2] | (q2[3] << 16);
+    thread const uint8_t * aux8 = (thread const uint8_t *)&aux32_g;
+    const float dl = d * (0.5f + (aux32_s >> 28)) * 0.25f;
+    constant uint8_t * grid = (constant uint8_t *)(iq2xxs_grid + aux8[2*il+0]);
+    uint8_t signs = ksigns_iq2xs[(aux32_s >> 14*il) & 127];
+    for (int i = 0; i < 8; ++i) {
+        reg[i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
+    }
+    grid = (constant uint8_t *)(iq2xxs_grid + aux8[2*il+1]);
+    signs = ksigns_iq2xs[(aux32_s >> (14*il+7)) & 127];
+    for (int i = 0; i < 8; ++i) {
+        reg[2+i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq2_xs(device const block_iq2_xs * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const float d = xb->d;
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    device const uint16_t * q2 = xb->qs + 4*ib32;
+    const float dl = d * (0.5f + ((xb->scales[ib32] >> 4*il) & 0xf)) * 0.25f;
+    constant uint8_t * grid = (constant uint8_t *)(iq2xs_grid + (q2[2*il+0] & 511));
+    uint8_t signs = ksigns_iq2xs[q2[2*il+0] >> 9];
+    for (int i = 0; i < 8; ++i) {
+        reg[i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
+    }
+    grid = (constant uint8_t *)(iq2xs_grid + (q2[2*il+1] & 511));
+    signs = ksigns_iq2xs[q2[2*il+1] >> 9];
+    for (int i = 0; i < 8; ++i) {
+        reg[2+i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq3_xxs(device const block_iq3_xxs * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const float d = xb->d;
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    device const uint8_t * q3 = xb->qs + 8*ib32;
+    device const uint16_t * gas = (device const uint16_t *)(xb->qs + QK_K/4) + 2*ib32;
+    const uint32_t aux32 = gas[0] | (gas[1] << 16);
+    const float dl = d * (0.5f + (aux32 >> 28)) * 0.5f;
+    constant uint8_t * grid1 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+0]);
+    constant uint8_t * grid2 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+1]);
+    uint8_t signs = ksigns_iq2xs[(aux32 >> 14*il) & 127];
+    for (int i = 0; i < 4; ++i) {
+        reg[0][i] = dl * grid1[i] * (signs & kmask_iq2xs[i+0] ? -1.f : 1.f);
+        reg[1][i] = dl * grid2[i] * (signs & kmask_iq2xs[i+4] ? -1.f : 1.f);
+    }
+    grid1 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+2]);
+    grid2 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+3]);
+    signs = ksigns_iq2xs[(aux32 >> (14*il+7)) & 127];
+    for (int i = 0; i < 4; ++i) {
+        reg[2][i] = dl * grid1[i] * (signs & kmask_iq2xs[i+0] ? -1.f : 1.f);
+        reg[3][i] = dl * grid2[i] * (signs & kmask_iq2xs[i+4] ? -1.f : 1.f);
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq3_s(device const block_iq3_s * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const float d = xb->d;
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    device const uint8_t * qs = xb->qs + 8*ib32;
+    device const uint8_t * signs = xb->signs + 4*ib32 + 2*il;
+    const uint8_t qh = xb->qh[ib32] >> 4*il;
+    const float dl = d * (0.5f + ((xb->scales[ib32/2] >> 4*(ib32%2)) & 0xf)) * 0.5f;
+    constant uint8_t * grid1 = (constant uint8_t *)(iq3xs_grid + (qs[4*il+0] | ((qh << 8) & 256)));
+    constant uint8_t * grid2 = (constant uint8_t *)(iq3xs_grid + (qs[4*il+1] | ((qh << 7) & 256)));
+    for (int i = 0; i < 4; ++i) {
+        reg[0][i] = dl * grid1[i] * select(1, -1, signs[0] & kmask_iq2xs[i+0]);
+        reg[1][i] = dl * grid2[i] * select(1, -1, signs[0] & kmask_iq2xs[i+4]);
+    }
+    grid1 = (constant uint8_t *)(iq3xs_grid + (qs[4*il+2] | ((qh << 6) & 256)));
+    grid2 = (constant uint8_t *)(iq3xs_grid + (qs[4*il+3] | ((qh << 5) & 256)));
+    for (int i = 0; i < 4; ++i) {
+        reg[2][i] = dl * grid1[i] * select(1, -1, signs[1] & kmask_iq2xs[i+0]);
+        reg[3][i] = dl * grid2[i] * select(1, -1, signs[1] & kmask_iq2xs[i+4]);
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq2_s(device const block_iq2_s * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const float d = xb->d;
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    device const uint8_t * qs = xb->qs + 4*ib32 + 2*il;
+    device const uint8_t * signs = qs + QK_K/8;
+    const uint8_t qh = xb->qh[ib32] >> 4*il;
+    const float dl = d * (0.5f + ((xb->scales[ib32] >> 4*il) & 0xf)) * 0.25f;
+    constant uint8_t * grid1 = (constant uint8_t *)(iq2s_grid + (qs[0] | ((qh << 8) & 0x300)));
+    constant uint8_t * grid2 = (constant uint8_t *)(iq2s_grid + (qs[1] | ((qh << 6) & 0x300)));
+    for (int i = 0; i < 8; ++i) {
+        reg[i/4+0][i%4] = dl * grid1[i] * select(1, -1, signs[0] & kmask_iq2xs[i]);
+        reg[i/4+2][i%4] = dl * grid2[i] * select(1, -1, signs[1] & kmask_iq2xs[i]);
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq1_s(device const block_iq1_s * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const float d = xb->d;
+    device const uint8_t * qs = xb->qs + 2*il;
+    device const uint8_t * sc = xb->scales + il;
+    const float dl1 = d * (2*(sc[0] & 7) + 1);
+    const float dl2 = d * (2*((sc[0] >> 4) & 7) + 1);
+    constant int8_t * grid1 = (constant int8_t *)(iq1s_grid + (qs[0] | ((sc[0] & 0x08) << 5)));
+    constant int8_t * grid2 = (constant int8_t *)(iq1s_grid + (qs[1] | ((sc[0] & 0x80) << 1)));
+    for (int i = 0; i < 8; ++i) {
+        reg[i/4+0][i%4] = dl1 * grid1[i];
+        reg[i/4+2][i%4] = dl2 * grid2[i];
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq4_nl(device const block_iq4_nl * xb, short il, thread type4x4 & reg) {
+    device const uint16_t * q4 = (device const uint16_t *)xb->qs;
+    const float d = xb->d;
+    uint32_t aux32;
+    thread const uint8_t * q8 = (thread const uint8_t *)&aux32;
+    for (int i = 0; i < 4; ++i) {
+        aux32 = ((q4[2*i] | (q4[2*i+1] << 16)) >> 4*il) & 0x0f0f0f0f;
+        reg[i][0] = d * kvalues_iq4nl_f[q8[0]];
+        reg[i][1] = d * kvalues_iq4nl_f[q8[1]];
+        reg[i][2] = d * kvalues_iq4nl_f[q8[2]];
+        reg[i][3] = d * kvalues_iq4nl_f[q8[3]];
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq4_xs(device const block_iq4_xs * xb, short il, thread type4x4 & reg) {
+#if QK_K == 64
+    dequantize_iq4_nl(xb, il, reg);
+#else
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    device const uint32_t * q4 = (device const uint32_t *)xb->qs + 4*ib32;
+    const int ls = ((xb->scales_l[ib32/2] >> 4*(ib32%2)) & 0xf) | (((xb->scales_h >> 2*ib32) & 3) << 4);
+    const float d = (float)xb->d * (ls - 32);
+    uint32_t aux32;
+    thread const uint8_t * q8 = (thread const uint8_t *)&aux32;
+    for (int i = 0; i < 4; ++i) {
+        aux32 = (q4[i] >> 4*il) & 0x0f0f0f0f;
+        reg[i][0] = d * kvalues_iq4nl_f[q8[0]];
+        reg[i][1] = d * kvalues_iq4nl_f[q8[1]];
+        reg[i][2] = d * kvalues_iq4nl_f[q8[2]];
+        reg[i][3] = d * kvalues_iq4nl_f[q8[3]];
+    }
+#endif
+}
+
+template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread float4x4 &)>
+kernel void kernel_get_rows(
+        device const  void * src0,
+        device const  char * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant   int64_t & ne10,
+        constant  uint64_t & nb10,
+        constant  uint64_t & nb11,
+        constant  uint64_t & nb1,
+        constant  uint64_t & nb2,
+        uint3                tgpig[[threadgroup_position_in_grid]],
+        uint                 tiitg[[thread_index_in_threadgroup]],
+        uint3                tptg [[threads_per_threadgroup]]) {
+    //const int64_t i = tgpig;
+    //const int64_t r = ((device int32_t *) src1)[i];
+
+    const int64_t i10 = tgpig.x;
+    const int64_t i11 = tgpig.y;
+
+    const int64_t r = ((device int32_t *) ((device char *) src1 + i11*nb11 + i10*nb10))[0];
+
+    const int64_t i02 = i11;
+
+    for (int64_t ind = tiitg; ind < ne00/16; ind += tptg.x) {
+        float4x4 temp;
+        dequantize_func(
+            ((device const block_q *) ((device char *) src0 + r*nb01 + i02*nb02)) + ind/nl, ind%nl, temp);
+        *(((device float4x4 *) ((device char *) dst + i11*nb2 + i10*nb1)) + ind) = temp;
+    }
+}
+
+kernel void kernel_get_rows_f32(
+        device const  void * src0,
+        device const  char * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant   int64_t & ne10,
+        constant  uint64_t & nb10,
+        constant  uint64_t & nb11,
+        constant  uint64_t & nb1,
+        constant  uint64_t & nb2,
+        uint3                tgpig[[threadgroup_position_in_grid]],
+        uint                 tiitg[[thread_index_in_threadgroup]],
+        uint3                tptg [[threads_per_threadgroup]]) {
+    const int64_t i10 = tgpig.x;
+    const int64_t i11 = tgpig.y;
+
+    const int64_t r = ((device int32_t *) ((device char *) src1 + i11*nb11 + i10*nb10))[0];
+
+    const int64_t i02 = i11;
+
+    for (int ind = tiitg; ind < ne00; ind += tptg.x) {
+        ((device float *) ((device char *) dst + i11*nb2 + i10*nb1))[ind] =
+            ((device float *) ((device char *) src0 + r*nb01 + i02*nb02))[ind];
+    }
+}
+
+kernel void kernel_get_rows_f16(
+        device const  void * src0,
+        device const  char * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant   int64_t & ne10,
+        constant  uint64_t & nb10,
+        constant  uint64_t & nb11,
+        constant  uint64_t & nb1,
+        constant  uint64_t & nb2,
+        uint3                tgpig[[threadgroup_position_in_grid]],
+        uint                 tiitg[[thread_index_in_threadgroup]],
+        uint3                tptg [[threads_per_threadgroup]]) {
+    const int64_t i10 = tgpig.x;
+    const int64_t i11 = tgpig.y;
+
+    const int64_t r = ((device int32_t *) ((device char *) src1 + i11*nb11 + i10*nb10))[0];
+
+    const int64_t i02 = i11;
+
+    for (int ind = tiitg; ind < ne00; ind += tptg.x) {
+        ((device float *) ((device char *) dst + i11*nb2 + i10*nb1))[ind] =
+            ((device half *) ((device char *) src0 + r*nb01 + i02*nb02))[ind];
+    }
+}
+
+kernel void kernel_get_rows_i32(
+        device const  void * src0,
+        device const  char * src1,
+        device     int32_t * dst,
+        constant   int64_t & ne00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant   int64_t & ne10,
+        constant  uint64_t & nb10,
+        constant  uint64_t & nb11,
+        constant  uint64_t & nb1,
+        constant  uint64_t & nb2,
+        uint3                tgpig[[threadgroup_position_in_grid]],
+        uint                 tiitg[[thread_index_in_threadgroup]],
+        uint3                tptg [[threads_per_threadgroup]]) {
+    const int64_t i10 = tgpig.x;
+    const int64_t i11 = tgpig.y;
+
+    const int64_t r = ((device int32_t *) ((device char *) src1 + i11*nb11 + i10*nb10))[0];
+
+    const int64_t i02 = i11;
+
+    for (int ind = tiitg; ind < ne00; ind += tptg.x) {
+        ((device int32_t *) ((device char *) dst + i11*nb2 + i10*nb1))[ind] =
+            ((device int32_t *) ((device char *) src0 + r*nb01 + i02*nb02))[ind];
+    }
+}
+
+
+#define BLOCK_SIZE_M 64 // 8 simdgroup matrices from matrix A
+#define BLOCK_SIZE_N 32 // 4 simdgroup matrices from matrix B
+#define BLOCK_SIZE_K 32
+#define THREAD_MAT_M 4 // each thread take 4 simdgroup matrices from matrix A
+#define THREAD_MAT_N 2 // each thread take 2 simdgroup matrices from matrix B
+#define THREAD_PER_BLOCK 128
+#define THREAD_PER_ROW 2 // 2 thread for each row in matrix A to load numbers
+#define THREAD_PER_COL 4 // 4 thread for each row in matrix B to load numbers
+#define SG_MAT_SIZE 64 // simdgroup matrix is of shape 8x8
+#define SG_MAT_ROW 8
+
+// each block_q contains 16*nl weights
+template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread half4x4 &)>
+void kernel_mul_mm_impl(device const  uchar * src0,
+                        device const  uchar * src1,
+                        device        float * dst,
+                        constant    int64_t & ne00,
+                        constant    int64_t & ne02,
+                        constant   uint64_t & nb01,
+                        constant   uint64_t & nb02,
+                        constant    int64_t & ne12,
+                        constant   uint64_t & nb10,
+                        constant   uint64_t & nb11,
+                        constant   uint64_t & nb12,
+                        constant    int64_t & ne0,
+                        constant    int64_t & ne1,
+                        constant       uint & r2,
+                        constant       uint & r3,
+                        threadgroup   uchar * shared_memory [[threadgroup(0)]],
+                        uint3                 tgpig[[threadgroup_position_in_grid]],
+                        uint                  tiitg[[thread_index_in_threadgroup]],
+                        uint                  sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    threadgroup half  * sa = (threadgroup half  *)(shared_memory);
+    threadgroup float * sb = (threadgroup float *)(shared_memory + 4096);
+
+    const uint r0 = tgpig.y;
+    const uint r1 = tgpig.x;
+    const uint im = tgpig.z;
+
+    // if this block is of 64x32 shape or smaller
+    short n_rows = (ne0 - r0 * BLOCK_SIZE_M < BLOCK_SIZE_M) ? (ne0 - r0 * BLOCK_SIZE_M) : BLOCK_SIZE_M;
+    short n_cols = (ne1 - r1 * BLOCK_SIZE_N < BLOCK_SIZE_N) ? (ne1 - r1 * BLOCK_SIZE_N) : BLOCK_SIZE_N;
+
+    // a thread shouldn't load data outside of the matrix
+    short thread_row = ((short)tiitg/THREAD_PER_ROW) < n_rows ? ((short)tiitg/THREAD_PER_ROW) : n_rows - 1;
+    short thread_col = ((short)tiitg/THREAD_PER_COL) < n_cols ? ((short)tiitg/THREAD_PER_COL) : n_cols - 1;
+
+    simdgroup_half8x8  ma[4];
+    simdgroup_float8x8 mb[2];
+    simdgroup_float8x8 c_res[8];
+    for (int i = 0; i < 8; i++){
+        c_res[i] = make_filled_simdgroup_matrix<float, 8>(0.f);
+    }
+
+    short il = (tiitg % THREAD_PER_ROW);
+
+    const uint i12 = im%ne12;
+    const uint i13 = im/ne12;
+
+    uint   offset0 = (i12/r2)*nb02 + (i13/r3)*(nb02*ne02);
+    ushort offset1 = il/nl;
+
+    device const block_q * x = (device const block_q *)(src0 + (r0 * BLOCK_SIZE_M + thread_row) * nb01 + offset0) + offset1;
+    device const float   * y = (device const float   *)(src1
+        + nb12 * im
+        + nb11 * (r1 * BLOCK_SIZE_N + thread_col)
+        + nb10 * (BLOCK_SIZE_K / THREAD_PER_COL * (tiitg % THREAD_PER_COL)));
+
+    for (int loop_k = 0; loop_k < ne00; loop_k += BLOCK_SIZE_K) {
+        // load data and store to threadgroup memory
+        half4x4 temp_a;
+        dequantize_func(x, il, temp_a);
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        #pragma unroll(16)
+        for (int i = 0; i < 16; i++) {
+            *(sa + SG_MAT_SIZE * ((tiitg / THREAD_PER_ROW / 8) \
+            +                     (tiitg % THREAD_PER_ROW) * 16 + (i / 8) * 8) \
+            +                     (tiitg / THREAD_PER_ROW) % 8  + (i & 7) * 8) = temp_a[i/4][i%4];
+        }
+
+        *(threadgroup float2x4 *)(sb + (tiitg % THREAD_PER_COL) * 8 * 32 + 8 * (tiitg / THREAD_PER_COL)) = *((device float2x4 *)y);
+
+        il = (il + 2 < nl) ? il + 2 : il % 2;
+        x  = (il < 2) ? x + (2+nl-1)/nl : x;
+        y += BLOCK_SIZE_K;
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        // load matrices from threadgroup memory and conduct outer products
+        threadgroup half  * lsma = (sa + THREAD_MAT_M * SG_MAT_SIZE * (sgitg % 2));
+        threadgroup float * lsmb = (sb + THREAD_MAT_N * SG_MAT_SIZE * (sgitg / 2));
+
+        #pragma unroll(4)
+        for (int ik = 0; ik < BLOCK_SIZE_K / 8; ik++) {
+            #pragma unroll(4)
+            for (int i = 0; i < 4; i++) {
+                simdgroup_load(ma[i],lsma + SG_MAT_SIZE * i);
+            }
+            simdgroup_barrier(mem_flags::mem_none);
+            #pragma unroll(2)
+            for (int i = 0; i < 2; i++) {
+                simdgroup_load(mb[i],lsmb + SG_MAT_SIZE * i);
+            }
+
+            lsma += BLOCK_SIZE_M / SG_MAT_ROW * SG_MAT_SIZE;
+            lsmb += BLOCK_SIZE_N / SG_MAT_ROW * SG_MAT_SIZE;
+
+            #pragma unroll(8)
+            for (int i = 0; i < 8; i++){
+                simdgroup_multiply_accumulate(c_res[i], mb[i/4], ma[i%4], c_res[i]);
+            }
+        }
+    }
+
+    if ((r0 + 1) * BLOCK_SIZE_M <= ne0 && (r1 + 1) * BLOCK_SIZE_N <= ne1) {
+        device float * C = dst + (BLOCK_SIZE_M * r0 + 32 * (sgitg &  1)) \
+                               + (BLOCK_SIZE_N * r1 + 16 * (sgitg >> 1)) * ne0 + im*ne1*ne0;
+        for (int i = 0; i < 8; i++) {
+            simdgroup_store(c_res[i], C + 8 * (i%4) + 8 * ne0 * (i/4), ne0);
+        }
+    } else {
+        // block is smaller than 64x32, we should avoid writing data outside of the matrix
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+        threadgroup float * temp_str = ((threadgroup float *)shared_memory) \
+                                      + 32 * (sgitg&1) + (16 * (sgitg>>1)) * BLOCK_SIZE_M;
+        for (int i = 0; i < 8; i++) {
+            simdgroup_store(c_res[i], temp_str + 8 * (i%4) + 8 * BLOCK_SIZE_M * (i/4), BLOCK_SIZE_M);
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        device float * C = dst + (BLOCK_SIZE_M * r0) + (BLOCK_SIZE_N * r1) * ne0 + im*ne1*ne0;
+        if (sgitg == 0) {
+            for (int i = 0; i < n_rows; i++) {
+                for (int j = tiitg; j < n_cols; j += BLOCK_SIZE_N) {
+                    *(C + i + j * ne0) = *(temp_str + i + j * BLOCK_SIZE_M);
+                }
+            }
+        }
+    }
+}
+
+// same as kernel_mul_mm_impl, but src1 and dst are accessed via indices stored in src1ids
+template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread half4x4 &)>
+void kernel_mul_mm_id_impl(
+        device const  uchar * src0,
+        device const  uchar * src1,
+        thread        short * src1ids,
+        device        float * dst,
+        constant    int64_t & ne00,
+        constant    int64_t & ne02,
+        constant   uint64_t & nb01,
+        constant   uint64_t & nb02,
+        constant    int64_t & ne12,
+        constant   uint64_t & nb10,
+        constant   uint64_t & nb11,
+        constant   uint64_t & nb12,
+        constant    int64_t & ne0,
+                    int64_t   ne1,
+        constant       uint & r2,
+        constant       uint & r3,
+        threadgroup   uchar * shared_memory,
+        uint3                 tgpig[[threadgroup_position_in_grid]],
+        uint                  tiitg[[thread_index_in_threadgroup]],
+        uint                  sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    threadgroup half  * sa = (threadgroup half  *)(shared_memory);
+    threadgroup float * sb = (threadgroup float *)(shared_memory + 4096);
+
+    const uint r0 = tgpig.y;
+    const uint r1 = tgpig.x;
+    const uint im = tgpig.z;
+
+    if (r1 * BLOCK_SIZE_N >= ne1) return;
+
+    // if this block is of 64x32 shape or smaller
+    short n_rows = (ne0 - r0 * BLOCK_SIZE_M < BLOCK_SIZE_M) ? (ne0 - r0 * BLOCK_SIZE_M) : BLOCK_SIZE_M;
+    short n_cols = (ne1 - r1 * BLOCK_SIZE_N < BLOCK_SIZE_N) ? (ne1 - r1 * BLOCK_SIZE_N) : BLOCK_SIZE_N;
+
+    // a thread shouldn't load data outside of the matrix
+    short thread_row = ((short)tiitg/THREAD_PER_ROW) < n_rows ? ((short)tiitg/THREAD_PER_ROW) : n_rows - 1;
+    short thread_col = ((short)tiitg/THREAD_PER_COL) < n_cols ? ((short)tiitg/THREAD_PER_COL) : n_cols - 1;
+
+    simdgroup_half8x8  ma[4];
+    simdgroup_float8x8 mb[2];
+    simdgroup_float8x8 c_res[8];
+    for (int i = 0; i < 8; i++){
+        c_res[i] = make_filled_simdgroup_matrix<float, 8>(0.f);
+    }
+
+    short il = (tiitg % THREAD_PER_ROW);
+
+    const uint i12 = im%ne12;
+    const uint i13 = im/ne12;
+
+    uint   offset0 = (i12/r2)*nb02 + (i13/r3)*(nb02*ne02);
+    ushort offset1 = il/nl;
+
+    device const block_q * x = (device const block_q *)(src0 + (r0 * BLOCK_SIZE_M + thread_row) * nb01 + offset0) + offset1;
+    device const float   * y = (device const float   *)(src1
+        + nb12 * im
+        + nb11 * src1ids[r1 * BLOCK_SIZE_N + thread_col]
+        + nb10 * (BLOCK_SIZE_K / THREAD_PER_COL * (tiitg % THREAD_PER_COL)));
+
+    for (int loop_k = 0; loop_k < ne00; loop_k += BLOCK_SIZE_K) {
+        // load data and store to threadgroup memory
+        half4x4 temp_a;
+        dequantize_func(x, il, temp_a);
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        for (int i = 0; i < 16; i++) {
+            *(sa + SG_MAT_SIZE * ((tiitg / THREAD_PER_ROW / 8) \
+            +                     (tiitg % THREAD_PER_ROW) * 16 + (i / 8) * 8) \
+            +                     (tiitg / THREAD_PER_ROW) % 8  + (i & 7) * 8) = temp_a[i/4][i%4];
+        }
+
+        *(threadgroup float2x4 *)(sb + (tiitg % THREAD_PER_COL) * 8 * 32 + 8 * (tiitg / THREAD_PER_COL)) = *((device float2x4 *)y);
+
+        il = (il + 2 < nl) ? il + 2 : il % 2;
+        x  = (il < 2) ? x + (2+nl-1)/nl : x;
+        y += BLOCK_SIZE_K;
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        // load matrices from threadgroup memory and conduct outer products
+        threadgroup half  * lsma = (sa + THREAD_MAT_M * SG_MAT_SIZE * (sgitg % 2));
+        threadgroup float * lsmb = (sb + THREAD_MAT_N * SG_MAT_SIZE * (sgitg / 2));
+
+        for (int ik = 0; ik < BLOCK_SIZE_K / 8; ik++) {
+            for (int i = 0; i < 4; i++) {
+                simdgroup_load(ma[i],lsma + SG_MAT_SIZE * i);
+            }
+            simdgroup_barrier(mem_flags::mem_none);
+            for (int i = 0; i < 2; i++) {
+                simdgroup_load(mb[i],lsmb + SG_MAT_SIZE * i);
+            }
+
+            lsma += BLOCK_SIZE_M / SG_MAT_ROW * SG_MAT_SIZE;
+            lsmb += BLOCK_SIZE_N / SG_MAT_ROW * SG_MAT_SIZE;
+
+            for (int i = 0; i < 8; i++){
+                simdgroup_multiply_accumulate(c_res[i], mb[i/4], ma[i%4], c_res[i]);
+            }
+        }
+    }
+
+    {
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+        threadgroup float * temp_str = ((threadgroup float *)shared_memory) \
+                                      + 32 * (sgitg&1) + (16 * (sgitg>>1)) * BLOCK_SIZE_M;
+        for (int i = 0; i < 8; i++) {
+            simdgroup_store(c_res[i], temp_str + 8 * (i%4) + 8 * BLOCK_SIZE_M * (i/4), BLOCK_SIZE_M);
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        device float * C = dst + (BLOCK_SIZE_M * r0) + im*ne1*ne0;
+        if (sgitg == 0) {
+            for (int i = 0; i < n_rows; i++) {
+                for (int j = tiitg; j < n_cols; j += BLOCK_SIZE_N) {
+                    *(C + i + src1ids[j + r1*BLOCK_SIZE_N] * ne0) = *(temp_str + i + j * BLOCK_SIZE_M);
+                }
+            }
+        }
+    }
+}
+
+template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread half4x4 &)>
+kernel void kernel_mul_mm(device const  uchar * src0,
+                          device const  uchar * src1,
+                          device        float * dst,
+                          constant    int64_t & ne00,
+                          constant    int64_t & ne02,
+                          constant   uint64_t & nb01,
+                          constant   uint64_t & nb02,
+                          constant    int64_t & ne12,
+                          constant   uint64_t & nb10,
+                          constant   uint64_t & nb11,
+                          constant   uint64_t & nb12,
+                          constant    int64_t & ne0,
+                          constant    int64_t & ne1,
+                          constant       uint & r2,
+                          constant       uint & r3,
+                          threadgroup   uchar * shared_memory [[threadgroup(0)]],
+                          uint3                 tgpig[[threadgroup_position_in_grid]],
+                          uint                  tiitg[[thread_index_in_threadgroup]],
+                          uint                  sgitg[[simdgroup_index_in_threadgroup]]) {
+    kernel_mul_mm_impl<block_q, nl, dequantize_func>(
+        src0,
+        src1,
+        dst,
+        ne00,
+        ne02,
+        nb01,
+        nb02,
+        ne12,
+        nb10,
+        nb11,
+        nb12,
+        ne0,
+        ne1,
+        r2,
+        r3,
+        shared_memory,
+        tgpig,
+        tiitg,
+        sgitg);
+}
+
+template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread half4x4 &)>
+kernel void kernel_mul_mm_id(
+        device const   uchar * ids,
+        device const   uchar * src1,
+        device         float * dst,
+        constant    uint64_t & nbi1,
+        constant     int64_t & ne00,
+        constant     int64_t & ne02,
+        constant    uint64_t & nb01,
+        constant    uint64_t & nb02,
+        constant     int64_t & ne12,
+        constant     int64_t & ne13,
+        constant    uint64_t & nb10,
+        constant    uint64_t & nb11,
+        constant    uint64_t & nb12,
+        constant     int64_t & ne0,
+        constant     int64_t & ne1,
+        constant    uint64_t & nb1,
+        constant        uint & r2,
+        constant        uint & r3,
+        constant         int & idx,
+        device const   uchar * src00,
+        device const   uchar * src01,
+        device const   uchar * src02,
+        device const   uchar * src03,
+        device const   uchar * src04,
+        device const   uchar * src05,
+        device const   uchar * src06,
+        device const   uchar * src07,
+        threadgroup    uchar * shared_memory [[threadgroup(0)]],
+        uint3                  tgpig[[threadgroup_position_in_grid]],
+        uint                   tiitg[[thread_index_in_threadgroup]],
+        uint                   sgitg[[simdgroup_index_in_threadgroup]]) {
+    device const uchar * src0s[8] = {src00, src01, src02, src03, src04, src05, src06, src07};
+
+    // expert id
+    const int32_t id = tgpig.z/(ne12*ne13);
+
+    tgpig.z = tgpig.z%(ne12*ne13);
+
+    // row indices of src1 for expert id
+    int64_t _ne1 = 0;
+    short src1ids[512];
+
+    for (int64_t i1 = 0; i1 < ne1; i1++) {
+        if (((device int32_t *) (ids + i1*nbi1))[idx] == id) {
+            src1ids[_ne1++] = i1;
+        }
+    }
+
+    kernel_mul_mm_id_impl<block_q, nl, dequantize_func>(
+        src0s[id],
+        src1,
+        src1ids,
+        dst,
+        ne00,
+        ne02,
+        nb01,
+        nb02,
+        ne12,
+        nb10,
+        nb11,
+        nb12,
+        ne0,
+        _ne1,
+        r2,
+        r3,
+        shared_memory,
+        tgpig,
+        tiitg,
+        sgitg);
+}
+
+#if QK_K == 256
+#define QK_NL 16
+#else
+#define QK_NL 4
+#endif
+
+//
+// get rows
+//
+
+typedef void (get_rows_t)(
+        device const void * src0,
+        device const char * src1,
+        device      float * dst,
+        constant  int64_t & ne00,
+        constant uint64_t & nb01,
+        constant uint64_t & nb02,
+        constant  int64_t & ne10,
+        constant uint64_t & nb10,
+        constant uint64_t & nb11,
+        constant uint64_t & nb1,
+        constant uint64_t & nb2,
+        uint3, uint, uint3);
+
+//template [[host_name("kernel_get_rows_f32")]]  kernel get_rows_t kernel_get_rows<float4x4,   1, dequantize_f32>;
+//template [[host_name("kernel_get_rows_f16")]]  kernel get_rows_t kernel_get_rows<half4x4,    1, dequantize_f16>;
+template [[host_name("kernel_get_rows_q4_0")]] kernel get_rows_t kernel_get_rows<block_q4_0, 2, dequantize_q4_0>;
+template [[host_name("kernel_get_rows_q4_1")]] kernel get_rows_t kernel_get_rows<block_q4_1, 2, dequantize_q4_1>;
+template [[host_name("kernel_get_rows_q5_0")]] kernel get_rows_t kernel_get_rows<block_q5_0, 2, dequantize_q5_0>;
+template [[host_name("kernel_get_rows_q5_1")]] kernel get_rows_t kernel_get_rows<block_q5_1, 2, dequantize_q5_1>;
+template [[host_name("kernel_get_rows_q8_0")]] kernel get_rows_t kernel_get_rows<block_q8_0, 2, dequantize_q8_0>;
+template [[host_name("kernel_get_rows_q2_K")]] kernel get_rows_t kernel_get_rows<block_q2_K, QK_NL, dequantize_q2_K>;
+template [[host_name("kernel_get_rows_q3_K")]] kernel get_rows_t kernel_get_rows<block_q3_K, QK_NL, dequantize_q3_K>;
+template [[host_name("kernel_get_rows_q4_K")]] kernel get_rows_t kernel_get_rows<block_q4_K, QK_NL, dequantize_q4_K>;
+template [[host_name("kernel_get_rows_q5_K")]] kernel get_rows_t kernel_get_rows<block_q5_K, QK_NL, dequantize_q5_K>;
+template [[host_name("kernel_get_rows_q6_K")]] kernel get_rows_t kernel_get_rows<block_q6_K, QK_NL, dequantize_q6_K>;
+template [[host_name("kernel_get_rows_iq2_xxs")]] kernel get_rows_t kernel_get_rows<block_iq2_xxs, QK_NL, dequantize_iq2_xxs>;
+template [[host_name("kernel_get_rows_iq2_xs")]]  kernel get_rows_t kernel_get_rows<block_iq2_xs,  QK_NL, dequantize_iq2_xs>;
+template [[host_name("kernel_get_rows_iq3_xxs")]] kernel get_rows_t kernel_get_rows<block_iq3_xxs, QK_NL, dequantize_iq3_xxs>;
+template [[host_name("kernel_get_rows_iq3_s")]]   kernel get_rows_t kernel_get_rows<block_iq3_s,   QK_NL, dequantize_iq3_s>;
+template [[host_name("kernel_get_rows_iq2_s")]]   kernel get_rows_t kernel_get_rows<block_iq2_s,   QK_NL, dequantize_iq2_s>;
+template [[host_name("kernel_get_rows_iq1_s")]]   kernel get_rows_t kernel_get_rows<block_iq1_s,   QK_NL, dequantize_iq1_s>;
+template [[host_name("kernel_get_rows_iq4_nl")]]  kernel get_rows_t kernel_get_rows<block_iq4_nl,  2,     dequantize_iq4_nl>;
+#if QK_K == 64
+template [[host_name("kernel_get_rows_iq4_xs")]]  kernel get_rows_t kernel_get_rows<block_iq4_xs,  2,     dequantize_iq4_xs>;
+#else
+template [[host_name("kernel_get_rows_iq4_xs")]]  kernel get_rows_t kernel_get_rows<block_iq4_xs,  QK_NL, dequantize_iq4_xs>;
+#endif
+
+//
+// matrix-matrix multiplication
+//
+
+typedef void (mat_mm_t)(
+        device const  uchar * src0,
+        device const  uchar * src1,
+        device        float * dst,
+        constant    int64_t & ne00,
+        constant    int64_t & ne02,
+        constant   uint64_t & nb01,
+        constant   uint64_t & nb02,
+        constant    int64_t & ne12,
+        constant   uint64_t & nb10,
+        constant   uint64_t & nb11,
+        constant   uint64_t & nb12,
+        constant    int64_t & ne0,
+        constant    int64_t & ne1,
+        constant       uint & r2,
+        constant       uint & r3,
+        threadgroup   uchar *,
+        uint3, uint, uint);
+
+template [[host_name("kernel_mul_mm_f32_f32")]]  kernel mat_mm_t kernel_mul_mm<float4x4,   1,     dequantize_f32>;
+template [[host_name("kernel_mul_mm_f16_f32")]]  kernel mat_mm_t kernel_mul_mm<half4x4,    1,     dequantize_f16>;
+template [[host_name("kernel_mul_mm_q4_0_f32")]] kernel mat_mm_t kernel_mul_mm<block_q4_0, 2,     dequantize_q4_0>;
+template [[host_name("kernel_mul_mm_q4_1_f32")]] kernel mat_mm_t kernel_mul_mm<block_q4_1, 2,     dequantize_q4_1>;
+template [[host_name("kernel_mul_mm_q5_0_f32")]] kernel mat_mm_t kernel_mul_mm<block_q5_0, 2,     dequantize_q5_0>;
+template [[host_name("kernel_mul_mm_q5_1_f32")]] kernel mat_mm_t kernel_mul_mm<block_q5_1, 2,     dequantize_q5_1>;
+template [[host_name("kernel_mul_mm_q8_0_f32")]] kernel mat_mm_t kernel_mul_mm<block_q8_0, 2,     dequantize_q8_0>;
+template [[host_name("kernel_mul_mm_q2_K_f32")]] kernel mat_mm_t kernel_mul_mm<block_q2_K, QK_NL, dequantize_q2_K>;
+template [[host_name("kernel_mul_mm_q3_K_f32")]] kernel mat_mm_t kernel_mul_mm<block_q3_K, QK_NL, dequantize_q3_K>;
+template [[host_name("kernel_mul_mm_q4_K_f32")]] kernel mat_mm_t kernel_mul_mm<block_q4_K, QK_NL, dequantize_q4_K>;
+template [[host_name("kernel_mul_mm_q5_K_f32")]] kernel mat_mm_t kernel_mul_mm<block_q5_K, QK_NL, dequantize_q5_K>;
+template [[host_name("kernel_mul_mm_q6_K_f32")]] kernel mat_mm_t kernel_mul_mm<block_q6_K, QK_NL, dequantize_q6_K>;
+template [[host_name("kernel_mul_mm_iq2_xxs_f32")]] kernel mat_mm_t kernel_mul_mm<block_iq2_xxs, QK_NL, dequantize_iq2_xxs>;
+template [[host_name("kernel_mul_mm_iq2_xs_f32")]]  kernel mat_mm_t kernel_mul_mm<block_iq2_xs,  QK_NL, dequantize_iq2_xs>;
+template [[host_name("kernel_mul_mm_iq3_xxs_f32")]] kernel mat_mm_t kernel_mul_mm<block_iq3_xxs, QK_NL, dequantize_iq3_xxs>;
+template [[host_name("kernel_mul_mm_iq3_s_f32")]]   kernel mat_mm_t kernel_mul_mm<block_iq3_s,   QK_NL, dequantize_iq3_s>;
+template [[host_name("kernel_mul_mm_iq2_s_f32")]]   kernel mat_mm_t kernel_mul_mm<block_iq2_s,   QK_NL, dequantize_iq2_s>;
+template [[host_name("kernel_mul_mm_iq1_s_f32")]]   kernel mat_mm_t kernel_mul_mm<block_iq1_s,   QK_NL, dequantize_iq1_s>;
+template [[host_name("kernel_mul_mm_iq4_nl_f32")]]  kernel mat_mm_t kernel_mul_mm<block_iq4_nl,  2,     dequantize_iq4_nl>;
+#if QK_K == 64
+template [[host_name("kernel_mul_mm_iq4_xs_f32")]]  kernel mat_mm_t kernel_mul_mm<block_iq4_nl,  2,     dequantize_iq4_xs>;
+#else
+template [[host_name("kernel_mul_mm_iq4_xs_f32")]]  kernel mat_mm_t kernel_mul_mm<block_iq4_xs,  QK_NL, dequantize_iq4_xs>;
+#endif
+
+//
+// indirect matrix-matrix multiplication
+//
+
+typedef void (mat_mm_id_t)(
+        device const   uchar * ids,
+        device const   uchar * src1,
+        device         float * dst,
+        constant    uint64_t & nbi1,
+        constant     int64_t & ne00,
+        constant     int64_t & ne02,
+        constant    uint64_t & nb01,
+        constant    uint64_t & nb02,
+        constant     int64_t & ne12,
+        constant     int64_t & ne13,
+        constant    uint64_t & nb10,
+        constant    uint64_t & nb11,
+        constant    uint64_t & nb12,
+        constant     int64_t & ne0,
+        constant     int64_t & ne1,
+        constant    uint64_t & nb1,
+        constant        uint & r2,
+        constant        uint & r3,
+        constant         int & idx,
+        device const   uchar * src00,
+        device const   uchar * src01,
+        device const   uchar * src02,
+        device const   uchar * src03,
+        device const   uchar * src04,
+        device const   uchar * src05,
+        device const   uchar * src06,
+        device const   uchar * src07,
+        threadgroup    uchar *,
+        uint3, uint, uint);
+
+template [[host_name("kernel_mul_mm_id_f32_f32")]]  kernel mat_mm_id_t kernel_mul_mm_id<float4x4,   1,     dequantize_f32>;
+template [[host_name("kernel_mul_mm_id_f16_f32")]]  kernel mat_mm_id_t kernel_mul_mm_id<half4x4,    1,     dequantize_f16>;
+template [[host_name("kernel_mul_mm_id_q4_0_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q4_0, 2,     dequantize_q4_0>;
+template [[host_name("kernel_mul_mm_id_q4_1_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q4_1, 2,     dequantize_q4_1>;
+template [[host_name("kernel_mul_mm_id_q5_0_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q5_0, 2,     dequantize_q5_0>;
+template [[host_name("kernel_mul_mm_id_q5_1_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q5_1, 2,     dequantize_q5_1>;
+template [[host_name("kernel_mul_mm_id_q8_0_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q8_0, 2,     dequantize_q8_0>;
+template [[host_name("kernel_mul_mm_id_q2_K_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q2_K, QK_NL, dequantize_q2_K>;
+template [[host_name("kernel_mul_mm_id_q3_K_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q3_K, QK_NL, dequantize_q3_K>;
+template [[host_name("kernel_mul_mm_id_q4_K_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q4_K, QK_NL, dequantize_q4_K>;
+template [[host_name("kernel_mul_mm_id_q5_K_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q5_K, QK_NL, dequantize_q5_K>;
+template [[host_name("kernel_mul_mm_id_q6_K_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q6_K, QK_NL, dequantize_q6_K>;
+template [[host_name("kernel_mul_mm_id_iq2_xxs_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_iq2_xxs, QK_NL, dequantize_iq2_xxs>;
+template [[host_name("kernel_mul_mm_id_iq2_xs_f32")]]  kernel mat_mm_id_t kernel_mul_mm_id<block_iq2_xs,  QK_NL, dequantize_iq2_xs>;
+template [[host_name("kernel_mul_mm_id_iq3_xxs_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_iq3_xxs, QK_NL, dequantize_iq3_xxs>;
+template [[host_name("kernel_mul_mm_id_iq3_s_f32")]]   kernel mat_mm_id_t kernel_mul_mm_id<block_iq3_s,   QK_NL, dequantize_iq3_s>;
+template [[host_name("kernel_mul_mm_id_iq2_s_f32")]]   kernel mat_mm_id_t kernel_mul_mm_id<block_iq2_s,   QK_NL, dequantize_iq2_s>;
+template [[host_name("kernel_mul_mm_id_iq1_s_f32")]]   kernel mat_mm_id_t kernel_mul_mm_id<block_iq1_s,   QK_NL, dequantize_iq1_s>;
+template [[host_name("kernel_mul_mm_id_iq4_nl_f32")]]  kernel mat_mm_id_t kernel_mul_mm_id<block_iq4_nl,  2,     dequantize_iq4_nl>;
+#if QK_K == 64
+template [[host_name("kernel_mul_mm_id_iq4_xs_f32")]]  kernel mat_mm_id_t kernel_mul_mm_id<block_iq4_xs,  2,     dequantize_iq4_xs>;
+#else
+template [[host_name("kernel_mul_mm_id_iq4_xs_f32")]]  kernel mat_mm_id_t kernel_mul_mm_id<block_iq4_xs,  QK_NL, dequantize_iq4_xs>;
+#endif
+
+//
+// matrix-vector multiplication
+//
+
+[[host_name("kernel_mul_mv_id_f32_f32")]]
+kernel void kernel_mul_mv_id_f32_f32(
+        device const    char * ids,
+        device const    char * src1,
+        device         float * dst,
+        constant    uint64_t & nbi1,
+        constant     int64_t & ne00,
+        constant     int64_t & ne01,
+        constant     int64_t & ne02,
+        constant    uint64_t & nb00,
+        constant    uint64_t & nb01,
+        constant    uint64_t & nb02,
+        constant     int64_t & ne10,
+        constant     int64_t & ne11,
+        constant     int64_t & ne12,
+        constant     int64_t & ne13,
+        constant    uint64_t & nb10,
+        constant    uint64_t & nb11,
+        constant    uint64_t & nb12,
+        constant     int64_t & ne0,
+        constant     int64_t & ne1,
+        constant    uint64_t & nb1,
+        constant        uint & r2,
+        constant        uint & r3,
+        constant         int & idx,
+        device const    char * src00,
+        device const    char * src01,
+        device const    char * src02,
+        device const    char * src03,
+        device const    char * src04,
+        device const    char * src05,
+        device const    char * src06,
+        device const    char * src07,
+        uint3                  tgpig[[threadgroup_position_in_grid]],
+        uint                   tiitg[[thread_index_in_threadgroup]],
+        uint                   tiisg[[thread_index_in_simdgroup]],
+        uint                   sgitg[[simdgroup_index_in_threadgroup]]) {
+    device const char * src0[8] = {src00, src01, src02, src03, src04, src05, src06, src07};
 
-    {
-        threadgroup_barrier(mem_flags::mem_threadgroup);
-        threadgroup float * temp_str = ((threadgroup float *)shared_memory) \
-                                      + 32 * (sgitg&1) + (16 * (sgitg>>1)) * BLOCK_SIZE_M;
-        for (int i = 0; i < 8; i++) {
-            simdgroup_store(c_res[i], temp_str + 8 * (i%4) + 8 * BLOCK_SIZE_M * (i/4), BLOCK_SIZE_M);
-        }
+    const int64_t bid = tgpig.z/(ne12*ne13);
 
-        threadgroup_barrier(mem_flags::mem_threadgroup);
+    tgpig.z = tgpig.z%(ne12*ne13);
 
-        device float * C = dst + (BLOCK_SIZE_M * r0) + im*ne1*ne0;
-        if (sgitg == 0) {
-            for (int i = 0; i < n_rows; i++) {
-                for (int j = tiitg; j < n_cols; j += BLOCK_SIZE_N) {
-                    *(C + i + src1ids[j + r1*BLOCK_SIZE_N] * ne0) = *(temp_str + i + j * BLOCK_SIZE_M);
-                }
-            }
-        }
-    }
-}
+    const int32_t id = ((device int32_t *) (ids + bid*nbi1))[idx];
 
-template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread half4x4 &)>
-kernel void kernel_mul_mm(device const  uchar * src0,
-                          device const  uchar * src1,
-                          device        float * dst,
-                          constant    int64_t & ne00,
-                          constant    int64_t & ne02,
-                          constant   uint64_t & nb01,
-                          constant   uint64_t & nb02,
-                          constant    int64_t & ne12,
-                          constant   uint64_t & nb10,
-                          constant   uint64_t & nb11,
-                          constant   uint64_t & nb12,
-                          constant    int64_t & ne0,
-                          constant    int64_t & ne1,
-                          constant       uint & r2,
-                          constant       uint & r3,
-                          threadgroup   uchar * shared_memory [[threadgroup(0)]],
-                          uint3                 tgpig[[threadgroup_position_in_grid]],
-                          uint                  tiitg[[thread_index_in_threadgroup]],
-                          uint                  sgitg[[simdgroup_index_in_threadgroup]]) {
-    kernel_mul_mm_impl<block_q, nl, dequantize_func>(
-        src0,
-        src1,
-        dst,
+    kernel_mul_mv_f32_f32_impl(
+        src0[id],
+        src1 + bid*nb11,
+        dst  + bid*ne0,
         ne00,
+        ne01,
         ne02,
+        nb00,
         nb01,
         nb02,
+        ne10,
+        ne11,
         ne12,
         nb10,
         nb11,
@@ -4942,22 +6503,24 @@ kernel void kernel_mul_mm(device const  uchar * src0,
         ne1,
         r2,
         r3,
-        shared_memory,
         tgpig,
-        tiitg,
-        sgitg);
+        tiisg);
 }
 
-template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread half4x4 &)>
-kernel void kernel_mul_mm_id(
-        device const   uchar * ids,
-        device const   uchar * src1,
+[[host_name("kernel_mul_mv_id_f16_f32")]]
+kernel void kernel_mul_mv_id_f16_f32(
+        device const    char * ids,
+        device const    char * src1,
         device         float * dst,
         constant    uint64_t & nbi1,
         constant     int64_t & ne00,
+        constant     int64_t & ne01,
         constant     int64_t & ne02,
+        constant    uint64_t & nb00,
         constant    uint64_t & nb01,
         constant    uint64_t & nb02,
+        constant     int64_t & ne10,
+        constant     int64_t & ne11,
         constant     int64_t & ne12,
         constant     int64_t & ne13,
         constant    uint64_t & nb10,
@@ -4969,150 +6532,190 @@ kernel void kernel_mul_mm_id(
         constant        uint & r2,
         constant        uint & r3,
         constant         int & idx,
-        device const   uchar * src00,
-        device const   uchar * src01,
-        device const   uchar * src02,
-        device const   uchar * src03,
-        device const   uchar * src04,
-        device const   uchar * src05,
-        device const   uchar * src06,
-        device const   uchar * src07,
-        threadgroup    uchar * shared_memory [[threadgroup(0)]],
+        device const    char * src00,
+        device const    char * src01,
+        device const    char * src02,
+        device const    char * src03,
+        device const    char * src04,
+        device const    char * src05,
+        device const    char * src06,
+        device const    char * src07,
         uint3                  tgpig[[threadgroup_position_in_grid]],
         uint                   tiitg[[thread_index_in_threadgroup]],
+        uint                   tiisg[[thread_index_in_simdgroup]],
         uint                   sgitg[[simdgroup_index_in_threadgroup]]) {
-    device const uchar * src0s[8] = {src00, src01, src02, src03, src04, src05, src06, src07};
+    device const char * src0[8] = {src00, src01, src02, src03, src04, src05, src06, src07};
 
-    // expert id
-    const int32_t id = tgpig.z/(ne12*ne13);
+    const int64_t bid = tgpig.z/(ne12*ne13);
 
     tgpig.z = tgpig.z%(ne12*ne13);
 
-    // row indices of src1 for expert id
-    int64_t _ne1 = 0;
-    short src1ids[512];
-
-    for (int64_t i1 = 0; i1 < ne1; i1++) {
-        if (((device int32_t *) (ids + i1*nbi1))[idx] == id) {
-            src1ids[_ne1++] = i1;
-        }
-    }
+    const int32_t id = ((device int32_t *) (ids + bid*nbi1))[idx];
 
-    kernel_mul_mm_id_impl<block_q, nl, dequantize_func>(
-        src0s[id],
-        src1,
-        src1ids,
-        dst,
+    kernel_mul_mv_f16_f32_impl(
+        src0[id],
+        src1 + bid*nb11,
+        dst  + bid*ne0,
         ne00,
+        ne01,
         ne02,
+        nb00,
         nb01,
         nb02,
+        ne10,
+        ne11,
         ne12,
         nb10,
         nb11,
         nb12,
         ne0,
-        _ne1,
+        ne1,
         r2,
         r3,
-        shared_memory,
         tgpig,
-        tiitg,
-        sgitg);
+        tiisg);
 }
 
-#if QK_K == 256
-#define QK_NL 16
-#else
-#define QK_NL 4
-#endif
+[[host_name("kernel_mul_mv_id_q8_0_f32")]]
+kernel void kernel_mul_mv_id_q8_0_f32(
+        device const    char * ids,
+        device const    char * src1,
+        device         float * dst,
+        constant    uint64_t & nbi1,
+        constant     int64_t & ne00,
+        constant     int64_t & ne01,
+        constant     int64_t & ne02,
+        constant    uint64_t & nb00,
+        constant    uint64_t & nb01,
+        constant    uint64_t & nb02,
+        constant     int64_t & ne10,
+        constant     int64_t & ne11,
+        constant     int64_t & ne12,
+        constant     int64_t & ne13,
+        constant    uint64_t & nb10,
+        constant    uint64_t & nb11,
+        constant    uint64_t & nb12,
+        constant     int64_t & ne0,
+        constant     int64_t & ne1,
+        constant    uint64_t & nb1,
+        constant        uint & r2,
+        constant        uint & r3,
+        constant         int & idx,
+        device const    char * src00,
+        device const    char * src01,
+        device const    char * src02,
+        device const    char * src03,
+        device const    char * src04,
+        device const    char * src05,
+        device const    char * src06,
+        device const    char * src07,
+        uint3                  tgpig[[threadgroup_position_in_grid]],
+        uint                   tiitg[[thread_index_in_threadgroup]],
+        uint                   tiisg[[thread_index_in_simdgroup]],
+        uint                   sgitg[[simdgroup_index_in_threadgroup]]) {
+    device const char * src0[8] = {src00, src01, src02, src03, src04, src05, src06, src07};
 
-//
-// get rows
-//
+    const int64_t bid = tgpig.z/(ne12*ne13);
 
-typedef void (get_rows_t)(
-        device const void * src0,
-        device const char * src1,
-        device      float * dst,
-        constant  int64_t & ne00,
-        constant uint64_t & nb01,
-        constant uint64_t & nb02,
-        constant  int64_t & ne10,
-        constant uint64_t & nb10,
-        constant uint64_t & nb11,
-        constant uint64_t & nb1,
-        constant uint64_t & nb2,
-        uint3, uint, uint3);
+    tgpig.z = tgpig.z%(ne12*ne13);
 
-//template [[host_name("kernel_get_rows_f32")]]  kernel get_rows_t kernel_get_rows<float4x4,   1, dequantize_f32>;
-//template [[host_name("kernel_get_rows_f16")]]  kernel get_rows_t kernel_get_rows<half4x4,    1, dequantize_f16>;
-template [[host_name("kernel_get_rows_q4_0")]] kernel get_rows_t kernel_get_rows<block_q4_0, 2, dequantize_q4_0>;
-template [[host_name("kernel_get_rows_q4_1")]] kernel get_rows_t kernel_get_rows<block_q4_1, 2, dequantize_q4_1>;
-template [[host_name("kernel_get_rows_q5_0")]] kernel get_rows_t kernel_get_rows<block_q5_0, 2, dequantize_q5_0>;
-template [[host_name("kernel_get_rows_q5_1")]] kernel get_rows_t kernel_get_rows<block_q5_1, 2, dequantize_q5_1>;
-template [[host_name("kernel_get_rows_q8_0")]] kernel get_rows_t kernel_get_rows<block_q8_0, 2, dequantize_q8_0>;
-template [[host_name("kernel_get_rows_q2_K")]] kernel get_rows_t kernel_get_rows<block_q2_K, QK_NL, dequantize_q2_K>;
-template [[host_name("kernel_get_rows_q3_K")]] kernel get_rows_t kernel_get_rows<block_q3_K, QK_NL, dequantize_q3_K>;
-template [[host_name("kernel_get_rows_q4_K")]] kernel get_rows_t kernel_get_rows<block_q4_K, QK_NL, dequantize_q4_K>;
-template [[host_name("kernel_get_rows_q5_K")]] kernel get_rows_t kernel_get_rows<block_q5_K, QK_NL, dequantize_q5_K>;
-template [[host_name("kernel_get_rows_q6_K")]] kernel get_rows_t kernel_get_rows<block_q6_K, QK_NL, dequantize_q6_K>;
-template [[host_name("kernel_get_rows_iq2_xxs")]] kernel get_rows_t kernel_get_rows<block_iq2_xxs, QK_NL, dequantize_iq2_xxs>;
-template [[host_name("kernel_get_rows_iq2_xs")]]  kernel get_rows_t kernel_get_rows<block_iq2_xs,  QK_NL, dequantize_iq2_xs>;
-template [[host_name("kernel_get_rows_iq3_xxs")]] kernel get_rows_t kernel_get_rows<block_iq3_xxs, QK_NL, dequantize_iq3_xxs>;
+    const int32_t id = ((device int32_t *) (ids + bid*nbi1))[idx];
+
+    kernel_mul_mv_q8_0_f32_impl(
+        src0[id],
+        (device const float *) (src1 + bid*nb11),
+        dst + bid*ne0,
+        ne00,
+        ne01,
+        ne02,
+        ne10,
+        ne12,
+        ne0,
+        ne1,
+        r2,
+        r3,
+        tgpig,
+        tiisg,
+        sgitg);
+}
+
+[[host_name("kernel_mul_mv_id_q4_0_f32")]]
+kernel void kernel_mul_mv_id_q4_0_f32(
+        device const    char * ids,
+        device const    char * src1,
+        device         float * dst,
+        constant    uint64_t & nbi1,
+        constant     int64_t & ne00,
+        constant     int64_t & ne01,
+        constant     int64_t & ne02,
+        constant    uint64_t & nb00,
+        constant    uint64_t & nb01,
+        constant    uint64_t & nb02,
+        constant     int64_t & ne10,
+        constant     int64_t & ne11,
+        constant     int64_t & ne12,
+        constant     int64_t & ne13,
+        constant    uint64_t & nb10,
+        constant    uint64_t & nb11,
+        constant    uint64_t & nb12,
+        constant     int64_t & ne0,
+        constant     int64_t & ne1,
+        constant    uint64_t & nb1,
+        constant        uint & r2,
+        constant        uint & r3,
+        constant         int & idx,
+        device const    char * src00,
+        device const    char * src01,
+        device const    char * src02,
+        device const    char * src03,
+        device const    char * src04,
+        device const    char * src05,
+        device const    char * src06,
+        device const    char * src07,
+        uint3                  tgpig[[threadgroup_position_in_grid]],
+        uint                   tiitg[[thread_index_in_threadgroup]],
+        uint                   tiisg[[thread_index_in_simdgroup]],
+        uint                   sgitg[[simdgroup_index_in_threadgroup]]) {
+    device const char * src0[8] = {src00, src01, src02, src03, src04, src05, src06, src07};
 
-//
-// matrix-matrix multiplication
-//
+    const int64_t bid = tgpig.z/(ne12*ne13);
 
-typedef void (mat_mm_t)(
-        device const  uchar * src0,
-        device const  uchar * src1,
-        device        float * dst,
-        constant    int64_t & ne00,
-        constant    int64_t & ne02,
-        constant   uint64_t & nb01,
-        constant   uint64_t & nb02,
-        constant    int64_t & ne12,
-        constant   uint64_t & nb10,
-        constant   uint64_t & nb11,
-        constant   uint64_t & nb12,
-        constant    int64_t & ne0,
-        constant    int64_t & ne1,
-        constant       uint & r2,
-        constant       uint & r3,
-        threadgroup   uchar *,
-        uint3, uint, uint);
+    tgpig.z = tgpig.z%(ne12*ne13);
 
-template [[host_name("kernel_mul_mm_f32_f32")]]  kernel mat_mm_t kernel_mul_mm<float4x4,   1,     dequantize_f32>;
-template [[host_name("kernel_mul_mm_f16_f32")]]  kernel mat_mm_t kernel_mul_mm<half4x4,    1,     dequantize_f16>;
-template [[host_name("kernel_mul_mm_q4_0_f32")]] kernel mat_mm_t kernel_mul_mm<block_q4_0, 2,     dequantize_q4_0>;
-template [[host_name("kernel_mul_mm_q4_1_f32")]] kernel mat_mm_t kernel_mul_mm<block_q4_1, 2,     dequantize_q4_1>;
-template [[host_name("kernel_mul_mm_q5_0_f32")]] kernel mat_mm_t kernel_mul_mm<block_q5_0, 2,     dequantize_q5_0>;
-template [[host_name("kernel_mul_mm_q5_1_f32")]] kernel mat_mm_t kernel_mul_mm<block_q5_1, 2,     dequantize_q5_1>;
-template [[host_name("kernel_mul_mm_q8_0_f32")]] kernel mat_mm_t kernel_mul_mm<block_q8_0, 2,     dequantize_q8_0>;
-template [[host_name("kernel_mul_mm_q2_K_f32")]] kernel mat_mm_t kernel_mul_mm<block_q2_K, QK_NL, dequantize_q2_K>;
-template [[host_name("kernel_mul_mm_q3_K_f32")]] kernel mat_mm_t kernel_mul_mm<block_q3_K, QK_NL, dequantize_q3_K>;
-template [[host_name("kernel_mul_mm_q4_K_f32")]] kernel mat_mm_t kernel_mul_mm<block_q4_K, QK_NL, dequantize_q4_K>;
-template [[host_name("kernel_mul_mm_q5_K_f32")]] kernel mat_mm_t kernel_mul_mm<block_q5_K, QK_NL, dequantize_q5_K>;
-template [[host_name("kernel_mul_mm_q6_K_f32")]] kernel mat_mm_t kernel_mul_mm<block_q6_K, QK_NL, dequantize_q6_K>;
-template [[host_name("kernel_mul_mm_iq2_xxs_f32")]] kernel mat_mm_t kernel_mul_mm<block_iq2_xxs, QK_NL, dequantize_iq2_xxs>;
-template [[host_name("kernel_mul_mm_iq2_xs_f32")]]  kernel mat_mm_t kernel_mul_mm<block_iq2_xs,  QK_NL, dequantize_iq2_xs>;
-template [[host_name("kernel_mul_mm_iq3_xxs_f32")]] kernel mat_mm_t kernel_mul_mm<block_iq3_xxs, QK_NL, dequantize_iq3_xxs>;
+    const int32_t id = ((device int32_t *) (ids + bid*nbi1))[idx];
 
-//
-// indirect matrix-matrix multiplication
-//
+    mul_vec_q_n_f32_impl<block_q4_0, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(
+        src0[id],
+        (device const float *) (src1 + bid*nb11),
+        dst + bid*ne0,
+        ne00,
+        ne01,
+        ne02,
+        ne10,
+        ne12,
+        ne0,
+        ne1,
+        r2,
+        r3,
+        tgpig,
+        tiisg,
+        sgitg);
+}
 
-typedef void (mat_mm_id_t)(
-        device const   uchar * ids,
-        device const   uchar * src1,
+[[host_name("kernel_mul_mv_id_q4_1_f32")]]
+kernel void kernel_mul_mv_id_q4_1_f32(
+        device const    char * ids,
+        device const    char * src1,
         device         float * dst,
         constant    uint64_t & nbi1,
         constant     int64_t & ne00,
+        constant     int64_t & ne01,
         constant     int64_t & ne02,
+        constant    uint64_t & nb00,
         constant    uint64_t & nb01,
         constant    uint64_t & nb02,
+        constant     int64_t & ne10,
+        constant     int64_t & ne11,
         constant     int64_t & ne12,
         constant     int64_t & ne13,
         constant    uint64_t & nb10,
@@ -5124,39 +6727,46 @@ typedef void (mat_mm_id_t)(
         constant        uint & r2,
         constant        uint & r3,
         constant         int & idx,
-        device const   uchar * src00,
-        device const   uchar * src01,
-        device const   uchar * src02,
-        device const   uchar * src03,
-        device const   uchar * src04,
-        device const   uchar * src05,
-        device const   uchar * src06,
-        device const   uchar * src07,
-        threadgroup    uchar *,
-        uint3, uint, uint);
+        device const    char * src00,
+        device const    char * src01,
+        device const    char * src02,
+        device const    char * src03,
+        device const    char * src04,
+        device const    char * src05,
+        device const    char * src06,
+        device const    char * src07,
+        uint3                  tgpig[[threadgroup_position_in_grid]],
+        uint                   tiitg[[thread_index_in_threadgroup]],
+        uint                   tiisg[[thread_index_in_simdgroup]],
+        uint                   sgitg[[simdgroup_index_in_threadgroup]]) {
+    device const char * src0[8] = {src00, src01, src02, src03, src04, src05, src06, src07};
 
-template [[host_name("kernel_mul_mm_id_f32_f32")]]  kernel mat_mm_id_t kernel_mul_mm_id<float4x4,   1,     dequantize_f32>;
-template [[host_name("kernel_mul_mm_id_f16_f32")]]  kernel mat_mm_id_t kernel_mul_mm_id<half4x4,    1,     dequantize_f16>;
-template [[host_name("kernel_mul_mm_id_q4_0_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q4_0, 2,     dequantize_q4_0>;
-template [[host_name("kernel_mul_mm_id_q4_1_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q4_1, 2,     dequantize_q4_1>;
-template [[host_name("kernel_mul_mm_id_q5_0_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q5_0, 2,     dequantize_q5_0>;
-template [[host_name("kernel_mul_mm_id_q5_1_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q5_1, 2,     dequantize_q5_1>;
-template [[host_name("kernel_mul_mm_id_q8_0_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q8_0, 2,     dequantize_q8_0>;
-template [[host_name("kernel_mul_mm_id_q2_K_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q2_K, QK_NL, dequantize_q2_K>;
-template [[host_name("kernel_mul_mm_id_q3_K_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q3_K, QK_NL, dequantize_q3_K>;
-template [[host_name("kernel_mul_mm_id_q4_K_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q4_K, QK_NL, dequantize_q4_K>;
-template [[host_name("kernel_mul_mm_id_q5_K_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q5_K, QK_NL, dequantize_q5_K>;
-template [[host_name("kernel_mul_mm_id_q6_K_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q6_K, QK_NL, dequantize_q6_K>;
-template [[host_name("kernel_mul_mm_id_iq2_xxs_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_iq2_xxs, QK_NL, dequantize_iq2_xxs>;
-template [[host_name("kernel_mul_mm_id_iq2_xs_f32")]]  kernel mat_mm_id_t kernel_mul_mm_id<block_iq2_xs,  QK_NL, dequantize_iq2_xs>;
-template [[host_name("kernel_mul_mm_id_iq3_xxs_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_iq3_xxs, QK_NL, dequantize_iq3_xxs>;
+    const int64_t bid = tgpig.z/(ne12*ne13);
 
-//
-// matrix-vector multiplication
-//
+    tgpig.z = tgpig.z%(ne12*ne13);
 
-[[host_name("kernel_mul_mv_id_f32_f32")]]
-kernel void kernel_mul_mv_id_f32_f32(
+    const int32_t id = ((device int32_t *) (ids + bid*nbi1))[idx];
+
+    mul_vec_q_n_f32_impl<block_q4_1, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(
+        src0[id],
+        (device const float *) (src1 + bid*nb11),
+        dst + bid*ne0,
+        ne00,
+        ne01,
+        ne02,
+        ne10,
+        ne12,
+        ne0,
+        ne1,
+        r2,
+        r3,
+        tgpig,
+        tiisg,
+        sgitg);
+}
+
+[[host_name("kernel_mul_mv_id_q5_0_f32")]]
+kernel void kernel_mul_mv_id_q5_0_f32(
         device const    char * ids,
         device const    char * src1,
         device         float * dst,
@@ -5200,32 +6810,26 @@ kernel void kernel_mul_mv_id_f32_f32(
 
     const int32_t id = ((device int32_t *) (ids + bid*nbi1))[idx];
 
-    kernel_mul_mv_f32_f32_impl(
+    mul_vec_q_n_f32_impl<block_q5_0, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(
         src0[id],
-        src1 + bid*nb11,
-        dst  + bid*ne0,
+        (device const float *) (src1 + bid*nb11),
+        dst + bid*ne0,
         ne00,
         ne01,
         ne02,
-        nb00,
-        nb01,
-        nb02,
         ne10,
-        ne11,
         ne12,
-        nb10,
-        nb11,
-        nb12,
         ne0,
         ne1,
         r2,
         r3,
         tgpig,
-        tiisg);
+        tiisg,
+        sgitg);
 }
 
-[[host_name("kernel_mul_mv_id_f16_f32")]]
-kernel void kernel_mul_mv_id_f16_f32(
+[[host_name("kernel_mul_mv_id_q5_1_f32")]]
+kernel void kernel_mul_mv_id_q5_1_f32(
         device const    char * ids,
         device const    char * src1,
         device         float * dst,
@@ -5269,32 +6873,26 @@ kernel void kernel_mul_mv_id_f16_f32(
 
     const int32_t id = ((device int32_t *) (ids + bid*nbi1))[idx];
 
-    kernel_mul_mv_f16_f32_impl(
+    mul_vec_q_n_f32_impl<block_q5_1, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(
         src0[id],
-        src1 + bid*nb11,
-        dst  + bid*ne0,
+        (device const float *) (src1 + bid*nb11),
+        dst + bid*ne0,
         ne00,
         ne01,
         ne02,
-        nb00,
-        nb01,
-        nb02,
         ne10,
-        ne11,
         ne12,
-        nb10,
-        nb11,
-        nb12,
         ne0,
         ne1,
         r2,
         r3,
         tgpig,
-        tiisg);
+        tiisg,
+        sgitg);
 }
 
-[[host_name("kernel_mul_mv_id_q8_0_f32")]]
-kernel void kernel_mul_mv_id_q8_0_f32(
+[[host_name("kernel_mul_mv_id_q2_K_f32")]]
+kernel void kernel_mul_mv_id_q2_K_f32(
         device const    char * ids,
         device const    char * src1,
         device         float * dst,
@@ -5338,7 +6936,7 @@ kernel void kernel_mul_mv_id_q8_0_f32(
 
     const int32_t id = ((device int32_t *) (ids + bid*nbi1))[idx];
 
-    kernel_mul_mv_q8_0_f32_impl(
+    kernel_mul_mv_q2_K_f32_impl(
         src0[id],
         (device const float *) (src1 + bid*nb11),
         dst + bid*ne0,
@@ -5356,8 +6954,8 @@ kernel void kernel_mul_mv_id_q8_0_f32(
         sgitg);
 }
 
-[[host_name("kernel_mul_mv_id_q4_0_f32")]]
-kernel void kernel_mul_mv_id_q4_0_f32(
+[[host_name("kernel_mul_mv_id_q3_K_f32")]]
+kernel void kernel_mul_mv_id_q3_K_f32(
         device const    char * ids,
         device const    char * src1,
         device         float * dst,
@@ -5401,7 +6999,7 @@ kernel void kernel_mul_mv_id_q4_0_f32(
 
     const int32_t id = ((device int32_t *) (ids + bid*nbi1))[idx];
 
-    mul_vec_q_n_f32_impl<block_q4_0, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(
+    kernel_mul_mv_q3_K_f32_impl(
         src0[id],
         (device const float *) (src1 + bid*nb11),
         dst + bid*ne0,
@@ -5419,8 +7017,8 @@ kernel void kernel_mul_mv_id_q4_0_f32(
         sgitg);
 }
 
-[[host_name("kernel_mul_mv_id_q4_1_f32")]]
-kernel void kernel_mul_mv_id_q4_1_f32(
+[[host_name("kernel_mul_mv_id_q4_K_f32")]]
+kernel void kernel_mul_mv_id_q4_K_f32(
         device const    char * ids,
         device const    char * src1,
         device         float * dst,
@@ -5464,7 +7062,7 @@ kernel void kernel_mul_mv_id_q4_1_f32(
 
     const int32_t id = ((device int32_t *) (ids + bid*nbi1))[idx];
 
-    mul_vec_q_n_f32_impl<block_q4_1, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(
+    kernel_mul_mv_q4_K_f32_impl(
         src0[id],
         (device const float *) (src1 + bid*nb11),
         dst + bid*ne0,
@@ -5482,8 +7080,8 @@ kernel void kernel_mul_mv_id_q4_1_f32(
         sgitg);
 }
 
-[[host_name("kernel_mul_mv_id_q5_0_f32")]]
-kernel void kernel_mul_mv_id_q5_0_f32(
+[[host_name("kernel_mul_mv_id_q5_K_f32")]]
+kernel void kernel_mul_mv_id_q5_K_f32(
         device const    char * ids,
         device const    char * src1,
         device         float * dst,
@@ -5527,7 +7125,7 @@ kernel void kernel_mul_mv_id_q5_0_f32(
 
     const int32_t id = ((device int32_t *) (ids + bid*nbi1))[idx];
 
-    mul_vec_q_n_f32_impl<block_q5_0, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(
+    kernel_mul_mv_q5_K_f32_impl(
         src0[id],
         (device const float *) (src1 + bid*nb11),
         dst + bid*ne0,
@@ -5545,8 +7143,8 @@ kernel void kernel_mul_mv_id_q5_0_f32(
         sgitg);
 }
 
-[[host_name("kernel_mul_mv_id_q5_1_f32")]]
-kernel void kernel_mul_mv_id_q5_1_f32(
+[[host_name("kernel_mul_mv_id_q6_K_f32")]]
+kernel void kernel_mul_mv_id_q6_K_f32(
         device const    char * ids,
         device const    char * src1,
         device         float * dst,
@@ -5590,7 +7188,7 @@ kernel void kernel_mul_mv_id_q5_1_f32(
 
     const int32_t id = ((device int32_t *) (ids + bid*nbi1))[idx];
 
-    mul_vec_q_n_f32_impl<block_q5_1, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(
+    kernel_mul_mv_q6_K_f32_impl(
         src0[id],
         (device const float *) (src1 + bid*nb11),
         dst + bid*ne0,
@@ -5608,8 +7206,8 @@ kernel void kernel_mul_mv_id_q5_1_f32(
         sgitg);
 }
 
-[[host_name("kernel_mul_mv_id_q2_K_f32")]]
-kernel void kernel_mul_mv_id_q2_K_f32(
+[[host_name("kernel_mul_mv_id_iq2_xxs_f32")]]
+kernel void kernel_mul_mv_id_iq2_xxs_f32(
         device const    char * ids,
         device const    char * src1,
         device         float * dst,
@@ -5641,6 +7239,7 @@ kernel void kernel_mul_mv_id_q2_K_f32(
         device const    char * src05,
         device const    char * src06,
         device const    char * src07,
+        threadgroup int8_t   * shared_values [[threadgroup(0)]],
         uint3                  tgpig[[threadgroup_position_in_grid]],
         uint                   tiitg[[thread_index_in_threadgroup]],
         uint                   tiisg[[thread_index_in_simdgroup]],
@@ -5653,7 +7252,7 @@ kernel void kernel_mul_mv_id_q2_K_f32(
 
     const int32_t id = ((device int32_t *) (ids + bid*nbi1))[idx];
 
-    kernel_mul_mv_q2_K_f32_impl(
+    kernel_mul_mv_iq2_xxs_f32_impl(
         src0[id],
         (device const float *) (src1 + bid*nb11),
         dst + bid*ne0,
@@ -5666,13 +7265,14 @@ kernel void kernel_mul_mv_id_q2_K_f32(
         ne1,
         r2,
         r3,
+        shared_values,
         tgpig,
         tiisg,
         sgitg);
 }
 
-[[host_name("kernel_mul_mv_id_q3_K_f32")]]
-kernel void kernel_mul_mv_id_q3_K_f32(
+[[host_name("kernel_mul_mv_id_iq2_xs_f32")]]
+kernel void kernel_mul_mv_id_iq2_xs_f32(
         device const    char * ids,
         device const    char * src1,
         device         float * dst,
@@ -5704,6 +7304,7 @@ kernel void kernel_mul_mv_id_q3_K_f32(
         device const    char * src05,
         device const    char * src06,
         device const    char * src07,
+        threadgroup int8_t   * shared_values [[threadgroup(0)]],
         uint3                  tgpig[[threadgroup_position_in_grid]],
         uint                   tiitg[[thread_index_in_threadgroup]],
         uint                   tiisg[[thread_index_in_simdgroup]],
@@ -5716,7 +7317,7 @@ kernel void kernel_mul_mv_id_q3_K_f32(
 
     const int32_t id = ((device int32_t *) (ids + bid*nbi1))[idx];
 
-    kernel_mul_mv_q3_K_f32_impl(
+    kernel_mul_mv_iq2_xs_f32_impl(
         src0[id],
         (device const float *) (src1 + bid*nb11),
         dst + bid*ne0,
@@ -5729,13 +7330,14 @@ kernel void kernel_mul_mv_id_q3_K_f32(
         ne1,
         r2,
         r3,
+        shared_values,
         tgpig,
         tiisg,
         sgitg);
 }
 
-[[host_name("kernel_mul_mv_id_q4_K_f32")]]
-kernel void kernel_mul_mv_id_q4_K_f32(
+[[host_name("kernel_mul_mv_id_iq3_xxs_f32")]]
+kernel void kernel_mul_mv_id_iq3_xxs_f32(
         device const    char * ids,
         device const    char * src1,
         device         float * dst,
@@ -5767,6 +7369,7 @@ kernel void kernel_mul_mv_id_q4_K_f32(
         device const    char * src05,
         device const    char * src06,
         device const    char * src07,
+        threadgroup int8_t   * shared_values [[threadgroup(0)]],
         uint3                  tgpig[[threadgroup_position_in_grid]],
         uint                   tiitg[[thread_index_in_threadgroup]],
         uint                   tiisg[[thread_index_in_simdgroup]],
@@ -5779,7 +7382,7 @@ kernel void kernel_mul_mv_id_q4_K_f32(
 
     const int32_t id = ((device int32_t *) (ids + bid*nbi1))[idx];
 
-    kernel_mul_mv_q4_K_f32_impl(
+    kernel_mul_mv_iq3_xxs_f32_impl(
         src0[id],
         (device const float *) (src1 + bid*nb11),
         dst + bid*ne0,
@@ -5792,13 +7395,14 @@ kernel void kernel_mul_mv_id_q4_K_f32(
         ne1,
         r2,
         r3,
+        shared_values,
         tgpig,
         tiisg,
         sgitg);
 }
 
-[[host_name("kernel_mul_mv_id_q5_K_f32")]]
-kernel void kernel_mul_mv_id_q5_K_f32(
+[[host_name("kernel_mul_mv_id_iq3_s_f32")]]
+kernel void kernel_mul_mv_id_iq3_s_f32(
         device const    char * ids,
         device const    char * src1,
         device         float * dst,
@@ -5830,6 +7434,7 @@ kernel void kernel_mul_mv_id_q5_K_f32(
         device const    char * src05,
         device const    char * src06,
         device const    char * src07,
+        threadgroup int8_t   * shared_values [[threadgroup(0)]],
         uint3                  tgpig[[threadgroup_position_in_grid]],
         uint                   tiitg[[thread_index_in_threadgroup]],
         uint                   tiisg[[thread_index_in_simdgroup]],
@@ -5842,7 +7447,7 @@ kernel void kernel_mul_mv_id_q5_K_f32(
 
     const int32_t id = ((device int32_t *) (ids + bid*nbi1))[idx];
 
-    kernel_mul_mv_q5_K_f32_impl(
+    kernel_mul_mv_iq3_s_f32_impl(
         src0[id],
         (device const float *) (src1 + bid*nb11),
         dst + bid*ne0,
@@ -5855,13 +7460,14 @@ kernel void kernel_mul_mv_id_q5_K_f32(
         ne1,
         r2,
         r3,
+        shared_values,
         tgpig,
         tiisg,
         sgitg);
 }
 
-[[host_name("kernel_mul_mv_id_q6_K_f32")]]
-kernel void kernel_mul_mv_id_q6_K_f32(
+[[host_name("kernel_mul_mv_id_iq2_s_f32")]]
+kernel void kernel_mul_mv_id_iq2_s_f32(
         device const    char * ids,
         device const    char * src1,
         device         float * dst,
@@ -5893,6 +7499,7 @@ kernel void kernel_mul_mv_id_q6_K_f32(
         device const    char * src05,
         device const    char * src06,
         device const    char * src07,
+        threadgroup int8_t   * shared_values [[threadgroup(0)]],
         uint3                  tgpig[[threadgroup_position_in_grid]],
         uint                   tiitg[[thread_index_in_threadgroup]],
         uint                   tiisg[[thread_index_in_simdgroup]],
@@ -5905,7 +7512,7 @@ kernel void kernel_mul_mv_id_q6_K_f32(
 
     const int32_t id = ((device int32_t *) (ids + bid*nbi1))[idx];
 
-    kernel_mul_mv_q6_K_f32_impl(
+    kernel_mul_mv_iq2_s_f32_impl(
         src0[id],
         (device const float *) (src1 + bid*nb11),
         dst + bid*ne0,
@@ -5918,13 +7525,14 @@ kernel void kernel_mul_mv_id_q6_K_f32(
         ne1,
         r2,
         r3,
+        shared_values,
         tgpig,
         tiisg,
         sgitg);
 }
 
-[[host_name("kernel_mul_mv_id_iq2_xxs_f32")]]
-kernel void kernel_mul_mv_id_iq2_xxs_f32(
+[[host_name("kernel_mul_mv_id_iq1_s_f32")]]
+kernel void kernel_mul_mv_id_iq1_s_f32(
         device const    char * ids,
         device const    char * src1,
         device         float * dst,
@@ -5956,7 +7564,6 @@ kernel void kernel_mul_mv_id_iq2_xxs_f32(
         device const    char * src05,
         device const    char * src06,
         device const    char * src07,
-        threadgroup int8_t   * shared_values [[threadgroup(0)]],
         uint3                  tgpig[[threadgroup_position_in_grid]],
         uint                   tiitg[[thread_index_in_threadgroup]],
         uint                   tiisg[[thread_index_in_simdgroup]],
@@ -5969,7 +7576,7 @@ kernel void kernel_mul_mv_id_iq2_xxs_f32(
 
     const int32_t id = ((device int32_t *) (ids + bid*nbi1))[idx];
 
-    kernel_mul_mv_iq2_xxs_f32_impl(
+    kernel_mul_mv_iq1_s_f32_impl(
         src0[id],
         (device const float *) (src1 + bid*nb11),
         dst + bid*ne0,
@@ -5982,14 +7589,13 @@ kernel void kernel_mul_mv_id_iq2_xxs_f32(
         ne1,
         r2,
         r3,
-        shared_values,
         tgpig,
         tiisg,
         sgitg);
 }
 
-[[host_name("kernel_mul_mv_id_iq2_xs_f32")]]
-kernel void kernel_mul_mv_id_iq2_xs_f32(
+[[host_name("kernel_mul_mv_id_iq4_nl_f32")]]
+kernel void kernel_mul_mv_id_iq4_nl_f32(
         device const    char * ids,
         device const    char * src1,
         device         float * dst,
@@ -6021,7 +7627,7 @@ kernel void kernel_mul_mv_id_iq2_xs_f32(
         device const    char * src05,
         device const    char * src06,
         device const    char * src07,
-        threadgroup int8_t   * shared_values [[threadgroup(0)]],
+        threadgroup float    * shared_values [[threadgroup(0)]],
         uint3                  tgpig[[threadgroup_position_in_grid]],
         uint                   tiitg[[thread_index_in_threadgroup]],
         uint                   tiisg[[thread_index_in_simdgroup]],
@@ -6034,7 +7640,7 @@ kernel void kernel_mul_mv_id_iq2_xs_f32(
 
     const int32_t id = ((device int32_t *) (ids + bid*nbi1))[idx];
 
-    kernel_mul_mv_iq2_xs_f32_impl(
+    kernel_mul_mv_iq4_nl_f32_impl(
         src0[id],
         (device const float *) (src1 + bid*nb11),
         dst + bid*ne0,
@@ -6053,8 +7659,8 @@ kernel void kernel_mul_mv_id_iq2_xs_f32(
         sgitg);
 }
 
-[[host_name("kernel_mul_mv_id_iq3_xxs_f32")]]
-kernel void kernel_mul_mv_id_iq3_xxs_f32(
+[[host_name("kernel_mul_mv_id_iq4_xs_f32")]]
+kernel void kernel_mul_mv_id_iq4_xs_f32(
         device const    char * ids,
         device const    char * src1,
         device         float * dst,
@@ -6086,7 +7692,7 @@ kernel void kernel_mul_mv_id_iq3_xxs_f32(
         device const    char * src05,
         device const    char * src06,
         device const    char * src07,
-        threadgroup int8_t   * shared_values [[threadgroup(0)]],
+        threadgroup float    * shared_values [[threadgroup(0)]],
         uint3                  tgpig[[threadgroup_position_in_grid]],
         uint                   tiitg[[thread_index_in_threadgroup]],
         uint                   tiisg[[thread_index_in_simdgroup]],
@@ -6099,7 +7705,11 @@ kernel void kernel_mul_mv_id_iq3_xxs_f32(
 
     const int32_t id = ((device int32_t *) (ids + bid*nbi1))[idx];
 
-    kernel_mul_mv_iq3_xxs_f32_impl(
+#if QK_K == 64
+    kernel_mul_mv_iq4_nl_f32_impl(
+#else
+    kernel_mul_mv_iq4_xs_f32_impl(
+#endif
         src0[id],
         (device const float *) (src1 + bid*nb11),
         dst + bid*ne0,