whisper.rn 0.4.0-rc.3 → 0.4.0-rc.4

package/cpp/{ggml-metal.metal → ggml-metal-whisper.metal}

@@ -13,23 +13,85 @@ typedef struct {
 
 #define QK4_1 32
 typedef struct {
-    half d;          // delta
-    half m;          // min
+    half d;                 // delta
+    half m;                 // min
     uint8_t qs[QK4_1 / 2];  // nibbles / quants
 } block_q4_1;
 
+#define QK5_0 32
+typedef struct {
+    half d;                // delta
+    uint8_t qh[4];         // 5-th bit of quants
+    uint8_t qs[QK5_0 / 2]; // nibbles / quants
+} block_q5_0;
+
+#define QK5_1 32
+typedef struct {
+    half d;                 // delta
+    half m;                 // min
+    uint8_t qh[4];          // 5-th bit of quants
+    uint8_t qs[QK5_1 / 2];  // nibbles / quants
+} block_q5_1;
+
 #define QK8_0 32
 typedef struct {
     half    d;         // delta
     int8_t  qs[QK8_0]; // quants
 } block_q8_0;
 
+// general-purpose kernel for addition of two tensors
+// pros: works for non-contiguous tensors, supports broadcast across dims 1, 2 and 3
+// cons: not very efficient
 kernel void kernel_add(
-        device const float4 * src0,
-        device const float4 * src1,
-        device       float4 * dst,
-        uint tpig[[thread_position_in_grid]]) {
-    dst[tpig] = src0[tpig] + src1[tpig];
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        constant  int64_t & ne00,
+        constant  int64_t & ne01,
+        constant  int64_t & ne02,
+        constant  int64_t & ne03,
+        constant  int64_t & nb00,
+        constant  int64_t & nb01,
+        constant  int64_t & nb02,
+        constant  int64_t & nb03,
+        constant  int64_t & ne10,
+        constant  int64_t & ne11,
+        constant  int64_t & ne12,
+        constant  int64_t & ne13,
+        constant  int64_t & nb10,
+        constant  int64_t & nb11,
+        constant  int64_t & nb12,
+        constant  int64_t & nb13,
+        constant  int64_t & ne0,
+        constant  int64_t & ne1,
+        constant  int64_t & ne2,
+        constant  int64_t & ne3,
+        constant  int64_t & nb0,
+        constant  int64_t & nb1,
+        constant  int64_t & nb2,
+        constant  int64_t & nb3,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    const int64_t i03 = tgpig.z;
+    const int64_t i02 = tgpig.y;
+    const int64_t i01 = tgpig.x;
+
+    const int64_t i13 = i03 % ne13;
+    const int64_t i12 = i02 % ne12;
+    const int64_t i11 = i01 % ne11;
+
+    device const char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01 + tpitg.x*nb00;
+    device const char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11 + tpitg.x*nb10;
+    device       char * dst_ptr  = dst  + i03*nb3  + i02*nb2  + i01*nb1  + tpitg.x*nb0;
+
+    for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
+        ((device float *)dst_ptr)[0] = ((device float *)src0_ptr)[0] + ((device float *)src1_ptr)[0];
+
+        src0_ptr += ntg.x*nb00;
+        src1_ptr += ntg.x*nb10;
+        dst_ptr  += ntg.x*nb0;
+    }
 }
 
 // assumption: src1 is a row
@@ -38,7 +100,7 @@ kernel void kernel_add_row(
         device const float4 * src0,
         device const float4 * src1,
         device       float4 * dst,
-        constant    int64_t & nb,
+        constant    int64_t & nb [[buffer(27)]],
         uint tpig[[thread_position_in_grid]]) {
     dst[tpig] = src0[tpig] + src1[tpig % nb];
 }
@@ -63,9 +125,17 @@ kernel void kernel_mul_row(
 }
 
 kernel void kernel_scale(
+        device const float * src0,
+        device       float * dst,
+        constant     float & scale,
+        uint tpig[[thread_position_in_grid]]) {
+    dst[tpig] = src0[tpig] * scale;
+}
+
+kernel void kernel_scale_4(
         device const float4 * src0,
         device       float4 * dst,
-        constant     float & scale,
+        constant     float  & scale,
         uint tpig[[thread_position_in_grid]]) {
     dst[tpig] = src0[tpig] * scale;
 }
@@ -85,6 +155,13 @@ kernel void kernel_relu(
     dst[tpig] = max(0.0f, src0[tpig]);
 }
 
+kernel void kernel_sqr(
+        device const float * src0,
+        device       float * dst,
+        uint tpig[[thread_position_in_grid]]) {
+    dst[tpig] = src0[tpig] * src0[tpig];
+}
+
 constant float GELU_COEF_A    = 0.044715f;
 constant float SQRT_2_OVER_PI = 0.79788456080286535587989211986876f;
 
@@ -107,36 +184,73 @@ kernel void kernel_soft_max(
         constant   int64_t & ne00,
         constant   int64_t & ne01,
         constant   int64_t & ne02,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]],
-        uint3   ntg[[threads_per_threadgroup]]) {
-    const int64_t i03 = tgpig[2];
-    const int64_t i02 = tgpig[1];
-    const int64_t i01 = tgpig[0];
+        threadgroup float  * buf [[threadgroup(0)]],
+        uint  tgpig[[threadgroup_position_in_grid]],
+        uint  tpitg[[thread_position_in_threadgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint    ntg[[threads_per_threadgroup]]) {
+    const int64_t i03 = (tgpig) / (ne02*ne01);
+    const int64_t i02 = (tgpig - i03*ne02*ne01) / ne01;
+    const int64_t i01 = (tgpig - i03*ne02*ne01 - i02*ne01);
 
     device const float * psrc0 = src0 + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
     device       float * pdst  = dst  + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
 
     // parallel max
-    float lmax = psrc0[tpitg[0]];
-    for (int i00 = tpitg[0] + ntg[0]; i00 < ne00; i00 += ntg[0]) {
+    float lmax = tpitg < ne00 ? psrc0[tpitg] : -INFINITY;
+
+    for (int i00 = tpitg + ntg; i00 < ne00; i00 += ntg) {
         lmax = MAX(lmax, psrc0[i00]);
     }
-    const float max = simd_max(lmax);
+
+    float max = simd_max(lmax);
+    if (tiisg == 0) {
+        buf[sgitg] = max;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    // broadcast, simd group number is ntg / 32
+    for (uint i = ntg / 32 / 2; i > 0; i /= 2) {
+       if (tpitg < i) {
+           buf[tpitg] = MAX(buf[tpitg], buf[tpitg + i]);
+       }
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    max = buf[0];
 
     // parallel sum
     float lsum = 0.0f;
-    for (int i00 = tpitg[0]; i00 < ne00; i00 += ntg[0]) {
+    for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
         const float exp_psrc0 = exp(psrc0[i00] - max);
         lsum += exp_psrc0;
         // Remember the result of exp here. exp is expensive, so we really do not
-        // whish to compute it twice.
+        // wish to compute it twice.
         pdst[i00] = exp_psrc0;
     }
 
-    const float sum = simd_sum(lsum);
+    float sum = simd_sum(lsum);
+    if (tiisg == 0) {
+        buf[sgitg] = sum;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    // broadcast, simd group number is ntg / 32
+    for (uint i = ntg / 32 / 2; i > 0; i /= 2) {
+       if (tpitg < i) {
+           buf[tpitg] += buf[tpitg + i];
+       }
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    sum = buf[0];
 
-    for (int i00 = tpitg[0]; i00 < ne00; i00 += ntg[0]) {
+    for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
         pdst[i00] /= sum;
     }
 }
@@ -147,37 +261,73 @@ kernel void kernel_soft_max_4(
         constant   int64_t & ne00,
         constant   int64_t & ne01,
         constant   int64_t & ne02,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]],
-        uint3   ntg[[threads_per_threadgroup]]) {
-    const int64_t i03 = tgpig[2];
-    const int64_t i02 = tgpig[1];
-    const int64_t i01 = tgpig[0];
+        threadgroup float  * buf [[threadgroup(0)]],
+        uint  tgpig[[threadgroup_position_in_grid]],
+        uint  tpitg[[thread_position_in_threadgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint    ntg[[threads_per_threadgroup]]) {
+    const int64_t i03 = (tgpig) / (ne02*ne01);
+    const int64_t i02 = (tgpig - i03*ne02*ne01) / ne01;
+    const int64_t i01 = (tgpig - i03*ne02*ne01 - i02*ne01);
 
     device const float4 * psrc4 = (device const float4 *)(src0 + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00);
     device       float4 * pdst4 = (device       float4 *)(dst  + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00);
 
     // parallel max
-    float4 lmax4 = psrc4[tpitg[0]];
-    for (int i00 = tpitg[0] + ntg[0]; i00 < ne00/4; i00 += ntg[0]) {
+    float4 lmax4 = tpitg < ne00/4 ? psrc4[tpitg] : -INFINITY;
+
+    for (int i00 = tpitg + ntg; i00 < ne00/4; i00 += ntg) {
         lmax4 = fmax(lmax4, psrc4[i00]);
     }
-    float lmax = MAX(MAX(lmax4[0], lmax4[1]), MAX(lmax4[2], lmax4[3]));
 
-    const float max = simd_max(lmax);
+    const float lmax = MAX(MAX(lmax4[0], lmax4[1]), MAX(lmax4[2], lmax4[3]));
+    float max = simd_max(lmax);
+    if (tiisg == 0) {
+        buf[sgitg] = max;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    // broadcast, simd group number is ntg / 32
+    for (uint i = ntg / 32 / 2; i > 0; i /= 2) {
+       if (tpitg < i) {
+           buf[tpitg] = MAX(buf[tpitg], buf[tpitg + i]);
+       }
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    max = buf[0];
 
     // parallel sum
     float4 lsum4 = 0.0f;
-    for (int i00 = tpitg[0]; i00 < ne00/4; i00 += ntg[0]) {
+    for (int i00 = tpitg; i00 < ne00/4; i00 += ntg) {
         const float4 exp_psrc4 = exp(psrc4[i00] - max);
         lsum4 += exp_psrc4;
         pdst4[i00] = exp_psrc4;
     }
-    float lsum = lsum4[0] + lsum4[1] + lsum4[2] + lsum4[3];
 
-    const float sum = simd_sum(lsum);
+    const float lsum = lsum4[0] + lsum4[1] + lsum4[2] + lsum4[3];
+    float sum = simd_sum(lsum);
+    if (tiisg == 0) {
+        buf[sgitg] = sum;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    // broadcast, simd group number is ntg / 32
+    for (uint i = ntg / 32 / 2; i > 0; i /= 2) {
+       if (tpitg < i) {
+           buf[tpitg] += buf[tpitg + i];
+       }
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    sum = buf[0];
 
-    for (int i00 = tpitg[0]; i00 < ne00/4; i00 += ntg[0]) {
+    for (int i00 = tpitg; i00 < ne00/4; i00 += ntg) {
         pdst4[i00] /= sum;
     }
 }
@@ -197,7 +347,7 @@ kernel void kernel_diag_mask_inf(
         dst[i02*ne01*ne00 + i01*ne00 + i00] = -INFINITY;
     } else {
         dst[i02*ne01*ne00 + i01*ne00 + i00] = src0[i02*ne01*ne00 + i01*ne00 + i00];
-     }
+    }
 }
 
 kernel void kernel_diag_mask_inf_8(
@@ -291,10 +441,11 @@ kernel void kernel_rms_norm(
         uint sgitg[[simdgroup_index_in_threadgroup]],
         uint tiisg[[thread_index_in_simdgroup]],
         uint   ntg[[threads_per_threadgroup]]) {
-    device const float4 * x = (device const float4 *) ((device const char *) src0 + tgpig*nb01);
-    device const float * x_scalar = (device const float *) x;
-    float4 sumf=0;
-    float all_sum=0;
+    device const float4 * x        = (device const float4 *) ((device const char *) src0 + tgpig*nb01);
+    device const float  * x_scalar = (device const float  *) x;
+
+    float4 sumf = 0;
+    float all_sum = 0;
 
     // parallel sum
     for (int i00 = tpitg; i00 < ne00/4; i00 += ntg) {
@@ -307,6 +458,7 @@ kernel void kernel_rms_norm(
     }
 
     threadgroup_barrier(mem_flags::mem_threadgroup);
+
     // broadcast, simd group number is ntg / 32
     for (uint i = ntg / 32 / 2; i > 0; i /= 2) {
        if (tpitg < i) {
@@ -314,7 +466,9 @@ kernel void kernel_rms_norm(
        }
     }
     if (tpitg == 0) {
-        for (int i = 4 * (ne00 / 4); i < ne00; i++) {sum[0] += x_scalar[i];}
+        for (int i = 4 * (ne00 / 4); i < ne00; i++) {
+            sum[0] += x_scalar[i];
+        }
         sum[0] /= ne00;
     }
 
@@ -329,7 +483,9 @@ kernel void kernel_rms_norm(
         y[i00] = x[i00] * scale;
     }
     if (tpitg == 0) {
-        for (int i00 = 4 * (ne00 / 4); i00 < ne00; i00++) {y_scalar[i00] = x_scalar[i00] * scale;}
+        for (int i00 = 4 * (ne00 / 4); i00 < ne00; i00++) {
+            y_scalar[i00] = x_scalar[i00] * scale;
+        }
     }
 }
 
@@ -339,8 +495,11 @@ kernel void kernel_rms_norm(
 // that corresponds to the missing bit shifts (1, 1/16, 1/256, 1/4096)
 inline float block_q_n_dot_y(device const block_q4_0 * qb_curr, float sumy, thread float * yl, int il) {
     float d = qb_curr->d;
+
     float2 acc = 0.f;
+
     device const uint16_t * qs = ((device const uint16_t *)qb_curr + 1 + il/2);
+
     for (int i = 0; i < 8; i+=2) {
         acc[0] += yl[i + 0] * (qs[i / 2] & 0x000F)
                 + yl[i + 1] * (qs[i / 2] & 0x0F00);
@@ -357,8 +516,11 @@ inline float block_q_n_dot_y(device const block_q4_0 * qb_curr, float sumy, thre
 inline float block_q_n_dot_y(device const block_q4_1 * qb_curr, float sumy, thread float * yl, int il) {
     float d = qb_curr->d;
     float m = qb_curr->m;
-    device const uint16_t * qs = ((device const uint16_t *)qb_curr + 2 + il/2);
+
     float2 acc = 0.f;
+
+    device const uint16_t * qs = ((device const uint16_t *)qb_curr + 2 + il/2);
+
     for (int i = 0; i < 8; i+=2) {
         acc[0] += yl[i + 0] * (qs[i / 2] & 0x000F)
                 + yl[i + 1] * (qs[i / 2] & 0x0F00);
@@ -368,9 +530,52 @@ inline float block_q_n_dot_y(device const block_q4_1 * qb_curr, float sumy, thre
     return d * (acc[0] + acc[1]) + sumy * m;
 }
 
+// function for calculate inner product between half a q5_0 block and 16 floats (yl), sumy is SUM(yl[i])
+// il indicates where the q5 quants begin (0 or QK5_0/4)
+// we assume that the yl's have been multiplied with the appropriate scale factor
+// that corresponds to the missing bit shifts (1, 1/16, 1/256, 1/4096)
+inline float block_q_n_dot_y(device const block_q5_0 * qb_curr, float sumy, thread float * yl, int il) {
+    float d = qb_curr->d;
+
+    float2 acc = 0.f;
+
+    device const uint16_t * qs =  ((device const uint16_t *)qb_curr + 3 + il/2);
+           const uint32_t   qh = *((device const uint32_t *)qb_curr->qh);
+
+    for (int i = 0; i < 8; i+=2) {
+        acc[0] += yl[i + 0] * ((qs[i / 2] & 0x000F) | ((qh >> (i+0+il        ) << 4 ) & 0x00010))
+                + yl[i + 1] * ((qs[i / 2] & 0x0F00) | ((qh >> (i+1+il        ) << 12) & 0x01000));
+        acc[1] += yl[i + 8] * ((qs[i / 2] & 0x00F0) | ((qh >> (i+0+il+QK5_0/2) << 8 ) & 0x00100))
+                + yl[i + 9] * ((qs[i / 2] & 0xF000) | ((qh >> (i+1+il+QK5_0/2) << 16) & 0x10000));
+    }
+    return d * (sumy * -16.f + acc[0] + acc[1]);
+}
+
+// function for calculate inner product between half a q5_1 block and 16 floats (yl), sumy is SUM(yl[i])
+// il indicates where the q5 quants begin (0 or QK5_1/4)
+// we assume that the yl's have been multiplied with the appropriate scale factor
+// that corresponds to the missing bit shifts (1, 1/16, 1/256, 1/4096)
+inline float block_q_n_dot_y(device const block_q5_1 * qb_curr, float sumy, thread float * yl, int il) {
+    float d = qb_curr->d;
+    float m = qb_curr->m;
+
+    float2 acc = 0.f;
+
+    device const uint16_t * qs =  ((device const uint16_t *)qb_curr + 4 + il/2);
+           const uint32_t   qh = *((device const uint32_t *)qb_curr->qh);
+
+    for (int i = 0; i < 8; i+=2) {
+        acc[0] += yl[i + 0] * ((qs[i / 2] & 0x000F) | ((qh >> (i+0+il        ) << 4 ) & 0x00010))
+                + yl[i + 1] * ((qs[i / 2] & 0x0F00) | ((qh >> (i+1+il        ) << 12) & 0x01000));
+        acc[1] += yl[i + 8] * ((qs[i / 2] & 0x00F0) | ((qh >> (i+0+il+QK5_0/2) << 8 ) & 0x00100))
+                + yl[i + 9] * ((qs[i / 2] & 0xF000) | ((qh >> (i+1+il+QK5_0/2) << 16) & 0x10000));
+    }
+    return d * (acc[0] + acc[1]) + sumy * m;
+}
+
 // putting them in the kernel cause a significant performance penalty
-#define N_DST 4 // each SIMD group works on 4 rows
-#define N_SIMDGROUP 2 // number of SIMD groups in a thread group
+#define N_DST 4        // each SIMD group works on 4 rows
+#define N_SIMDGROUP 2  // number of SIMD groups in a thread group
 #define N_SIMDWIDTH 32 // assuming SIMD group size is 32
 //Note: This is a template, but strictly speaking it only applies to
 //      quantizations where the block size is 32. It also does not
@@ -381,18 +586,23 @@ void mul_vec_q_n_f32(device const void * src0, device const float * src1, device
                     int64_t ne00, int64_t ne01, int64_t ne02, int64_t ne10, int64_t ne12, int64_t ne0, int64_t ne1, uint gqa,
                     uint3 tgpig, uint tiisg, uint sgitg) {
     const int nb = ne00/QK4_0;
+
     const int r0 = tgpig.x;
     const int r1 = tgpig.y;
     const int im = tgpig.z;
+
     const int first_row = (r0 * nsg + sgitg) * nr;
+
     const uint offset0 = first_row * nb + im/gqa*(nb*ne0);
+
     device const block_q_type * x = (device const block_q_type *) src0 + offset0;
     device const float        * y = (device const float        *) src1 + r1*ne10 + im*ne00*ne1;
-    float yl[16];       // src1 vector cache
-    float sumf[nr]={0.f};
 
-    const int ix = tiisg/2;
-    const int il = 8*(tiisg%2);
+    float yl[16]; // src1 vector cache
+    float sumf[nr] = {0.f};
+
+    const int ix = (tiisg/2);
+    const int il = (tiisg%2)*8;
 
     device const float * yb = y + ix * QK4_0 + il;
 
@@ -403,6 +613,7 @@ void mul_vec_q_n_f32(device const void * src0, device const float * src1, device
             sumy += yb[i] + yb[i+1];
             yl[i+0] = yb[i+ 0];
             yl[i+1] = yb[i+ 1]/256.f;
+
             sumy += yb[i+16] + yb[i+17];
             yl[i+8] = yb[i+16]/16.f;
             yl[i+9] = yb[i+17]/4096.f;
@@ -418,12 +629,12 @@ void mul_vec_q_n_f32(device const void * src0, device const float * src1, device
     for (int row = 0; row < nr; ++row) {
         const float tot = simd_sum(sumf[row]);
         if (tiisg == 0 && first_row + row < ne01) {
-            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = tot;
+            dst[im*ne0*ne1 + r1*ne0 + first_row + row] = tot;
         }
     }
 }
 
-kernel void kernel_mul_mat_q4_0_f32(
+kernel void kernel_mul_mv_q4_0_f32(
         device const  void * src0,
         device const float * src1,
         device       float * dst,
@@ -436,12 +647,12 @@ kernel void kernel_mul_mat_q4_0_f32(
         constant   int64_t & ne1[[buffer(16)]],
         constant   uint    & gqa[[buffer(17)]],
         uint3 tgpig[[threadgroup_position_in_grid]],
-        uint tiisg[[thread_index_in_simdgroup]],
-        uint sgitg[[simdgroup_index_in_threadgroup]]) {
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
     mul_vec_q_n_f32<block_q4_0, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(src0,src1,dst,ne00,ne01,ne02,ne10,ne12,ne0,ne1,gqa,tgpig,tiisg,sgitg);
 }
 
-kernel void kernel_mul_mat_q4_1_f32(
+kernel void kernel_mul_mv_q4_1_f32(
         device const  void * src0,
         device const float * src1,
         device       float * dst,
@@ -459,9 +670,46 @@ kernel void kernel_mul_mat_q4_1_f32(
      mul_vec_q_n_f32<block_q4_1, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(src0,src1,dst,ne00,ne01,ne02,ne10,ne12,ne0,ne1,gqa,tgpig,tiisg,sgitg);
 }
 
+kernel void kernel_mul_mv_q5_0_f32(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01[[buffer(4)]],
+        constant   int64_t & ne02[[buffer(5)]],
+        constant   int64_t & ne10[[buffer(9)]],
+        constant   int64_t & ne12[[buffer(11)]],
+        constant   int64_t & ne0[[buffer(15)]],
+        constant   int64_t & ne1[[buffer(16)]],
+        constant   uint    & gqa[[buffer(17)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+    mul_vec_q_n_f32<block_q5_0, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(src0,src1,dst,ne00,ne01,ne02,ne10,ne12,ne0,ne1,gqa,tgpig,tiisg,sgitg);
+}
+
+kernel void kernel_mul_mv_q5_1_f32(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01[[buffer(4)]],
+        constant   int64_t & ne02[[buffer(5)]],
+        constant   int64_t & ne10[[buffer(9)]],
+        constant   int64_t & ne12[[buffer(11)]],
+        constant   int64_t & ne0[[buffer(15)]],
+        constant   int64_t & ne1[[buffer(16)]],
+        constant   uint    & gqa[[buffer(17)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+    mul_vec_q_n_f32<block_q5_1, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(src0,src1,dst,ne00,ne01,ne02,ne10,ne12,ne0,ne1,gqa,tgpig,tiisg,sgitg);
+}
+
+
 #define NB_Q8_0 8
 
-kernel void kernel_mul_mat_q8_0_f32(
+kernel void kernel_mul_mv_q8_0_f32(
         device const  void * src0,
         device const float * src1,
         device       float * dst,
@@ -525,7 +773,7 @@ kernel void kernel_mul_mat_q8_0_f32(
 
 #define N_F32_F32 4
 
-kernel void kernel_mul_mat_f32_f32(
+kernel void kernel_mul_mv_f32_f32(
         device const  char * src0,
         device const  char * src1,
         device       float * dst,
@@ -596,7 +844,7 @@ kernel void kernel_mul_mat_f32_f32(
     }
 }
 
-kernel void kernel_mul_mat_f16_f32_1row(
+kernel void kernel_mul_mv_f16_f32_1row(
         device const  char * src0,
         device const  char * src1,
         device       float * dst,
@@ -615,7 +863,7 @@ kernel void kernel_mul_mat_f16_f32_1row(
         constant   int64_t & ne0,
         constant   int64_t & ne1,
         uint3 tgpig[[threadgroup_position_in_grid]],
-        uint tiisg[[thread_index_in_simdgroup]]) {
+        uint  tiisg[[thread_index_in_simdgroup]]) {
 
     const int64_t r0 = tgpig.x;
     const int64_t r1 = tgpig.y;
@@ -650,7 +898,7 @@ kernel void kernel_mul_mat_f16_f32_1row(
 
 #define N_F16_F32 4
 
-kernel void kernel_mul_mat_f16_f32(
+kernel void kernel_mul_mv_f16_f32(
         device const  char * src0,
         device const  char * src1,
         device       float * dst,
@@ -722,7 +970,7 @@ kernel void kernel_mul_mat_f16_f32(
 }
 
 // Assumes row size (ne00) is a multiple of 4
-kernel void kernel_mul_mat_f16_f32_l4(
+kernel void kernel_mul_mv_f16_f32_l4(
         device const  char * src0,
         device const  char * src1,
         device       float * dst,
@@ -783,7 +1031,9 @@ kernel void kernel_alibi_f32(
         constant  uint64_t & nb1,
         constant  uint64_t & nb2,
         constant  uint64_t & nb3,
-        constant      float & m0,
+        constant     float & m0,
+        constant     float & m1,
+        constant       int & n_heads_log2_floor,
         uint3 tgpig[[threadgroup_position_in_grid]],
         uint3 tpitg[[thread_position_in_threadgroup]],
         uint3   ntg[[threads_per_threadgroup]]) {
@@ -799,37 +1049,122 @@ kernel void kernel_alibi_f32(
     const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0);
 
     device float * dst_data = (device float *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
-    float m_k = pow(m0, i2 + 1);
+    float m_k;
+    if (i2 < n_heads_log2_floor) {
+        m_k = pow(m0, i2 + 1);
+    } else {
+        m_k = pow(m1, 2 * (i2 - n_heads_log2_floor) + 1);
+    }
     for (int64_t i00 = tpitg.x; i00 < ne00; i00 += ntg.x) {
         device const float * src = (device float *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
         dst_data[i00] = src[0] + m_k * (i00 - ne00 + 1);
     }
 }
 
+static float rope_yarn_ramp(const float low, const float high, const int i0) {
+    const float y = (i0 / 2 - low) / max(0.001f, high - low);
+    return 1.0f - min(1.0f, max(0.0f, y));
+}
+
+// YaRN algorithm based on LlamaYaRNScaledRotaryEmbedding.py from https://github.com/jquesnelle/yarn
+// MIT licensed. Copyright (c) 2023 Jeffrey Quesnelle and Bowen Peng.
+static void rope_yarn(
+    float theta_extrap, float freq_scale, float corr_dims[2], int64_t i0, float ext_factor, float mscale,
+    thread float * cos_theta, thread float * sin_theta
+) {
+    // Get n-d rotational scaling corrected for extrapolation
+    float theta_interp = freq_scale * theta_extrap;
+    float theta = theta_interp;
+    if (ext_factor != 0.0f) {
+        float ramp_mix = rope_yarn_ramp(corr_dims[0], corr_dims[1], i0) * ext_factor;
+        theta = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix;
+
+        // Get n-d magnitude scaling corrected for interpolation
+        mscale *= 1.0f + 0.1f * log(1.0f / freq_scale);
+    }
+    *cos_theta = cos(theta) * mscale;
+    *sin_theta = sin(theta) * mscale;
+}
+
+// Apparently solving `n_rot = 2pi * x * base^((2 * max_pos_emb) / n_dims)` for x, we get
+// `corr_fac(n_rot) = n_dims * log(max_pos_emb / (n_rot * 2pi)) / (2 * log(base))`
+static float rope_yarn_corr_factor(int n_dims, int n_orig_ctx, float n_rot, float base) {
+    return n_dims * log(n_orig_ctx / (n_rot * 2 * M_PI_F)) / (2 * log(base));
+}
+
+static void rope_yarn_corr_dims(
+    int n_dims, int n_orig_ctx, float freq_base, float beta_fast, float beta_slow, float dims[2]
+) {
+    // start and end correction dims
+    dims[0] = max(0.0f,         floor(rope_yarn_corr_factor(n_dims, n_orig_ctx, beta_fast, freq_base)));
+    dims[1] = min(n_dims - 1.0f, ceil(rope_yarn_corr_factor(n_dims, n_orig_ctx, beta_slow, freq_base)));
+}
+
+typedef void (rope_t)(
+        device const    void * src0,
+        device const int32_t * src1,
+        device         float * dst,
+        constant     int64_t & ne00,
+        constant     int64_t & ne01,
+        constant     int64_t & ne02,
+        constant     int64_t & ne03,
+        constant    uint64_t & nb00,
+        constant    uint64_t & nb01,
+        constant    uint64_t & nb02,
+        constant    uint64_t & nb03,
+        constant     int64_t & ne0,
+        constant     int64_t & ne1,
+        constant     int64_t & ne2,
+        constant     int64_t & ne3,
+        constant    uint64_t & nb0,
+        constant    uint64_t & nb1,
+        constant    uint64_t & nb2,
+        constant    uint64_t & nb3,
+        constant         int & n_past,
+        constant         int & n_dims,
+        constant         int & mode,
+        constant         int & n_orig_ctx,
+        constant       float & freq_base,
+        constant       float & freq_scale,
+        constant       float & ext_factor,
+        constant       float & attn_factor,
+        constant       float & beta_fast,
+        constant       float & beta_slow,
+        uint  tiitg[[thread_index_in_threadgroup]],
+        uint3 tptg[[threads_per_threadgroup]],
+        uint3 tgpig[[threadgroup_position_in_grid]]);
+
+template<typename T>
 kernel void kernel_rope(
-        device const  void * src0,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant   int64_t & ne03,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   int64_t & ne2,
-        constant   int64_t & ne3,
-        constant  uint64_t & nb0,
-        constant  uint64_t & nb1,
-        constant  uint64_t & nb2,
-        constant  uint64_t & nb3,
-        constant       int & n_past,
-        constant       int & n_dims,
-        constant       int & mode,
-        constant     float & freq_base,
-        constant     float & freq_scale,
+        device const    void * src0,
+        device const int32_t * src1,
+        device         float * dst,
+        constant     int64_t & ne00,
+        constant     int64_t & ne01,
+        constant     int64_t & ne02,
+        constant     int64_t & ne03,
+        constant    uint64_t & nb00,
+        constant    uint64_t & nb01,
+        constant    uint64_t & nb02,
+        constant    uint64_t & nb03,
+        constant     int64_t & ne0,
+        constant     int64_t & ne1,
+        constant     int64_t & ne2,
+        constant     int64_t & ne3,
+        constant    uint64_t & nb0,
+        constant    uint64_t & nb1,
+        constant    uint64_t & nb2,
+        constant    uint64_t & nb3,
+        constant         int & n_past,
+        constant         int & n_dims,
+        constant         int & mode,
+        constant         int & n_orig_ctx,
+        constant       float & freq_base,
+        constant       float & freq_scale,
+        constant       float & ext_factor,
+        constant       float & attn_factor,
+        constant       float & beta_fast,
+        constant       float & beta_slow,
         uint  tiitg[[thread_index_in_threadgroup]],
         uint3 tptg[[threads_per_threadgroup]],
         uint3 tgpig[[threadgroup_position_in_grid]]) {
@@ -839,23 +1174,28 @@ kernel void kernel_rope(
 
     const bool is_neox = mode & 2;
 
-    const int64_t p = ((mode & 1) == 0 ? n_past + i2 : i2);
+    float corr_dims[2];
+    rope_yarn_corr_dims(n_dims, n_orig_ctx, freq_base, beta_fast, beta_slow, corr_dims);
 
-    const float theta_0 = freq_scale * (float)p;
+    device const int32_t * pos = src1;
+
+    const int64_t p = pos[i2];
+
+    const float theta_0 = (float)p;
     const float inv_ndims = -1.f/n_dims;
 
     if (!is_neox) {
         for (int64_t i0 = 2*tiitg; i0 < ne0; i0 += 2*tptg.x) {
 
             const float theta = theta_0 * pow(freq_base, inv_ndims*i0);
-            const float cos_theta = cos(theta);
-            const float sin_theta = sin(theta);
+            float cos_theta, sin_theta;
+            rope_yarn(theta, freq_scale, corr_dims, i0, ext_factor, attn_factor, &cos_theta, &sin_theta);
 
-            device const float * const src = (device float *)((device char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
-            device       float * dst_data  = (device float *)((device char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+            device const T * const src = (device T *)((device char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+            device       T * dst_data  = (device T *)((device char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
 
-            const float x0 = src[0];
-            const float x1 = src[1];
+            const T x0 = src[0];
+            const T x1 = src[1];
 
             dst_data[0] = x0*cos_theta - x1*sin_theta;
             dst_data[1] = x0*sin_theta + x1*cos_theta;
@@ -864,14 +1204,17 @@ kernel void kernel_rope(
         for (int64_t ib = 0; ib < ne0/n_dims; ++ib) {
             for (int64_t ic = 2*tiitg; ic < n_dims; ic += 2*tptg.x) {
 
-                const float theta = theta_0 * pow(freq_base, inv_ndims*ic - ib);
-                const float cos_theta = cos(theta);
-                const float sin_theta = sin(theta);
+                // simplified from `(ib * n_dims + ic) * inv_ndims`
+                const float cur_rot = inv_ndims*ic - ib;
+
+                const float theta = theta_0 * pow(freq_base, cur_rot);
+                float cos_theta, sin_theta;
+                rope_yarn(theta, freq_scale, corr_dims, cur_rot, ext_factor, attn_factor, &cos_theta, &sin_theta);
 
                 const int64_t i0 = ib*n_dims + ic/2;
 
-                device const float * const src = (device float *)((device char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
-                device       float * dst_data  = (device float *)((device char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+                device const T * const src = (device T *)((device char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+                device       T * dst_data  = (device T *)((device char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
 
                 const float x0 = src[0];
                 const float x1 = src[n_dims/2];
@@ -883,6 +1226,9 @@ kernel void kernel_rope(
     }
 }
 
+template [[host_name("kernel_rope_f32")]] kernel rope_t kernel_rope<float>;
+template [[host_name("kernel_rope_f16")]] kernel rope_t kernel_rope<half>;
+
 kernel void kernel_cpy_f16_f16(
         device const half * src0,
         device       half * dst,
@@ -1008,6 +1354,62 @@ kernel void kernel_cpy_f32_f32(
     }
 }
 
+kernel void kernel_concat(
+    device const char * src0,
+    device const char * src1,
+    device       char * dst,
+    constant   int64_t & ne00,
+    constant   int64_t & ne01,
+    constant   int64_t & ne02,
+    constant   int64_t & ne03,
+    constant  uint64_t & nb00,
+    constant  uint64_t & nb01,
+    constant  uint64_t & nb02,
+    constant  uint64_t & nb03,
+    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  uint64_t & nb13,
+    constant   int64_t & ne0,
+    constant   int64_t & ne1,
+    constant   int64_t & ne2,
+    constant   int64_t & ne3,
+    constant  uint64_t & nb0,
+    constant  uint64_t & nb1,
+    constant  uint64_t & nb2,
+    constant  uint64_t & nb3,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]]) {
+
+    const int64_t i03 = tgpig.z;
+    const int64_t i02 = tgpig.y;
+    const int64_t i01 = tgpig.x;
+
+    const int64_t i13 = i03 % ne13;
+    const int64_t i12 = i02 % ne12;
+    const int64_t i11 = i01 % ne11;
+
+    device const char * src0_ptr = src0 + i03 * nb03 + i02 * nb02 + i01 * nb01 + tpitg.x*nb00;
+    device const char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11 + tpitg.x*nb10;
+    device       char * dst_ptr  = dst  + i03*nb3  + i02*nb2  + i01*nb1  + tpitg.x*nb0;
+
+    for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
+        if (i02 < ne02) {
+            ((device float *)dst_ptr)[0] = ((device float *)src0_ptr)[0];
+            src0_ptr += ntg.x*nb00;
+        } else {
+            ((device float *)dst_ptr)[0] = ((device float *)src1_ptr)[0];
+            src1_ptr += ntg.x*nb10;
+        }
+        dst_ptr += ntg.x*nb0;
+    }
+}
+
 //============================================ k-quants ======================================================
 
 #ifndef QK_K
@@ -1100,7 +1502,7 @@ static inline uchar4 get_scale_min_k4(int j, device const uint8_t * q) {
 
 //====================================== dot products =========================
 
-kernel void kernel_mul_mat_q2_K_f32(
+kernel void kernel_mul_mv_q2_K_f32(
         device const  void * src0,
         device const float * src1,
         device       float * dst,
@@ -1244,7 +1646,7 @@ kernel void kernel_mul_mat_q2_K_f32(
 }
 
 #if QK_K == 256
-kernel void kernel_mul_mat_q3_K_f32(
+kernel void kernel_mul_mv_q3_K_f32(
         device const  void * src0,
         device const float * src1,
         device       float * dst,
@@ -1273,8 +1675,8 @@ kernel void kernel_mul_mat_q3_K_f32(
 
     float yl[32];
 
-    const uint16_t kmask1 = 0x3030;
-    const uint16_t kmask2 = 0x0f0f;
+    //const uint16_t kmask1 = 0x3030;
+    //const uint16_t kmask2 = 0x0f0f;
 
     const int tid = tiisg/4;
     const int ix  = tiisg%4;
@@ -1396,7 +1798,7 @@ kernel void kernel_mul_mat_q3_K_f32(
     }
 }
 #else
-kernel void kernel_mul_mat_q3_K_f32(
+kernel void kernel_mul_mv_q3_K_f32(
         device const  void * src0,
         device const float * src1,
         device       float * dst,
@@ -1467,7 +1869,7 @@ kernel void kernel_mul_mat_q3_K_f32(
 #endif
 
 #if QK_K == 256
-kernel void kernel_mul_mat_q4_K_f32(
+kernel void kernel_mul_mv_q4_K_f32(
         device const  void * src0,
         device const float * src1,
         device       float * dst,
@@ -1573,7 +1975,7 @@ kernel void kernel_mul_mat_q4_K_f32(
     }
 }
 #else
-kernel void kernel_mul_mat_q4_K_f32(
+kernel void kernel_mul_mv_q4_K_f32(
         device const  void * src0,
         device const float * src1,
         device       float * dst,
@@ -1662,7 +2064,7 @@ kernel void kernel_mul_mat_q4_K_f32(
 }
 #endif
 
-kernel void kernel_mul_mat_q5_K_f32(
+kernel void kernel_mul_mv_q5_K_f32(
         device const  void * src0,
         device const float * src1,
         device       float * dst,
@@ -1835,7 +2237,7 @@ kernel void kernel_mul_mat_q5_K_f32(
 
 }
 
-kernel void kernel_mul_mat_q6_K_f32(
+kernel void kernel_mul_mv_q6_K_f32(
         device const  void * src0,
         device const float * src1,
         device       float * dst,
@@ -1984,6 +2386,62 @@ void dequantize_q4_1(device const block_q4_1 *xb, short il, thread type4x4 & reg
     }
 }
 
+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);
@@ -2173,7 +2631,7 @@ kernel void kernel_get_rows(
 }
 
 #define BLOCK_SIZE_M 64 // 8 simdgroup matrices from matrix A
-#define BLOCK_SIZE_N 32 // 4 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
@@ -2210,9 +2668,11 @@ kernel void kernel_mul_mm(device const  uchar * src0,
     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;
@@ -2236,26 +2696,30 @@ kernel void kernel_mul_mm(device const  uchar * src0,
         + 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
+        // 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) \
-            + 16 * (tiitg % THREAD_PER_ROW) + 8 * (i / 8)) \
-            + (tiitg / THREAD_PER_ROW) % 8 + (i & 7) * 8) = temp_a[i/4][i%4];
+            +                     (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);
+
+        *(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
+
+        // 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)
@@ -2270,6 +2734,7 @@ kernel void kernel_mul_mm(device const  uchar * src0,
 
             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]);
@@ -2278,25 +2743,26 @@ kernel void kernel_mul_mm(device const  uchar * src0,
     }
 
     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;
+        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) \
+        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) {
+
+        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) {
+                for (int j = tiitg; j < n_cols; j += BLOCK_SIZE_N) {
                     *(C + i + j * ne0) = *(temp_str + i + j * BLOCK_SIZE_M);
                 }
             }
@@ -2317,6 +2783,8 @@ template [[host_name("kernel_get_rows_f32")]]  kernel get_rows_t kernel_get_rows
 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>;
@@ -2345,6 +2813,8 @@ template [[host_name("kernel_mul_mm_f32_f32")]]  kernel mat_mm_t kernel_mul_mm<f
 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>;