llama_cpp 0.10.0 → 0.10.2

data/ext/llama_cpp/src/ggml-metal.metal

@@ -79,6 +79,7 @@ kernel void kernel_add(
         constant  int64_t & nb1,
         constant  int64_t & nb2,
         constant  int64_t & nb3,
+        constant  int64_t & offs,
         uint3 tgpig[[threadgroup_position_in_grid]],
         uint3 tpitg[[thread_position_in_threadgroup]],
         uint3   ntg[[threads_per_threadgroup]]) {
@@ -90,9 +91,9 @@ kernel void kernel_add(
     const int64_t i12 = i02 % ne12;
     const int64_t i11 = i01 % ne11;
 
-    device const char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01;
+    device const char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01 + offs;
     device const char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11;
-    device       char * dst_ptr  = dst  + i03*nb3  + i02*nb2  + i01*nb1;
+    device       char * dst_ptr  = dst  + i03*nb3  + i02*nb2  + i01*nb1  + offs;
 
     for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
         const int i10 = i0 % ne10;
@@ -204,7 +205,7 @@ kernel void kernel_add_row(
         device const float4 * src0,
         device const float4 * src1,
         device       float4 * dst,
-        constant    int64_t & nb [[buffer(27)]],
+        constant    int64_t & nb [[buffer(28)]],
         uint tpig[[thread_position_in_grid]]) {
     dst[tpig] = src0[tpig] + src1[tpig % nb];
 }
@@ -213,7 +214,7 @@ kernel void kernel_mul_row(
         device const float4 * src0,
         device const float4 * src1,
         device       float4 * dst,
-        constant    int64_t & nb  [[buffer(27)]],
+        constant    int64_t & nb  [[buffer(28)]],
         uint tpig[[thread_position_in_grid]]) {
     dst[tpig] = src0[tpig] * src1[tpig % nb];
 }
@@ -222,7 +223,7 @@ kernel void kernel_div_row(
         device const float4 * src0,
         device const float4 * src1,
         device       float4 * dst,
-        constant    int64_t & nb  [[buffer(27)]],
+        constant    int64_t & nb  [[buffer(28)]],
         uint tpig[[thread_position_in_grid]]) {
     dst[tpig] = src0[tpig] / src1[tpig % nb];
 }
@@ -243,19 +244,53 @@ kernel void kernel_scale_4(
     dst[tpig] = src0[tpig] * scale;
 }
 
-kernel void kernel_silu(
-        device const float4 * src0,
-        device       float4 * dst,
+kernel void kernel_relu(
+        device const float * src0,
+        device       float * dst,
         uint tpig[[thread_position_in_grid]]) {
-    device const float4 & x = src0[tpig];
-    dst[tpig] = x / (1.0f + exp(-x));
+    dst[tpig] = max(0.0f, src0[tpig]);
 }
 
-kernel void kernel_relu(
+kernel void kernel_tanh(
         device const float * src0,
         device       float * dst,
         uint tpig[[thread_position_in_grid]]) {
-    dst[tpig] = max(0.0f, src0[tpig]);
+    device const float & x = src0[tpig];
+    dst[tpig] = precise::tanh(x);
+}
+
+constant float GELU_COEF_A     = 0.044715f;
+constant float GELU_QUICK_COEF = -1.702f;
+constant float SQRT_2_OVER_PI  = 0.79788456080286535587989211986876f;
+
+kernel void kernel_gelu(
+    device const float4 * src0,
+    device       float4 * dst,
+    uint tpig[[thread_position_in_grid]]) {
+    device const float4 & x = src0[tpig];
+
+    // BEWARE !!!
+    // Simply using "tanh" instead of "precise::tanh" will sometimes results in NaNs!
+    // This was observed with Falcon 7B and 40B models
+    //
+    dst[tpig] = 0.5f*x*(1.0f + precise::tanh(SQRT_2_OVER_PI*x*(1.0f + GELU_COEF_A*x*x)));
+}
+
+kernel void kernel_gelu_quick(
+    device const float4 * src0,
+    device       float4 * dst,
+    uint tpig[[thread_position_in_grid]]) {
+    device const float4 & x = src0[tpig];
+
+    dst[tpig] = x*(1.0f/(1.0f+exp(GELU_QUICK_COEF*x)));
+}
+
+kernel void kernel_silu(
+        device const float4 * src0,
+        device       float4 * dst,
+        uint tpig[[thread_position_in_grid]]) {
+    device const float4 & x = src0[tpig];
+    dst[tpig] = x / (1.0f + exp(-x));
 }
 
 kernel void kernel_sqr(
@@ -313,22 +348,6 @@ kernel void kernel_sum_rows(
     dst_row[0] = row_sum;
 }
 
-constant float GELU_COEF_A    = 0.044715f;
-constant float SQRT_2_OVER_PI = 0.79788456080286535587989211986876f;
-
-kernel void kernel_gelu(
-    device const float4 * src0,
-    device       float4 * dst,
-    uint tpig[[thread_position_in_grid]]) {
-    device const float4 & x = src0[tpig];
-
-    // BEWARE !!!
-    // Simply using "tanh" instead of "precise::tanh" will sometimes results in NaNs!
-    // This was observed with Falcon 7B and 40B models
-    //
-    dst[tpig] = 0.5f*x*(1.0f + precise::tanh(SQRT_2_OVER_PI*x*(1.0f + GELU_COEF_A*x*x)));
-}
-
 kernel void kernel_soft_max(
         device const float * src0,
         device const float * src1,
@@ -347,9 +366,9 @@ kernel void kernel_soft_max(
     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 const float * pmask = src1 ? src1                                      + i01*ne00 : nullptr;
-    device       float * pdst  =        dst  + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+    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       float * pdst  =         dst  + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
 
     // parallel max
     float lmax = -INFINITY;
@@ -385,7 +404,12 @@ kernel void kernel_soft_max(
         pdst[i00] = exp_psrc0;
     }
 
+    // This barrier fixes a failing test
+    // ref: https://github.com/ggerganov/ggml/pull/621#discussion_r1425156335
+    threadgroup_barrier(mem_flags::mem_none);
+
     float sum = simd_sum(lsum);
+
     if (ntg > N_SIMDWIDTH) {
         if (sgitg == 0) {
             buf[tiisg] = 0.0f;
@@ -428,9 +452,9 @@ kernel void kernel_soft_max_4(
     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 const float4 * pmask = src1 ? (device const float4 *)(src1 +                                      i01*ne00) : nullptr;
-    device       float4 * pdst4 =        (device       float4 *)(dst  + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00);
+    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       float4 * pdst4 =                (device       float4 *)(dst  + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00);
 
     // parallel max
     float4 lmax4 = -INFINITY;
@@ -468,7 +492,13 @@ kernel void kernel_soft_max_4(
     }
 
     const float lsum = lsum4[0] + lsum4[1] + lsum4[2] + lsum4[3];
+
+    // This barrier fixes a failing test
+    // ref: https://github.com/ggerganov/ggml/pull/621#discussion_r1425156335
+    threadgroup_barrier(mem_flags::mem_none);
+
     float sum = simd_sum(lsum);
+
     if (ntg > N_SIMDWIDTH) {
         if (sgitg == 0) {
             buf[tiisg] = 0.0f;
@@ -639,6 +669,94 @@ kernel void kernel_rms_norm(
     }
 }
 
+kernel void kernel_group_norm(
+        device const float * src0,
+        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   int32_t & n_groups,
+        constant     float & eps,
+        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 ne = ne00*ne01*ne02;
+    const int64_t gs = ne00*ne01*((ne02 + n_groups - 1) / n_groups);
+
+    int start = tgpig * gs;
+    int end   = start + gs;
+
+    start += tpitg;
+
+    if (end >= ne) {
+        end = ne;
+    }
+
+    float tmp = 0.0f; // partial sum for thread in warp
+
+    for (int j = start; j < end; j += ntg) {
+        tmp += src0[j];
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+    tmp = simd_sum(tmp);
+    if (ntg > N_SIMDWIDTH) {
+        if (sgitg == 0) {
+            buf[tiisg] = 0.0f;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (tiisg == 0) {
+            buf[sgitg] = tmp;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        tmp = buf[tiisg];
+        tmp = simd_sum(tmp);
+    }
+
+    const float mean = tmp / gs;
+    tmp = 0.0f;
+
+    for (int j = start; j < end; j += ntg) {
+        float xi = src0[j] - mean;
+        dst[j] = xi;
+        tmp += xi * xi;
+    }
+
+    tmp = simd_sum(tmp);
+    if (ntg > N_SIMDWIDTH) {
+        if (sgitg == 0) {
+            buf[tiisg] = 0.0f;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (tiisg == 0) {
+            buf[sgitg] = tmp;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        tmp = buf[tiisg];
+        tmp = simd_sum(tmp);
+    }
+
+    const float variance = tmp / gs;
+    const float scale = 1.0f/sqrt(variance + eps);
+    for (int j = start; j < end; j += ntg) {
+        dst[j] *= scale;
+    }
+}
+
 // function for calculate inner product between half a q4_0 block and 16 floats (yl), sumy is SUM(yl[i])
 // il indicates where the q4 quants begin (0 or QK4_0/4)
 // we assume that the yl's have been multiplied with the appropriate scale factor
@@ -731,7 +849,7 @@ inline float block_q_n_dot_y(device const block_q5_1 * qb_curr, float sumy, thre
 //      giard against the number of rows not being divisible by
 //      N_DST, so this is another explicit assumption of the implementation.
 template<typename block_q_type, int nr, int nsg, int nw>
-void mul_vec_q_n_f32(
+void mul_vec_q_n_f32_impl(
         device const void  * src0,
         device const float * src1,
         device       float * dst,
@@ -813,7 +931,7 @@ kernel void kernel_mul_mv_q4_0_f32(
         uint3 tgpig[[threadgroup_position_in_grid]],
         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,r2,r3,tgpig,tiisg,sgitg);
+    mul_vec_q_n_f32_impl<block_q4_0, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(src0,src1,dst,ne00,ne01,ne02,ne10,ne12,ne0,ne1,r2,r3,tgpig,tiisg,sgitg);
 }
 
 kernel void kernel_mul_mv_q4_1_f32(
@@ -832,7 +950,7 @@ kernel void kernel_mul_mv_q4_1_f32(
         uint3 tgpig[[threadgroup_position_in_grid]],
         uint tiisg[[thread_index_in_simdgroup]],
         uint sgitg[[simdgroup_index_in_threadgroup]]) {
-     mul_vec_q_n_f32<block_q4_1, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(src0,src1,dst,ne00,ne01,ne02,ne10,ne12,ne0,ne1,r2,r3,tgpig,tiisg,sgitg);
+     mul_vec_q_n_f32_impl<block_q4_1, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(src0,src1,dst,ne00,ne01,ne02,ne10,ne12,ne0,ne1,r2,r3,tgpig,tiisg,sgitg);
 }
 
 kernel void kernel_mul_mv_q5_0_f32(
@@ -851,7 +969,7 @@ kernel void kernel_mul_mv_q5_0_f32(
         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,r2,r3,tgpig,tiisg,sgitg);
+    mul_vec_q_n_f32_impl<block_q5_0, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(src0,src1,dst,ne00,ne01,ne02,ne10,ne12,ne0,ne1,r2,r3,tgpig,tiisg,sgitg);
 }
 
 kernel void kernel_mul_mv_q5_1_f32(
@@ -870,28 +988,28 @@ kernel void kernel_mul_mv_q5_1_f32(
         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,r2,r3,tgpig,tiisg,sgitg);
+    mul_vec_q_n_f32_impl<block_q5_1, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(src0,src1,dst,ne00,ne01,ne02,ne10,ne12,ne0,ne1,r2,r3,tgpig,tiisg,sgitg);
 }
 
 
 #define NB_Q8_0 8
 
-kernel void kernel_mul_mv_q8_0_f32(
+void kernel_mul_mv_q8_0_f32_impl(
         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    & r2  [[buffer(17)]],
-        constant   uint    & r3  [[buffer(18)]],
+        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,
         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]]) {
     const int nr  = N_DST;
     const int nsg = N_SIMDGROUP;
     const int nw  = N_SIMDWIDTH;
@@ -945,9 +1063,29 @@ kernel void kernel_mul_mv_q8_0_f32(
     }
 }
 
+[[host_name("kernel_mul_mv_q8_0_f32")]]
+kernel void kernel_mul_mv_q8_0_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   int64_t & ne10,
+        constant   int64_t & ne12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   uint    & r2   [[buffer(17)]],
+        constant   uint    & r3   [[buffer(18)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+    kernel_mul_mv_q8_0_f32_impl(src0,src1,dst,ne00,ne01,ne02,ne10,ne12,ne0,ne1,r2,r3,tgpig,tiisg,sgitg);
+}
+
 #define N_F32_F32 4
 
-kernel void kernel_mul_mv_f32_f32(
+void kernel_mul_mv_f32_f32_impl(
         device const  char * src0,
         device const  char * src1,
         device       float * dst,
@@ -965,8 +1103,8 @@ kernel void kernel_mul_mv_f32_f32(
         constant  uint64_t & nb12,
         constant   int64_t & ne0,
         constant   int64_t & ne1,
-        constant   uint    & r2   [[buffer(17)]],
-        constant   uint    & r3   [[buffer(18)]],
+        constant   uint    & r2,
+        constant   uint    & r3,
         uint3 tgpig[[threadgroup_position_in_grid]],
         uint  tiisg[[thread_index_in_simdgroup]]) {
 
@@ -1025,6 +1163,32 @@ kernel void kernel_mul_mv_f32_f32(
     }
 }
 
+[[host_name("kernel_mul_mv_f32_f32")]]
+kernel void kernel_mul_mv_f32_f32(
+        device const  char * src0,
+        device const  char * 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   [[buffer(17)]],
+        constant   uint    & r3   [[buffer(18)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]]) {
+    kernel_mul_mv_f32_f32_impl(src0, src1, dst, ne00, ne01, ne02, nb00, nb01, nb02, ne10, ne11, ne12, nb10, nb11, nb12, ne0, ne1, r2, r3, tgpig, tiisg);
+}
+
 #define N_F16_F16 4
 
 kernel void kernel_mul_mv_f16_f16(
@@ -1105,7 +1269,7 @@ kernel void kernel_mul_mv_f16_f16(
     }
 }
 
-kernel void kernel_mul_mv_f16_f32_1row(
+void kernel_mul_mv_f16_f32_1row_impl(
         device const  char * src0,
         device const  char * src1,
         device       float * dst,
@@ -1123,8 +1287,8 @@ kernel void kernel_mul_mv_f16_f32_1row(
         constant  uint64_t & nb12,
         constant   int64_t & ne0,
         constant   int64_t & ne1,
-        constant   uint    & r2   [[buffer(17)]],
-        constant   uint    & r3   [[buffer(18)]],
+        constant   uint    & r2,
+        constant   uint    & r3,
         uint3 tgpig[[threadgroup_position_in_grid]],
         uint  tiisg[[thread_index_in_simdgroup]]) {
 
@@ -1161,12 +1325,37 @@ kernel void kernel_mul_mv_f16_f32_1row(
             dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
         }
     }
+}
 
+[[host_name("kernel_mul_mv_f16_f32_1row")]]
+kernel void kernel_mul_mv_f16_f32_1row(
+        device const  char * src0,
+        device const  char * 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   [[buffer(17)]],
+        constant   uint    & r3   [[buffer(18)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]]) {
+    kernel_mul_mv_f16_f32_1row_impl(src0, src1, dst, ne00, ne01, ne02, nb00, nb01, nb02, ne10, ne11, ne12, nb10, nb11, nb12, ne0, ne1, r2, r3, tgpig, tiisg);
 }
 
 #define N_F16_F32 4
 
-kernel void kernel_mul_mv_f16_f32(
+void kernel_mul_mv_f16_f32_impl(
         device const  char * src0,
         device const  char * src1,
         device       float * dst,
@@ -1184,8 +1373,8 @@ kernel void kernel_mul_mv_f16_f32(
         constant  uint64_t & nb12,
         constant   int64_t & ne0,
         constant   int64_t & ne1,
-        constant   uint    & r2   [[buffer(17)]],
-        constant   uint    & r3   [[buffer(18)]],
+        constant   uint    & r2,
+        constant   uint    & r3,
         uint3 tgpig[[threadgroup_position_in_grid]],
         uint tiisg[[thread_index_in_simdgroup]]) {
 
@@ -1244,6 +1433,32 @@ kernel void kernel_mul_mv_f16_f32(
     }
 }
 
+[[host_name("kernel_mul_mv_f16_f32")]]
+kernel void kernel_mul_mv_f16_f32(
+        device const  char * src0,
+        device const  char * 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   [[buffer(17)]],
+        constant   uint    & r3   [[buffer(18)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint tiisg[[thread_index_in_simdgroup]]) {
+    kernel_mul_mv_f16_f32_impl(src0, src1, dst, ne00, ne01, ne02, nb00, nb01, nb02, ne10, ne11, ne12, nb10, nb11, nb12, ne0, ne1, r2, r3, tgpig, tiisg);
+}
+
 // Assumes row size (ne00) is a multiple of 4
 kernel void kernel_mul_mv_f16_f32_l4(
         device const  char * src0,
@@ -1487,8 +1702,9 @@ kernel void kernel_rope(
             dst_data[1] = x0*sin_theta + x1*cos_theta;
         }
     } else {
-        for (int64_t ib = 0; ib < ne0/n_dims; ++ib) {
-            for (int64_t ic = 2*tiitg; ic < n_dims; ic += 2*tptg.x) {
+        for (int64_t ic = 2*tiitg; ic < ne0; ic += 2*tptg.x) {
+            if (ic < n_dims) {
+                const int64_t ib = 0;
 
                 // simplified from `(ib * n_dims + ic) * inv_ndims`
                 const float cur_rot = inv_ndims*ic - ib;
@@ -1507,6 +1723,14 @@ kernel void kernel_rope(
 
                 dst_data[0]        = x0*cos_theta - x1*sin_theta;
                 dst_data[n_dims/2] = x0*sin_theta + x1*cos_theta;
+            } else {
+                const int64_t i0 = ic;
+
+                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);
+
+                dst_data[0] = src[0];
+                dst_data[1] = src[1];
             }
         }
     }
@@ -1548,21 +1772,112 @@ kernel void kernel_im2col_f16(
     }
 }
 
-// bitonic sort implementation following the CUDA kernels as reference
-typedef void (argsort_t)(
-        device const float * x,
-        device     int32_t * dst,
-        constant   int64_t & ncols,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]]);
-
-template<ggml_sort_order order>
-kernel void kernel_argsort_f32_i32(
-        device const float   * x,
-        device       int32_t * dst,
-        constant     int64_t & ncols,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]]) {
+kernel void kernel_upscale_f32(
+    device  const char * src0,
+    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 & 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   int32_t & sf,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]]) {
+
+    const int64_t i3 = tgpig.z;
+    const int64_t i2 = tgpig.y;
+    const int64_t i1 = tgpig.x;
+
+    const int64_t i03 = i3;
+    const int64_t i02 = i2;
+    const int64_t i01 = i1/sf;
+
+    device const float * src0_ptr = (device const float *) (src0 + i03*nb03 + i02*nb02 + i01*nb01);
+    device       float * dst_ptr  = (device       float *) (dst  +  i3*nb3  +  i2*nb2  +  i1*nb1);
+
+    for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
+        dst_ptr[i0] = src0_ptr[i0/sf];
+    }
+}
+
+kernel void kernel_pad_f32(
+    device  const char * src0,
+    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 & 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 i3 = tgpig.z;
+    const int64_t i2 = tgpig.y;
+    const int64_t i1 = tgpig.x;
+
+    const int64_t i03 = i3;
+    const int64_t i02 = i2;
+    const int64_t i01 = i1;
+
+    device const float * src0_ptr = (device const float *) (src0 + i03*nb03 + i02*nb02 + i01*nb01);
+    device       float * dst_ptr  = (device       float *) (dst  +  i3*nb3  +  i2*nb2  +  i1*nb1);
+
+    if (i1 < ne01 && i2 < ne02 && i3 < ne03) {
+        for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
+            if (i0 < ne00) {
+                dst_ptr[i0] = src0_ptr[i0];
+            } else {
+                dst_ptr[i0] = 0.0f;
+            }
+        }
+
+        return;
+    }
+
+    for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
+        dst_ptr[i0] = 0.0f;
+    }
+}
+
+// bitonic sort implementation following the CUDA kernels as reference
+typedef void (argsort_t)(
+        device const float * x,
+        device     int32_t * dst,
+        constant   int64_t & ncols,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]]);
+
+template<ggml_sort_order order>
+kernel void kernel_argsort_f32_i32(
+        device const float   * x,
+        device       int32_t * dst,
+        constant     int64_t & ncols,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]]) {
     // bitonic sort
     int col = tpitg[0];
     int row = tgpig[1];
@@ -1600,9 +1915,17 @@ kernel void kernel_argsort_f32_i32(
 template [[host_name("kernel_argsort_f32_i32_asc")]]  kernel argsort_t kernel_argsort_f32_i32<GGML_SORT_ASC>;
 template [[host_name("kernel_argsort_f32_i32_desc")]] kernel argsort_t kernel_argsort_f32_i32<GGML_SORT_DESC>;
 
+kernel void kernel_leaky_relu_f32(
+        device const float * src0,
+        device       float * dst,
+        constant     float & slope,
+        uint tpig[[thread_position_in_grid]]) {
+    dst[tpig] = src0[tpig] > 0.0f ? src0[tpig] : src0[tpig] * slope;
+}
+
 kernel void kernel_cpy_f16_f16(
-        device const half * src0,
-        device       half * dst,
+        device  const half * src0,
+        device        half * dst,
         constant   int64_t & ne00,
         constant   int64_t & ne01,
         constant   int64_t & ne02,
@@ -1641,6 +1964,47 @@ kernel void kernel_cpy_f16_f16(
     }
 }
 
+kernel void kernel_cpy_f16_f32(
+        device  const half * 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,
+        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];
+
+    const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+
+    const int64_t i3 = n / (ne2*ne1*ne0);
+    const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
+    const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
+    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);
+
+    for (int64_t i00 = tpitg.x; i00 < ne00; i00 += ntg.x) {
+        device const half * src = (device half *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
+        dst_data[i00] = src[0];
+    }
+}
+
 kernel void kernel_cpy_f32_f16(
         device const float * src0,
         device        half * dst,
@@ -1917,9 +2281,9 @@ kernel void kernel_cpy_f32_q4_1(
 }
 
 kernel void kernel_concat(
-    device const char * src0,
-    device const char * src1,
-    device       char * dst,
+    device  const char * src0,
+    device  const char * src1,
+    device        char * dst,
     constant   int64_t & ne00,
     constant   int64_t & ne01,
     constant   int64_t & ne02,
@@ -1956,7 +2320,7 @@ kernel void kernel_concat(
     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 * 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;
 
@@ -2064,19 +2428,19 @@ static inline uchar4 get_scale_min_k4(int j, device const uint8_t * q) {
 
 //====================================== dot products =========================
 
-kernel void kernel_mul_mv_q2_K_f32(
+void kernel_mul_mv_q2_K_f32_impl(
         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    & r2  [[buffer(17)]],
-        constant   uint    & r3  [[buffer(18)]],
+        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,
         uint3 tgpig[[threadgroup_position_in_grid]],
         uint  tiisg[[thread_index_in_simdgroup]],
         uint  sgitg[[simdgroup_index_in_threadgroup]]) {
@@ -2214,8 +2578,8 @@ kernel void kernel_mul_mv_q2_K_f32(
     }
 }
 
-#if QK_K == 256
-kernel void kernel_mul_mv_q3_K_f32(
+[[host_name("kernel_mul_mv_q2_K_f32")]]
+kernel void kernel_mul_mv_q2_K_f32(
         device const  void * src0,
         device const float * src1,
         device       float * dst,
@@ -2229,8 +2593,29 @@ kernel void kernel_mul_mv_q3_K_f32(
         constant   uint    & r2  [[buffer(17)]],
         constant   uint    & r3  [[buffer(18)]],
         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]]) {
+
+    kernel_mul_mv_q2_K_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, tgpig, tiisg, sgitg);
+}
+
+#if QK_K == 256
+void kernel_mul_mv_q3_K_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,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
 
     const int nb = ne00/QK_K;
 
@@ -2373,19 +2758,19 @@ kernel void kernel_mul_mv_q3_K_f32(
     }
 }
 #else
-kernel void kernel_mul_mv_q3_K_f32(
+void kernel_mul_mv_q3_K_f32_impl(
         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    & r2  [[buffer(17)]],
-        constant   uint    & r3  [[buffer(18)]],
+        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,
         uint3 tgpig[[threadgroup_position_in_grid]],
         uint  tiisg[[thread_index_in_simdgroup]],
         uint  sgitg[[simdgroup_index_in_threadgroup]]) {
@@ -2450,20 +2835,41 @@ kernel void kernel_mul_mv_q3_K_f32(
 }
 #endif
 
+[[host_name("kernel_mul_mv_q3_K_f32")]]
+kernel void kernel_mul_mv_q3_K_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    & r2  [[buffer(17)]],
+        constant   uint    & r3  [[buffer(18)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    kernel_mul_mv_q3_K_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, tgpig, tiisg, sgitg);
+}
+
 #if QK_K == 256
-kernel void kernel_mul_mv_q4_K_f32(
+void kernel_mul_mv_q4_K_f32_impl(
         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    & r2   [[buffer(17)]],
-        constant   uint    & r3   [[buffer(18)]],
+        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,
         uint3 tgpig[[threadgroup_position_in_grid]],
         uint  tiisg[[thread_index_in_simdgroup]],
         uint  sgitg[[simdgroup_index_in_threadgroup]]) {
@@ -2564,19 +2970,19 @@ kernel void kernel_mul_mv_q4_K_f32(
     }
 }
 #else
-kernel void kernel_mul_mv_q4_K_f32(
+void kernel_mul_mv_q4_K_f32_impl(
         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    & r2  [[buffer(17)]],
-        constant   uint    & r3  [[buffer(18)]],
+        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,
         uint3 tgpig[[threadgroup_position_in_grid]],
         uint tiisg[[thread_index_in_simdgroup]],
         uint sgitg[[simdgroup_index_in_threadgroup]]) {
@@ -2660,7 +3066,8 @@ kernel void kernel_mul_mv_q4_K_f32(
 }
 #endif
 
-kernel void kernel_mul_mv_q5_K_f32(
+[[host_name("kernel_mul_mv_q4_K_f32")]]
+kernel void kernel_mul_mv_q4_K_f32(
         device const  void * src0,
         device const float * src1,
         device       float * dst,
@@ -2677,6 +3084,26 @@ kernel void kernel_mul_mv_q5_K_f32(
         uint tiisg[[thread_index_in_simdgroup]],
         uint sgitg[[simdgroup_index_in_threadgroup]]) {
 
+    kernel_mul_mv_q4_K_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, tgpig, tiisg, sgitg);
+}
+
+void kernel_mul_mv_q5_K_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,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+
     const int nb = ne00/QK_K;
 
     const int64_t r0 = tgpig.x;
@@ -2836,10 +3263,10 @@ kernel void kernel_mul_mv_q5_K_f32(
             dst[r1*ne0 + im*ne0*ne1 + first_row + row] = tot;
         }
     }
-
 }
 
-kernel void kernel_mul_mv_q6_K_f32(
+[[host_name("kernel_mul_mv_q5_K_f32")]]
+kernel void kernel_mul_mv_q5_K_f32(
         device const  void * src0,
         device const float * src1,
         device       float * dst,
@@ -2853,21 +3280,41 @@ kernel void kernel_mul_mv_q6_K_f32(
         constant   uint    & r2  [[buffer(17)]],
         constant   uint    & r3  [[buffer(18)]],
         uint3 tgpig[[threadgroup_position_in_grid]],
-        uint tiisg[[thread_index_in_simdgroup]],
-        uint sgitg[[simdgroup_index_in_threadgroup]]) {
-
-    const uint8_t kmask1 = 0x03;
-    const uint8_t kmask2 = 0x0C;
-    const uint8_t kmask3 = 0x30;
-    const uint8_t kmask4 = 0xC0;
-
-    const int nb = ne00/QK_K;
-
-    const int64_t r0 = tgpig.x;
-    const int64_t r1 = tgpig.y;
-    const int     im = tgpig.z;
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
 
-    const int row = 2 * r0 + sgitg;
+    kernel_mul_mv_q5_K_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, tgpig, tiisg, sgitg);
+}
+
+void kernel_mul_mv_q6_K_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,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    const uint8_t kmask1 = 0x03;
+    const uint8_t kmask2 = 0x0C;
+    const uint8_t kmask3 = 0x30;
+    const uint8_t kmask4 = 0xC0;
+
+    const int nb = ne00/QK_K;
+
+    const int64_t r0 = tgpig.x;
+    const int64_t r1 = tgpig.y;
+    const int     im = tgpig.z;
+
+    const int row = 2 * r0 + sgitg;
 
     const uint i12 = im%ne12;
     const uint i13 = im/ne12;
@@ -2945,6 +3392,27 @@ kernel void kernel_mul_mv_q6_K_f32(
     }
 }
 
+[[host_name("kernel_mul_mv_q6_K_f32")]]
+kernel void kernel_mul_mv_q6_K_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    & r2  [[buffer(17)]],
+        constant   uint    & r3  [[buffer(18)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    kernel_mul_mv_q6_K_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, tgpig, tiisg, sgitg);
+}
+
 //============================= templates and their specializations =============================
 
 // NOTE: this is not dequantizing - we are simply fitting the template
@@ -3062,10 +3530,10 @@ void dequantize_q8_0(device const block_q8_0 *xb, short il, thread type4x4 & reg
 
 template <typename type4x4>
 void dequantize_q2_K(device const block_q2_K *xb, short il, thread type4x4 & reg) {
-    const half d = xb->d;
-    const half min = xb->dmin;
+    const float d = xb->d;
+    const float min = xb->dmin;
     device const uint8_t * q = (device const uint8_t *)xb->qs;
-    half dl, ml;
+    float dl, ml;
     uint8_t sc = xb->scales[il];
 
 #if QK_K == 256
@@ -3135,10 +3603,10 @@ void dequantize_q4_K(device const block_q4_K *xb, short il, thread type4x4 & reg
     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 half d   = il < 2 ? xb->d : xb->d / 16.h;
-    const half min = xb->dmin;
-    const half dl = d * sc[0];
-    const half ml = min * sc[1];
+    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;
@@ -3165,13 +3633,13 @@ void dequantize_q5_K(device const block_q5_K *xb, short il, thread type4x4 & reg
     uint8_t ul = 1 << (il/2);
     il = il & 3;
     const uchar2 sc = get_scale_min_k4_just2(is, il/2, xb->scales);
-    const half d = il < 2 ? xb->d : xb->d / 16.h;
-    const half min = xb->dmin;
-    const half dl = d * sc[0];
-    const half ml = min * sc[1];
+    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];
 
-    const ushort mask = il<2 ? 0x0F : 0xF0;
-    const half qh_val = il<2 ? 16.h : 256.h;
+    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;
     }
@@ -3219,22 +3687,90 @@ void dequantize_q6_K(device const block_q6_K *xb, short il, thread type4x4 & reg
 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   int * src1,
+        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,
-        uint                 tgpig[[threadgroup_position_in_grid]],
+        constant  uint64_t & nb2,
+        uint3                tgpig[[threadgroup_position_in_grid]],
         uint                 tiitg[[thread_index_in_threadgroup]],
-        uint                 tptg[[threads_per_threadgroup]]) {
-    const int i = tgpig;
-    const int r = ((device int32_t *) src1)[i];
+        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];
 
-    for (int ind = tiitg; ind < ne00/16; ind += tptg) {
+    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)) + ind/nl, ind%nl, temp);
-        *(((device float4x4 *) ((device char *) dst + i*nb1)) + ind) = temp;
+            ((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];
     }
 }
 
@@ -3426,19 +3962,22 @@ kernel void kernel_mul_mm(device const  uchar * src0,
 
 template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread half4x4 &)>
 kernel void kernel_mul_mm_id(
-        device const int32_t * ids,
+        device const   uchar * ids,
         device const   uchar * src1,
-        device         float * dst,
+        device         uchar * dst,
+        constant     int64_t & nbi1,
         constant     int64_t & ne00,
         constant     int64_t & ne02,
         constant     int64_t & nb01,
         constant     int64_t & nb02,
         constant     int64_t & ne12,
+        constant     int64_t & ne13,
         constant     int64_t & nb10,
         constant     int64_t & nb11,
         constant     int64_t & nb12,
         constant     int64_t & ne0,
         constant     int64_t & ne1,
+        constant     int64_t & nb1,
         constant        uint & r2,
         constant        uint & r3,
         constant         int & idx,
@@ -3456,10 +3995,16 @@ kernel void kernel_mul_mm_id(
         uint                   sgitg[[simdgroup_index_in_threadgroup]]) {
     device const uchar * src0[8] = {src00, src01, src02, src03, src04, src05, src06, src07};
 
+    const int64_t bid = tgpig.z/(ne12*ne13);
+
+    tgpig.z = tgpig.z%(ne12*ne13);
+
+    const int32_t id = ((device int32_t *) (ids + bid*nbi1))[idx];
+
     kernel_mul_mm_impl<block_q, nl, dequantize_func>(
-        src0[ids[idx]],
-        src1,
-        dst,
+        src0[id],
+        src1 + bid*nb11,
+        (device float *) (dst + bid*nb1),
         ne00,
         ne02,
         nb01,
@@ -3484,17 +4029,26 @@ kernel void kernel_mul_mm_id(
 #define QK_NL 4
 #endif
 
+//
+// get rows
+//
+
 typedef void (get_rows_t)(
         device const void * src0,
-        device const  int * src1,
+        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,
-        uint, uint, uint);
+        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_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>;
@@ -3506,6 +4060,10 @@ template [[host_name("kernel_get_rows_q4_K")]] kernel get_rows_t kernel_get_rows
 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>;
 
+//
+// matrix-matrix multiplication
+//
+
 typedef void (mat_mm_t)(
         device const  uchar * src0,
         device const  uchar * src1,
@@ -3538,20 +4096,27 @@ template [[host_name("kernel_mul_mm_q4_K_f32")]] kernel mat_mm_t kernel_mul_mm<b
 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>;
 
+//
+// indirect matrix-matrix multiplication
+//
+
 typedef void (mat_mm_id_t)(
-        device const int32_t * ids,
+        device const   uchar * ids,
         device const   uchar * src1,
-        device         float * dst,
+        device         uchar * dst,
+        constant     int64_t & nbi1,
         constant     int64_t & ne00,
         constant     int64_t & ne02,
         constant     int64_t & nb01,
         constant     int64_t & nb02,
         constant     int64_t & ne12,
+        constant     int64_t & ne13,
         constant     int64_t & nb10,
         constant     int64_t & nb11,
         constant     int64_t & nb12,
         constant     int64_t & ne0,
         constant     int64_t & ne1,
+        constant     int64_t & nb1,
         constant        uint & r2,
         constant        uint & r3,
         constant         int & idx,
@@ -3578,3 +4143,775 @@ template [[host_name("kernel_mul_mm_id_q3_K_f32")]] kernel mat_mm_id_t kernel_mu
 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>;
+
+//
+// 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         uchar * dst,
+        constant     int64_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     int64_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};
+
+    const int64_t bid = tgpig.z/(ne12*ne13);
+
+    tgpig.z = tgpig.z%(ne12*ne13);
+
+    const int32_t id = ((device int32_t *) (ids + bid*nbi1))[idx];
+
+    kernel_mul_mv_f32_f32_impl(
+        src0[id],
+        src1 + bid*nb11,
+        (device float *) (dst + bid*nb1),
+        ne00,
+        ne01,
+        ne02,
+        nb00,
+        nb01,
+        nb02,
+        ne10,
+        ne11,
+        ne12,
+        nb10,
+        nb11,
+        nb12,
+        ne0,
+        ne1,
+        r2,
+        r3,
+        tgpig,
+        tiisg);
+}
+
+[[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         uchar * dst,
+        constant     int64_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     int64_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};
+
+    const int64_t bid = tgpig.z/(ne12*ne13);
+
+    tgpig.z = tgpig.z%(ne12*ne13);
+
+    const int32_t id = ((device int32_t *) (ids + bid*nbi1))[idx];
+
+    kernel_mul_mv_f16_f32_impl(
+        src0[id],
+        src1 + bid*nb11,
+        (device float *) (dst + bid*nb1),
+        ne00,
+        ne01,
+        ne02,
+        nb00,
+        nb01,
+        nb02,
+        ne10,
+        ne11,
+        ne12,
+        nb10,
+        nb11,
+        nb12,
+        ne0,
+        ne1,
+        r2,
+        r3,
+        tgpig,
+        tiisg);
+}
+
+[[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         uchar * dst,
+        constant     int64_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     int64_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};
+
+    const int64_t bid = tgpig.z/(ne12*ne13);
+
+    tgpig.z = tgpig.z%(ne12*ne13);
+
+    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),
+        (device       float *) ( dst + bid*nb1),
+        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         uchar * dst,
+        constant     int64_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     int64_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};
+
+    const int64_t bid = tgpig.z/(ne12*ne13);
+
+    tgpig.z = tgpig.z%(ne12*ne13);
+
+    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>(
+        src0[id],
+        (device const float *) (src1 + bid*nb11),
+        (device       float *) ( dst + bid*nb1),
+        ne00,
+        ne01,
+        ne02,
+        ne10,
+        ne12,
+        ne0,
+        ne1,
+        r2,
+        r3,
+        tgpig,
+        tiisg,
+        sgitg);
+}
+
+[[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         uchar * dst,
+        constant     int64_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     int64_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};
+
+    const int64_t bid = tgpig.z/(ne12*ne13);
+
+    tgpig.z = tgpig.z%(ne12*ne13);
+
+    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),
+        (device       float *) ( dst + bid*nb1),
+        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         uchar * dst,
+        constant     int64_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     int64_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};
+
+    const int64_t bid = tgpig.z/(ne12*ne13);
+
+    tgpig.z = tgpig.z%(ne12*ne13);
+
+    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>(
+        src0[id],
+        (device const float *) (src1 + bid*nb11),
+        (device       float *) ( dst + bid*nb1),
+        ne00,
+        ne01,
+        ne02,
+        ne10,
+        ne12,
+        ne0,
+        ne1,
+        r2,
+        r3,
+        tgpig,
+        tiisg,
+        sgitg);
+}
+
+[[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         uchar * dst,
+        constant     int64_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     int64_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};
+
+    const int64_t bid = tgpig.z/(ne12*ne13);
+
+    tgpig.z = tgpig.z%(ne12*ne13);
+
+    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>(
+        src0[id],
+        (device const float *) (src1 + bid*nb11),
+        (device       float *) ( dst + bid*nb1),
+        ne00,
+        ne01,
+        ne02,
+        ne10,
+        ne12,
+        ne0,
+        ne1,
+        r2,
+        r3,
+        tgpig,
+        tiisg,
+        sgitg);
+}
+
+[[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         uchar * dst,
+        constant     int64_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     int64_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};
+
+    const int64_t bid = tgpig.z/(ne12*ne13);
+
+    tgpig.z = tgpig.z%(ne12*ne13);
+
+    const int32_t id = ((device int32_t *) (ids + bid*nbi1))[idx];
+
+    kernel_mul_mv_q2_K_f32_impl(
+        src0[id],
+        (device const float *) (src1 + bid*nb11),
+        (device       float *) ( dst + bid*nb1),
+        ne00,
+        ne01,
+        ne02,
+        ne10,
+        ne12,
+        ne0,
+        ne1,
+        r2,
+        r3,
+        tgpig,
+        tiisg,
+        sgitg);
+}
+
+[[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         uchar * dst,
+        constant     int64_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     int64_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};
+
+    const int64_t bid = tgpig.z/(ne12*ne13);
+
+    tgpig.z = tgpig.z%(ne12*ne13);
+
+    const int32_t id = ((device int32_t *) (ids + bid*nbi1))[idx];
+
+    kernel_mul_mv_q3_K_f32_impl(
+        src0[id],
+        (device const float *) (src1 + bid*nb11),
+        (device       float *) ( dst + bid*nb1),
+        ne00,
+        ne01,
+        ne02,
+        ne10,
+        ne12,
+        ne0,
+        ne1,
+        r2,
+        r3,
+        tgpig,
+        tiisg,
+        sgitg);
+}
+
+[[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         uchar * dst,
+        constant     int64_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     int64_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};
+
+    const int64_t bid = tgpig.z/(ne12*ne13);
+
+    tgpig.z = tgpig.z%(ne12*ne13);
+
+    const int32_t id = ((device int32_t *) (ids + bid*nbi1))[idx];
+
+    kernel_mul_mv_q4_K_f32_impl(
+        src0[id],
+        (device const float *) (src1 + bid*nb11),
+        (device       float *) ( dst + bid*nb1),
+        ne00,
+        ne01,
+        ne02,
+        ne10,
+        ne12,
+        ne0,
+        ne1,
+        r2,
+        r3,
+        tgpig,
+        tiisg,
+        sgitg);
+}
+
+[[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         uchar * dst,
+        constant     int64_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     int64_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};
+
+    const int64_t bid = tgpig.z/(ne12*ne13);
+
+    tgpig.z = tgpig.z%(ne12*ne13);
+
+    const int32_t id = ((device int32_t *) (ids + bid*nbi1))[idx];
+
+    kernel_mul_mv_q5_K_f32_impl(
+        src0[id],
+        (device const float *) (src1 + bid*nb11),
+        (device       float *) ( dst + bid*nb1),
+        ne00,
+        ne01,
+        ne02,
+        ne10,
+        ne12,
+        ne0,
+        ne1,
+        r2,
+        r3,
+        tgpig,
+        tiisg,
+        sgitg);
+}
+
+[[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         uchar * dst,
+        constant     int64_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     int64_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};
+
+    const int64_t bid = tgpig.z/(ne12*ne13);
+
+    tgpig.z = tgpig.z%(ne12*ne13);
+
+    const int32_t id = ((device int32_t *) (ids + bid*nbi1))[idx];
+
+    kernel_mul_mv_q6_K_f32_impl(
+        src0[id],
+        (device const float *) (src1 + bid*nb11),
+        (device       float *) ( dst + bid*nb1),
+        ne00,
+        ne01,
+        ne02,
+        ne10,
+        ne12,
+        ne0,
+        ne1,
+        r2,
+        r3,
+        tgpig,
+        tiisg,
+        sgitg);
+}