whispercpp 1.3.4 → 1.3.6

data/ext/sources/ggml/src/ggml-metal/ggml-metal-ops.cpp

@@ -10,6 +10,8 @@
 
 #include <cassert>
 #include <algorithm>
+#include <limits>
+#include <cmath>
 
 static ggml_metal_buffer_id ggml_metal_get_buffer_id(const ggml_tensor * t) {
     if (!t) {
@@ -201,6 +203,10 @@ static int ggml_metal_op_encode_impl(ggml_metal_op_t ctx, int idx) {
         GGML_ABORT("unsupported op");
     }
 
+    if ((node->flags & GGML_TENSOR_FLAG_COMPUTE) == 0) {
+        return 1;
+    }
+
     int n_fuse = 1;
 
     // check if the current node can run concurrently with other nodes before it
@@ -219,13 +225,17 @@ static int ggml_metal_op_encode_impl(ggml_metal_op_t ctx, int idx) {
         }
 
         if (ctx->debug_graph > 0) {
-            GGML_LOG_DEBUG("%s: node[%5d] - %-12s %s\n", __func__, idx, ggml_op_name(node->op), is_concurrent ? "(concurrent)" : "");
+            GGML_LOG_DEBUG("%s: node[%5d] - %-12s %-12s %s\n", __func__, idx, ggml_op_name(node->op), ggml_get_name(node), is_concurrent ? "(concurrent)" : "");
         }
         if (ctx->debug_graph > 1) {
             GGML_TENSOR_LOCALS( int64_t, ne0, node->src[0], ne);
             GGML_TENSOR_LOCALS(uint64_t, nb0, node->src[0], nb);
             GGML_TENSOR_LOCALS( int64_t, ne1, node->src[1], ne);
             GGML_TENSOR_LOCALS(uint64_t, nb1, node->src[1], nb);
+            GGML_TENSOR_LOCALS( int64_t, ne2, node->src[2], ne);
+            GGML_TENSOR_LOCALS(uint64_t, nb2, node->src[2], nb);
+            GGML_TENSOR_LOCALS( int64_t, ne3, node->src[3], ne);
+            GGML_TENSOR_LOCALS(uint64_t, nb3, node->src[3], nb);
             GGML_TENSOR_LOCALS( int64_t, ne,  node,         ne);
             GGML_TENSOR_LOCALS(uint64_t, nb,  node,         nb);
 
@@ -237,6 +247,14 @@ static int ggml_metal_op_encode_impl(ggml_metal_op_t ctx, int idx) {
                 GGML_LOG_DEBUG("%s: src1 - %4s [%5lld, %5lld, %5lld, %5lld] [%5lld, %5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(node->src[1]->type), ne10, ne11, ne12, ne13, nb10, nb11, nb12, nb13,
                         ggml_is_contiguous(node->src[1]), node->src[1]->name);
             }
+            if (node->src[2]) {
+                GGML_LOG_DEBUG("%s: src2 - %4s [%5lld, %5lld, %5lld, %5lld] [%5lld, %5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(node->src[2]->type), ne20, ne21, ne22, ne23, nb20, nb21, nb22, nb23,
+                        ggml_is_contiguous(node->src[2]), node->src[2]->name);
+            }
+            if (node->src[3]) {
+                GGML_LOG_DEBUG("%s: src3 - %4s [%5lld, %5lld, %5lld, %5lld] [%5lld, %5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(node->src[3]->type), ne30, ne31, ne32, ne33, nb30, nb31, nb32, nb33,
+                        ggml_is_contiguous(node->src[3]), node->src[3]->name);
+            }
             if (node) {
                 GGML_LOG_DEBUG("%s: node  - %4s [%5lld, %5lld, %5lld, %5lld] [%5lld, %5lld, %5lld, %5lld], 1, %s\n", __func__, ggml_type_name(node->type), ne0, ne1, ne2, ne3, nb0, nb1, nb2, nb3,
                         node->name);
@@ -269,13 +287,9 @@ static int ggml_metal_op_encode_impl(ggml_metal_op_t ctx, int idx) {
                 n_fuse = ggml_metal_op_acc(ctx, idx);
             } break;
         case GGML_OP_SCALE:
-            {
-                n_fuse = ggml_metal_op_scale(ctx, idx);
-            } break;
+        case GGML_OP_FILL:
         case GGML_OP_CLAMP:
-            {
-                n_fuse = ggml_metal_op_clamp(ctx, idx);
-            } break;
+        case GGML_OP_LEAKY_RELU:
         case GGML_OP_SQR:
         case GGML_OP_SQRT:
         case GGML_OP_SIN:
@@ -289,11 +303,19 @@ static int ggml_metal_op_encode_impl(ggml_metal_op_t ctx, int idx) {
             {
                 n_fuse = ggml_metal_op_glu(ctx, idx);
             } break;
+        case GGML_OP_SUM:
+            {
+                n_fuse = ggml_metal_op_sum(ctx, idx);
+            } break;
         case GGML_OP_SUM_ROWS:
         case GGML_OP_MEAN:
             {
                 n_fuse = ggml_metal_op_sum_rows(ctx, idx);
             } break;
+        case GGML_OP_CUMSUM:
+            {
+                n_fuse = ggml_metal_op_cumsum(ctx, idx);
+            } break;
         case GGML_OP_SOFT_MAX:
             {
                 n_fuse = ggml_metal_op_soft_max(ctx, idx);
@@ -311,6 +333,14 @@ static int ggml_metal_op_encode_impl(ggml_metal_op_t ctx, int idx) {
             {
                 n_fuse = ggml_metal_op_rwkv(ctx, idx);
             } break;
+        case GGML_OP_GATED_DELTA_NET:
+            {
+                n_fuse = ggml_metal_op_gated_delta_net(ctx, idx);
+            } break;
+        case GGML_OP_SOLVE_TRI:
+            {
+                n_fuse = ggml_metal_op_solve_tri(ctx, idx);
+            } break;
         case GGML_OP_MUL_MAT:
             {
                 n_fuse = ggml_metal_op_mul_mat(ctx, idx);
@@ -327,6 +357,10 @@ static int ggml_metal_op_encode_impl(ggml_metal_op_t ctx, int idx) {
             {
                 n_fuse = ggml_metal_op_set_rows(ctx, idx);
             } break;
+        case GGML_OP_DIAG:
+            {
+                n_fuse = ggml_metal_op_diag(ctx, idx);
+            } break;
         case GGML_OP_L2_NORM:
             {
                 n_fuse = ggml_metal_op_l2_norm(ctx, idx);
@@ -348,10 +382,18 @@ static int ggml_metal_op_encode_impl(ggml_metal_op_t ctx, int idx) {
             {
                 n_fuse = ggml_metal_op_im2col(ctx, idx);
             } break;
+        case GGML_OP_CONV_2D:
+            {
+                n_fuse = ggml_metal_op_conv_2d(ctx, idx);
+            } break;
         case GGML_OP_CONV_TRANSPOSE_1D:
             {
                 n_fuse = ggml_metal_op_conv_transpose_1d(ctx, idx);
             } break;
+        case GGML_OP_CONV_TRANSPOSE_2D:
+            {
+                n_fuse = ggml_metal_op_conv_transpose_2d(ctx, idx);
+            } break;
         case GGML_OP_UPSCALE:
             {
                 n_fuse = ggml_metal_op_upscale(ctx, idx);
@@ -376,20 +418,32 @@ static int ggml_metal_op_encode_impl(ggml_metal_op_t ctx, int idx) {
             {
                 n_fuse = ggml_metal_op_argsort(ctx, idx);
             } break;
-        case GGML_OP_LEAKY_RELU:
+        case GGML_OP_TOP_K:
+            {
+                n_fuse = ggml_metal_op_top_k(ctx, idx);
+            } break;
+        case GGML_OP_TRI:
             {
-                n_fuse = ggml_metal_op_leaky_relu(ctx, idx);
+                n_fuse = ggml_metal_op_tri(ctx, idx);
             } break;
         case GGML_OP_FLASH_ATTN_EXT:
             {
                 n_fuse = ggml_metal_op_flash_attn_ext(ctx, idx);
             } break;
+        case GGML_OP_SET:
+            {
+                n_fuse = ggml_metal_op_set(ctx, idx);
+            } break;
         case GGML_OP_DUP:
         case GGML_OP_CPY:
         case GGML_OP_CONT:
             {
                 n_fuse = ggml_metal_op_cpy(ctx, idx);
             } break;
+        case GGML_OP_POOL_1D:
+            {
+                n_fuse = ggml_metal_op_pool_1d(ctx, idx);
+            } break;
         case GGML_OP_POOL_2D:
             {
                 n_fuse = ggml_metal_op_pool_2d(ctx, idx);
@@ -398,7 +452,19 @@ static int ggml_metal_op_encode_impl(ggml_metal_op_t ctx, int idx) {
             {
                 n_fuse = ggml_metal_op_argmax(ctx, idx);
             } break;
-       default:
+        case GGML_OP_OPT_STEP_ADAMW:
+            {
+                n_fuse = ggml_metal_op_opt_step_adamw(ctx, idx);
+            } break;
+        case GGML_OP_OPT_STEP_SGD:
+            {
+                n_fuse = ggml_metal_op_opt_step_sgd(ctx, idx);
+            } break;
+        case GGML_OP_COUNT_EQUAL:
+            {
+                n_fuse = ggml_metal_op_count_equal(ctx, idx);
+            } break;
+        default:
             {
                 GGML_LOG_ERROR("%s: error: node %3d, op = %8s not implemented\n", __func__, idx, ggml_op_name(node->op));
                 GGML_ABORT("fatal error");
@@ -482,7 +548,7 @@ int ggml_metal_op_concat(ggml_metal_op_t ctx, int idx) {
         /*.dim  =*/ dim,
     };
 
-    ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_base(lib, GGML_OP_CONCAT);
+    auto pipeline = ggml_metal_library_get_pipeline_base(lib, GGML_OP_CONCAT);
 
     ggml_metal_encoder_set_pipeline(enc, pipeline);
     ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
@@ -506,9 +572,9 @@ int ggml_metal_op_repeat(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
-    ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_repeat(lib, op->type);
+    auto pipeline = ggml_metal_library_get_pipeline_repeat(lib, op->type);
 
     ggml_metal_kargs_repeat args = {
         /*.ne00 =*/ ne00,
@@ -552,14 +618,14 @@ int ggml_metal_op_acc(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     GGML_ASSERT(op->src[0]->type == GGML_TYPE_F32);
     GGML_ASSERT(op->src[1]->type == GGML_TYPE_F32);
     GGML_ASSERT(op->type         == GGML_TYPE_F32);
 
-    GGML_ASSERT(ggml_is_contiguous(op->src[0]));
-    GGML_ASSERT(ggml_is_contiguous(op->src[1]));
+    GGML_ASSERT(ggml_is_contiguous_rows(op->src[0]));
+    GGML_ASSERT(ggml_is_contiguous_rows(op->src[1]));
 
     const size_t pnb1 = ((const int32_t *) op->op_params)[0];
     const size_t pnb2 = ((const int32_t *) op->op_params)[1];
@@ -569,14 +635,15 @@ int ggml_metal_op_acc(ggml_metal_op_t ctx, int idx) {
     const bool inplace = (bool) ((const int32_t *) op->op_params)[4];
 
     if (!inplace) {
-        // run a separete kernel to cpy src->dst
+        // run a separate kernel to cpy src->dst
         // not sure how to avoid this
         // TODO: make a simpler cpy_bytes kernel
 
         //const id<MTLComputePipelineState> pipeline = ctx->pipelines[GGML_METAL_PIPELINE_TYPE_CPY_F32_F32].obj;
-        ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_cpy(lib, op->src[0]->type, op->type);
+        auto pipeline = ggml_metal_library_get_pipeline_cpy(lib, op->src[0]->type, op->type);
 
         ggml_metal_kargs_cpy args = {
+            /*.nk0  =*/ ne00,
             /*.ne00 =*/ ne00,
             /*.ne01 =*/ ne01,
             /*.ne02 =*/ ne02,
@@ -608,10 +675,10 @@ int ggml_metal_op_acc(ggml_metal_op_t ctx, int idx) {
     }
 
     ggml_metal_kargs_bin args = {
-        /*.ne00 =*/ ne00,
-        /*.ne01 =*/ ne01,
-        /*.ne02 =*/ ne02,
-        /*.ne03 =*/ ne03,
+        /*.ne00 =*/ ne10,
+        /*.ne01 =*/ ne11,
+        /*.ne02 =*/ ne12,
+        /*.ne03 =*/ ne13,
         /*.nb00 =*/ nb00,
         /*.nb01 =*/ pnb1,
         /*.nb02 =*/ pnb2,
@@ -624,10 +691,10 @@ int ggml_metal_op_acc(ggml_metal_op_t ctx, int idx) {
         /*.nb11 =*/ nb11,
         /*.nb12 =*/ nb12,
         /*.nb13 =*/ nb13,
-        /*.ne0  =*/ ne0,
-        /*.ne1  =*/ ne1,
-        /*.ne2  =*/ ne2,
-        /*.ne3  =*/ ne3,
+        /*.ne0  =*/ ne10,
+        /*.ne1  =*/ ne11,
+        /*.ne2  =*/ ne12,
+        /*.ne3  =*/ ne13,
         /*.nb0  =*/ nb0,
         /*.nb1  =*/ pnb1,
         /*.nb2  =*/ pnb2,
@@ -636,7 +703,7 @@ int ggml_metal_op_acc(ggml_metal_op_t ctx, int idx) {
         /*.o1   =*/ { 0 },
     };
 
-    ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_bin(lib, GGML_OP_ADD, 1, false);
+    auto pipeline = ggml_metal_library_get_pipeline_bin_one(lib, GGML_OP_ADD);
 
     ggml_metal_encoder_set_pipeline(enc, pipeline);
     ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
@@ -644,14 +711,20 @@ int ggml_metal_op_acc(ggml_metal_op_t ctx, int idx) {
     ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[1]), 2);
     ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op),         3);
 
-    const int nth = std::min(ggml_metal_pipeline_max_theads_per_threadgroup(pipeline), ne00);
+    const int nth_max = MIN(256, ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));
+
+    int nth = 1;
+
+    while (2*nth < args.ne0 && nth < nth_max) {
+        nth *= 2;
+    }
 
     ggml_metal_encoder_dispatch_threadgroups(enc, ne11, ne12, ne13, nth, 1, 1);
 
     return 1;
 }
 
-int ggml_metal_op_scale(ggml_metal_op_t ctx, int idx) {
+int ggml_metal_op_unary(ggml_metal_op_t ctx, int idx) {
     ggml_tensor * op = ctx->node(idx);
 
     ggml_metal_library_t lib = ctx->lib;
@@ -660,100 +733,82 @@ int ggml_metal_op_scale(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
-    float scale;
-    float bias;
-    memcpy(&scale, ((const int32_t *) op->op_params) + 0, sizeof(float));
-    memcpy(&bias,  ((const int32_t *) op->op_params) + 1, sizeof(float));
+    GGML_ASSERT(ggml_is_contiguous_rows(op->src[0]));
 
-    ggml_metal_kargs_scale args = {
-        /*.scale =*/ scale,
-        /*.bias  =*/ bias,
-    };
+    ggml_metal_buffer_id bid_src0 = ggml_metal_get_buffer_id(op->src[0]);
+    ggml_metal_buffer_id bid_dst  = ggml_metal_get_buffer_id(op);
 
-    int64_t n = ggml_nelements(op);
+    ggml_metal_kargs_unary args = {
+        /*.ne00  =*/ ne00,
+        /*.ne01  =*/ ne01,
+        /*.ne02  =*/ ne02,
+        /*.ne03  =*/ ne03,
+        /*.nb00  =*/ nb00,
+        /*.nb01  =*/ nb01,
+        /*.nb02  =*/ nb02,
+        /*.nb03  =*/ nb03,
+        /*.ne0   =*/ ne0,
+        /*.ne1   =*/ ne1,
+        /*.ne2   =*/ ne2,
+        /*.ne3   =*/ ne3,
+        /*.nb0   =*/ nb0,
+        /*.nb1   =*/ nb1,
+        /*.nb2   =*/ nb2,
+        /*.nb3   =*/ nb3,
+        /*.slope =*/ 0.0,
+        /*.scale =*/ 0.0,
+        /*.bias  =*/ 0.0,
+        /*.val   =*/ 0.0,
+        /*.min   =*/ 0.0,
+        /*.max   =*/ 0.0,
+    };
 
-    if (n % 4 == 0) {
-        n /= 4;
+    if (op->op == GGML_OP_LEAKY_RELU) {
+        args.slope = ggml_get_op_params_f32(op, 0);
     }
 
-    ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_unary(lib, op);
-
-    ggml_metal_encoder_set_pipeline(enc, pipeline);
-    ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
-    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[0]), 1);
-    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op),         2);
-
-    ggml_metal_encoder_dispatch_threadgroups(enc, n, 1, 1, 1, 1, 1);
-
-    return 1;
-}
-
-int ggml_metal_op_clamp(ggml_metal_op_t ctx, int idx) {
-    ggml_tensor * op = ctx->node(idx);
-
-    ggml_metal_library_t lib = ctx->lib;
-    ggml_metal_encoder_t enc = ctx->enc;
-
-    GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
-    GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
-    GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    if (op->op == GGML_OP_SCALE) {
+        args.scale = ggml_get_op_params_f32(op, 0);
+        args.bias  = ggml_get_op_params_f32(op, 1);
+    }
 
-    float min;
-    float max;
-    memcpy(&min, ((const int32_t *) op->op_params) + 0, sizeof(float));
-    memcpy(&max, ((const int32_t *) op->op_params) + 1, sizeof(float));
+    if (op->op == GGML_OP_FILL) {
+        args.val = ggml_get_op_params_f32(op, 0);
+    }
 
-    ggml_metal_kargs_clamp args = {
-        /*.min =*/ min,
-        /*.max =*/ max,
-    };
+    if (op->op == GGML_OP_CLAMP) {
+        args.min = ggml_get_op_params_f32(op, 0);
+        args.max = ggml_get_op_params_f32(op, 1);
+    }
 
-    int64_t n = ggml_nelements(op);
+    auto pipeline = ggml_metal_library_get_pipeline_unary(lib, op);
 
-    if (n % 4 == 0) {
-        n /= 4;
+    if (pipeline.c4) {
+        args.ne00 = ne00/4;
+        args.ne0  = ne0/4;
     }
 
-    ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_unary(lib, op);
-
     ggml_metal_encoder_set_pipeline(enc, pipeline);
     ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
-    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[0]), 1);
-    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op),         2);
-
-    ggml_metal_encoder_dispatch_threadgroups(enc, n, 1, 1, 1, 1, 1);
-
-    return 1;
-}
+    ggml_metal_encoder_set_buffer  (enc, bid_src0, 1);
+    ggml_metal_encoder_set_buffer  (enc, bid_dst,  2);
 
-int ggml_metal_op_unary(ggml_metal_op_t ctx, int idx) {
-    ggml_tensor * op = ctx->node(idx);
+    if (pipeline.cnt) {
+        const int n = pipeline.c4 ? ggml_nelements(op)/4 : ggml_nelements(op);
 
-    ggml_metal_library_t lib = ctx->lib;
-    ggml_metal_encoder_t enc = ctx->enc;
+        ggml_metal_encoder_dispatch_threadgroups(enc, n, 1, 1, 1, 1, 1);
+    } else {
+        const int nth_max = MIN(256, ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));
 
-    GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
-    GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
-    GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+        const int nth = MIN(args.ne00, nth_max);
 
-    int64_t n = ggml_nelements(op);
+        const int nk0 = (args.ne00 + nth - 1)/nth;
 
-    if (n % 4 == 0) {
-        n /= 4;
+        ggml_metal_encoder_dispatch_threadgroups(enc, nk0*ne01, ne02, ne03, nth, 1, 1);
     }
 
-    ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_unary(lib, op);
-
-    ggml_metal_encoder_set_pipeline(enc, pipeline);
-    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[0]), 0);
-    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op),         1);
-
-    ggml_metal_encoder_dispatch_threadgroups(enc, n, 1, 1, 1, 1, 1);
-
     return 1;
 }
 
@@ -768,13 +823,13 @@ int ggml_metal_op_glu(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     if (op->src[1]) {
         GGML_ASSERT(ggml_are_same_shape(op->src[0], op->src[1]));
     }
 
-    ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_glu(lib, op);
+    auto pipeline = ggml_metal_library_get_pipeline_glu(lib, op);
 
     const int32_t swp = ggml_get_op_params_i32(op, 1);
     const float alpha = ggml_get_op_params_f32(op, 2);
@@ -800,18 +855,6 @@ int ggml_metal_op_glu(ggml_metal_op_t ctx, int idx) {
 
     const int32_t nth = std::min(ggml_metal_pipeline_max_theads_per_threadgroup(pipeline), ne00/2);
 
-    //[encoder setComputePipelineState:pipeline];
-    //[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-    //if (src1) {
-    //    [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
-    //} else {
-    //    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1];
-    //}
-    //[encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
-    //[encoder setBytes:&args length:sizeof(args) atIndex:3];
-
-    //[encoder dispatchThreadgroups:MTLSizeMake(nrows, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
-
     ggml_metal_encoder_set_pipeline(enc, pipeline);
     ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
     ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[0]), 1);
@@ -827,6 +870,43 @@ int ggml_metal_op_glu(ggml_metal_op_t ctx, int idx) {
     return 1;
 }
 
+int ggml_metal_op_sum(ggml_metal_op_t ctx, int idx) {
+    ggml_tensor * op  = ctx->node(idx);
+
+    ggml_metal_library_t lib = ctx->lib;
+    ggml_metal_encoder_t enc = ctx->enc;
+
+    const uint64_t n = (uint64_t) ggml_nelements(op->src[0]);
+
+    ggml_metal_kargs_sum args = {
+        /*.np =*/ n,
+    };
+
+    auto pipeline = ggml_metal_library_get_pipeline_sum(lib, op);
+
+    int nth = 32; // SIMD width
+
+    while (nth < (int) n && nth < ggml_metal_pipeline_max_theads_per_threadgroup(pipeline)) {
+        nth *= 2;
+    }
+
+    nth = std::min(nth, ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));
+    nth = std::min(nth, (int) n);
+
+    const int nsg = (nth + 31) / 32;
+
+    ggml_metal_encoder_set_pipeline(enc, pipeline);
+    ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[0]), 1);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op),         2);
+
+    ggml_metal_encoder_set_threadgroup_memory_size(enc, nsg * sizeof(float), 0);
+
+    ggml_metal_encoder_dispatch_threadgroups(enc, 1, 1, 1, nth, 1, 1);
+
+    return 1;
+}
+
 int ggml_metal_op_sum_rows(ggml_metal_op_t ctx, int idx) {
     ggml_tensor * op = ctx->node(idx);
 
@@ -836,7 +916,12 @@ int ggml_metal_op_sum_rows(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
+
+    GGML_ASSERT(ggml_is_contiguous_rows(op->src[0]));
+
+    ggml_metal_buffer_id bid_src0 = ggml_metal_get_buffer_id(op->src[0]);
+    ggml_metal_buffer_id bid_dst  = ggml_metal_get_buffer_id(op);
 
     ggml_metal_kargs_sum_rows args = {
         /*.ne00 =*/ ne00,
@@ -857,31 +942,28 @@ int ggml_metal_op_sum_rows(ggml_metal_op_t ctx, int idx) {
         /*.nb3  =*/ nb3,
     };
 
-    ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_sum_rows(lib, op);
+    auto pipeline = ggml_metal_library_get_pipeline_sum_rows(lib, op);
+
+    if (pipeline.c4) {
+        args.ne00 = ne00/4;
+        args.ne0  = ne0/4;
+    }
 
     int nth = 32; // SIMD width
 
-    while (nth < ne00 && nth < ggml_metal_pipeline_max_theads_per_threadgroup(pipeline)) {
+    while (nth < args.ne00 && nth < ggml_metal_pipeline_max_theads_per_threadgroup(pipeline)) {
         nth *= 2;
     }
 
     nth = std::min(nth, ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));
-    nth = std::min(nth, ne00);
-
-    const size_t smem = ggml_metal_pipeline_get_smem(pipeline);
+    nth = std::min(nth, (int) args.ne00);
 
-    //[encoder setComputePipelineState:pipeline];
-    //[encoder setBytes:&args length:sizeof(args) atIndex:0];
-    //[encoder setBuffer:id_src0 offset:offs_src0 atIndex:1];
-    //[encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
-    //[encoder setThreadgroupMemoryLength:32*sizeof(float) atIndex:0];
-
-    //[encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+    const size_t smem = pipeline.smem;
 
     ggml_metal_encoder_set_pipeline(enc, pipeline);
     ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
-    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[0]), 1);
-    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op),         2);
+    ggml_metal_encoder_set_buffer  (enc, bid_src0, 1);
+    ggml_metal_encoder_set_buffer  (enc, bid_dst,  2);
 
     ggml_metal_encoder_set_threadgroup_memory_size(enc, smem, 0);
 
@@ -890,6 +972,149 @@ int ggml_metal_op_sum_rows(ggml_metal_op_t ctx, int idx) {
     return 1;
 }
 
+int ggml_metal_op_cumsum(ggml_metal_op_t ctx, int idx) {
+    ggml_tensor * op = ctx->node(idx);
+
+    ggml_metal_library_t lib = ctx->lib;
+    ggml_metal_encoder_t enc = ctx->enc;
+
+    GGML_ASSERT(ggml_is_contiguous_rows(op->src[0]));
+
+    GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
+
+    auto pipeline_blk = ggml_metal_library_get_pipeline_cumsum_blk(lib, op);
+
+    int nth = 1;
+    while (nth < ne00 && 2*nth <= ggml_metal_pipeline_max_theads_per_threadgroup(pipeline_blk)) {
+        nth *= 2;
+    }
+
+    GGML_ASSERT(ne00 <= nth*nth);
+
+    const int64_t net0 = (ne00 + nth - 1) / nth;
+    const int64_t net1 = ne01;
+    const int64_t net2 = ne02;
+    const int64_t net3 = ne03;
+
+    const uint64_t nbt0 = sizeof(float);
+    const uint64_t nbt1 = net0*nbt0;
+    const uint64_t nbt2 = net1*nbt1;
+    const uint64_t nbt3 = net2*nbt2;
+
+    const size_t smem = GGML_PAD(32*sizeof(float), 16);
+
+    ggml_metal_buffer_id bid_src0 = ggml_metal_get_buffer_id(op->src[0]);
+    ggml_metal_buffer_id bid_dst  = ggml_metal_get_buffer_id(op);
+
+    ggml_metal_buffer_id bid_tmp = bid_dst;
+    bid_tmp.offs += ggml_nbytes(op);
+
+    {
+        ggml_metal_kargs_cumsum_blk args = {
+            /*.ne00 =*/ ne00,
+            /*.ne01 =*/ ne01,
+            /*.ne02 =*/ ne02,
+            /*.ne03 =*/ ne03,
+            /*.nb00 =*/ nb00,
+            /*.nb01 =*/ nb01,
+            /*.nb02 =*/ nb02,
+            /*.nb03 =*/ nb03,
+            /*.net0 =*/ net0,
+            /*.net1 =*/ net1,
+            /*.net2 =*/ net2,
+            /*.net3 =*/ net3,
+            /*.nbt0 =*/ nbt0,
+            /*.nbt1 =*/ nbt1,
+            /*.nbt2 =*/ nbt2,
+            /*.nbt3 =*/ nbt3,
+            /*.outb =*/ ne00 > nth,
+        };
+
+        ggml_metal_encoder_set_pipeline(enc, pipeline_blk);
+        ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+        ggml_metal_encoder_set_buffer  (enc, bid_src0, 1);
+        ggml_metal_encoder_set_buffer  (enc, bid_tmp,  2);
+        ggml_metal_encoder_set_buffer  (enc, bid_dst,  3);
+
+        ggml_metal_encoder_set_threadgroup_memory_size(enc, smem, 0);
+
+        ggml_metal_encoder_dispatch_threadgroups(enc, net0*ne01, ne02, ne03, nth, 1, 1);
+    }
+
+    if (ne00 > nth) {
+        ggml_metal_op_concurrency_reset(ctx);
+
+        {
+            ggml_metal_kargs_cumsum_blk args = {
+                /*.ne00 =*/ net0,
+                /*.ne01 =*/ net1,
+                /*.ne02 =*/ net2,
+                /*.ne03 =*/ net3,
+                /*.nb00 =*/ nbt0,
+                /*.nb01 =*/ nbt1,
+                /*.nb02 =*/ nbt2,
+                /*.nb03 =*/ nbt3,
+                /*.net0 =*/ net0,
+                /*.net1 =*/ net1,
+                /*.net2 =*/ net2,
+                /*.net3 =*/ net3,
+                /*.nbt0 =*/ nbt0,
+                /*.nbt1 =*/ nbt1,
+                /*.nbt2 =*/ nbt2,
+                /*.nbt3 =*/ nbt3,
+                /*.outb =*/ false,
+            };
+
+            ggml_metal_encoder_set_pipeline(enc, pipeline_blk);
+            ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+            ggml_metal_encoder_set_buffer  (enc, bid_tmp, 1);
+            ggml_metal_encoder_set_buffer  (enc, bid_tmp, 2);
+            ggml_metal_encoder_set_buffer  (enc, bid_tmp, 3);
+
+            ggml_metal_encoder_set_threadgroup_memory_size(enc, smem, 0);
+
+            ggml_metal_encoder_dispatch_threadgroups(enc, net1, net2, net3, nth, 1, 1);
+        }
+
+        ggml_metal_op_concurrency_reset(ctx);
+
+        {
+            auto pipeline_add = ggml_metal_library_get_pipeline_cumsum_add(lib, op);
+
+            ggml_metal_kargs_cumsum_add args = {
+                /*.ne00 =*/ ne00,
+                /*.ne01 =*/ ne01,
+                /*.ne02 =*/ ne02,
+                /*.ne03 =*/ ne03,
+                /*.nb00 =*/ nb00,
+                /*.nb01 =*/ nb01,
+                /*.nb02 =*/ nb02,
+                /*.nb03 =*/ nb03,
+                /*.net0 =*/ net0,
+                /*.net1 =*/ net1,
+                /*.net2 =*/ net2,
+                /*.net3 =*/ net3,
+                /*.nbt0 =*/ nbt0,
+                /*.nbt1 =*/ nbt1,
+                /*.nbt2 =*/ nbt2,
+                /*.nbt3 =*/ nbt3,
+            };
+
+            ggml_metal_encoder_set_pipeline(enc, pipeline_add);
+            ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+            ggml_metal_encoder_set_buffer  (enc, bid_tmp, 1);
+            ggml_metal_encoder_set_buffer  (enc, bid_dst, 2);
+
+            ggml_metal_encoder_dispatch_threadgroups(enc, net0*ne01, ne02, ne03, nth, 1, 1);
+        }
+    }
+
+    return 1;
+}
+
 int ggml_metal_op_get_rows(ggml_metal_op_t ctx, int idx) {
     ggml_tensor * op = ctx->node(idx);
 
@@ -901,28 +1126,36 @@ int ggml_metal_op_get_rows(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
-    ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_get_rows(lib, op->src[0]->type);
+    auto pipeline = ggml_metal_library_get_pipeline_get_rows(lib, op->src[0]->type);
 
     ggml_metal_kargs_get_rows args = {
-        /*.ne00 =*/ ne00,
-        /*.nb01 =*/ nb01,
-        /*.nb02 =*/ nb02,
-        /*.ne10 =*/ ne10,
-        /*.nb10 =*/ nb10,
-        /*.nb11 =*/ nb11,
-        /*.nb1  =*/ nb1,
-        /*.nb2  =*/ nb2,
+        /*.ne00t =*/ ggml_is_quantized(op->src[0]->type) ? ne00/16 : ne00,
+        /*.ne00  =*/ ne00,
+        /*.nb01  =*/ nb01,
+        /*.nb02  =*/ nb02,
+        /*.nb03  =*/ nb03,
+        /*.ne10  =*/ ne10,
+        /*.nb10  =*/ nb10,
+        /*.nb11  =*/ nb11,
+        /*.nb12  =*/ nb12,
+        /*.nb1   =*/ nb1,
+        /*.nb2   =*/ nb2,
+        /*.nb3   =*/ nb3,
     };
 
+    const int nth = std::min(args.ne00t, ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));
+
+    const int nw0 = (args.ne00t + nth - 1)/nth;
+
     ggml_metal_encoder_set_pipeline(enc, pipeline);
     ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
     ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[0]), 1);
     ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[1]), 2);
     ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op),         3);
 
-    ggml_metal_encoder_dispatch_threadgroups(enc, ne10, ne11, ne12, 32, 1, 1);
+    ggml_metal_encoder_dispatch_threadgroups(enc, nw0*ne10, ne11, ne12, nth, 1, 1);
 
     return 1;
 }
@@ -938,9 +1171,9 @@ int ggml_metal_op_set_rows(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
-    ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_set_rows(lib, op->src[1]->type, op->type);
+    auto pipeline = ggml_metal_library_get_pipeline_set_rows(lib, op->src[1]->type, op->type);
 
     const int32_t nk0 = ne0/ggml_blck_size(op->type);
 
@@ -989,6 +1222,48 @@ int ggml_metal_op_set_rows(ggml_metal_op_t ctx, int idx) {
     return 1;
 }
 
+int ggml_metal_op_diag(ggml_metal_op_t ctx, int idx) {
+    ggml_tensor * op = ctx->node(idx);
+
+    ggml_metal_library_t lib = ctx->lib;
+    ggml_metal_encoder_t enc = ctx->enc;
+
+    GGML_TENSOR_LOCALS(int32_t,  ne0, op->src[0], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    GGML_TENSOR_LOCALS(int32_t,  ne, op, ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb, op, nb);
+
+    ggml_metal_kargs_diag args = {
+        /*.ne00 =*/ne00,
+        /*.ne01 =*/ne01,
+        /*.ne02 =*/ne02,
+        /*.ne03 =*/ne03,
+        /*.nb00 =*/nb00,
+        /*.nb01 =*/nb01,
+        /*.nb02 =*/nb02,
+        /*.nb03 =*/nb03,
+        /*.ne0  =*/ne0,
+        /*.ne1  =*/ne1,
+        /*.ne2  =*/ne2,
+        /*.ne3  =*/ne3,
+        /*.nb0  =*/nb0,
+        /*.nb1  =*/nb1,
+        /*.nb2  =*/nb2,
+        /*.nb3  =*/nb3,
+    };
+
+    auto pipeline = ggml_metal_library_get_pipeline_diag(lib, op);
+
+    ggml_metal_encoder_set_pipeline(enc, pipeline);
+    ggml_metal_encoder_set_bytes(enc, &args, sizeof(args), 0);
+    ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op->src[0]), 1);
+    ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op),         2);
+
+    ggml_metal_encoder_dispatch_threadgroups(enc, ne1, ne2, ne3, 32, 1, 1);
+
+    return 1;
+}
+
 int ggml_metal_op_soft_max(ggml_metal_op_t ctx, int idx) {
     ggml_tensor * op = ctx->node(idx);
 
@@ -1002,7 +1277,7 @@ int ggml_metal_op_soft_max(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne2, op->src[2], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb2, op->src[2], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     float scale;
     float max_bias;
@@ -1041,7 +1316,7 @@ int ggml_metal_op_soft_max(ggml_metal_op_t ctx, int idx) {
         /*.n_head_log2 =*/ n_head_log2,
     };
 
-    ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_soft_max(lib, op);
+    auto pipeline = ggml_metal_library_get_pipeline_soft_max(lib, op);
 
     int nth = 32; // SIMD width
 
@@ -1055,7 +1330,7 @@ int ggml_metal_op_soft_max(ggml_metal_op_t ctx, int idx) {
         }
     }
 
-    const size_t smem = ggml_metal_pipeline_get_smem(pipeline);
+    const size_t smem = pipeline.smem;
 
     ggml_metal_encoder_set_pipeline(enc, pipeline);
     ggml_metal_encoder_set_bytes(enc, &args, sizeof(args), 0);
@@ -1090,7 +1365,7 @@ int ggml_metal_op_ssm_conv(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     ggml_metal_kargs_ssm_conv args = {
         /*.ne00 =*/ ne00,
@@ -1111,18 +1386,46 @@ int ggml_metal_op_ssm_conv(ggml_metal_op_t ctx, int idx) {
         /*.nb2  =*/ nb2,
     };
 
-    ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_ssm_conv(lib, op);
-
-    ggml_metal_encoder_set_pipeline(enc, pipeline);
-    ggml_metal_encoder_set_bytes(enc, &args, sizeof(args), 0);
-    ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op->src[0]), 1);
-    ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op->src[1]), 2);
-    ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op), 3);
+    // Use batched kernel for prefill (ne1 > 1) to reduce threadgroup dispatch overhead
+    const bool use_batched = (ne1 > 1);
 
-    ggml_metal_encoder_dispatch_threadgroups(enc, ne01, ne1, ne02, 1, 1, 1);
+    if (use_batched) {
+        // Determine the smallest power of 2 that's >= ne1, but <= 256
+        int BATCH_SIZE;
+        if      (ne1 > 128) BATCH_SIZE = 256;
+        else if (ne1 > 64 ) BATCH_SIZE = 128;
+        else if (ne1 > 32 ) BATCH_SIZE = 64;
+        else if (ne1 > 16 ) BATCH_SIZE = 32;
+        else if (ne1 > 8  ) BATCH_SIZE = 16;
+        else if (ne1 > 4  ) BATCH_SIZE = 8;
+        else                BATCH_SIZE = 2;
 
-    return 1;
-}
+        auto pipeline = ggml_metal_library_get_pipeline_ssm_conv_batched(lib, op, BATCH_SIZE);
+
+        ggml_metal_encoder_set_pipeline(enc, pipeline);
+        ggml_metal_encoder_set_bytes(enc, &args, sizeof(args), 0);
+        ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op->src[0]), 1);
+        ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op->src[1]), 2);
+        ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op),         3);
+
+        // Dispatch: ne01 rows, ceil(ne1/BATCH_SIZE) token batches, ne02 sequences
+        // Each threadgroup has BATCH_SIZE threads, each handling one token
+        const int n_token_batches = (ne1 + BATCH_SIZE - 1) / BATCH_SIZE;
+        ggml_metal_encoder_dispatch_threadgroups(enc, ne01, n_token_batches, ne02, BATCH_SIZE, 1, 1);
+    } else {
+        auto pipeline = ggml_metal_library_get_pipeline_ssm_conv(lib, op);
+
+        ggml_metal_encoder_set_pipeline(enc, pipeline);
+        ggml_metal_encoder_set_bytes(enc, &args, sizeof(args), 0);
+        ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op->src[0]), 1);
+        ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op->src[1]), 2);
+        ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op),         3);
+
+        ggml_metal_encoder_dispatch_threadgroups(enc, ne01, ne1, ne02, 1, 1, 1);
+    }
+
+    return 1;
+}
 
 int ggml_metal_op_ssm_scan(ggml_metal_op_t ctx, int idx) {
     ggml_tensor * op = ctx->node(idx);
@@ -1145,7 +1448,7 @@ int ggml_metal_op_ssm_scan(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne6, op->src[6], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb6, op->src[6], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     const ggml_tensor * src3 = op->src[3];
     const ggml_tensor * src4 = op->src[4];
@@ -1172,26 +1475,37 @@ int ggml_metal_op_ssm_scan(ggml_metal_op_t ctx, int idx) {
         /*.n_seq_tokens =*/ n_seq_tokens,
         /*.n_seqs       =*/ n_seqs,
         /*.s_off        =*/ ggml_nelements(op->src[1]) * sizeof(float),
+        /*.nb00         =*/ nb00,
         /*.nb01         =*/ nb01,
         /*.nb02         =*/ nb02,
         /*.nb03         =*/ nb03,
+        /*.nb10         =*/ nb10,
         /*.nb11         =*/ nb11,
         /*.nb12         =*/ nb12,
+        /*.ns12         =*/ nb12/nb10,
         /*.nb13         =*/ nb13,
+        /*.nb20         =*/ nb20,
         /*.nb21         =*/ nb21,
+        /*.ns21         =*/ nb21/nb20,
         /*.nb22         =*/ nb22,
+        /*.ne30         =*/ ne30,
         /*.nb31         =*/ nb31,
         /*.nb41         =*/ nb41,
         /*.nb42         =*/ nb42,
+        /*.ns42         =*/ nb42/nb40,
         /*.nb43         =*/ nb43,
         /*.nb51         =*/ nb51,
         /*.nb52         =*/ nb52,
+        /*.ns52         =*/ nb52/nb50,
         /*.nb53         =*/ nb53,
+        /*.nb0          =*/ nb0,
     };
 
-    ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_ssm_scan(lib, op);
+    auto pipeline = ggml_metal_library_get_pipeline_ssm_scan(lib, op);
 
-    const size_t sms = ggml_metal_pipeline_get_smem(pipeline);
+    GGML_ASSERT(d_state <= ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));
+
+    const size_t smem = pipeline.smem;
 
     ggml_metal_encoder_set_pipeline(enc, pipeline);
     ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
@@ -1204,15 +1518,9 @@ int ggml_metal_op_ssm_scan(ggml_metal_op_t ctx, int idx) {
     ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[6]), 7);
     ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op),         8);
 
-    ggml_metal_encoder_set_threadgroup_memory_size(enc, sms, 0);
+    ggml_metal_encoder_set_threadgroup_memory_size(enc, smem, 0);
 
-    if (ne30 == 1) {
-        // Mamba-2
-        ggml_metal_encoder_dispatch_threadgroups(enc, d_inner, n_head, n_seqs, d_state, 1, 1);
-    } else {
-        GGML_ASSERT(d_inner == 1);
-        ggml_metal_encoder_dispatch_threadgroups(enc, n_head, n_seqs, 1, d_state, 1, 1);
-    }
+    ggml_metal_encoder_dispatch_threadgroups(enc, d_inner, n_head, n_seqs, d_state, 1, 1);
 
     return 1;
 }
@@ -1226,14 +1534,14 @@ int ggml_metal_op_rwkv(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     const int64_t B = op->op == GGML_OP_RWKV_WKV6 ? op->src[5]->ne[1] : op->src[6]->ne[1];
     const int64_t T = op->src[0]->ne[2];
     const int64_t C = op->ne[0];
     const int64_t H = op->src[0]->ne[1];
 
-    ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_rwkv(lib, op);
+    auto pipeline = ggml_metal_library_get_pipeline_rwkv(lib, op);
 
     int ida = 0;
 
@@ -1258,41 +1566,298 @@ int ggml_metal_op_rwkv(ggml_metal_op_t ctx, int idx) {
     return 1;
 }
 
-int ggml_metal_op_cpy(ggml_metal_op_t ctx, int idx) {
+int ggml_metal_op_gated_delta_net(ggml_metal_op_t ctx, int idx) {
     ggml_tensor * op = ctx->node(idx);
 
     ggml_metal_library_t lib = ctx->lib;
     ggml_metal_encoder_t enc = ctx->enc;
 
+
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
+    GGML_TENSOR_LOCALS( int32_t, ne2, op->src[2], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb2, op->src[2], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
-    ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_cpy(lib, op->src[0]->type, op->type);
+    auto pipeline = ggml_metal_library_get_pipeline_gated_delta_net(lib, op);
 
-    GGML_ASSERT(ne00 % ggml_blck_size(op->src[0]->type) == 0);
+    int ida = 0;
+
+    ggml_metal_kargs_gated_delta_net args = {
+        /*.ne00 =*/ ne00,
+        /*.ne01 =*/ ne01,
+        /*.ne02 =*/ ne02,
+        /*.ne03 =*/ ne03,
+        /*.nb00 =*/ nb00,
+        /*.nb01 =*/ nb01,
+        /*.nb02 =*/ nb02,
+        /*.nb03 =*/ nb03,
+        /*.ne10 =*/ ne10,
+        /*.ne11 =*/ ne11,
+        /*.ne12 =*/ ne12,
+        /*.ne13 =*/ ne13,
+        /*.nb10 =*/ nb10,
+        /*.nb11 =*/ nb11,
+        /*.nb12 =*/ nb12,
+        /*.nb13 =*/ nb13,
+        /*.ne20 =*/ ne20,
+        /*.ne21 =*/ ne21,
+        /*.ne22 =*/ ne22,
+        /*.ne23 =*/ ne23,
+        /*.nb20 =*/ nb20,
+        /*.nb21 =*/ nb21,
+        /*.nb22 =*/ nb22,
+        /*.nb23 =*/ nb23,
+        /*.ns02 =*/ (int32_t) (nb02/sizeof(float)),
+        /*.ns12 =*/ (int32_t) (nb12/sizeof(float)),
+        /*.ns22 =*/ (int32_t) (nb22/sizeof(float)),
+        /*.ne0  =*/ ne0,
+        /*.ne1  =*/ ne1,
+        /*.ne2  =*/ ne2,
+        /*.ne3  =*/ ne3,
+        /*.nb0  =*/ nb0,
+        /*.nb1  =*/ nb1,
+        /*.nb2  =*/ nb2,
+        /*.nb3  =*/ nb3,
+    };
 
-    // TODO: support
-    //const int32_t nk00 = ne00/ggml_blck_size(op->type);
-    const int32_t nk00 = ne00;
+    ggml_metal_encoder_set_pipeline(enc, pipeline);
+    ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args),                  ida++);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[0]), ida++); // q
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[1]), ida++); // k
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[2]), ida++); // v
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[3]), ida++); // gate
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[4]), ida++); // beta
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[5]), ida++); // state
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op),         ida++); // dst
 
-    int nth = 32; // SIMD width
+    const int nsg = pipeline.nsg;
 
-    while (nth < nk00 && nth < ggml_metal_pipeline_max_theads_per_threadgroup(pipeline)) {
-        nth *= 2;
+    ggml_metal_encoder_dispatch_threadgroups(enc, op->src[2]->ne[0]/nsg, op->src[2]->ne[1], op->src[2]->ne[3], 32, nsg, 1);
+
+    return 1;
+}
+
+int ggml_metal_op_solve_tri(ggml_metal_op_t ctx, int idx) {
+    ggml_tensor * op = ctx->node(idx);
+
+    ggml_metal_library_t lib = ctx->lib;
+    ggml_metal_encoder_t enc = ctx->enc;
+
+    GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
+    GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
+
+    ggml_metal_kargs_solve_tri args = {
+        /*.ne00 =*/ ne00,
+        /*.ne01 =*/ ne01,
+        /*.ne02 =*/ ne02,
+        /*.ne03 =*/ ne03,
+        /*.nb00 =*/ nb00,
+        /*.nb01 =*/ nb01,
+        /*.nb02 =*/ nb02,
+        /*.nb03 =*/ nb03,
+        /*.ne10 =*/ ne10,
+        /*.ne11 =*/ ne11,
+        /*.ne12 =*/ ne12,
+        /*.ne13 =*/ ne13,
+        /*.nb10 =*/ nb10,
+        /*.nb11 =*/ nb11,
+        /*.nb12 =*/ nb12,
+        /*.nb13 =*/ nb13,
+        /*.ne0  =*/ ne0,
+        /*.ne1  =*/ ne1,
+        /*.ne2  =*/ ne2,
+        /*.ne3  =*/ ne3,
+        /*.nb0  =*/ nb0,
+        /*.nb1  =*/ nb1,
+        /*.nb2  =*/ nb2,
+        /*.nb3  =*/ nb3,
+    };
+
+    auto pipeline = ggml_metal_library_get_pipeline_solve_tri(lib, op);
+
+    ggml_metal_encoder_set_pipeline(enc, pipeline);
+    ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[0]), 1);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[1]), 2);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op),         3);
+
+    const int nsg = pipeline.nsg;
+
+    ggml_metal_encoder_set_threadgroup_memory_size(enc, pipeline.smem, 0);
+
+    ggml_metal_encoder_dispatch_threadgroups(enc, (ne10 + nsg - 1)/nsg, ne02, ne03, 32, nsg, 1);
+
+    return 1;
+}
+
+int ggml_metal_op_set(ggml_metal_op_t ctx, int idx) {
+    ggml_tensor * op = ctx->node(idx);
+
+    ggml_metal_library_t lib = ctx->lib;
+    ggml_metal_encoder_t enc = ctx->enc;
+
+    GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
+    GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
+
+    ggml_metal_buffer_id bid_src0 = ggml_metal_get_buffer_id(op->src[0]);
+    ggml_metal_buffer_id bid_src1 = ggml_metal_get_buffer_id(op->src[1]);
+    ggml_metal_buffer_id bid_dst  = ggml_metal_get_buffer_id(op);
+
+    const size_t pnb1 = ((const int32_t *) op->op_params)[0];
+    const size_t pnb2 = ((const int32_t *) op->op_params)[1];
+    const size_t pnb3 = ((const int32_t *) op->op_params)[2];
+    const size_t offs = ((const int32_t *) op->op_params)[3];
+
+    const bool inplace = (bool) ((const int32_t *) op->op_params)[4];
+
+    if (!inplace) {
+        // run a separate kernel to cpy src->dst
+        // not sure how to avoid this
+        // TODO: make a simpler cpy_bytes kernel
+
+        //const id<MTLComputePipelineState> pipeline = ctx->pipelines[GGML_METAL_PIPELINE_TYPE_CPY_F32_F32].obj;
+        auto pipeline = ggml_metal_library_get_pipeline_cpy(lib, op->src[0]->type, op->type);
+
+        ggml_metal_kargs_cpy args = {
+            /*.nk0  =*/ ne00,
+            /*.ne00 =*/ ne00,
+            /*.ne01 =*/ ne01,
+            /*.ne02 =*/ ne02,
+            /*.ne03 =*/ ne03,
+            /*.nb00 =*/ nb00,
+            /*.nb01 =*/ nb01,
+            /*.nb02 =*/ nb02,
+            /*.nb03 =*/ nb03,
+            /*.ne0  =*/ ne0,
+            /*.ne1  =*/ ne1,
+            /*.ne2  =*/ ne2,
+            /*.ne3  =*/ ne3,
+            /*.nb0  =*/ nb0,
+            /*.nb1  =*/ nb1,
+            /*.nb2  =*/ nb2,
+            /*.nb3  =*/ nb3,
+        };
+
+        ggml_metal_encoder_set_pipeline(enc, pipeline);
+        ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+        ggml_metal_encoder_set_buffer  (enc, bid_src0, 1);
+        ggml_metal_encoder_set_buffer  (enc, bid_dst,  2);
+
+        const int nth = std::min(ggml_metal_pipeline_max_theads_per_threadgroup(pipeline), ne00);
+
+        ggml_metal_encoder_dispatch_threadgroups(enc, ne01, ne02, ne03, nth, 1, 1);
+
+        ggml_metal_op_concurrency_reset(ctx);
     }
 
-    nth = std::min(nth, ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));
+    auto pipeline = ggml_metal_library_get_pipeline_cpy(lib, op->src[1]->type, op->type);
+
+    GGML_ASSERT(ne10 % ggml_blck_size(op->src[1]->type) == 0);
+
+    int64_t nk0 = ne10;
+    if (ggml_is_quantized(op->src[1]->type)) {
+        nk0 = ne10/16;
+    } else if (ggml_is_quantized(op->type)) {
+        nk0 = ne10/ggml_blck_size(op->type);
+    }
+
+    int nth = std::min<int>(nk0, ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));
+
+    // when rows are small, we can batch them together in a single threadgroup
+    int nrptg = 1;
+
+    // TODO: relax this constraint in the future
+    if (ggml_blck_size(op->src[1]->type) == 1 && ggml_blck_size(op->type) == 1) {
+        if (nth > nk0) {
+            nrptg = (nth + nk0 - 1)/nk0;
+            nth   = nk0;
+
+            if (nrptg*nth > ggml_metal_pipeline_max_theads_per_threadgroup(pipeline)) {
+                nrptg--;
+            }
+        }
+    }
+
+    nth = std::min<int>(nth, nk0);
+
+    ggml_metal_kargs_cpy args = {
+        /*.nk0  =*/ nk0,
+        /*.ne00 =*/ ne10,
+        /*.ne01 =*/ ne11,
+        /*.ne02 =*/ ne12,
+        /*.ne03 =*/ ne13,
+        /*.nb00 =*/ nb10,
+        /*.nb01 =*/ nb11,
+        /*.nb02 =*/ nb12,
+        /*.nb03 =*/ nb13,
+        /*.ne0  =*/ ne10,
+        /*.ne1  =*/ ne11,
+        /*.ne2  =*/ ne12,
+        /*.ne3  =*/ ne13,
+        /*.nb0  =*/ ggml_element_size(op),
+        /*.nb1  =*/ pnb1,
+        /*.nb2  =*/ pnb2,
+        /*.nb3  =*/ pnb3,
+    };
+
+    const int nw0 = nrptg == 1 ? (nk0 + nth - 1)/nth : 1;
+
+    bid_dst.offs += offs;
+
+    ggml_metal_encoder_set_pipeline(enc, pipeline);
+    ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+    ggml_metal_encoder_set_buffer  (enc, bid_src1, 1);
+    ggml_metal_encoder_set_buffer  (enc, bid_dst,  2);
+
+    ggml_metal_encoder_dispatch_threadgroups(enc, nw0*(ne11 + nrptg - 1)/nrptg, ne12, ne13, nth, nrptg, 1);
+
+    return 1;
+}
+
+int ggml_metal_op_cpy(ggml_metal_op_t ctx, int idx) {
+    ggml_tensor * op = ctx->node(idx);
+
+    ggml_metal_library_t lib = ctx->lib;
+    ggml_metal_encoder_t enc = ctx->enc;
+
+    GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
+
+    auto pipeline = ggml_metal_library_get_pipeline_cpy(lib, op->src[0]->type, op->type);
+
+    GGML_ASSERT(ne00 % ggml_blck_size(op->src[0]->type) == 0);
+
+    int64_t nk0 = ne00;
+    if (ggml_is_quantized(op->src[0]->type)) {
+        nk0 = ne00/16;
+    } else if (ggml_is_quantized(op->type)) {
+        nk0 = ne00/ggml_blck_size(op->type);
+    }
+
+    int nth = std::min<int>(nk0, ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));
 
     // when rows are small, we can batch them together in a single threadgroup
     int nrptg = 1;
 
     // TODO: relax this constraint in the future
     if (ggml_blck_size(op->src[0]->type) == 1 && ggml_blck_size(op->type) == 1) {
-        if (nth > nk00) {
-            nrptg = (nth + nk00 - 1)/nk00;
-            nth   = nk00;
+        if (nth > nk0) {
+            nrptg = (nth + nk0 - 1)/nk0;
+            nth   = nk0;
 
             if (nrptg*nth > ggml_metal_pipeline_max_theads_per_threadgroup(pipeline)) {
                 nrptg--;
@@ -1300,10 +1865,11 @@ int ggml_metal_op_cpy(ggml_metal_op_t ctx, int idx) {
         }
     }
 
-    nth = std::min(nth, nk00);
+    nth = std::min<int>(nth, nk0);
 
     ggml_metal_kargs_cpy args = {
-        /*.ne00 =*/ nk00,
+        /*.nk0  =*/ nk0,
+        /*.ne00 =*/ ne00,
         /*.ne01 =*/ ne01,
         /*.ne02 =*/ ne02,
         /*.ne03 =*/ ne03,
@@ -1321,16 +1887,66 @@ int ggml_metal_op_cpy(ggml_metal_op_t ctx, int idx) {
         /*.nb3  =*/ nb3,
     };
 
+    const int nw0 = nrptg == 1 ? (nk0 + nth - 1)/nth : 1;
+
     ggml_metal_encoder_set_pipeline(enc, pipeline);
     ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
     ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[0]), 1);
     ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op),         2);
 
-    ggml_metal_encoder_dispatch_threadgroups(enc, ne01, ne02, ne03, nth, nrptg, 1);
+    ggml_metal_encoder_dispatch_threadgroups(enc, nw0*(ne01 + nrptg - 1)/nrptg, ne02, ne03, nth, nrptg, 1);
+
+    return 1;
+}
+
+int ggml_metal_op_pool_1d(ggml_metal_op_t ctx, int idx) {
+    ggml_tensor * op = ctx->node(idx);
+
+    ggml_metal_library_t lib = ctx->lib;
+    ggml_metal_encoder_t enc = ctx->enc;
+
+    GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
+
+    const int32_t * opts = op->op_params;
+    ggml_op_pool op_pool = (ggml_op_pool) opts[0];
+
+    const int32_t k0 = opts[1];
+    const int32_t s0 = opts[2];
+    const int32_t p0 = opts[3];
+
+    const int64_t IW = op->src[0]->ne[0];
+    const int64_t OW = op->ne[0];
+
+    const int64_t np = ggml_nelements(op);
+
+    ggml_metal_kargs_pool_1d args_pool_1d = {
+        /* .k0 = */  k0,
+        /* .s0 = */  s0,
+        /* .p0 = */  p0,
+        /* .IW = */  IW,
+        /* .OW = */  OW,
+        /* .np = */  np
+    };
+
+    auto pipeline = ggml_metal_library_get_pipeline_pool_1d(lib, op, op_pool);
+
+    const int nth = std::min(ggml_metal_pipeline_max_theads_per_threadgroup(pipeline), (int) np);
+    const int ntg = (np + nth - 1) / nth;
+
+    ggml_metal_encoder_set_pipeline(enc, pipeline);
+    ggml_metal_encoder_set_bytes   (enc, &args_pool_1d, sizeof(args_pool_1d),  0);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[0]), 1);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op),         2);
+
+    ggml_metal_encoder_dispatch_threadgroups(enc, ntg, 1, 1, nth, 1, 1);
 
     return 1;
 }
 
+
 int ggml_metal_op_pool_2d(ggml_metal_op_t ctx, int idx) {
     ggml_tensor * op = ctx->node(idx);
 
@@ -1340,7 +1956,7 @@ int ggml_metal_op_pool_2d(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     const int32_t * opts = op->op_params;
     ggml_op_pool op_pool = (ggml_op_pool) opts[0];
@@ -1376,7 +1992,7 @@ int ggml_metal_op_pool_2d(ggml_metal_op_t ctx, int idx) {
         /* .np = */ np
     };
 
-    ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_pool_2d(lib, op, op_pool);
+    auto pipeline = ggml_metal_library_get_pipeline_pool_2d(lib, op, op_pool);
 
     const int nth = std::min(ggml_metal_pipeline_max_theads_per_threadgroup(pipeline), (int) np);
     const int ntg = (np + nth - 1) / nth;
@@ -1404,7 +2020,7 @@ int ggml_metal_op_mul_mat(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     GGML_ASSERT(ne00 == ne10);
 
@@ -1426,6 +2042,7 @@ int ggml_metal_op_mul_mat(ggml_metal_op_t ctx, int idx) {
           (
            op->src[0]->type == GGML_TYPE_F32  || // TODO: helper function
            op->src[0]->type == GGML_TYPE_F16  ||
+           op->src[0]->type == GGML_TYPE_BF16 ||
            op->src[0]->type == GGML_TYPE_Q4_0 ||
            op->src[0]->type == GGML_TYPE_Q4_1 ||
            op->src[0]->type == GGML_TYPE_Q5_0 ||
@@ -1440,6 +2057,8 @@ int ggml_metal_op_mul_mat(ggml_metal_op_t ctx, int idx) {
            op->src[0]->type == GGML_TYPE_Q4_K ||
            op->src[0]->type == GGML_TYPE_Q5_K ||
            op->src[0]->type == GGML_TYPE_Q6_K ||
+           op->src[0]->type == GGML_TYPE_Q2_K ||
+           op->src[0]->type == GGML_TYPE_Q3_K ||
            false) && (ne11 >= 4 && ne11 <= 8)
          )
         )
@@ -1468,7 +2087,7 @@ int ggml_metal_op_mul_mat(ggml_metal_op_t ctx, int idx) {
         const int16_t r0ptg  = nypsg*nsg;         // num src0 rows per threadgroup
               int16_t r1ptg  = 4;                 // num src1 rows per threadgroup
 
-        // note: not sure how optimal are those across all different hardware. there might be someting cleverer
+        // note: not sure how optimal are those across all different hardware. there might be something cleverer
         switch (ne11) {
             case 2:
                 r1ptg = 2; break;
@@ -1485,7 +2104,7 @@ int ggml_metal_op_mul_mat(ggml_metal_op_t ctx, int idx) {
                 GGML_ABORT("unsupported ne11");
         };
 
-        ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_mul_mv_ext(lib, op->src[0]->type, op->src[1]->type, nsg, nxpsg, r1ptg);
+        auto pipeline = ggml_metal_library_get_pipeline_mul_mv_ext(lib, op->src[0]->type, op->src[1]->type, nsg, nxpsg, r1ptg);
 
         ggml_metal_kargs_mul_mv_ext args = {
             /*.ne00  =*/ ne00,
@@ -1520,9 +2139,8 @@ int ggml_metal_op_mul_mat(ggml_metal_op_t ctx, int idx) {
         !ggml_is_transposed(op->src[1]) &&
         // for now the matrix-matrix multiplication kernel only works on A14+/M1+ SoCs
         // AMD GPU and older A-chips will reuse matrix-vector multiplication kernel
-        props_dev->has_simdgroup_mm && ne00 >= 64 &&
-        (ne11 > ne11_mm_min || (ggml_is_quantized(op->src[0]->type) && ne12 > 1))) {
-        //printf("matrix: ne00 = %6d, ne01 = %6d, ne02 = %6d, ne11 = %6d, ne12 = %6d\n", ne00, ne01, ne02, ne11, ne12);
+        props_dev->has_simdgroup_mm && ne00 >= 64 && ne11 > ne11_mm_min) {
+        //GGML_LOG_INFO("matrix: ne00 = %6d, ne01 = %6d, ne02 = %6d, ne11 = %6d, ne12 = %6d\n", ne00, ne01, ne02, ne11, ne12);
 
         // some Metal matrix data types require aligned pointers
         // ref: https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf (Table 2.5)
@@ -1533,7 +2151,7 @@ int ggml_metal_op_mul_mat(ggml_metal_op_t ctx, int idx) {
         //    default: break;
         //}
 
-        ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_mul_mm(lib, op);
+        auto pipeline = ggml_metal_library_get_pipeline_mul_mm(lib, op);
 
         ggml_metal_kargs_mul_mm args = {
             /*.ne00 =*/ ne00,
@@ -1558,18 +2176,18 @@ int ggml_metal_op_mul_mat(ggml_metal_op_t ctx, int idx) {
         ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[1]), 2);
         ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op),         3);
 
-        const size_t smem = ggml_metal_pipeline_get_smem(pipeline);
+        const size_t smem = pipeline.smem;
 
         ggml_metal_encoder_set_threadgroup_memory_size(enc, smem, 0);
         ggml_metal_encoder_dispatch_threadgroups(enc, ((ne11 + 31)/32), ((ne01 + 63)/64), ne12*ne13, 128, 1, 1);
     } else {
-        ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_mul_mv(lib, op);
+        auto pipeline = ggml_metal_library_get_pipeline_mul_mv(lib, op);
 
-        const int nr0 = ggml_metal_pipeline_get_nr0(pipeline);
-        const int nr1 = ggml_metal_pipeline_get_nr1(pipeline);
-        const int nsg = ggml_metal_pipeline_get_nsg(pipeline);
+        const int nr0 = pipeline.nr0;
+        const int nr1 = pipeline.nr1;
+        const int nsg = pipeline.nsg;
 
-        const size_t smem = ggml_metal_pipeline_get_smem(pipeline);
+        const size_t smem = pipeline.smem;
 
         ggml_metal_kargs_mul_mv args = {
             /*.ne00 =*/ ne00,
@@ -1646,7 +2264,7 @@ int ggml_metal_op_mul_mat_id(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne2, op->src[2], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb2, op->src[2], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     // src2 = ids
     GGML_ASSERT(op->src[2]->type == GGML_TYPE_I32);
@@ -1700,9 +2318,9 @@ int ggml_metal_op_mul_mat_id(ggml_metal_op_t ctx, int idx) {
                 nb21,
             };
 
-            ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_mul_mm_id_map0(lib, ne02, ne20);
+            auto pipeline = ggml_metal_library_get_pipeline_mul_mm_id_map0(lib, ne02, ne20);
 
-            const size_t smem = ggml_metal_pipeline_get_smem(pipeline);
+            const size_t smem = pipeline.smem;
 
             GGML_ASSERT(ne02 <= ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));
 
@@ -1723,7 +2341,7 @@ int ggml_metal_op_mul_mat_id(ggml_metal_op_t ctx, int idx) {
         ggml_metal_op_concurrency_reset(ctx);
 
         {
-            ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_mul_mm_id(lib, op);
+            auto pipeline = ggml_metal_library_get_pipeline_mul_mm_id(lib, op);
 
             ggml_metal_kargs_mul_mm_id args = {
                 /*.ne00  =*/ ne00,
@@ -1752,20 +2370,20 @@ int ggml_metal_op_mul_mat_id(ggml_metal_op_t ctx, int idx) {
             ggml_metal_encoder_set_buffer  (enc, bid_ids,  4);
             ggml_metal_encoder_set_buffer  (enc, bid_dst,  5);
 
-            const size_t smem = ggml_metal_pipeline_get_smem(pipeline);
+            const size_t smem = pipeline.smem;
 
             ggml_metal_encoder_set_threadgroup_memory_size(enc, smem, 0);
 
             ggml_metal_encoder_dispatch_threadgroups(enc, (ne21 + 31)/32, (ne01 + 63)/64, ne02, 128, 1, 1);
         }
     } else {
-        ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_mul_mv_id(lib, op);
+        auto pipeline = ggml_metal_library_get_pipeline_mul_mv_id(lib, op);
 
-        const int nr0 = ggml_metal_pipeline_get_nr0(pipeline);
-        const int nr1 = ggml_metal_pipeline_get_nr1(pipeline);
-        const int nsg = ggml_metal_pipeline_get_nsg(pipeline);
+        const int nr0 = pipeline.nr0;
+        const int nr1 = pipeline.nr1;
+        const int nsg = pipeline.nsg;
 
-        const size_t smem = ggml_metal_pipeline_get_smem(pipeline);
+        const size_t smem = pipeline.smem;
 
         ggml_metal_kargs_mul_mv_id args = {
             /*.nei0 =*/ ne20,
@@ -1849,7 +2467,7 @@ int ggml_metal_op_add_id(ggml_metal_op_t ctx, int idx) {
         /*.nb21 =*/ nb21,
     };
 
-    ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_base(lib, GGML_OP_ADD_ID);
+    auto pipeline = ggml_metal_library_get_pipeline_base(lib, GGML_OP_ADD_ID);
 
     ggml_metal_encoder_set_pipeline(enc, pipeline);
     ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
@@ -1875,20 +2493,118 @@ bool ggml_metal_op_flash_attn_ext_use_vec(const ggml_tensor * op) {
     return (ne01 < 20) && (ne00 % 32 == 0);
 }
 
+size_t ggml_metal_op_flash_attn_ext_extra_pad(const ggml_tensor * op) {
+    assert(op->op == GGML_OP_FLASH_ATTN_EXT);
+
+    GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
+    GGML_TENSOR_LOCALS( int32_t, ne2, op->src[2], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb2, op->src[2], nb);
+    GGML_TENSOR_LOCALS( int32_t, ne3, op->src[3], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb3, op->src[3], nb);
+
+    size_t res = 0;
+
+    const bool has_mask = op->src[3] != nullptr;
+
+    // note: the non-vec kernel requires more extra memory, so always reserve for it
+    GGML_ASSERT(OP_FLASH_ATTN_EXT_NCPSG >= OP_FLASH_ATTN_EXT_VEC_NCPSG);
+
+    //if (ggml_metal_op_flash_attn_ext_use_vec(op)) {
+    if (false) {
+        // note: always reserve the padding space to avoid graph reallocations
+        //const bool has_kvpad = ne11 % OP_FLASH_ATTN_EXT_VEC_NCPSG != 0;
+        const bool has_kvpad = true;
+
+        if (has_kvpad) {
+            res += OP_FLASH_ATTN_EXT_VEC_NCPSG*(
+                nb11*ne12*ne13 +
+                nb21*ne22*ne23 +
+                (has_mask ? ggml_type_size(GGML_TYPE_F16)*ne31*ne32*ne33 : 0));
+        }
+    } else {
+        //const bool has_kvpad = ne11 % OP_FLASH_ATTN_EXT_NCPSG != 0;
+        const bool has_kvpad = true;
+
+        if (has_kvpad) {
+            res += OP_FLASH_ATTN_EXT_NCPSG*(
+                nb11*ne12*ne13 +
+                nb21*ne22*ne23 +
+                (has_mask ? ggml_type_size(GGML_TYPE_F16)*ne31*ne32*ne33 : 0));
+        }
+    }
+
+    return res;
+}
+
+size_t ggml_metal_op_flash_attn_ext_extra_blk(const ggml_tensor * op) {
+    assert(op->op == GGML_OP_FLASH_ATTN_EXT);
+
+    GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+  //GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+  //GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
+  //GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
+  //GGML_TENSOR_LOCALS( int32_t, ne2, op->src[2], ne);
+  //GGML_TENSOR_LOCALS(uint64_t, nb2, op->src[2], nb);
+    GGML_TENSOR_LOCALS( int32_t, ne3, op->src[3], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb3, op->src[3], nb);
+
+    size_t res = 0;
+
+    const bool has_mask = op->src[3] != nullptr;
+
+    if (!has_mask) {
+        return res;
+    }
+
+    const bool is_vec = ggml_metal_op_flash_attn_ext_use_vec(op);
+
+    // this optimization is not useful for the vector kernels
+    // note: always reserve the blk buffer to avoid graph reallocations
+    //if (is_vec) {
+    //    return res;
+    //}
+
+    const int nqptg = is_vec ? OP_FLASH_ATTN_EXT_VEC_NQPSG : OP_FLASH_ATTN_EXT_NQPSG;
+    const int ncpsg = is_vec ? OP_FLASH_ATTN_EXT_VEC_NCPSG : OP_FLASH_ATTN_EXT_NCPSG;
+
+    const int64_t ne1 = (ne01 + nqptg - 1)/nqptg;
+    const int64_t ne0 = (ne30 + ncpsg - 1)/ncpsg;
+
+    res += GGML_PAD(ggml_type_size(GGML_TYPE_I8)*ne0*ne1*ne32*ne33, 32);
+
+    return res;
+}
+
 size_t ggml_metal_op_flash_attn_ext_extra_tmp(const ggml_tensor * op) {
     assert(op->op == GGML_OP_FLASH_ATTN_EXT);
 
-    const int64_t nwg = 32;
+    GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+  //GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
+  //GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
+    GGML_TENSOR_LOCALS( int32_t, ne2, op->src[2], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb2, op->src[2], nb);
+  //GGML_TENSOR_LOCALS( int32_t, ne3, op->src[3], ne);
+  //GGML_TENSOR_LOCALS(uint64_t, nb3, op->src[3], nb);
+
+    size_t res = 0;
 
-    const int64_t ne01 = op->src[0]->ne[1];
-    const int64_t ne02 = op->src[0]->ne[2];
-    const int64_t ne03 = op->src[0]->ne[3];
-    const int64_t ne20 = op->src[2]->ne[0];
+    // note: always reserve the temp buffer to avoid graph reallocations
+    //if (ggml_metal_op_flash_attn_ext_use_vec(op)) {
+    if (true) {
+        const int64_t nwg = 32;
+        const int64_t ne01_max = std::min(ne01, 32);
 
-    // temp buffer for writing the results from each workgroup
-    // - ne20: the size of the Value head
-    // -  + 2: the S and M values for each intermediate result
-    return ggml_type_size(GGML_TYPE_F32)*(ne01*ne02*ne03*nwg*(ne20 + 2));
+        // temp buffer for writing the results from each workgroup
+        // - ne20: the size of the Value head
+        // -  + 2: the S and M values for each intermediate result
+        res += ggml_type_size(GGML_TYPE_F32)*(ne01_max*ne02*ne03*nwg*(ne20 + 2));
+    }
+
+    return res;
 }
 
 int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
@@ -1910,8 +2626,7 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
     GGML_TENSOR_LOCALS( int32_t, nb,  op,         nb);
 
-    GGML_ASSERT(ne00 % 4  == 0);
-    GGML_ASSERT(ne11 % 32 == 0);
+    GGML_ASSERT(ne00 % 4 == 0);
 
     GGML_ASSERT(op->src[0]->type == GGML_TYPE_F32);
     GGML_ASSERT(op->src[1]->type == op->src[2]->type);
@@ -1921,8 +2636,8 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
     GGML_ASSERT(ne12 == ne22);
 
     GGML_ASSERT(!op->src[3] || op->src[3]->type == GGML_TYPE_F16);
-    GGML_ASSERT(!op->src[3] || op->src[3]->ne[1] >= GGML_PAD(op->src[0]->ne[1], 8) &&
-            "the Flash-Attention Metal kernel requires the mask to be padded to 8 and at least n_queries big");
+    GGML_ASSERT(!op->src[3] || op->src[3]->ne[1] >= op->src[0]->ne[1] &&
+            "the Flash-Attention Metal kernel requires the mask to be at least n_queries big");
 
     float scale;
     float max_bias;
@@ -1949,15 +2664,107 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
 
     GGML_ASSERT(ne01 < 65536);
 
+    ggml_metal_buffer_id bid_src0 = ggml_metal_get_buffer_id(op->src[0]);
+    ggml_metal_buffer_id bid_src1 = ggml_metal_get_buffer_id(op->src[1]);
+    ggml_metal_buffer_id bid_src2 = ggml_metal_get_buffer_id(op->src[2]);
+    ggml_metal_buffer_id bid_src3 = has_mask  ? ggml_metal_get_buffer_id(op->src[3]) : bid_src0;
+    ggml_metal_buffer_id bid_src4 = has_sinks ? ggml_metal_get_buffer_id(op->src[4]) : bid_src0;
+
+    ggml_metal_buffer_id bid_dst = ggml_metal_get_buffer_id(op);
+
+    ggml_metal_buffer_id bid_pad = bid_dst;
+    bid_pad.offs += ggml_nbytes(op);
+
+    ggml_metal_buffer_id bid_blk = bid_pad;
+    bid_blk.offs += ggml_metal_op_flash_attn_ext_extra_pad(op);
+
+    ggml_metal_buffer_id bid_tmp = bid_blk;
+    bid_tmp.offs += ggml_metal_op_flash_attn_ext_extra_blk(op);
+
     if (!ggml_metal_op_flash_attn_ext_use_vec(op)) {
         // half8x8 kernel
-        const int64_t nqptg = 8;  // queries per threadgroup    !! sync with kernel template arguments !!
-        const int64_t ncpsg = 64; // cache values per simdgroup !! sync with kernel template arguments !!
+        const int nqptg = OP_FLASH_ATTN_EXT_NQPSG; // queries per threadgroup
+        const int ncpsg = OP_FLASH_ATTN_EXT_NCPSG; // cache values per simdgroup
 
         GGML_ASSERT(nqptg <= 32);
         GGML_ASSERT(nqptg  % 8  == 0);
         GGML_ASSERT(ncpsg  % 32 == 0);
 
+        bool need_sync = false;
+
+        const bool has_kvpad = ne11 % ncpsg != 0;
+
+        if (has_kvpad) {
+            assert(ggml_metal_op_flash_attn_ext_extra_pad(op) != 0);
+
+            ggml_metal_kargs_flash_attn_ext_pad args0 = {
+                /*.ne11    =*/ne11,
+                /*.ne_12_2 =*/ne12,
+                /*.ne_12_3 =*/ne13,
+                /*.nb11    =*/nb11,
+                /*.nb12    =*/nb12,
+                /*.nb13    =*/nb13,
+                /*.nb21    =*/nb21,
+                /*.nb22    =*/nb22,
+                /*.nb23    =*/nb23,
+                /*.ne31    =*/ne31,
+                /*.ne32    =*/ne32,
+                /*.ne33    =*/ne33,
+                /*.nb31    =*/nb31,
+                /*.nb32    =*/nb32,
+                /*.nb33    =*/nb33,
+            };
+
+            auto pipeline0 = ggml_metal_library_get_pipeline_flash_attn_ext_pad(lib, op, has_mask, ncpsg);
+
+            ggml_metal_encoder_set_pipeline(enc, pipeline0);
+            ggml_metal_encoder_set_bytes   (enc, &args0, sizeof(args0), 0);
+            ggml_metal_encoder_set_buffer  (enc, bid_src1, 1);
+            ggml_metal_encoder_set_buffer  (enc, bid_src2, 2);
+            ggml_metal_encoder_set_buffer  (enc, bid_src3, 3);
+            ggml_metal_encoder_set_buffer  (enc, bid_pad,  4);
+
+            assert(ne12 == ne22);
+            assert(ne13 == ne23);
+
+            ggml_metal_encoder_dispatch_threadgroups(enc, ncpsg, std::max(ne12, ne32), std::max(ne13, ne33), 32, 1, 1);
+
+            need_sync = true;
+        }
+
+        if (has_mask) {
+            assert(ggml_metal_op_flash_attn_ext_extra_blk(op) != 0);
+
+            ggml_metal_kargs_flash_attn_ext_blk args0 = {
+                /*.ne01 =*/ ne01,
+                /*.ne30 =*/ ne30,
+                /*.ne31 =*/ ne31,
+                /*.ne32 =*/ ne32,
+                /*.ne33 =*/ ne33,
+                /*.nb31 =*/ nb31,
+                /*.nb32 =*/ nb32,
+                /*.nb33 =*/ nb33,
+            };
+
+            auto pipeline0 = ggml_metal_library_get_pipeline_flash_attn_ext_blk(lib, op, nqptg, ncpsg);
+
+            ggml_metal_encoder_set_pipeline(enc, pipeline0);
+            ggml_metal_encoder_set_bytes   (enc, &args0, sizeof(args0), 0);
+            ggml_metal_encoder_set_buffer  (enc, bid_src3, 1);
+            ggml_metal_encoder_set_buffer  (enc, bid_blk,  2);
+
+            const int32_t nblk1 = ((ne01 + nqptg - 1)/nqptg);
+            const int32_t nblk0 = ((ne30 + ncpsg - 1)/ncpsg);
+
+            ggml_metal_encoder_dispatch_threadgroups(enc, nblk0, nblk1, ne32*ne33, 32, 1, 1);
+
+            need_sync = true;
+        }
+
+        if (need_sync) {
+            ggml_metal_op_concurrency_reset(ctx);
+        }
+
         const int is_q = ggml_is_quantized(op->src[1]->type) ? 1 : 0;
 
         // 2*(2*ncpsg)
@@ -1985,7 +2792,7 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
 
         // simdgroups per threadgroup (a.k.a. warps)
         //nsg = ne01 <= nqptg ? MAX(4, MIN(nsgmax, MIN(ne11/ncpsg, (int64_t) pipeline.maxTotalThreadsPerThreadgroup/32))) : 4;
-        int32_t nsg = 4;
+        int32_t nsg = ne00 >= 512 ? 8 : 4;
 
         const size_t smem = FATTN_SMEM(nsg);
 
@@ -2007,6 +2814,7 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
             /*.nb21          =*/ nb21,
             /*.nb22          =*/ nb22,
             /*.nb23          =*/ nb23,
+            /*.ne31          =*/ ne31,
             /*.ne32          =*/ ne32,
             /*.ne33          =*/ ne33,
             /*.nb31          =*/ nb31,
@@ -2023,24 +2831,18 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
             /*.logit_softcap =*/ logit_softcap,
         };
 
-        ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_flash_attn_ext(lib, op, has_mask, has_sinks, has_bias, has_scap, nsg);
+        auto pipeline = ggml_metal_library_get_pipeline_flash_attn_ext(lib, op, has_mask, has_sinks, has_bias, has_scap, has_kvpad, nsg);
 
         ggml_metal_encoder_set_pipeline(enc, pipeline);
         ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
-        ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[0]), 1);
-        ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[1]), 2);
-        ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[2]), 3);
-        if (op->src[3]) {
-            ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op->src[3]), 4);
-        } else {
-            ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op->src[0]), 4);
-        }
-        if (op->src[4]) {
-            ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op->src[4]), 5);
-        } else {
-            ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op->src[0]), 5);
-        }
-        ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op),         6);
+        ggml_metal_encoder_set_buffer  (enc, bid_src0, 1);
+        ggml_metal_encoder_set_buffer  (enc, bid_src1, 2);
+        ggml_metal_encoder_set_buffer  (enc, bid_src2, 3);
+        ggml_metal_encoder_set_buffer  (enc, bid_src3, 4);
+        ggml_metal_encoder_set_buffer  (enc, bid_src4, 5);
+        ggml_metal_encoder_set_buffer  (enc, bid_pad,  6);
+        ggml_metal_encoder_set_buffer  (enc, bid_blk,  7);
+        ggml_metal_encoder_set_buffer  (enc, bid_dst,  8);
 
         ggml_metal_encoder_set_threadgroup_memory_size(enc, smem, 0);
 
@@ -2048,14 +2850,63 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
 #undef FATTN_SMEM
     } else {
         // half4x4 kernel
-        const int64_t nqptg = 1;  // queries per threadgroup    !! sync with kernel template arguments !!
-        const int64_t ncpsg = 32; // cache values per simdgroup !! sync with kernel template arguments !!
-        const int64_t nkpsg = 1*ncpsg;
+        const int nqptg = OP_FLASH_ATTN_EXT_VEC_NQPSG; // queries per threadgroup
+        const int ncpsg = OP_FLASH_ATTN_EXT_VEC_NCPSG; // cache values per simdgroup !! sync with kernel template arguments !!
+        const int nhptg = 1;                           // heads per threadgroup
 
         GGML_ASSERT(nqptg <= 32);
         GGML_ASSERT(nqptg  % 1  == 0);
         GGML_ASSERT(ncpsg  % 32 == 0);
 
+        bool need_sync = false;
+
+        const bool has_kvpad = ne11 % ncpsg != 0;
+
+        if (has_kvpad) {
+            assert(ggml_metal_op_flash_attn_ext_extra_pad(op) != 0);
+
+            ggml_metal_kargs_flash_attn_ext_pad args0 = {
+                /*.ne11    =*/ne11,
+                /*.ne_12_2 =*/ne12,
+                /*.ne_12_3 =*/ne13,
+                /*.nb11    =*/nb11,
+                /*.nb12    =*/nb12,
+                /*.nb13    =*/nb13,
+                /*.nb21    =*/nb21,
+                /*.nb22    =*/nb22,
+                /*.nb23    =*/nb23,
+                /*.ne31    =*/ne31,
+                /*.ne32    =*/ne32,
+                /*.ne33    =*/ne33,
+                /*.nb31    =*/nb31,
+                /*.nb32    =*/nb32,
+                /*.nb33    =*/nb33,
+            };
+
+            auto pipeline0 = ggml_metal_library_get_pipeline_flash_attn_ext_pad(lib, op, has_mask, ncpsg);
+
+            ggml_metal_encoder_set_pipeline(enc, pipeline0);
+            ggml_metal_encoder_set_bytes   (enc, &args0, sizeof(args0), 0);
+            ggml_metal_encoder_set_buffer  (enc, bid_src1, 1);
+            ggml_metal_encoder_set_buffer  (enc, bid_src2, 2);
+            ggml_metal_encoder_set_buffer  (enc, bid_src3, 3);
+            ggml_metal_encoder_set_buffer  (enc, bid_pad,  4);
+
+            assert(ne12 == ne22);
+            assert(ne13 == ne23);
+
+            ggml_metal_encoder_dispatch_threadgroups(enc, ncpsg, std::max(ne12, ne32), std::max(ne13, ne33), 32, 1, 1);
+
+            need_sync = true;
+        }
+
+        if (need_sync) {
+            ggml_metal_op_concurrency_reset(ctx);
+        }
+
+        // note: for simplicity assume the K is larger or equal than V
+        GGML_ASSERT(ne10 >= ne20);
+
         // ne00 + 2*ncpsg*(nsg)
         // for each query, we load it as f16 in shared memory (ne00)
         // and store the soft_max values and the mask
@@ -2063,28 +2914,9 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
         // ne20*(nsg)
         // each simdgroup has a full f32 head vector in shared mem to accumulate results
         //
-#define FATTN_SMEM(nsg) (GGML_PAD((nqptg*(GGML_PAD(ne00, 128) + 4*ncpsg*(nsg)) + 2*GGML_PAD(ne20, 128)*(nsg))*(sizeof(float)/2), 16))
-
-        int64_t nsgmax = 2;
-        while (true) {
-            const size_t smem = FATTN_SMEM(nsgmax);
-            // avoid using more than half of the threadgroup memory - can cause slow downs especially for large head sizes
-            if (smem > props_dev->max_theadgroup_memory_size/2) {
-                break;
-            }
-            nsgmax *= 2;
-        }
-        nsgmax /= 2;
-
-        // simdgroups per threadgroup (a.k.a. warps)
-        //const int64_t nsgt = MAX(2, MIN(nsgmax, MIN((ne11 + nkpsg - 1)/(nkpsg), (int64_t) pipeline.maxTotalThreadsPerThreadgroup/32)));
-        const int64_t nsgt = MAX(2, MIN(nsgmax, MIN((ne11 + nkpsg - 1)/(nkpsg), (int64_t) 1024/32)));
+#define FATTN_SMEM(nsg) (GGML_PAD(((GGML_PAD(ne00, 128) + 4*ncpsg + 2*GGML_PAD(ne20, 128))*(nsg))*(sizeof(float)/2), 16))
 
         int64_t nsg = 1;
-        while (nsg <= nsgt) {
-            nsg *= 2;
-        }
-        nsg /= 2;
 
         // workgroups
         // each workgroup handles nsg*nkpsg cache values
@@ -2097,7 +2929,7 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
         } else {
             nwg = 32;
             nsg = 1;
-            while (2*nwg*nsg*nkpsg < ne11 && nsg < 4) {
+            while (2*nwg*nsg*ncpsg < ne11 && nsg < 4) {
                 nsg *= 2;
             }
         }
@@ -2120,6 +2952,7 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
             /*.nb21          =*/ nb21,
             /*.nb22          =*/ nb22,
             /*.nb23          =*/ nb23,
+            /*.ne31          =*/ ne31,
             /*.ne32          =*/ ne32,
             /*.ne33          =*/ ne33,
             /*.nb31          =*/ nb31,
@@ -2136,25 +2969,17 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
             /*.logit_softcap =*/ logit_softcap,
         };
 
-        ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_flash_attn_ext_vec(lib, op, has_mask, has_sinks, has_bias, has_scap, nsg, nwg);
+        auto pipeline = ggml_metal_library_get_pipeline_flash_attn_ext_vec(lib, op, has_mask, has_sinks, has_bias, has_scap, has_kvpad, nsg, nwg);
 
         GGML_ASSERT(nsg*32 <= ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));
 
         ggml_metal_encoder_set_pipeline(enc, pipeline);
         ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
-        ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[0]), 1);
-        ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[1]), 2);
-        ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[2]), 3);
-        if (op->src[3]) {
-            ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op->src[3]), 4);
-        } else {
-            ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op->src[0]), 4);
-        }
-        if (op->src[4]) {
-            ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op->src[4]), 5);
-        } else {
-            ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op->src[0]), 5);
-        }
+        ggml_metal_encoder_set_buffer  (enc, bid_src0, 1);
+        ggml_metal_encoder_set_buffer  (enc, bid_src1, 2);
+        ggml_metal_encoder_set_buffer  (enc, bid_src2, 3);
+        ggml_metal_encoder_set_buffer  (enc, bid_src3, 4);
+        ggml_metal_encoder_set_buffer  (enc, bid_src4, 5);
 
         const size_t smem = FATTN_SMEM(nsg);
 
@@ -2162,26 +2987,28 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
         GGML_ASSERT(smem <= props_dev->max_theadgroup_memory_size);
 
         if (nwg == 1) {
+            assert(ggml_metal_op_flash_attn_ext_extra_tmp(op) == 0);
+
             // using 1 workgroup -> write the result directly into dst
-            ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op), 6);
+            ggml_metal_encoder_set_buffer(enc, bid_pad, 6);
+            ggml_metal_encoder_set_buffer(enc, bid_dst, 7);
 
             ggml_metal_encoder_set_threadgroup_memory_size(enc, smem, 0);
 
-            ggml_metal_encoder_dispatch_threadgroups(enc, (ne01 + nqptg - 1)/nqptg, ne02, ne03*nwg, 32, nsg, 1);
+            ggml_metal_encoder_dispatch_threadgroups(enc, (ne01 + nqptg - 1)/nqptg, (ne02 + nhptg - 1)/nhptg, ne03*nwg, 32, nsg, 1);
         } else {
             // sanity checks
+            assert(ggml_metal_op_flash_attn_ext_extra_tmp(op) != 0);
+
             GGML_ASSERT(ne01*ne02*ne03 == ne1*ne2*ne3);
             GGML_ASSERT((uint64_t)ne1*ne2*ne3 <= (1u << 31));
 
-            ggml_metal_buffer_id bid_dst = ggml_metal_get_buffer_id(op);
-
             // write the results from each workgroup into a temp buffer
-            ggml_metal_buffer_id bid_tmp = bid_dst;
-            bid_tmp.offs += ggml_nbytes(op);
-            ggml_metal_encoder_set_buffer(enc, bid_tmp, 6);
+            ggml_metal_encoder_set_buffer(enc, bid_pad, 6);
+            ggml_metal_encoder_set_buffer(enc, bid_tmp, 7);
 
             ggml_metal_encoder_set_threadgroup_memory_size(enc, smem, 0);
-            ggml_metal_encoder_dispatch_threadgroups(enc, (ne01 + nqptg - 1)/nqptg, ne02, ne03*nwg, 32, nsg, 1);
+            ggml_metal_encoder_dispatch_threadgroups(enc, (ne01 + nqptg - 1)/nqptg, (ne02 + nhptg - 1)/nhptg, ne03*nwg, 32, nsg, 1);
 
             // sync the 2 kernels
             ggml_metal_op_concurrency_reset(ctx);
@@ -2194,7 +3021,7 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
                     nrows,
                 };
 
-                ggml_metal_pipeline_t pipeline0 = ggml_metal_library_get_pipeline_flash_attn_ext_vec_reduce(lib, op, ne20, nwg);
+                auto pipeline0 = ggml_metal_library_get_pipeline_flash_attn_ext_vec_reduce(lib, op, ne20, nwg);
 
                 ggml_metal_encoder_set_pipeline(enc, pipeline0);
                 ggml_metal_encoder_set_bytes   (enc, &args0, sizeof(args0), 0);
@@ -2233,8 +3060,6 @@ int ggml_metal_op_bin(ggml_metal_op_t ctx, int idx) {
     GGML_ASSERT(ggml_is_contiguous_rows(op->src[0]));
     GGML_ASSERT(ggml_is_contiguous_rows(op->src[1]));
 
-    bool bcast_row = false;
-
     ggml_metal_buffer_id bid_src0 = ggml_metal_get_buffer_id(op->src[0]);
     ggml_metal_buffer_id bid_src1 = ggml_metal_get_buffer_id(op->src[1]);
     ggml_metal_buffer_id bid_dst  = ggml_metal_get_buffer_id(op);
@@ -2326,20 +3151,9 @@ int ggml_metal_op_bin(ggml_metal_op_t ctx, int idx) {
     // the offsets of src1 and all fused buffers are relative to the start of the src1 buffer
     bid_src1.offs = 0;
 
-    ggml_metal_pipeline_t pipeline = nullptr;
-
-    if (ggml_nelements(op->src[1]) == ne10 && ggml_is_contiguous(op->src[1]) && ne00 % 4 == 0 && ne10 % 4 == 0) {
-        GGML_ASSERT(ggml_is_contiguous(op->src[0]));
+    struct ggml_metal_pipeline_with_params pipeline;
 
-        // src1 is a row
-        GGML_ASSERT(ne11 == 1);
-
-        pipeline = ggml_metal_library_get_pipeline_bin(lib, op->op, n_fuse, true);
-
-        bcast_row = true;
-    } else {
-        pipeline = ggml_metal_library_get_pipeline_bin(lib, op->op, n_fuse, false);
-    }
+    pipeline = ggml_metal_library_get_pipeline_bin(lib, op, n_fuse);
 
     if (n_fuse > 1) {
         bid_dst = ggml_metal_get_buffer_id(ctx->node(idx + n_fuse - 1));
@@ -2353,20 +3167,26 @@ int ggml_metal_op_bin(ggml_metal_op_t ctx, int idx) {
         }
     }
 
+    if (pipeline.c4) {
+        args.ne00 = ne00/4;
+        args.ne10 = ne10/4;
+        args.ne0  = ne0/4;
+    }
+
     ggml_metal_encoder_set_pipeline(enc, pipeline);
     ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
     ggml_metal_encoder_set_buffer  (enc, bid_src0, 1);
     ggml_metal_encoder_set_buffer  (enc, bid_src1, 2);
     ggml_metal_encoder_set_buffer  (enc, bid_dst,  3);
 
-    if (bcast_row) {
-        const int64_t n = ggml_nelements(op)/4;
-
-        ggml_metal_encoder_dispatch_threadgroups(enc, n, 1, 1, 1, 1, 1);
+    if (pipeline.cnt) {
+        ggml_metal_encoder_dispatch_threadgroups(enc, args.ne0, ggml_nrows(op), 1, 1, 1, 1);
     } else {
-        int nth = 32;
+        const int nth_max = MIN(256, ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));
+
+        int nth = 1;
 
-        while (16*nth < ne0 && nth < ggml_metal_pipeline_max_theads_per_threadgroup(pipeline)) {
+        while (2*nth < args.ne0 && nth < nth_max) {
             nth *= 2;
         }
 
@@ -2385,41 +3205,61 @@ int ggml_metal_op_l2_norm(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
+
+    GGML_ASSERT(ggml_is_contiguous_rows(op->src[0]));
+
+    ggml_metal_buffer_id bid_src0 = ggml_metal_get_buffer_id(op->src[0]);
+    ggml_metal_buffer_id bid_dst  = ggml_metal_get_buffer_id(op);
 
     float eps;
     memcpy(&eps, op->op_params, sizeof(float));
 
-    int nth = 32; // SIMD width
-
     ggml_metal_kargs_l2_norm args = {
-        /*.ne00   =*/ ne00,
-        /*.ne00_4 =*/ ne00/4,
-        /*.nb01   =*/ nb01,
-        /*.eps    =*/ eps,
+        /*.ne00  =*/ ne00,
+        /*.ne01  =*/ ne01,
+        /*.ne02  =*/ ne02,
+        /*.ne03  =*/ ne03,
+        /*.nb00  =*/ nb00,
+        /*.nb01  =*/ nb01,
+        /*.nb02  =*/ nb02,
+        /*.nb03  =*/ nb03,
+        /*.ne0   =*/ ne0,
+        /*.ne1   =*/ ne1,
+        /*.ne2   =*/ ne2,
+        /*.ne3   =*/ ne3,
+        /*.nb0   =*/ nb0,
+        /*.nb1   =*/ nb1,
+        /*.nb2   =*/ nb2,
+        /*.nb3   =*/ nb3,
+        /*.eps   =*/ eps,
     };
 
-    ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_l2_norm(lib, op);
+    auto pipeline = ggml_metal_library_get_pipeline_l2_norm(lib, op);
+
+    if (pipeline.c4) {
+        args.ne00 = ne00/4;
+        args.ne0  = ne0/4;
+    }
+
+    int nth = 32; // SIMD width
 
-    while (nth < ne00/4 && nth < ggml_metal_pipeline_max_theads_per_threadgroup(pipeline)) {
+    while (nth < ne00 && nth < ggml_metal_pipeline_max_theads_per_threadgroup(pipeline)) {
         nth *= 2;
     }
 
     nth = std::min(nth, ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));
-    nth = std::min(nth, ne00/4);
 
-    const size_t smem = ggml_metal_pipeline_get_smem(pipeline);
-
-    const int64_t nrows = ggml_nrows(op->src[0]);
+    const size_t smem = pipeline.smem;
 
     ggml_metal_encoder_set_pipeline(enc, pipeline);
     ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
-    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[0]), 1);
-    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op),         2);
+    ggml_metal_encoder_set_buffer  (enc, bid_src0, 1);
+    ggml_metal_encoder_set_buffer  (enc, bid_dst,  2);
 
     ggml_metal_encoder_set_threadgroup_memory_size(enc, smem, 0);
 
-    ggml_metal_encoder_dispatch_threadgroups(enc, nrows, 1, 1, nth, 1, 1);
+    ggml_metal_encoder_dispatch_threadgroups(enc, ne01, ne02, ne03, nth, 1, 1);
 
     return 1;
 }
@@ -2433,7 +3273,7 @@ int ggml_metal_op_group_norm(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     const int32_t ngrp = ((const int32_t *) op->op_params)[0];
 
@@ -2451,7 +3291,7 @@ int ggml_metal_op_group_norm(ggml_metal_op_t ctx, int idx) {
         /*.eps  =*/ eps,
     };
 
-    ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_group_norm(lib, op);
+    auto pipeline = ggml_metal_library_get_pipeline_group_norm(lib, op);
 
     int nth = 32; // SIMD width
     //while (nth < ne00/4 && nth < ggml_metal_pipeline_max_theads_per_threadgroup(pipeline)) {
@@ -2461,7 +3301,7 @@ int ggml_metal_op_group_norm(ggml_metal_op_t ctx, int idx) {
     //nth = std::min(nth, ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));
     //nth = std::min(nth, ne00/4);
 
-    const size_t smem = ggml_metal_pipeline_get_smem(pipeline);
+    const size_t smem = pipeline.smem;
 
     ggml_metal_encoder_set_pipeline(enc, pipeline);
     ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
@@ -2488,7 +3328,7 @@ int ggml_metal_op_norm(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     float eps;
     memcpy(&eps, op->op_params, sizeof(float));
@@ -2586,7 +3426,7 @@ int ggml_metal_op_norm(ggml_metal_op_t ctx, int idx) {
         }
     }
 
-    ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_norm(lib, op, n_fuse);
+    auto pipeline = ggml_metal_library_get_pipeline_norm(lib, op, n_fuse);
 
     int nth = 32; // SIMD width
 
@@ -2597,7 +3437,7 @@ int ggml_metal_op_norm(ggml_metal_op_t ctx, int idx) {
     nth = std::min(nth, ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));
     nth = std::min(nth, args.ne00_t);
 
-    const size_t smem = ggml_metal_pipeline_get_smem(pipeline);
+    const size_t smem = pipeline.smem;
 
     ggml_metal_encoder_set_pipeline(enc, pipeline);
     ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
@@ -2624,7 +3464,7 @@ int ggml_metal_op_rope(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     // make sure we have one or more position id(ne10) per token(ne02)
     GGML_ASSERT(ne10 % ne02 == 0);
@@ -2688,9 +3528,10 @@ int ggml_metal_op_rope(ggml_metal_op_t ctx, int idx) {
         /* sect_1      =*/ sect_1,
         /* sect_2      =*/ sect_2,
         /* sect_3      =*/ sect_3,
+        /* src2        =*/ op->src[2] != nullptr,
     };
 
-    ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_rope(lib, op);
+    auto pipeline = ggml_metal_library_get_pipeline_rope(lib, op);
 
     ggml_metal_encoder_set_pipeline(enc, pipeline);
     ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
@@ -2717,7 +3558,7 @@ int ggml_metal_op_im2col(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     const int32_t s0 = ((const int32_t *)(op->op_params))[0];
     const int32_t s1 = ((const int32_t *)(op->op_params))[1];
@@ -2762,7 +3603,7 @@ int ggml_metal_op_im2col(ggml_metal_op_t ctx, int idx) {
         /*.KHW  =*/ KH * KW,
     };
 
-    ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_im2col(lib, op);
+    auto pipeline = ggml_metal_library_get_pipeline_im2col(lib, op);
 
     GGML_ASSERT(KH*KW <= ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));
 
@@ -2770,15 +3611,138 @@ int ggml_metal_op_im2col(ggml_metal_op_t ctx, int idx) {
 
     ggml_metal_encoder_set_pipeline(enc, pipeline);
     ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
-    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[1]), 1);
-    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op),         2);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[1]), 1);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op),         2);
+
+    ggml_metal_encoder_dispatch_threadgroups(enc, IC, OH, OW, ntptg0, KH, KW);
+
+    return 1;
+}
+
+int ggml_metal_op_conv_2d(ggml_metal_op_t ctx, int idx) {
+    ggml_tensor * op = ctx->node(idx);
+
+    ggml_metal_library_t lib = ctx->lib;
+    ggml_metal_encoder_t enc = ctx->enc;
+
+    GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
+    GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
+
+    GGML_ASSERT(ggml_is_contiguous(op->src[0]));
+    GGML_ASSERT(op->src[1]->type == GGML_TYPE_F32);
+    GGML_ASSERT(op->type == GGML_TYPE_F32);
+    GGML_ASSERT(op->src[0]->type == GGML_TYPE_F16 || op->src[0]->type == GGML_TYPE_F32);
+
+    const int32_t s0 = ((const int32_t *) op->op_params)[0];
+    const int32_t s1 = ((const int32_t *) op->op_params)[1];
+    const int32_t p0 = ((const int32_t *) op->op_params)[2];
+    const int32_t p1 = ((const int32_t *) op->op_params)[3];
+    const int32_t d0 = ((const int32_t *) op->op_params)[4];
+    const int32_t d1 = ((const int32_t *) op->op_params)[5];
+
+    ggml_metal_kargs_conv_2d args = {
+        /*.nb00 =*/ nb00,
+        /*.nb01 =*/ nb01,
+        /*.nb02 =*/ nb02,
+        /*.nb03 =*/ nb03,
+        /*.nb10 =*/ nb10,
+        /*.nb11 =*/ nb11,
+        /*.nb12 =*/ nb12,
+        /*.nb13 =*/ nb13,
+        /*.nb0  =*/ nb0,
+        /*.nb1  =*/ nb1,
+        /*.nb2  =*/ nb2,
+        /*.nb3  =*/ nb3,
+        /*.IW   =*/ ne10,
+        /*.IH   =*/ ne11,
+        /*.KW   =*/ ne00,
+        /*.KH   =*/ ne01,
+        /*.IC   =*/ ne02,
+        /*.OC   =*/ ne03,
+        /*.OW   =*/ ne0,
+        /*.OH   =*/ ne1,
+        /*.N    =*/ ne3,
+        /*.s0   =*/ s0,
+        /*.s1   =*/ s1,
+        /*.p0   =*/ p0,
+        /*.p1   =*/ p1,
+        /*.d0   =*/ d0,
+        /*.d1   =*/ d1,
+    };
+
+    auto pipeline = ggml_metal_library_get_pipeline_conv_2d(lib, op);
+
+    int nth = ggml_metal_pipeline_max_theads_per_threadgroup(pipeline);
+    nth = std::min(nth, 256);
+    nth = std::max(nth, 1);
+
+    const uint64_t n_out = ggml_nelements(op);
+
+    uint64_t tg = (n_out + nth - 1)/nth;
+    tg = std::max<uint64_t>(tg, 1);
+    tg = std::min<uint64_t>(tg, (uint64_t) std::numeric_limits<int>::max());
+
+    ggml_metal_encoder_set_pipeline(enc, pipeline);
+    ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[0]), 1);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[1]), 2);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op),         3);
+
+    ggml_metal_encoder_dispatch_threadgroups(enc, tg, 1, 1, nth, 1, 1);
+
+    return 1;
+}
+
+int ggml_metal_op_conv_transpose_1d(ggml_metal_op_t ctx, int idx) {
+    ggml_tensor * op = ctx->node(idx);
+
+    ggml_metal_library_t lib = ctx->lib;
+    ggml_metal_encoder_t enc = ctx->enc;
+
+    GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
+    GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
+
+    const int32_t s0 = ((const int32_t *)(op->op_params))[0];
+
+    const int32_t IC = op->src[1]->ne[1];
+    const int32_t IL = op->src[1]->ne[0];
+
+    const int32_t K  = op->src[0]->ne[0];
+
+    const int32_t OL = op->ne[0];
+    const int32_t OC = op->ne[1];
+
+    ggml_metal_kargs_conv_transpose_1d args = {
+        /*.IC  =*/ IC,
+        /*.IL  =*/ IL,
+        /*.K   =*/ K,
+        /*.s0  =*/ s0,
+        /*.nb0 =*/ nb0,
+        /*.nb1 =*/ nb1,
+    };
+
+    auto pipeline = ggml_metal_library_get_pipeline_conv_transpose_1d(lib, op);
+
+    ggml_metal_encoder_set_pipeline(enc, pipeline);
+    ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[0]), 1);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[1]), 2);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op),         3);
 
-    ggml_metal_encoder_dispatch_threadgroups(enc, IC, OH, OW, ntptg0, KH, KW);
+    ggml_metal_encoder_dispatch_threadgroups(enc, OL, OC, 1, 1, 1, 1);
 
     return 1;
 }
 
-int ggml_metal_op_conv_transpose_1d(ggml_metal_op_t ctx, int idx) {
+int ggml_metal_op_conv_transpose_2d(ggml_metal_op_t ctx, int idx) {
     ggml_tensor * op = ctx->node(idx);
 
     ggml_metal_library_t lib = ctx->lib;
@@ -2789,28 +3753,35 @@ int ggml_metal_op_conv_transpose_1d(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     const int32_t s0 = ((const int32_t *)(op->op_params))[0];
 
-    const int32_t IC = op->src[1]->ne[1];
-    const int32_t IL = op->src[1]->ne[0];
+    const int32_t IC = op->src[1]->ne[2];
+    const int32_t IH = op->src[1]->ne[1];
+    const int32_t IW = op->src[1]->ne[0];
 
-    const int32_t K  = op->src[0]->ne[0];
+    const int32_t KH = op->src[0]->ne[1];
+    const int32_t KW = op->src[0]->ne[0];
 
-    const int32_t OL = op->ne[0];
-    const int32_t OC = op->ne[1];
+    const int32_t OW = op->ne[0];
+    const int32_t OH = op->ne[1];
+    const int32_t OC = op->ne[2];
 
-    ggml_metal_kargs_conv_transpose_1d args = {
+    ggml_metal_kargs_conv_transpose_2d args = {
         /*.IC  =*/ IC,
-        /*.IL  =*/ IL,
-        /*.K   =*/ K,
+        /*.IH  =*/ IH,
+        /*.IW  =*/ IW,
+        /*.KH  =*/ KH,
+        /*.KW  =*/ KW,
+        /*.OC  =*/ OC,
         /*.s0  =*/ s0,
         /*.nb0 =*/ nb0,
         /*.nb1 =*/ nb1,
+        /*.nb2 =*/ nb2,
     };
 
-    ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_conv_transpose_1d(lib, op);
+    auto pipeline = ggml_metal_library_get_pipeline_conv_transpose_2d(lib, op);
 
     ggml_metal_encoder_set_pipeline(enc, pipeline);
     ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
@@ -2818,7 +3789,11 @@ int ggml_metal_op_conv_transpose_1d(ggml_metal_op_t ctx, int idx) {
     ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[1]), 2);
     ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op),         3);
 
-    ggml_metal_encoder_dispatch_threadgroups(enc, OL, OC, 1, 1, 1, 1);
+    // Metal requires buffer size to be multiple of 16 bytes
+    const size_t smem = GGML_PAD(KW * KH * sizeof(float), 16);
+    ggml_metal_encoder_set_threadgroup_memory_size(enc, smem, 0);
+
+    ggml_metal_encoder_dispatch_threadgroups(enc, OW, OH, OC, KW, KH, 1);
 
     return 1;
 }
@@ -2832,37 +3807,48 @@ int ggml_metal_op_upscale(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
+
+    float sf0 = (float)ne0/op->src[0]->ne[0];
+    float sf1 = (float)ne1/op->src[0]->ne[1];
+    float sf2 = (float)ne2/op->src[0]->ne[2];
+    float sf3 = (float)ne3/op->src[0]->ne[3];
+
+    const int32_t mode_flags = ggml_get_op_params_i32(op, 0);
 
-    const float sf0 = (float)ne0/op->src[0]->ne[0];
-    const float sf1 = (float)ne1/op->src[0]->ne[1];
-    const float sf2 = (float)ne2/op->src[0]->ne[2];
-    const float sf3 = (float)ne3/op->src[0]->ne[3];
+    float poffs = 0.5f;
+
+    if (mode_flags & GGML_SCALE_FLAG_ALIGN_CORNERS) {
+        poffs = 0.0f;
+        sf0 = ne0 > 1 && ne00 > 1 ? (float)(ne0 - 1) / (ne00 - 1) : sf0;
+        sf1 = ne1 > 1 && ne01 > 1 ? (float)(ne1 - 1) / (ne01 - 1) : sf1;
+    }
 
     ggml_metal_kargs_upscale args = {
-        /*.ne00 =*/ ne00,
-        /*.ne01 =*/ ne01,
-        /*.ne02 =*/ ne02,
-        /*.ne03 =*/ ne03,
-        /*.nb00 =*/ nb00,
-        /*.nb01 =*/ nb01,
-        /*.nb02 =*/ nb02,
-        /*.nb03 =*/ nb03,
-        /*.ne0 =*/ ne0,
-        /*.ne1 =*/ ne1,
-        /*.ne2 =*/ ne2,
-        /*.ne3 =*/ ne3,
-        /*.nb0 =*/ nb0,
-        /*.nb1 =*/ nb1,
-        /*.nb2 =*/ nb2,
-        /*.nb3 =*/ nb3,
-        /*.sf0 =*/ sf0,
-        /*.sf1 =*/ sf1,
-        /*.sf2 =*/ sf2,
-        /*.sf3 =*/ sf3
+        /*.ne00  =*/ ne00,
+        /*.ne01  =*/ ne01,
+        /*.ne02  =*/ ne02,
+        /*.ne03  =*/ ne03,
+        /*.nb00  =*/ nb00,
+        /*.nb01  =*/ nb01,
+        /*.nb02  =*/ nb02,
+        /*.nb03  =*/ nb03,
+        /*.ne0   =*/ ne0,
+        /*.ne1   =*/ ne1,
+        /*.ne2   =*/ ne2,
+        /*.ne3   =*/ ne3,
+        /*.nb0   =*/ nb0,
+        /*.nb1   =*/ nb1,
+        /*.nb2   =*/ nb2,
+        /*.nb3   =*/ nb3,
+        /*.sf0   =*/ sf0,
+        /*.sf1   =*/ sf1,
+        /*.sf2   =*/ sf2,
+        /*.sf3   =*/ sf3,
+        /*.poffs =*/ poffs,
     };
 
-    ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_upscale(lib, op);
+    auto pipeline = ggml_metal_library_get_pipeline_upscale(lib, op);
 
     const int nth = std::min(ggml_metal_pipeline_max_theads_per_threadgroup(pipeline), ne0);
 
@@ -2885,7 +3871,7 @@ int ggml_metal_op_pad(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     ggml_metal_kargs_pad args = {
         /*.ne00 =*/ ne00,
@@ -2906,7 +3892,7 @@ int ggml_metal_op_pad(ggml_metal_op_t ctx, int idx) {
         /*.nb3  =*/ nb3
     };
 
-    ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_pad(lib, op);
+    auto pipeline = ggml_metal_library_get_pipeline_pad(lib, op);
 
     const int nth = std::min(1024, ne0);
 
@@ -2929,7 +3915,7 @@ int ggml_metal_op_pad_reflect_1d(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     ggml_metal_kargs_pad_reflect_1d args = {
         /*.ne00 =*/ ne00,
@@ -2952,7 +3938,7 @@ int ggml_metal_op_pad_reflect_1d(ggml_metal_op_t ctx, int idx) {
         /*.p1 =*/ ((const int32_t *)(op->op_params))[1]
     };
 
-    ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_pad_reflect_1d(lib, op);
+    auto pipeline = ggml_metal_library_get_pipeline_pad_reflect_1d(lib, op);
 
     const int nth = std::min(1024, ne0);
 
@@ -2973,7 +3959,7 @@ int ggml_metal_op_arange(ggml_metal_op_t ctx, int idx) {
     ggml_metal_encoder_t enc = ctx->enc;
 
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     float start;
     float step;
@@ -2989,13 +3975,7 @@ int ggml_metal_op_arange(ggml_metal_op_t ctx, int idx) {
 
     const int nth = std::min(1024, ne0);
 
-    ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_arange(lib, op);
-
-    //[encoder setComputePipelineState:pipeline];
-    //[encoder setBuffer:id_dst  offset:offs_dst  atIndex:0];
-    //[encoder setBytes:&args length:sizeof(args) atIndex:1];
-
-    //[encoder dispatchThreadgroups:MTLSizeMake(1, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+    auto pipeline = ggml_metal_library_get_pipeline_arange(lib, op);
 
     ggml_metal_encoder_set_pipeline(enc, pipeline);
     ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
@@ -3015,7 +3995,7 @@ int ggml_metal_op_timestep_embedding(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     const int dim        = op->op_params[0];
     const int max_period = op->op_params[1];
@@ -3026,7 +4006,7 @@ int ggml_metal_op_timestep_embedding(ggml_metal_op_t ctx, int idx) {
         /*.max_period =*/ max_period,
     };
 
-    ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_timestep_embedding(lib, op);
+    auto pipeline = ggml_metal_library_get_pipeline_timestep_embedding(lib, op);
 
     const int nth = std::max(1, std::min(1024, dim/2));
 
@@ -3049,14 +4029,14 @@ int ggml_metal_op_argmax(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     ggml_metal_kargs_argmax args = {
         /*.ne00 = */ ne00,
         /*.nb01 = */ nb01,
     };
 
-    ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_argmax(lib, op);
+    auto pipeline = ggml_metal_library_get_pipeline_argmax(lib, op);
 
     const int64_t nrows = ggml_nrows(op->src[0]);
 
@@ -3065,7 +4045,7 @@ int ggml_metal_op_argmax(ggml_metal_op_t ctx, int idx) {
         nth *= 2;
     }
 
-    const size_t smem = ggml_metal_pipeline_get_smem(pipeline);
+    const size_t smem = pipeline.smem;
 
     ggml_metal_encoder_set_pipeline(enc, pipeline);
     ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
@@ -3085,74 +4065,397 @@ int ggml_metal_op_argsort(ggml_metal_op_t ctx, int idx) {
     ggml_metal_library_t lib = ctx->lib;
     ggml_metal_encoder_t enc = ctx->enc;
 
+    GGML_ASSERT(ggml_is_contiguous_rows(op->src[0]));
+
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
+
+    auto pipeline = ggml_metal_library_get_pipeline_argsort(lib, op);
 
     // bitonic sort requires the number of elements to be power of 2
-    int64_t ne00_padded = 1;
-    while (ne00_padded < ne00) {
-        ne00_padded *= 2;
+    int nth = 1;
+    while (nth < ne00 && 2*nth <= ggml_metal_pipeline_max_theads_per_threadgroup(pipeline)) {
+        nth *= 2;
     }
 
-    ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_argsort(lib, op);
-
-    const int64_t nrows = ggml_nrows(op->src[0]);
+    const int npr = (ne00 + nth - 1)/nth;
 
     // Metal kernels require the buffer size to be multiple of 16 bytes
     // https://developer.apple.com/documentation/metal/mtlcomputecommandencoder/1443142-setthreadgroupmemorylength
-    const size_t smem = GGML_PAD(ne00_padded*sizeof(int32_t), 16);
+    const size_t smem = GGML_PAD(nth*sizeof(int32_t), 16);
+
+    ggml_metal_buffer_id bid_src0 = ggml_metal_get_buffer_id(op->src[0]);
+    ggml_metal_buffer_id bid_dst  = ggml_metal_get_buffer_id(op);
+
+    ggml_metal_buffer_id bid_tmp = bid_dst;
+    bid_tmp.offs += ggml_nbytes(op);
+
+    if ((int) ceil(std::log(npr) / std::log(2)) % 2 == 1) {
+        std::swap(bid_dst, bid_tmp);
+    }
 
     ggml_metal_kargs_argsort args = {
-        /*.ncols =*/ ne00,
-        /*.ncols_pad =*/ ne00_padded
+        /*.ne00  =*/ ne00,
+        /*.ne01  =*/ ne01,
+        /*.ne02  =*/ ne02,
+        /*.ne03  =*/ ne03,
+        /*.nb00  =*/ nb00,
+        /*.nb01  =*/ nb01,
+        /*.nb02  =*/ nb02,
+        /*.nb03  =*/ nb03,
+        /*.ne0   =*/ ne0,
+        /*.ne1   =*/ ne1,
+        /*.ne2   =*/ ne2,
+        /*.ne3   =*/ ne3,
+        /*.top_k =*/ nth,
     };
 
     ggml_metal_encoder_set_pipeline(enc, pipeline);
     ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
-    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[0]), 1);
-    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op),         2);
+    ggml_metal_encoder_set_buffer  (enc, bid_src0, 1);
+    ggml_metal_encoder_set_buffer  (enc, bid_dst,  2);
 
     ggml_metal_encoder_set_threadgroup_memory_size(enc, smem, 0);
 
-    ggml_metal_encoder_dispatch_threadgroups(enc, 1, nrows, 1, ne00_padded, 1, 1);
+    ggml_metal_encoder_dispatch_threadgroups(enc, npr*ne01, ne02, ne03, nth, 1, 1);
+
+    auto pipeline_merge = ggml_metal_library_get_pipeline_argsort_merge(lib, op);
+
+    int len = nth;
+
+    while (len < ne00) {
+        ggml_metal_op_concurrency_reset(ctx);
+
+        ggml_metal_kargs_argsort_merge args_merge = {
+            /*.ne00  =*/ ne00,
+            /*.ne01  =*/ ne01,
+            /*.ne02  =*/ ne02,
+            /*.ne03  =*/ ne03,
+            /*.nb00  =*/ nb00,
+            /*.nb01  =*/ nb01,
+            /*.nb02  =*/ nb02,
+            /*.nb03  =*/ nb03,
+            /*.ne0   =*/ ne0,
+            /*.ne1   =*/ ne1,
+            /*.ne2   =*/ ne2,
+            /*.ne3   =*/ ne3,
+            /*.top_k =*/ ne00,
+            /*.len   =*/ len,
+        };
+
+        // merges per row
+        const int nm = (ne00 + 2*len - 1) / (2*len);
+
+        const int nth = std::min(512, ggml_metal_pipeline_max_theads_per_threadgroup(pipeline_merge));
+
+        ggml_metal_encoder_set_pipeline(enc, pipeline_merge);
+        ggml_metal_encoder_set_bytes   (enc, &args_merge, sizeof(args_merge), 0);
+        ggml_metal_encoder_set_buffer  (enc, bid_src0, 1);
+        ggml_metal_encoder_set_buffer  (enc, bid_dst,  2);
+        ggml_metal_encoder_set_buffer  (enc, bid_tmp,  3);
+
+        ggml_metal_encoder_dispatch_threadgroups(enc, nm*ne01, ne02, ne03, nth, 1, 1);
+
+        std::swap(bid_dst, bid_tmp);
+
+        len <<= 1;
+    }
 
     return 1;
 }
 
-int ggml_metal_op_leaky_relu(ggml_metal_op_t ctx, int idx) {
+int ggml_metal_op_top_k(ggml_metal_op_t ctx, int idx) {
     ggml_tensor * op = ctx->node(idx);
 
     ggml_metal_library_t lib = ctx->lib;
     ggml_metal_encoder_t enc = ctx->enc;
 
+    GGML_ASSERT(ggml_is_contiguous_rows(op->src[0]));
+
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
+
+    auto pipeline = ggml_metal_library_get_pipeline_top_k(lib, op);
+
+    // bitonic sort requires the number of elements to be power of 2
+    int nth = 1;
+    while (nth < ne00 && 2*nth <= ggml_metal_pipeline_max_theads_per_threadgroup(pipeline)) {
+        nth *= 2;
+    }
+
+    // blocks per row
+    const int npr = (ne00 + nth - 1)/nth;
+
+    const size_t smem = GGML_PAD(nth*sizeof(int32_t), 16);
+
+    ggml_metal_buffer_id bid_src0 = ggml_metal_get_buffer_id(op->src[0]);
+    ggml_metal_buffer_id bid_dst  = ggml_metal_get_buffer_id(op);
+
+    ggml_metal_buffer_id bid_tmp = bid_dst;
+    bid_tmp.offs += sizeof(int32_t)*ggml_nelements(op->src[0]);
+
+    if ((int) ceil(std::log(npr) / std::log(2)) % 2 == 1) {
+        std::swap(bid_dst, bid_tmp);
+    }
+
+    const int top_k = ne0;
+
+    ggml_metal_kargs_argsort args = {
+        /*.ne00  =*/ ne00,
+        /*.ne01  =*/ ne01,
+        /*.ne02  =*/ ne02,
+        /*.ne03  =*/ ne03,
+        /*.nb00  =*/ nb00,
+        /*.nb01  =*/ nb01,
+        /*.nb02  =*/ nb02,
+        /*.nb03  =*/ nb03,
+        /*.ne0   =*/ ne0,
+        /*.ne1   =*/ ne1,
+        /*.ne2   =*/ ne2,
+        /*.ne3   =*/ ne3,
+        /*.top_k =*/ std::min(nth, top_k), // for each block, keep just the top_k indices
+    };
+
+    if (npr > 1) {
+        args.ne0 = (npr - 1)*args.top_k + std::min(ne00 - (npr - 1)*nth, args.top_k);
+    }
+
+    ggml_metal_encoder_set_pipeline(enc, pipeline);
+    ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+    ggml_metal_encoder_set_buffer  (enc, bid_src0, 1);
+    ggml_metal_encoder_set_buffer  (enc, bid_dst,  2);
+
+    ggml_metal_encoder_set_threadgroup_memory_size(enc, smem, 0);
+
+    ggml_metal_encoder_dispatch_threadgroups(enc, npr*ne01, ne02, ne03, nth, 1, 1);
+
+    auto pipeline_merge = ggml_metal_library_get_pipeline_top_k_merge(lib, op);
+
+    int len = args.top_k;
+
+    while (len < args.ne0) {
+        ggml_metal_op_concurrency_reset(ctx);
+
+        // merges per row
+        const int nm = (args.ne0 + 2*len - 1) / (2*len);
+
+        const int nth = std::min(512, std::min(len, ggml_metal_pipeline_max_theads_per_threadgroup(pipeline_merge)));
+
+        ggml_metal_kargs_argsort_merge args_merge = {
+            /*.ne00  =*/ ne00,
+            /*.ne01  =*/ ne01,
+            /*.ne02  =*/ ne02,
+            /*.ne03  =*/ ne03,
+            /*.nb00  =*/ nb00,
+            /*.nb01  =*/ nb01,
+            /*.nb02  =*/ nb02,
+            /*.nb03  =*/ nb03,
+            /*.ne0   =*/ args.ne0,
+            /*.ne1   =*/ ne1,
+            /*.ne2   =*/ ne2,
+            /*.ne3   =*/ ne3,
+            /*.top_k =*/ nm == 1 ? top_k : args.ne0, // the final merge outputs top_k elements
+            /*.len   =*/ len,
+        };
+
+        ggml_metal_encoder_set_pipeline(enc, pipeline_merge);
+        ggml_metal_encoder_set_bytes   (enc, &args_merge, sizeof(args_merge), 0);
+        ggml_metal_encoder_set_buffer  (enc, bid_src0, 1);
+        ggml_metal_encoder_set_buffer  (enc, bid_dst,  2);
+        ggml_metal_encoder_set_buffer  (enc, bid_tmp,  3);
+
+        ggml_metal_encoder_dispatch_threadgroups(enc, nm*ne01, ne02, ne03, nth, 1, 1);
+
+        std::swap(bid_dst, bid_tmp);
+
+        len <<= 1;
+    }
+
+    return 1;
+}
+
+int ggml_metal_op_tri(ggml_metal_op_t ctx, int idx) {
+    ggml_tensor * op = ctx->node(idx);
 
-    float slope;
-    memcpy(&slope, op->op_params, sizeof(float));
+    ggml_metal_library_t lib = ctx->lib;
+    ggml_metal_encoder_t enc = ctx->enc;
+
+    GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
-    ggml_metal_kargs_leaky_relu args = {
-        /*.slope =*/ slope
+    ggml_metal_kargs_tri args = {
+        /*.ne00  =*/ ne00,
+        /*.ne01  =*/ ne01,
+        /*.ne02  =*/ ne02,
+        /*.ne03  =*/ ne03,
+        /*.nb00  =*/ nb00,
+        /*.nb01  =*/ nb01,
+        /*.nb02  =*/ nb02,
+        /*.nb03  =*/ nb03,
+        /*.ne0   =*/ ne0,
+        /*.ne1   =*/ ne1,
+        /*.ne2   =*/ ne2,
+        /*.ne3   =*/ ne3,
+        /*.nb0   =*/ nb0,
+        /*.nb1   =*/ nb1,
+        /*.nb2   =*/ nb2,
+        /*.nb3   =*/ nb3,
     };
 
-    ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_unary(lib, op);
+    auto pipeline = ggml_metal_library_get_pipeline_tri(lib, op);
 
-    int64_t n = ggml_nelements(op);
+    int nth = 32; // SIMD width
 
-    if (n % 4 == 0) {
-        n /= 4;
+    while (nth < ne00 && nth < ggml_metal_pipeline_max_theads_per_threadgroup(pipeline)) {
+        nth *= 2;
     }
 
+    nth = std::min(nth, ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));
+    nth = std::min(nth, ne00);
+
     ggml_metal_encoder_set_pipeline(enc, pipeline);
     ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
     ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[0]), 1);
     ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op),         2);
 
-    ggml_metal_encoder_dispatch_threadgroups(enc, n, 1, 1, 1, 1, 1);
+    ggml_metal_encoder_dispatch_threadgroups(enc, ne01, ne02, ne03, nth, 1, 1);
+
+    return 1;
+}
+
+int ggml_metal_op_opt_step_adamw(ggml_metal_op_t ctx, int idx) {
+    ggml_tensor * op = ctx->node(idx);
+
+    ggml_metal_library_t lib = ctx->lib;
+    ggml_metal_encoder_t enc = ctx->enc;
+
+    GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
+
+    auto pipeline = ggml_metal_library_get_pipeline_opt_step_adamw(lib, op);
+
+    const int64_t np = ggml_nelements(op->src[0]);
+    ggml_metal_kargs_opt_step_adamw args = {
+        /*.np =*/ np,
+    };
+
+    int ida = 0;
+
+    ggml_metal_encoder_set_pipeline(enc, pipeline);
+    ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), ida++);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[0]), ida++);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[1]), ida++);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[2]), ida++);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[3]), ida++);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[4]), ida++);
+
+    const int nth = std::min(ggml_metal_pipeline_max_theads_per_threadgroup(pipeline), ne0);
+    const int64_t n = (np + nth - 1) / nth;
+
+    ggml_metal_encoder_dispatch_threadgroups(enc, n, 1, 1, nth, 1, 1);
+
+    return 1;
+}
+
+int ggml_metal_op_opt_step_sgd(ggml_metal_op_t ctx, int idx) {
+    ggml_tensor * op = ctx->node(idx);
+
+    ggml_metal_library_t lib = ctx->lib;
+    ggml_metal_encoder_t enc = ctx->enc;
+
+    GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
+
+    auto pipeline = ggml_metal_library_get_pipeline_opt_step_sgd(lib, op);
+
+    const int64_t np = ggml_nelements(op->src[0]);
+    ggml_metal_kargs_opt_step_sgd args = {
+        /*.np =*/ np,
+    };
+
+    int ida = 0;
+
+    ggml_metal_encoder_set_pipeline(enc, pipeline);
+    ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), ida++);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[0]), ida++);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[1]), ida++);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[2]), ida++);
+
+    const int nth = std::min(ggml_metal_pipeline_max_theads_per_threadgroup(pipeline), ne0);
+    const int64_t n = (np + nth - 1) / nth;
+
+    ggml_metal_encoder_dispatch_threadgroups(enc, n, 1, 1, nth, 1, 1);
+
+    return 1;
+}
+
+int ggml_metal_op_count_equal(ggml_metal_op_t ctx, int idx) {
+    ggml_tensor * op = ctx->node(idx);
+
+    ggml_metal_library_t lib = ctx->lib;
+    ggml_metal_encoder_t enc = ctx->enc;
+
+    GGML_TENSOR_LOCALS(int32_t,  ne0, op->src[0], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
+
+    {
+        ggml_metal_kargs_memset args = { /*.val =*/ 0 };
+
+        auto pipeline = ggml_metal_library_get_pipeline_memset(lib, op);
+
+        ggml_metal_encoder_set_pipeline(enc, pipeline);
+        ggml_metal_encoder_set_bytes(enc, &args, sizeof(args), 0);
+        ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op), 1);
+
+        ggml_metal_encoder_dispatch_threadgroups(enc, 1, 1, 1, 1, 1, 1);
+    }
+
+    ggml_metal_op_concurrency_reset(ctx);
+
+    {
+        ggml_metal_kargs_count_equal args = {
+            /*.ne00 =*/ ne00,
+            /*.ne01 =*/ ne01,
+            /*.ne02 =*/ ne02,
+            /*.ne03 =*/ ne03,
+            /*.nb00 =*/ nb00,
+            /*.nb01 =*/ nb01,
+            /*.nb02 =*/ nb02,
+            /*.nb03 =*/ nb03,
+            /*.nb10 =*/ nb10,
+            /*.nb11 =*/ nb11,
+            /*.nb12 =*/ nb12,
+            /*.nb13 =*/ nb13,
+        };
+
+        auto pipeline = ggml_metal_library_get_pipeline_count_equal(lib, op);
+
+        const size_t smem = pipeline.smem;
+
+        const int nth = 32*pipeline.nsg;
+
+        GGML_ASSERT(nth <= ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));
+
+        ggml_metal_encoder_set_pipeline(enc, pipeline);
+        ggml_metal_encoder_set_bytes(enc, &args, sizeof(args), 0);
+        ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op->src[0]), 1);
+        ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op->src[1]), 2);
+        ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op), 3);
+
+        ggml_metal_encoder_set_threadgroup_memory_size(enc, smem, 0);
+        ggml_metal_encoder_dispatch_threadgroups(enc, ne01, ne02, ne03, nth, 1, 1);
+    }
 
     return 1;
 }