whisper.rn 0.5.0 → 0.5.2

package/cpp/ggml-metal/ggml-metal-ops.cpp

@@ -0,0 +1,3583 @@
+#include "ggml-metal-ops.h"
+
+#include "ggml.h"
+#include "ggml-impl.h"
+#include "ggml-backend-impl.h"
+
+#include "ggml-metal-impl.h"
+#include "ggml-metal-common.h"
+#include "ggml-metal-device.h"
+
+#include <cassert>
+#include <algorithm>
+
+static wsp_ggml_metal_buffer_id wsp_ggml_metal_get_buffer_id(const wsp_ggml_tensor * t) {
+    if (!t) {
+        return { nullptr, 0 };
+    }
+
+    wsp_ggml_backend_buffer_t buffer = t->view_src ? t->view_src->buffer : t->buffer;
+
+    wsp_ggml_metal_buffer_t ctx = (wsp_ggml_metal_buffer_t) buffer->context;
+
+    return wsp_ggml_metal_buffer_get_id(ctx, t);
+}
+
+struct wsp_ggml_metal_op {
+    wsp_ggml_metal_op(
+        wsp_ggml_metal_device_t dev,
+        wsp_ggml_metal_cmd_buf_t cmd_buf,
+        wsp_ggml_cgraph * gf,
+        int  idx_start,
+        int  idx_end,
+        bool use_fusion,
+        bool use_concurrency,
+        bool use_capture,
+        int  debug_graph,
+        int  debug_fusion) {
+        this->dev             = dev;
+        this->lib             = wsp_ggml_metal_device_get_library(dev);
+        this->enc             = wsp_ggml_metal_encoder_init(cmd_buf, use_concurrency);
+        this->mem_ranges      = wsp_ggml_mem_ranges_init(debug_graph);
+        this->idx_start       = idx_start;
+        this->idx_end         = idx_end;
+        this->use_fusion      = use_fusion;
+        this->use_concurrency = use_concurrency;
+        this->use_capture     = use_capture;
+        this->debug_graph     = debug_graph;
+        this->debug_fusion    = debug_fusion;
+        this->gf              = gf;
+
+        idxs.reserve(gf->n_nodes);
+
+        // filter empty nodes
+        // TODO: this can be removed when the allocator starts filtering them earlier
+        //       https://github.com/ggml-org/llama.cpp/pull/16130#issuecomment-3327905830
+        for (int i = idx_start; i < idx_end; i++) {
+            if (!wsp_ggml_op_is_empty(gf->nodes[i]->op) && !wsp_ggml_is_empty(gf->nodes[i])) {
+                idxs.push_back(i);
+            }
+        }
+    }
+
+    ~wsp_ggml_metal_op() {
+        wsp_ggml_metal_encoder_end_encoding(this->enc);
+        wsp_ggml_metal_encoder_free(this->enc);
+        wsp_ggml_mem_ranges_free(this->mem_ranges);
+    }
+
+    int n_nodes() const {
+        return idxs.size();
+    }
+
+    wsp_ggml_tensor * node(int i) const {
+        assert(i >= 0 && i < (int) idxs.size());
+        return wsp_ggml_graph_node(gf, idxs[i]);
+    }
+
+    bool can_fuse(int i0, const wsp_ggml_op * ops, int n_ops) const {
+        assert(use_fusion);
+        assert(i0 >= 0 && i0 < n_nodes());
+
+        if (i0 + n_ops > n_nodes()) {
+            return false;
+        }
+
+        return wsp_ggml_can_fuse_ext(gf, idxs.data() + i0, ops, n_ops);
+    }
+
+    wsp_ggml_metal_device_t  dev;
+    wsp_ggml_metal_library_t lib;
+    wsp_ggml_metal_encoder_t enc;
+    wsp_ggml_mem_ranges_t    mem_ranges;
+
+    bool use_fusion;
+    bool use_concurrency;
+    bool use_capture;
+
+    int debug_graph;
+    int debug_fusion;
+
+private:
+    wsp_ggml_cgraph * gf;
+
+    int idx_start;
+    int idx_end;
+
+    // non-empty node indices
+    std::vector<int> idxs;
+};
+
+wsp_ggml_metal_op_t wsp_ggml_metal_op_init(
+        wsp_ggml_metal_device_t dev,
+        wsp_ggml_metal_cmd_buf_t cmd_buf,
+        wsp_ggml_cgraph * gf,
+        int idx_start,
+        int idx_end,
+        bool use_fusion,
+        bool use_concurrency,
+        bool use_capture,
+        int debug_graph,
+        int debug_fusion) {
+    wsp_ggml_metal_op_t res = new wsp_ggml_metal_op(
+        dev,
+        cmd_buf,
+        gf,
+        idx_start,
+        idx_end,
+        use_fusion,
+        use_concurrency,
+        use_capture,
+        debug_graph,
+        debug_fusion);
+
+    return res;
+}
+
+void wsp_ggml_metal_op_free(wsp_ggml_metal_op_t ctx) {
+    delete ctx;
+}
+
+int wsp_ggml_metal_op_n_nodes(wsp_ggml_metal_op_t ctx) {
+    return ctx->n_nodes();
+}
+
+static bool wsp_ggml_metal_op_concurrency_reset(wsp_ggml_metal_op_t ctx) {
+    if (!ctx->mem_ranges) {
+        return true;
+    }
+
+    wsp_ggml_metal_encoder_memory_barrier(ctx->enc);
+
+    wsp_ggml_mem_ranges_reset(ctx->mem_ranges);
+
+    return true;
+}
+
+static bool wsp_ggml_metal_op_concurrency_check(wsp_ggml_metal_op_t ctx, const wsp_ggml_tensor * node) {
+    if (!ctx->mem_ranges) {
+        return false;
+    }
+
+    return wsp_ggml_mem_ranges_check(ctx->mem_ranges, node);
+}
+
+static bool wsp_ggml_metal_op_concurrency_add(wsp_ggml_metal_op_t ctx, const wsp_ggml_tensor * node) {
+    if (!ctx->mem_ranges) {
+        return true;
+    }
+
+    return wsp_ggml_mem_ranges_add(ctx->mem_ranges, node);
+}
+
+static int wsp_ggml_metal_op_encode_impl(wsp_ggml_metal_op_t ctx, int idx) {
+    struct wsp_ggml_tensor * node = ctx->node(idx);
+
+    //WSP_GGML_LOG_INFO("%s: encoding node %3d, op = %8s\n", __func__, idx, wsp_ggml_op_name(node->op));
+
+    if (wsp_ggml_is_empty(node)) {
+        return 1;
+    }
+
+    switch (node->op) {
+        case WSP_GGML_OP_NONE:
+        case WSP_GGML_OP_RESHAPE:
+        case WSP_GGML_OP_VIEW:
+        case WSP_GGML_OP_TRANSPOSE:
+        case WSP_GGML_OP_PERMUTE:
+            {
+                // noop -> next node
+                if (ctx->debug_graph > 0) {
+                    WSP_GGML_LOG_DEBUG("%s: node[%5d] - %-12s %s\n", __func__, idx, wsp_ggml_op_name(node->op), "(noop)");
+                }
+            } return 1;
+        default:
+            {
+            } break;
+    }
+
+    if (!wsp_ggml_metal_device_supports_op(ctx->dev, node)) {
+        WSP_GGML_LOG_ERROR("%s: error: unsupported op '%s'\n", __func__, wsp_ggml_op_desc(node));
+        WSP_GGML_ABORT("unsupported op");
+    }
+
+    int n_fuse = 1;
+
+    // check if the current node can run concurrently with other nodes before it
+    // the condition is that:
+    //  - the current node cannot write to any previous src or dst ranges
+    //  - the current node cannot read from any previous dst ranges
+    //
+    // if the condition is not satisfied, we put a memory barrier and clear all ranges
+    // otherwise, we add the new ranges to the encoding context and process the node concurrently
+    //
+    {
+        const bool is_concurrent = wsp_ggml_metal_op_concurrency_check(ctx, node);
+
+        if (!is_concurrent) {
+            wsp_ggml_metal_op_concurrency_reset(ctx);
+        }
+
+        if (ctx->debug_graph > 0) {
+            WSP_GGML_LOG_DEBUG("%s: node[%5d] - %-12s %s\n", __func__, idx, wsp_ggml_op_name(node->op), is_concurrent ? "(concurrent)" : "");
+        }
+        if (ctx->debug_graph > 1) {
+            WSP_GGML_TENSOR_LOCALS( int64_t, ne0, node->src[0], ne);
+            WSP_GGML_TENSOR_LOCALS(uint64_t, nb0, node->src[0], nb);
+            WSP_GGML_TENSOR_LOCALS( int64_t, ne1, node->src[1], ne);
+            WSP_GGML_TENSOR_LOCALS(uint64_t, nb1, node->src[1], nb);
+            WSP_GGML_TENSOR_LOCALS( int64_t, ne2, node->src[2], ne);
+            WSP_GGML_TENSOR_LOCALS(uint64_t, nb2, node->src[2], nb);
+            WSP_GGML_TENSOR_LOCALS( int64_t, ne3, node->src[3], ne);
+            WSP_GGML_TENSOR_LOCALS(uint64_t, nb3, node->src[3], nb);
+            WSP_GGML_TENSOR_LOCALS( int64_t, ne,  node,         ne);
+            WSP_GGML_TENSOR_LOCALS(uint64_t, nb,  node,         nb);
+
+            if (node->src[0]) {
+                WSP_GGML_LOG_DEBUG("%s: src0 - %4s [%5lld, %5lld, %5lld, %5lld] [%5lld, %5lld, %5lld, %5lld], %d, %s\n", __func__, wsp_ggml_type_name(node->src[0]->type), ne00, ne01, ne02, ne03, nb00, nb01, nb02, nb03,
+                        wsp_ggml_is_contiguous(node->src[0]), node->src[0]->name);
+            }
+            if (node->src[1]) {
+                WSP_GGML_LOG_DEBUG("%s: src1 - %4s [%5lld, %5lld, %5lld, %5lld] [%5lld, %5lld, %5lld, %5lld], %d, %s\n", __func__, wsp_ggml_type_name(node->src[1]->type), ne10, ne11, ne12, ne13, nb10, nb11, nb12, nb13,
+                        wsp_ggml_is_contiguous(node->src[1]), node->src[1]->name);
+            }
+            if (node->src[2]) {
+                WSP_GGML_LOG_DEBUG("%s: src2 - %4s [%5lld, %5lld, %5lld, %5lld] [%5lld, %5lld, %5lld, %5lld], %d, %s\n", __func__, wsp_ggml_type_name(node->src[2]->type), ne20, ne21, ne22, ne23, nb20, nb21, nb22, nb23,
+                        wsp_ggml_is_contiguous(node->src[2]), node->src[2]->name);
+            }
+            if (node->src[3]) {
+                WSP_GGML_LOG_DEBUG("%s: src3 - %4s [%5lld, %5lld, %5lld, %5lld] [%5lld, %5lld, %5lld, %5lld], %d, %s\n", __func__, wsp_ggml_type_name(node->src[3]->type), ne30, ne31, ne32, ne33, nb30, nb31, nb32, nb33,
+                        wsp_ggml_is_contiguous(node->src[3]), node->src[3]->name);
+            }
+            if (node) {
+                WSP_GGML_LOG_DEBUG("%s: node  - %4s [%5lld, %5lld, %5lld, %5lld] [%5lld, %5lld, %5lld, %5lld], 1, %s\n", __func__, wsp_ggml_type_name(node->type), ne0, ne1, ne2, ne3, nb0, nb1, nb2, nb3,
+                        node->name);
+            }
+        }
+    }
+
+    switch (node->op) {
+        case WSP_GGML_OP_CONCAT:
+            {
+                n_fuse = wsp_ggml_metal_op_concat(ctx, idx);
+            } break;
+        case WSP_GGML_OP_ADD:
+        case WSP_GGML_OP_SUB:
+        case WSP_GGML_OP_MUL:
+        case WSP_GGML_OP_DIV:
+            {
+                n_fuse = wsp_ggml_metal_op_bin(ctx, idx);
+            } break;
+        case WSP_GGML_OP_ADD_ID:
+            {
+                n_fuse = wsp_ggml_metal_op_add_id(ctx, idx);
+            } break;
+        case WSP_GGML_OP_REPEAT:
+            {
+                n_fuse = wsp_ggml_metal_op_repeat(ctx, idx);
+            } break;
+        case WSP_GGML_OP_ACC:
+            {
+                n_fuse = wsp_ggml_metal_op_acc(ctx, idx);
+            } break;
+        case WSP_GGML_OP_SCALE:
+            {
+                n_fuse = wsp_ggml_metal_op_scale(ctx, idx);
+            } break;
+        case WSP_GGML_OP_CLAMP:
+            {
+                n_fuse = wsp_ggml_metal_op_clamp(ctx, idx);
+            } break;
+        case WSP_GGML_OP_SQR:
+        case WSP_GGML_OP_SQRT:
+        case WSP_GGML_OP_SIN:
+        case WSP_GGML_OP_COS:
+        case WSP_GGML_OP_LOG:
+        case WSP_GGML_OP_UNARY:
+            {
+                n_fuse = wsp_ggml_metal_op_unary(ctx, idx);
+            } break;
+        case WSP_GGML_OP_GLU:
+            {
+                n_fuse = wsp_ggml_metal_op_glu(ctx, idx);
+            } break;
+        case WSP_GGML_OP_SUM:
+            {
+                n_fuse = wsp_ggml_metal_op_sum(ctx, idx);
+            } break;
+        case WSP_GGML_OP_SUM_ROWS:
+        case WSP_GGML_OP_MEAN:
+            {
+                n_fuse = wsp_ggml_metal_op_sum_rows(ctx, idx);
+            } break;
+        case WSP_GGML_OP_SOFT_MAX:
+            {
+                n_fuse = wsp_ggml_metal_op_soft_max(ctx, idx);
+            } break;
+        case WSP_GGML_OP_SSM_CONV:
+            {
+                n_fuse = wsp_ggml_metal_op_ssm_conv(ctx, idx);
+            } break;
+        case WSP_GGML_OP_SSM_SCAN:
+            {
+                n_fuse = wsp_ggml_metal_op_ssm_scan(ctx, idx);
+            } break;
+        case WSP_GGML_OP_RWKV_WKV6:
+        case WSP_GGML_OP_RWKV_WKV7:
+            {
+                n_fuse = wsp_ggml_metal_op_rwkv(ctx, idx);
+            } break;
+        case WSP_GGML_OP_MUL_MAT:
+            {
+                n_fuse = wsp_ggml_metal_op_mul_mat(ctx, idx);
+            } break;
+        case WSP_GGML_OP_MUL_MAT_ID:
+            {
+                n_fuse = wsp_ggml_metal_op_mul_mat_id(ctx, idx);
+            } break;
+        case WSP_GGML_OP_GET_ROWS:
+            {
+                n_fuse = wsp_ggml_metal_op_get_rows(ctx, idx);
+            } break;
+        case WSP_GGML_OP_SET_ROWS:
+            {
+                n_fuse = wsp_ggml_metal_op_set_rows(ctx, idx);
+            } break;
+        case WSP_GGML_OP_L2_NORM:
+            {
+                n_fuse = wsp_ggml_metal_op_l2_norm(ctx, idx);
+            } break;
+        case WSP_GGML_OP_GROUP_NORM:
+            {
+                n_fuse = wsp_ggml_metal_op_group_norm(ctx, idx);
+            } break;
+        case WSP_GGML_OP_NORM:
+        case WSP_GGML_OP_RMS_NORM:
+            {
+                n_fuse = wsp_ggml_metal_op_norm(ctx, idx);
+            } break;
+        case WSP_GGML_OP_ROPE:
+            {
+                n_fuse = wsp_ggml_metal_op_rope(ctx, idx);
+            } break;
+        case WSP_GGML_OP_IM2COL:
+            {
+                n_fuse = wsp_ggml_metal_op_im2col(ctx, idx);
+            } break;
+        case WSP_GGML_OP_CONV_TRANSPOSE_1D:
+            {
+                n_fuse = wsp_ggml_metal_op_conv_transpose_1d(ctx, idx);
+            } break;
+        case WSP_GGML_OP_CONV_TRANSPOSE_2D:
+            {
+                n_fuse = wsp_ggml_metal_op_conv_transpose_2d(ctx, idx);
+            } break;
+        case WSP_GGML_OP_UPSCALE:
+            {
+                n_fuse = wsp_ggml_metal_op_upscale(ctx, idx);
+            } break;
+        case WSP_GGML_OP_PAD:
+            {
+                n_fuse = wsp_ggml_metal_op_pad(ctx, idx);
+            } break;
+        case WSP_GGML_OP_PAD_REFLECT_1D:
+            {
+                n_fuse = wsp_ggml_metal_op_pad_reflect_1d(ctx, idx);
+            } break;
+        case WSP_GGML_OP_ARANGE:
+            {
+                n_fuse = wsp_ggml_metal_op_arange(ctx, idx);
+            } break;
+        case WSP_GGML_OP_TIMESTEP_EMBEDDING:
+            {
+                n_fuse = wsp_ggml_metal_op_timestep_embedding(ctx, idx);
+            } break;
+        case WSP_GGML_OP_ARGSORT:
+            {
+                n_fuse = wsp_ggml_metal_op_argsort(ctx, idx);
+            } break;
+        case WSP_GGML_OP_LEAKY_RELU:
+            {
+                n_fuse = wsp_ggml_metal_op_leaky_relu(ctx, idx);
+            } break;
+        case WSP_GGML_OP_FLASH_ATTN_EXT:
+            {
+                n_fuse = wsp_ggml_metal_op_flash_attn_ext(ctx, idx);
+            } break;
+        case WSP_GGML_OP_DUP:
+        case WSP_GGML_OP_CPY:
+        case WSP_GGML_OP_CONT:
+            {
+                n_fuse = wsp_ggml_metal_op_cpy(ctx, idx);
+            } break;
+        case WSP_GGML_OP_POOL_2D:
+            {
+                n_fuse = wsp_ggml_metal_op_pool_2d(ctx, idx);
+            } break;
+        case WSP_GGML_OP_ARGMAX:
+            {
+                n_fuse = wsp_ggml_metal_op_argmax(ctx, idx);
+            } break;
+        case WSP_GGML_OP_OPT_STEP_ADAMW:
+            {
+                n_fuse = wsp_ggml_metal_op_opt_step_adamw(ctx, idx);
+            } break;
+        case WSP_GGML_OP_OPT_STEP_SGD:
+            {
+                n_fuse = wsp_ggml_metal_op_opt_step_sgd(ctx, idx);
+            } break;
+       default:
+            {
+                WSP_GGML_LOG_ERROR("%s: error: node %3d, op = %8s not implemented\n", __func__, idx, wsp_ggml_op_name(node->op));
+                WSP_GGML_ABORT("fatal error");
+            }
+    }
+
+    if (ctx->debug_graph > 0) {
+        if (n_fuse > 1) {
+            WSP_GGML_LOG_DEBUG("%s:               fuse %d ops\n", __func__, n_fuse);
+        }
+    }
+
+    // update the mem ranges in the encoding context
+    for (int i = 0; i < n_fuse; ++i) {
+        if (!wsp_ggml_metal_op_concurrency_add(ctx, ctx->node(idx + i))) {
+            wsp_ggml_metal_op_concurrency_reset(ctx);
+        }
+    }
+
+    return n_fuse;
+}
+
+int wsp_ggml_metal_op_encode(wsp_ggml_metal_op_t ctx, int idx) {
+    if (ctx->use_capture) {
+        wsp_ggml_metal_encoder_debug_group_push(ctx->enc, wsp_ggml_op_desc(ctx->node(idx)));
+    }
+
+    int res = wsp_ggml_metal_op_encode_impl(ctx, idx);
+    if (idx + res > ctx->n_nodes()) {
+        WSP_GGML_ABORT("fusion error: nodes spanning multiple encoders have been fused. this indicates a bug in the fusion logic %s",
+                "https://github.com/ggml-org/llama.cpp/pull/14849");
+    }
+
+    if (ctx->use_capture) {
+        wsp_ggml_metal_encoder_debug_group_pop(ctx->enc);
+    }
+
+    return res;
+}
+
+int wsp_ggml_metal_op_concat(wsp_ggml_metal_op_t ctx, int idx) {
+    wsp_ggml_tensor * op = ctx->node(idx);
+
+    wsp_ggml_metal_library_t lib = ctx->lib;
+    wsp_ggml_metal_encoder_t enc = ctx->enc;
+
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
+
+    const int32_t dim = ((const int32_t *) op->op_params)[0];
+
+    wsp_ggml_metal_kargs_concat 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,
+        /*.dim  =*/ dim,
+    };
+
+    wsp_ggml_metal_pipeline_t pipeline = wsp_ggml_metal_library_get_pipeline_base(lib, WSP_GGML_OP_CONCAT);
+
+    wsp_ggml_metal_encoder_set_pipeline(enc, pipeline);
+    wsp_ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[0]), 1);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[1]), 2);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op),         3);
+
+    const int nth = std::min(1024, ne0);
+
+    wsp_ggml_metal_encoder_dispatch_threadgroups(enc, ne1, ne2, ne3, nth, 1, 1);
+
+    return 1;
+}
+
+int wsp_ggml_metal_op_repeat(wsp_ggml_metal_op_t ctx, int idx) {
+    wsp_ggml_tensor * op = ctx->node(idx);
+
+    wsp_ggml_metal_library_t lib = ctx->lib;
+    wsp_ggml_metal_encoder_t enc = ctx->enc;
+
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
+    WSP_GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+
+    wsp_ggml_metal_pipeline_t pipeline = wsp_ggml_metal_library_get_pipeline_repeat(lib, op->type);
+
+    wsp_ggml_metal_kargs_repeat 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,
+    };
+
+    wsp_ggml_metal_encoder_set_pipeline(enc, pipeline);
+    wsp_ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[0]), 1);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op),         2);
+
+    const int nth = std::min(wsp_ggml_metal_pipeline_max_theads_per_threadgroup(pipeline), ne0);
+
+    wsp_ggml_metal_encoder_dispatch_threadgroups(enc, ne1, ne2, ne3, nth, 1, 1);
+
+    return 1;
+}
+
+int wsp_ggml_metal_op_acc(wsp_ggml_metal_op_t ctx, int idx) {
+    wsp_ggml_tensor * op = ctx->node(idx);
+
+    wsp_ggml_metal_library_t lib = ctx->lib;
+    wsp_ggml_metal_encoder_t enc = ctx->enc;
+
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
+    WSP_GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+
+    WSP_GGML_ASSERT(op->src[0]->type == WSP_GGML_TYPE_F32);
+    WSP_GGML_ASSERT(op->src[1]->type == WSP_GGML_TYPE_F32);
+    WSP_GGML_ASSERT(op->type         == WSP_GGML_TYPE_F32);
+
+    WSP_GGML_ASSERT(wsp_ggml_is_contiguous(op->src[0]));
+    WSP_GGML_ASSERT(wsp_ggml_is_contiguous(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];
+    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 separete kernel to cpy src->dst
+        // not sure how to avoid this
+        // TODO: make a simpler cpy_bytes kernel
+
+        //const id<MTLComputePipelineState> pipeline = ctx->pipelines[WSP_GGML_METAL_PIPELINE_TYPE_CPY_F32_F32].obj;
+        wsp_ggml_metal_pipeline_t pipeline = wsp_ggml_metal_library_get_pipeline_cpy(lib, op->src[0]->type, op->type);
+
+        wsp_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,
+        };
+
+        wsp_ggml_metal_encoder_set_pipeline(enc, pipeline);
+        wsp_ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+        wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[0]), 1);
+        wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op),         2);
+
+        const int nth = std::min(wsp_ggml_metal_pipeline_max_theads_per_threadgroup(pipeline), ne00);
+
+        wsp_ggml_metal_encoder_dispatch_threadgroups(enc, ne01, ne02, ne03, nth, 1, 1);
+
+        wsp_ggml_metal_op_concurrency_reset(ctx);
+    }
+
+    wsp_ggml_metal_kargs_bin args = {
+        /*.ne00 =*/ ne00,
+        /*.ne01 =*/ ne01,
+        /*.ne02 =*/ ne02,
+        /*.ne03 =*/ ne03,
+        /*.nb00 =*/ nb00,
+        /*.nb01 =*/ pnb1,
+        /*.nb02 =*/ pnb2,
+        /*.nb03 =*/ pnb3,
+        /*.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  =*/ pnb1,
+        /*.nb2  =*/ pnb2,
+        /*.nb3  =*/ pnb3,
+        /*.offs =*/ offs,
+        /*.o1   =*/ { 0 },
+    };
+
+    wsp_ggml_metal_pipeline_t pipeline = wsp_ggml_metal_library_get_pipeline_bin(lib, WSP_GGML_OP_ADD, 1, false);
+
+    wsp_ggml_metal_encoder_set_pipeline(enc, pipeline);
+    wsp_ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[0]), 1);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[1]), 2);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op),         3);
+
+    const int nth = std::min(wsp_ggml_metal_pipeline_max_theads_per_threadgroup(pipeline), ne00);
+
+    wsp_ggml_metal_encoder_dispatch_threadgroups(enc, ne11, ne12, ne13, nth, 1, 1);
+
+    return 1;
+}
+
+int wsp_ggml_metal_op_scale(wsp_ggml_metal_op_t ctx, int idx) {
+    wsp_ggml_tensor * op = ctx->node(idx);
+
+    wsp_ggml_metal_library_t lib = ctx->lib;
+    wsp_ggml_metal_encoder_t enc = ctx->enc;
+
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
+    WSP_GGML_TENSOR_LOCALS(uint32_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));
+
+    wsp_ggml_metal_kargs_scale args = {
+        /*.scale =*/ scale,
+        /*.bias  =*/ bias,
+    };
+
+    int64_t n = wsp_ggml_nelements(op);
+
+    if (n % 4 == 0) {
+        n /= 4;
+    }
+
+    wsp_ggml_metal_pipeline_t pipeline = wsp_ggml_metal_library_get_pipeline_unary(lib, op);
+
+    wsp_ggml_metal_encoder_set_pipeline(enc, pipeline);
+    wsp_ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[0]), 1);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op),         2);
+
+    wsp_ggml_metal_encoder_dispatch_threadgroups(enc, n, 1, 1, 1, 1, 1);
+
+    return 1;
+}
+
+int wsp_ggml_metal_op_clamp(wsp_ggml_metal_op_t ctx, int idx) {
+    wsp_ggml_tensor * op = ctx->node(idx);
+
+    wsp_ggml_metal_library_t lib = ctx->lib;
+    wsp_ggml_metal_encoder_t enc = ctx->enc;
+
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
+    WSP_GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+
+    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));
+
+    wsp_ggml_metal_kargs_clamp args = {
+        /*.min =*/ min,
+        /*.max =*/ max,
+    };
+
+    int64_t n = wsp_ggml_nelements(op);
+
+    if (n % 4 == 0) {
+        n /= 4;
+    }
+
+    wsp_ggml_metal_pipeline_t pipeline = wsp_ggml_metal_library_get_pipeline_unary(lib, op);
+
+    wsp_ggml_metal_encoder_set_pipeline(enc, pipeline);
+    wsp_ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[0]), 1);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op),         2);
+
+    wsp_ggml_metal_encoder_dispatch_threadgroups(enc, n, 1, 1, 1, 1, 1);
+
+    return 1;
+}
+
+int wsp_ggml_metal_op_unary(wsp_ggml_metal_op_t ctx, int idx) {
+    wsp_ggml_tensor * op = ctx->node(idx);
+
+    wsp_ggml_metal_library_t lib = ctx->lib;
+    wsp_ggml_metal_encoder_t enc = ctx->enc;
+
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
+    WSP_GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+
+    int64_t n = wsp_ggml_nelements(op);
+
+    if (n % 4 == 0) {
+        n /= 4;
+    }
+
+    wsp_ggml_metal_pipeline_t pipeline = wsp_ggml_metal_library_get_pipeline_unary(lib, op);
+
+    wsp_ggml_metal_encoder_set_pipeline(enc, pipeline);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[0]), 0);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op),         1);
+
+    wsp_ggml_metal_encoder_dispatch_threadgroups(enc, n, 1, 1, 1, 1, 1);
+
+    return 1;
+}
+
+int wsp_ggml_metal_op_glu(wsp_ggml_metal_op_t ctx, int idx) {
+    wsp_ggml_tensor * op = ctx->node(idx);
+
+    wsp_ggml_metal_library_t lib = ctx->lib;
+    wsp_ggml_metal_encoder_t enc = ctx->enc;
+
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
+    WSP_GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+
+    if (op->src[1]) {
+        WSP_GGML_ASSERT(wsp_ggml_are_same_shape(op->src[0], op->src[1]));
+    }
+
+    wsp_ggml_metal_pipeline_t pipeline = wsp_ggml_metal_library_get_pipeline_glu(lib, op);
+
+    const int32_t swp = wsp_ggml_get_op_params_i32(op, 1);
+    const float alpha = wsp_ggml_get_op_params_f32(op, 2);
+    const float limit = wsp_ggml_get_op_params_f32(op, 3);
+
+    const int32_t i00 = swp ? ne0 : 0;
+    const int32_t i10 = swp ? 0 : ne0;
+
+    wsp_ggml_metal_kargs_glu args = {
+        /*.ne00 =*/ ne00,
+        /*.nb01 =*/ nb01,
+        /*.ne10 =*/ op->src[1] ? ne10 : ne00,
+        /*.nb11 =*/ op->src[1] ? nb11 : nb01,
+        /*.ne0  =*/ ne0,
+        /*.nb1  =*/ nb1,
+        /*.i00  =*/ op->src[1] ? 0 : i00,
+        /*.i10  =*/ op->src[1] ? 0 : i10,
+        /*.alpha=*/ alpha,
+        /*.limit=*/ limit
+    };
+
+    const int64_t nrows = wsp_ggml_nrows(op->src[0]);
+
+    const int32_t nth = std::min(wsp_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)];
+
+    wsp_ggml_metal_encoder_set_pipeline(enc, pipeline);
+    wsp_ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[0]), 1);
+    if (op->src[1]) {
+        wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[1]), 2);
+    } else {
+        wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[0]), 2);
+    }
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op),         3);
+
+    wsp_ggml_metal_encoder_dispatch_threadgroups(enc, nrows, 1, 1, nth, 1, 1);
+
+    return 1;
+}
+
+int wsp_ggml_metal_op_sum(wsp_ggml_metal_op_t ctx, int idx) {
+    wsp_ggml_tensor * op  = ctx->node(idx);
+
+    wsp_ggml_metal_library_t lib = ctx->lib;
+    wsp_ggml_metal_encoder_t enc = ctx->enc;
+
+    const uint64_t n = (uint64_t) wsp_ggml_nelements(op->src[0]);
+
+    wsp_ggml_metal_kargs_sum args = {
+        /*.np =*/ n,
+    };
+
+    wsp_ggml_metal_pipeline_t pipeline = wsp_ggml_metal_library_get_pipeline_sum(lib, op);
+
+    int nth = 32; // SIMD width
+
+    while (nth < (int) n && nth < wsp_ggml_metal_pipeline_max_theads_per_threadgroup(pipeline)) {
+        nth *= 2;
+    }
+
+    nth = std::min(nth, wsp_ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));
+    nth = std::min(nth, (int) n);
+
+    const int nsg = (nth + 31) / 32;
+
+    wsp_ggml_metal_encoder_set_pipeline(enc, pipeline);
+    wsp_ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[0]), 1);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op),         2);
+
+    wsp_ggml_metal_encoder_set_threadgroup_memory_size(enc, nsg * sizeof(float), 0);
+
+    wsp_ggml_metal_encoder_dispatch_threadgroups(enc, 1, 1, 1, nth, 1, 1);
+
+    return 1;
+}
+
+int wsp_ggml_metal_op_sum_rows(wsp_ggml_metal_op_t ctx, int idx) {
+    wsp_ggml_tensor * op = ctx->node(idx);
+
+    wsp_ggml_metal_library_t lib = ctx->lib;
+    wsp_ggml_metal_encoder_t enc = ctx->enc;
+
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
+    WSP_GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+
+    wsp_ggml_metal_kargs_sum_rows 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,
+    };
+
+    wsp_ggml_metal_pipeline_t pipeline = wsp_ggml_metal_library_get_pipeline_sum_rows(lib, op);
+
+    int nth = 32; // SIMD width
+
+    while (nth < ne00 && nth < wsp_ggml_metal_pipeline_max_theads_per_threadgroup(pipeline)) {
+        nth *= 2;
+    }
+
+    nth = std::min(nth, wsp_ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));
+    nth = std::min(nth, ne00);
+
+    const size_t smem = wsp_ggml_metal_pipeline_get_smem(pipeline);
+
+    //[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)];
+
+    wsp_ggml_metal_encoder_set_pipeline(enc, pipeline);
+    wsp_ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[0]), 1);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op),         2);
+
+    wsp_ggml_metal_encoder_set_threadgroup_memory_size(enc, smem, 0);
+
+    wsp_ggml_metal_encoder_dispatch_threadgroups(enc, ne01, ne02, ne03, nth, 1, 1);
+
+    return 1;
+}
+
+int wsp_ggml_metal_op_get_rows(wsp_ggml_metal_op_t ctx, int idx) {
+    wsp_ggml_tensor * op = ctx->node(idx);
+
+    wsp_ggml_metal_library_t lib = ctx->lib;
+    wsp_ggml_metal_encoder_t enc = ctx->enc;
+
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
+    WSP_GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+
+    wsp_ggml_metal_pipeline_t pipeline = wsp_ggml_metal_library_get_pipeline_get_rows(lib, op->src[0]->type);
+
+    wsp_ggml_metal_kargs_get_rows args = {
+        /*.ne00t =*/ wsp_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, wsp_ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));
+
+    const int nw0 = (args.ne00t + nth - 1)/nth;
+
+    wsp_ggml_metal_encoder_set_pipeline(enc, pipeline);
+    wsp_ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[0]), 1);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[1]), 2);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op),         3);
+
+    wsp_ggml_metal_encoder_dispatch_threadgroups(enc, nw0*ne10, ne11, ne12, nth, 1, 1);
+
+    return 1;
+}
+
+int wsp_ggml_metal_op_set_rows(wsp_ggml_metal_op_t ctx, int idx) {
+    wsp_ggml_tensor * op = ctx->node(idx);
+
+    wsp_ggml_metal_library_t lib = ctx->lib;
+    wsp_ggml_metal_encoder_t enc = ctx->enc;
+
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
+    WSP_GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+
+    wsp_ggml_metal_pipeline_t pipeline = wsp_ggml_metal_library_get_pipeline_set_rows(lib, op->src[1]->type, op->type);
+
+    const int32_t nk0 = ne0/wsp_ggml_blck_size(op->type);
+
+    int nth = 32; // SIMD width
+
+    while (nth < nk0 && nth < wsp_ggml_metal_pipeline_max_theads_per_threadgroup(pipeline)) {
+        nth *= 2;
+    }
+
+    int nrptg = 1;
+    if (nth > nk0) {
+        nrptg = (nth + nk0 - 1)/nk0;
+        nth   = nk0;
+
+        if (nrptg*nth > wsp_ggml_metal_pipeline_max_theads_per_threadgroup(pipeline)) {
+            nrptg--;
+        }
+    }
+
+    nth = std::min(nth, nk0);
+
+    wsp_ggml_metal_kargs_set_rows args = {
+        /*.nk0  =*/ nk0,
+        /*.ne01 =*/ ne01,
+        /*.nb01 =*/ nb01,
+        /*.nb02 =*/ nb02,
+        /*.nb03 =*/ nb03,
+        /*.ne11 =*/ ne11,
+        /*.ne12 =*/ ne12,
+        /*.nb10 =*/ nb10,
+        /*.nb11 =*/ nb11,
+        /*.nb12 =*/ nb12,
+        /*.nb1  =*/ nb1,
+        /*.nb2  =*/ nb2,
+        /*.nb3  =*/ nb3,
+    };
+
+    wsp_ggml_metal_encoder_set_pipeline(enc, pipeline);
+    wsp_ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[0]), 1);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[1]), 2);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op),         3);
+
+    wsp_ggml_metal_encoder_dispatch_threadgroups(enc, (ne01 + nrptg - 1)/nrptg, ne02, ne03, nth, nrptg, 1);
+
+    return 1;
+}
+
+int wsp_ggml_metal_op_soft_max(wsp_ggml_metal_op_t ctx, int idx) {
+    wsp_ggml_tensor * op = ctx->node(idx);
+
+    wsp_ggml_metal_library_t lib = ctx->lib;
+    wsp_ggml_metal_encoder_t enc = ctx->enc;
+
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne2, op->src[2], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb2, op->src[2], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
+    WSP_GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+
+    float scale;
+    float max_bias;
+
+    memcpy(&scale,    ((const int32_t *) op->op_params) + 0, sizeof(scale));
+    memcpy(&max_bias, ((const int32_t *) op->op_params) + 1, sizeof(max_bias));
+
+    const uint32_t n_head      = op->src[0]->ne[2];
+    const  int32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head));
+
+    const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
+    const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
+
+    // softmax
+
+    wsp_ggml_metal_kargs_soft_max args = {
+        /*.ne00        =*/ ne00,
+        /*.ne01        =*/ ne01,
+        /*.ne02        =*/ ne02,
+        /*.nb01        =*/ nb01,
+        /*.nb02        =*/ nb02,
+        /*.nb03        =*/ nb03,
+        /*.ne11        =*/ ne11,
+        /*.ne12        =*/ ne12,
+        /*.ne13        =*/ ne13,
+        /*.nb11        =*/ nb11,
+        /*.nb12        =*/ nb12,
+        /*.nb13        =*/ nb13,
+        /*.nb1         =*/ nb1,
+        /*.nb2         =*/ nb2,
+        /*.nb3         =*/ nb3,
+        /*.scale       =*/ scale,
+        /*.max_bias    =*/ max_bias,
+        /*.m0          =*/ m0,
+        /*.m1          =*/ m1,
+        /*.n_head_log2 =*/ n_head_log2,
+    };
+
+    wsp_ggml_metal_pipeline_t pipeline = wsp_ggml_metal_library_get_pipeline_soft_max(lib, op);
+
+    int nth = 32; // SIMD width
+
+    if (ne00%4 == 0) {
+        while (nth < ne00/4 && nth*ne01*ne02*ne03 < 256) {
+            nth *= 2;
+        }
+    } else {
+        while (nth < ne00 && nth*ne01*ne02*ne03 < 256) {
+            nth *= 2;
+        }
+    }
+
+    const size_t smem = wsp_ggml_metal_pipeline_get_smem(pipeline);
+
+    wsp_ggml_metal_encoder_set_pipeline(enc, pipeline);
+    wsp_ggml_metal_encoder_set_bytes(enc, &args, sizeof(args), 0);
+    wsp_ggml_metal_encoder_set_buffer(enc, wsp_ggml_metal_get_buffer_id(op->src[0]), 1);
+    if (op->src[1]) {
+        wsp_ggml_metal_encoder_set_buffer(enc, wsp_ggml_metal_get_buffer_id(op->src[1]), 2);
+    } else {
+        wsp_ggml_metal_encoder_set_buffer(enc, wsp_ggml_metal_get_buffer_id(op->src[0]), 2);
+    }
+    if (op->src[2]) {
+        wsp_ggml_metal_encoder_set_buffer(enc, wsp_ggml_metal_get_buffer_id(op->src[2]), 3);
+    } else {
+        wsp_ggml_metal_encoder_set_buffer(enc, wsp_ggml_metal_get_buffer_id(op->src[0]), 3);
+    }
+    wsp_ggml_metal_encoder_set_buffer(enc, wsp_ggml_metal_get_buffer_id(op), 4);
+
+    wsp_ggml_metal_encoder_set_threadgroup_memory_size(enc, smem, 0);
+
+    wsp_ggml_metal_encoder_dispatch_threadgroups(enc, ne01, ne02, ne03, nth, 1, 1);
+
+    return 1;
+}
+
+int wsp_ggml_metal_op_ssm_conv(wsp_ggml_metal_op_t ctx, int idx) {
+    wsp_ggml_tensor * op = ctx->node(idx);
+
+    wsp_ggml_metal_library_t lib = ctx->lib;
+    wsp_ggml_metal_encoder_t enc = ctx->enc;
+
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
+    WSP_GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+
+    wsp_ggml_metal_kargs_ssm_conv args = {
+        /*.ne00 =*/ ne00,
+        /*.ne01 =*/ ne01,
+        /*.ne02 =*/ ne02,
+        /*.nb00 =*/ nb00,
+        /*.nb01 =*/ nb01,
+        /*.nb02 =*/ nb02,
+        /*.ne10 =*/ ne10,
+        /*.ne11 =*/ ne11,
+        /*.nb10 =*/ nb10,
+        /*.nb11 =*/ nb11,
+        /*.ne0  =*/ ne0,
+        /*.ne1  =*/ ne1,
+        /*.ne2  =*/ ne2,
+        /*.nb0  =*/ nb0,
+        /*.nb1  =*/ nb1,
+        /*.nb2  =*/ nb2,
+    };
+
+    wsp_ggml_metal_pipeline_t pipeline = wsp_ggml_metal_library_get_pipeline_ssm_conv(lib, op);
+
+    wsp_ggml_metal_encoder_set_pipeline(enc, pipeline);
+    wsp_ggml_metal_encoder_set_bytes(enc, &args, sizeof(args), 0);
+    wsp_ggml_metal_encoder_set_buffer(enc, wsp_ggml_metal_get_buffer_id(op->src[0]), 1);
+    wsp_ggml_metal_encoder_set_buffer(enc, wsp_ggml_metal_get_buffer_id(op->src[1]), 2);
+    wsp_ggml_metal_encoder_set_buffer(enc, wsp_ggml_metal_get_buffer_id(op),         3);
+
+    wsp_ggml_metal_encoder_dispatch_threadgroups(enc, ne01, ne1, ne02, 1, 1, 1);
+
+    return 1;
+}
+
+int wsp_ggml_metal_op_ssm_scan(wsp_ggml_metal_op_t ctx, int idx) {
+    wsp_ggml_tensor * op = ctx->node(idx);
+
+    wsp_ggml_metal_library_t lib = ctx->lib;
+    wsp_ggml_metal_encoder_t enc = ctx->enc;
+
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne2, op->src[2], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb2, op->src[2], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne3, op->src[3], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb3, op->src[3], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne4, op->src[4], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb4, op->src[4], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne5, op->src[5], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb5, op->src[5], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne6, op->src[6], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb6, op->src[6], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
+    WSP_GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+
+    const wsp_ggml_tensor * src3 = op->src[3];
+    const wsp_ggml_tensor * src4 = op->src[4];
+    const wsp_ggml_tensor * src5 = op->src[5];
+    const wsp_ggml_tensor * src6 = op->src[6];
+
+    WSP_GGML_ASSERT(src3);
+    WSP_GGML_ASSERT(src4);
+    WSP_GGML_ASSERT(src5);
+    WSP_GGML_ASSERT(src6);
+
+    const int64_t d_state      = ne00;
+    const int64_t d_inner      = ne01;
+    const int64_t n_head       = ne02;
+    const int64_t n_group      = ne41;
+    const int64_t n_seq_tokens = ne12;
+    const int64_t n_seqs       = ne13;
+
+    wsp_ggml_metal_kargs_ssm_scan args = {
+        /*.d_state      =*/ d_state,
+        /*.d_inner      =*/ d_inner,
+        /*.n_head       =*/ n_head,
+        /*.n_group      =*/ n_group,
+        /*.n_seq_tokens =*/ n_seq_tokens,
+        /*.n_seqs       =*/ n_seqs,
+        /*.s_off        =*/ wsp_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,
+    };
+
+    wsp_ggml_metal_pipeline_t pipeline = wsp_ggml_metal_library_get_pipeline_ssm_scan(lib, op);
+
+    WSP_GGML_ASSERT(d_state <= wsp_ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));
+
+    const size_t sms = wsp_ggml_metal_pipeline_get_smem(pipeline);
+
+    wsp_ggml_metal_encoder_set_pipeline(enc, pipeline);
+    wsp_ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[0]), 1);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[1]), 2);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[2]), 3);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[3]), 4);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[4]), 5);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[5]), 6);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[6]), 7);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op),         8);
+
+    wsp_ggml_metal_encoder_set_threadgroup_memory_size(enc, sms, 0);
+
+    wsp_ggml_metal_encoder_dispatch_threadgroups(enc, d_inner, n_head, n_seqs, d_state, 1, 1);
+
+    return 1;
+}
+
+int wsp_ggml_metal_op_rwkv(wsp_ggml_metal_op_t ctx, int idx) {
+    wsp_ggml_tensor * op = ctx->node(idx);
+
+    wsp_ggml_metal_library_t lib = ctx->lib;
+    wsp_ggml_metal_encoder_t enc = ctx->enc;
+
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
+    WSP_GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+
+    const int64_t B = op->op == WSP_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];
+
+    wsp_ggml_metal_pipeline_t pipeline = wsp_ggml_metal_library_get_pipeline_rwkv(lib, op);
+
+    int ida = 0;
+
+    wsp_ggml_metal_encoder_set_pipeline(enc, pipeline);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[0]), ida++);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[1]), ida++);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[2]), ida++);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[3]), ida++);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[4]), ida++);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[5]), ida++);
+    if (op->op == WSP_GGML_OP_RWKV_WKV7) {
+        wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[6]), ida++);
+    }
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op),         ida++);
+    wsp_ggml_metal_encoder_set_bytes   (enc, (void *) &B, sizeof(B), ida++);
+    wsp_ggml_metal_encoder_set_bytes   (enc, (void *) &T, sizeof(T), ida++);
+    wsp_ggml_metal_encoder_set_bytes   (enc, (void *) &C, sizeof(C), ida++);
+    wsp_ggml_metal_encoder_set_bytes   (enc, (void *) &H, sizeof(H), ida++);
+
+    wsp_ggml_metal_encoder_dispatch_threadgroups(enc, B * H, 1, 1, C/H, 1, 1);
+
+    return 1;
+}
+
+int wsp_ggml_metal_op_cpy(wsp_ggml_metal_op_t ctx, int idx) {
+    wsp_ggml_tensor * op = ctx->node(idx);
+
+    wsp_ggml_metal_library_t lib = ctx->lib;
+    wsp_ggml_metal_encoder_t enc = ctx->enc;
+
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
+    WSP_GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+
+    wsp_ggml_metal_pipeline_t pipeline = wsp_ggml_metal_library_get_pipeline_cpy(lib, op->src[0]->type, op->type);
+
+    WSP_GGML_ASSERT(ne00 % wsp_ggml_blck_size(op->src[0]->type) == 0);
+
+    int64_t nk0 = ne00;
+    if (wsp_ggml_is_quantized(op->src[0]->type)) {
+        nk0 = ne00/16;
+    } else if (wsp_ggml_is_quantized(op->type)) {
+        nk0 = ne00/wsp_ggml_blck_size(op->type);
+    }
+
+    int nth = std::min<int>(nk0, wsp_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 (wsp_ggml_blck_size(op->src[0]->type) == 1 && wsp_ggml_blck_size(op->type) == 1) {
+        if (nth > nk0) {
+            nrptg = (nth + nk0 - 1)/nk0;
+            nth   = nk0;
+
+            if (nrptg*nth > wsp_ggml_metal_pipeline_max_theads_per_threadgroup(pipeline)) {
+                nrptg--;
+            }
+        }
+    }
+
+    nth = std::min<int>(nth, nk0);
+
+    wsp_ggml_metal_kargs_cpy args = {
+        /*.nk0  =*/ nk0,
+        /*.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,
+    };
+
+    const int nw0 = nrptg == 1 ? (nk0 + nth - 1)/nth : 1;
+
+    wsp_ggml_metal_encoder_set_pipeline(enc, pipeline);
+    wsp_ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[0]), 1);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op),         2);
+
+    wsp_ggml_metal_encoder_dispatch_threadgroups(enc, nw0*(ne01 + nrptg - 1)/nrptg, ne02, ne03, nth, nrptg, 1);
+
+    return 1;
+}
+
+int wsp_ggml_metal_op_pool_2d(wsp_ggml_metal_op_t ctx, int idx) {
+    wsp_ggml_tensor * op = ctx->node(idx);
+
+    wsp_ggml_metal_library_t lib = ctx->lib;
+    wsp_ggml_metal_encoder_t enc = ctx->enc;
+
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
+    WSP_GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+
+    const int32_t * opts = op->op_params;
+    wsp_ggml_op_pool op_pool = (wsp_ggml_op_pool) opts[0];
+
+    const int32_t k0 = opts[1];
+    const int32_t k1 = opts[2];
+    const int32_t s0 = opts[3];
+    const int32_t s1 = opts[4];
+    const int32_t p0 = opts[5];
+    const int32_t p1 = opts[6];
+
+    const int64_t IH = op->src[0]->ne[1];
+    const int64_t IW = op->src[0]->ne[0];
+
+    const int64_t N  = op->ne[3];
+    const int64_t OC = op->ne[2];
+    const int64_t OH = op->ne[1];
+    const int64_t OW = op->ne[0];
+
+    const int64_t np = N * OC * OH * OW;
+
+    wsp_ggml_metal_kargs_pool_2d args_pool_2d = {
+        /* .k0 = */ k0,
+        /* .k1 = */ k1,
+        /* .s0 = */ s0,
+        /* .s1 = */ s1,
+        /* .p0 = */ p0,
+        /* .p1 = */ p1,
+        /* .IH = */ IH,
+        /* .IW = */ IW,
+        /* .OH = */ OH,
+        /* .OW = */ OW,
+        /* .np = */ np
+    };
+
+    wsp_ggml_metal_pipeline_t pipeline = wsp_ggml_metal_library_get_pipeline_pool_2d(lib, op, op_pool);
+
+    const int nth = std::min(wsp_ggml_metal_pipeline_max_theads_per_threadgroup(pipeline), (int) np);
+    const int ntg = (np + nth - 1) / nth;
+
+    wsp_ggml_metal_encoder_set_pipeline(enc, pipeline);
+    wsp_ggml_metal_encoder_set_bytes   (enc, &args_pool_2d, sizeof(args_pool_2d), 0);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[0]), 1);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op),         2);
+
+    wsp_ggml_metal_encoder_dispatch_threadgroups(enc, ntg, 1, 1, nth, 1, 1);
+
+    return 1;
+}
+
+int wsp_ggml_metal_op_mul_mat(wsp_ggml_metal_op_t ctx, int idx) {
+    wsp_ggml_tensor * op = ctx->node(idx);
+
+    wsp_ggml_metal_library_t lib = ctx->lib;
+    wsp_ggml_metal_encoder_t enc = ctx->enc;
+
+    const wsp_ggml_metal_device_props * props_dev = wsp_ggml_metal_device_get_props(ctx->dev);
+
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
+    WSP_GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+
+    WSP_GGML_ASSERT(ne00 == ne10);
+
+    WSP_GGML_ASSERT(ne12 % ne02 == 0);
+    WSP_GGML_ASSERT(ne13 % ne03 == 0);
+
+    const int16_t r2 = ne12/ne02;
+    const int16_t r3 = ne13/ne03;
+
+    // find the break-even point where the matrix-matrix kernel becomes more efficient compared
+    // to the matrix-vector kernel
+    const int ne11_mm_min = 8;
+
+    // first try to use small-batch mat-mv kernels
+    // these should be efficient for BS [2, ~8]
+    if (op->src[1]->type == WSP_GGML_TYPE_F32 && (ne00%128 == 0) &&
+        (
+         (
+          (
+           op->src[0]->type == WSP_GGML_TYPE_F32  || // TODO: helper function
+           op->src[0]->type == WSP_GGML_TYPE_F16  ||
+           op->src[0]->type == WSP_GGML_TYPE_Q4_0 ||
+           op->src[0]->type == WSP_GGML_TYPE_Q4_1 ||
+           op->src[0]->type == WSP_GGML_TYPE_Q5_0 ||
+           op->src[0]->type == WSP_GGML_TYPE_Q5_1 ||
+           op->src[0]->type == WSP_GGML_TYPE_Q8_0 ||
+           op->src[0]->type == WSP_GGML_TYPE_MXFP4 ||
+           op->src[0]->type == WSP_GGML_TYPE_IQ4_NL ||
+           false) && (ne11 >= 2 && ne11 <= 8)
+         ) ||
+         (
+          (
+           op->src[0]->type == WSP_GGML_TYPE_Q4_K ||
+           op->src[0]->type == WSP_GGML_TYPE_Q5_K ||
+           op->src[0]->type == WSP_GGML_TYPE_Q6_K ||
+           false) && (ne11 >= 4 && ne11 <= 8)
+         )
+        )
+       ) {
+        // TODO: determine the optimal parameters based on grid utilization
+        //       I still don't know why we should not always use the maximum available threads:
+        //
+        //       nsg = pipeline.maxTotalThreadsPerThreadgroup / 32
+        //
+        //       my current hypothesis is that the work grid is not evenly divisible for different nsg
+        //       values and there can be some tail effects when nsg is high. need to confirm this
+        //
+        const int nsg    = 2;                 // num simdgroups per threadgroup
+
+        // num threads along row per simdgroup
+        int16_t nxpsg = 0;
+        if (ne00 % 256 == 0 && ne11 < 3) {
+            nxpsg = 16;
+        } else if (ne00 % 128 == 0) {
+            nxpsg = 8;
+        } else {
+            nxpsg = 4;
+        }
+
+        const int16_t nypsg  = 32/nxpsg;          // num threads along col per simdgroup (i.e. a simdgroup processes that many src0 rows at a time)
+        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
+        switch (ne11) {
+            case 2:
+                r1ptg = 2; break;
+            case 3:
+            case 6:
+                r1ptg = 3; break;
+            case 4:
+            case 7:
+            case 8:
+                r1ptg = 4; break;
+            case 5:
+                r1ptg = 5; break;
+            default:
+                WSP_GGML_ABORT("unsupported ne11");
+        };
+
+        wsp_ggml_metal_pipeline_t pipeline = wsp_ggml_metal_library_get_pipeline_mul_mv_ext(lib, op->src[0]->type, op->src[1]->type, nsg, nxpsg, r1ptg);
+
+        wsp_ggml_metal_kargs_mul_mv_ext args = {
+            /*.ne00  =*/ ne00,
+            /*.ne01  =*/ ne01,
+            /*.ne02  =*/ ne02,
+            /*.nb00  =*/ nb00,
+            /*.nb01  =*/ nb01,
+            /*.nb02  =*/ nb02,
+            /*.nb03  =*/ nb03,
+            /*.ne10  =*/ ne10,
+            /*.ne11  =*/ ne11,
+            /*.ne12  =*/ ne12,
+            /*.nb10  =*/ nb10,
+            /*.nb11  =*/ nb11,
+            /*.nb12  =*/ nb12,
+            /*.nb13  =*/ nb13,
+            /*.ne0   =*/ ne0,
+            /*.ne1   =*/ ne1,
+            /*.r2    =*/ r2,
+            /*.r3    =*/ r3,
+        };
+
+        wsp_ggml_metal_encoder_set_pipeline(enc, pipeline);
+        wsp_ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+        wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[0]), 1);
+        wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[1]), 2);
+        wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op),         3);
+
+        wsp_ggml_metal_encoder_dispatch_threadgroups(enc, ((ne01 + r0ptg - 1)/r0ptg), ((ne11 + r1ptg - 1)/r1ptg), ne12*ne13, 32, nsg, 1);
+    } else if (
+        !wsp_ggml_is_transposed(op->src[0]) &&
+        !wsp_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) {
+        //WSP_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)
+        //switch (op->src[0]->type) {
+        //    case WSP_GGML_TYPE_F32:  WSP_GGML_ASSERT(nb01 % 16 == 0); break;
+        //    case WSP_GGML_TYPE_F16:  WSP_GGML_ASSERT(nb01 % 8  == 0); break;
+        //    case WSP_GGML_TYPE_BF16: WSP_GGML_ASSERT(nb01 % 8  == 0); break;
+        //    default: break;
+        //}
+
+        wsp_ggml_metal_pipeline_t pipeline = wsp_ggml_metal_library_get_pipeline_mul_mm(lib, op);
+
+        wsp_ggml_metal_kargs_mul_mm args = {
+            /*.ne00 =*/ ne00,
+            /*.ne02 =*/ ne02,
+            /*.nb01 =*/ nb01,
+            /*.nb02 =*/ nb02,
+            /*.nb03 =*/ nb03,
+            /*.ne12 =*/ ne12,
+            /*.nb10 =*/ nb10,
+            /*.nb11 =*/ nb11,
+            /*.nb12 =*/ nb12,
+            /*.nb13 =*/ nb13,
+            /*.ne0  =*/ ne0,
+            /*.ne1  =*/ ne1,
+            /*.r2   =*/ r2,
+            /*.r3   =*/ r3,
+        };
+
+        wsp_ggml_metal_encoder_set_pipeline(enc, pipeline);
+        wsp_ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+        wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[0]), 1);
+        wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[1]), 2);
+        wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op),         3);
+
+        const size_t smem = wsp_ggml_metal_pipeline_get_smem(pipeline);
+
+        wsp_ggml_metal_encoder_set_threadgroup_memory_size(enc, smem, 0);
+        wsp_ggml_metal_encoder_dispatch_threadgroups(enc, ((ne11 + 31)/32), ((ne01 + 63)/64), ne12*ne13, 128, 1, 1);
+    } else {
+        wsp_ggml_metal_pipeline_t pipeline = wsp_ggml_metal_library_get_pipeline_mul_mv(lib, op);
+
+        const int nr0 = wsp_ggml_metal_pipeline_get_nr0(pipeline);
+        const int nr1 = wsp_ggml_metal_pipeline_get_nr1(pipeline);
+        const int nsg = wsp_ggml_metal_pipeline_get_nsg(pipeline);
+
+        const size_t smem = wsp_ggml_metal_pipeline_get_smem(pipeline);
+
+        wsp_ggml_metal_kargs_mul_mv args = {
+            /*.ne00 =*/ ne00,
+            /*.ne01 =*/ ne01,
+            /*.ne02 =*/ ne02,
+            /*.nb00 =*/ nb00,
+            /*.nb01 =*/ nb01,
+            /*.nb02 =*/ nb02,
+            /*.nb03 =*/ nb03,
+            /*.ne10 =*/ ne10,
+            /*.ne11 =*/ ne11,
+            /*.ne12 =*/ ne12,
+            /*.nb10 =*/ nb10,
+            /*.nb11 =*/ nb11,
+            /*.nb12 =*/ nb12,
+            /*.nb13 =*/ nb13,
+            /*.ne0  =*/ ne0,
+            /*.ne1  =*/ ne1,
+            /*.nr0  =*/ nr0,
+            /*.r2   =*/ r2,
+            /*.r3   =*/ r3,
+        };
+
+        wsp_ggml_metal_encoder_set_pipeline(enc, pipeline);
+        wsp_ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+        wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[0]), 1);
+        wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[1]), 2);
+        wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op),         3);
+
+        wsp_ggml_metal_encoder_set_threadgroup_memory_size(enc, smem, 0);
+
+        if (op->src[0]->type == WSP_GGML_TYPE_F32 ||
+            op->src[0]->type == WSP_GGML_TYPE_F16 ||
+            op->src[0]->type == WSP_GGML_TYPE_BF16 ||
+            op->src[0]->type == WSP_GGML_TYPE_Q8_0) {
+            wsp_ggml_metal_encoder_dispatch_threadgroups(enc, ((ne01 + nr0 - 1)/(nr0)), ((ne11 + nr1 - 1)/nr1), ne12*ne13, 32, nsg, 1);
+        } else {
+            wsp_ggml_metal_encoder_dispatch_threadgroups(enc, ((ne01 + nr0*nsg - 1)/(nr0*nsg)), ((ne11 + nr1 - 1)/nr1), ne12*ne13, 32, nsg, 1);
+        }
+    }
+
+    return 1;
+}
+
+size_t wsp_ggml_metal_op_mul_mat_id_extra_tpe(const wsp_ggml_tensor * op) {
+    assert(op->op == WSP_GGML_OP_MUL_MAT_ID);
+
+    const int64_t ne02 = op->src[0]->ne[2]; // n_expert
+
+    return wsp_ggml_type_size(WSP_GGML_TYPE_I32)*ne02;
+}
+
+size_t wsp_ggml_metal_op_mul_mat_id_extra_ids(const wsp_ggml_tensor * op) {
+    assert(op->op == WSP_GGML_OP_MUL_MAT_ID);
+
+    const int64_t ne02 = op->src[0]->ne[2]; // n_expert
+    const int64_t ne21 = op->src[2]->ne[1]; // n_token
+
+    return wsp_ggml_type_size(WSP_GGML_TYPE_I32)*ne02*ne21;
+}
+
+int wsp_ggml_metal_op_mul_mat_id(wsp_ggml_metal_op_t ctx, int idx) {
+    wsp_ggml_tensor * op = ctx->node(idx);
+
+    wsp_ggml_metal_library_t lib = ctx->lib;
+    wsp_ggml_metal_encoder_t enc = ctx->enc;
+
+    const wsp_ggml_metal_device_props * props_dev = wsp_ggml_metal_device_get_props(ctx->dev);
+
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne2, op->src[2], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb2, op->src[2], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
+    WSP_GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+
+    // src2 = ids
+    WSP_GGML_ASSERT(op->src[2]->type == WSP_GGML_TYPE_I32);
+
+    WSP_GGML_ASSERT(!wsp_ggml_is_transposed(op->src[0]));
+    WSP_GGML_ASSERT(!wsp_ggml_is_transposed(op->src[1]));
+
+    WSP_GGML_ASSERT(ne03 == 1);
+    WSP_GGML_ASSERT(ne13 == 1);
+
+    wsp_ggml_metal_buffer_id bid_src0 = wsp_ggml_metal_get_buffer_id(op->src[0]);
+    wsp_ggml_metal_buffer_id bid_src1 = wsp_ggml_metal_get_buffer_id(op->src[1]);
+    wsp_ggml_metal_buffer_id bid_src2 = wsp_ggml_metal_get_buffer_id(op->src[2]);
+    wsp_ggml_metal_buffer_id bid_dst  = wsp_ggml_metal_get_buffer_id(op);
+
+    const uint32_t r2 = 1;
+    const uint32_t r3 = 1;
+
+    // find the break-even point where the matrix-matrix kernel becomes more efficient compared
+    // to the matrix-vector kernel
+    // ne20 = n_used_experts
+    // ne21 = n_rows (batch size)
+    const int ne21_mm_id_min = 32;
+
+    if (props_dev->has_simdgroup_mm && ne00 >= 64 && (ne21 >= ne21_mm_id_min)) {
+        // some Metal matrix data types require aligned pointers
+        // ref: https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf (Table 2.5)
+        //switch (op->src[0]->type) {
+        //    case WSP_GGML_TYPE_F32:  WSP_GGML_ASSERT(nb01 % 16 == 0); break;
+        //    case WSP_GGML_TYPE_F16:  WSP_GGML_ASSERT(nb01 % 8  == 0); break;
+        //    case WSP_GGML_TYPE_BF16: WSP_GGML_ASSERT(nb01 % 8  == 0); break;
+        //    default: break;
+        //}
+
+        // extra buffers for intermediate id mapping
+        wsp_ggml_metal_buffer_id bid_tpe = bid_dst;
+        bid_tpe.offs += wsp_ggml_nbytes(op);
+
+        wsp_ggml_metal_buffer_id bid_ids = bid_tpe;
+        bid_ids.offs += wsp_ggml_metal_op_mul_mat_id_extra_tpe(op);
+
+        {
+            wsp_ggml_metal_kargs_mul_mm_id_map0 args = {
+                ne02,
+                ne10,
+                ne11, // n_expert_used (bcast)
+                nb11,
+                nb12,
+                ne21, // n_tokens
+                ne20, // n_expert_used
+                nb21,
+            };
+
+            wsp_ggml_metal_pipeline_t pipeline = wsp_ggml_metal_library_get_pipeline_mul_mm_id_map0(lib, ne02, ne20);
+
+            const size_t smem = wsp_ggml_metal_pipeline_get_smem(pipeline);
+
+            WSP_GGML_ASSERT(ne02 <= wsp_ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));
+
+            WSP_GGML_ASSERT(smem <= props_dev->max_theadgroup_memory_size);
+
+            wsp_ggml_metal_encoder_set_pipeline(enc, pipeline);
+            wsp_ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+            wsp_ggml_metal_encoder_set_buffer  (enc, bid_src2, 1);
+            wsp_ggml_metal_encoder_set_buffer  (enc, bid_tpe,  2);
+            wsp_ggml_metal_encoder_set_buffer  (enc, bid_ids,  3);
+
+            wsp_ggml_metal_encoder_set_threadgroup_memory_size(enc, smem, 0);
+
+            wsp_ggml_metal_encoder_dispatch_threadgroups(enc, 1, 1, 1, ne02, 1, 1);
+        }
+
+        // this barrier is always needed because the next kernel has to wait for the id maps to be computed
+        wsp_ggml_metal_op_concurrency_reset(ctx);
+
+        {
+            wsp_ggml_metal_pipeline_t pipeline = wsp_ggml_metal_library_get_pipeline_mul_mm_id(lib, op);
+
+            wsp_ggml_metal_kargs_mul_mm_id args = {
+                /*.ne00  =*/ ne00,
+                /*.ne02  =*/ ne02,
+                /*.nb01  =*/ nb01,
+                /*.nb02  =*/ nb02,
+                /*.nb03  =*/ nb03,
+                /*.ne11  =*/ ne11, // n_expert_used (bcast)
+                /*.nb10  =*/ nb10,
+                /*.nb11  =*/ nb11,
+                /*.nb12  =*/ nb12,
+                /*.nb13  =*/ nb13,
+                /*.ne20  =*/ ne20, // n_expert_used
+                /*.ne21  =*/ ne21, // n_tokens
+                /*.ne0   =*/ ne0,
+                /*.ne1   =*/ ne1,
+                /*.r2    =*/ r2,
+                /*.r3    =*/ r3,
+            };
+
+            wsp_ggml_metal_encoder_set_pipeline(enc, pipeline);
+            wsp_ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+            wsp_ggml_metal_encoder_set_buffer  (enc, bid_src0, 1);
+            wsp_ggml_metal_encoder_set_buffer  (enc, bid_src1, 2);
+            wsp_ggml_metal_encoder_set_buffer  (enc, bid_tpe,  3);
+            wsp_ggml_metal_encoder_set_buffer  (enc, bid_ids,  4);
+            wsp_ggml_metal_encoder_set_buffer  (enc, bid_dst,  5);
+
+            const size_t smem = wsp_ggml_metal_pipeline_get_smem(pipeline);
+
+            wsp_ggml_metal_encoder_set_threadgroup_memory_size(enc, smem, 0);
+
+            wsp_ggml_metal_encoder_dispatch_threadgroups(enc, (ne21 + 31)/32, (ne01 + 63)/64, ne02, 128, 1, 1);
+        }
+    } else {
+        wsp_ggml_metal_pipeline_t pipeline = wsp_ggml_metal_library_get_pipeline_mul_mv_id(lib, op);
+
+        const int nr0 = wsp_ggml_metal_pipeline_get_nr0(pipeline);
+        const int nr1 = wsp_ggml_metal_pipeline_get_nr1(pipeline);
+        const int nsg = wsp_ggml_metal_pipeline_get_nsg(pipeline);
+
+        const size_t smem = wsp_ggml_metal_pipeline_get_smem(pipeline);
+
+        wsp_ggml_metal_kargs_mul_mv_id args = {
+            /*.nei0 =*/ ne20,
+            /*.nei1 =*/ ne21,
+            /*.nbi1 =*/ nb21,
+            /*.ne00 =*/ ne00,
+            /*.ne01 =*/ ne01,
+            /*.ne02 =*/ ne02,
+            /*.nb00 =*/ nb00,
+            /*.nb01 =*/ nb01,
+            /*.nb02 =*/ nb02,
+            /*.ne10 =*/ ne10,
+            /*.ne11 =*/ ne11,
+            /*.ne12 =*/ ne12,
+            /*.ne13 =*/ ne13,
+            /*.nb10 =*/ nb10,
+            /*.nb11 =*/ nb11,
+            /*.nb12 =*/ nb12,
+            /*.ne0  =*/ ne0,
+            /*.ne1  =*/ ne1,
+            /*.nb1  =*/ nb1,
+            /*.nr0  =*/ nr0,
+        };
+
+        if (wsp_ggml_is_quantized(op->src[0]->type)) {
+            WSP_GGML_ASSERT(ne00 >= nsg*nr0);
+        }
+
+        wsp_ggml_metal_encoder_set_pipeline(enc, pipeline);
+        wsp_ggml_metal_encoder_set_bytes(enc, &args, sizeof(args), 0);
+        wsp_ggml_metal_encoder_set_buffer(enc, bid_src0, 1);
+        wsp_ggml_metal_encoder_set_buffer(enc, bid_src1, 2);
+        wsp_ggml_metal_encoder_set_buffer(enc, bid_dst,  3);
+        wsp_ggml_metal_encoder_set_buffer(enc, bid_src2, 4);
+
+        const int64_t _ne1 = 1;
+        const int64_t ne123 = ne20*ne21;
+
+        wsp_ggml_metal_encoder_set_threadgroup_memory_size(enc, smem, 0);
+
+        if (op->src[0]->type == WSP_GGML_TYPE_F32 ||
+            op->src[0]->type == WSP_GGML_TYPE_F16 ||
+            op->src[0]->type == WSP_GGML_TYPE_BF16 ||
+            op->src[0]->type == WSP_GGML_TYPE_Q8_0) {
+            wsp_ggml_metal_encoder_dispatch_threadgroups(enc, (ne01 + nr0 - 1)/(nr0), (_ne1 + nr1 - 1)/nr1, ne123, 32, nsg, 1);
+        } else {
+            wsp_ggml_metal_encoder_dispatch_threadgroups(enc, (ne01 + nr0*nsg - 1)/(nr0*nsg), (_ne1 + nr1 - 1)/nr1, ne123, 32, nsg, 1);
+        }
+    }
+
+    return 1;
+}
+
+int wsp_ggml_metal_op_add_id(wsp_ggml_metal_op_t ctx, int idx) {
+    wsp_ggml_tensor * op = ctx->node(idx);
+
+    wsp_ggml_metal_library_t lib = ctx->lib;
+    wsp_ggml_metal_encoder_t enc = ctx->enc;
+
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne2, op->src[2], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb2, op->src[2], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
+
+    WSP_GGML_ASSERT(op->src[0]->type == WSP_GGML_TYPE_F32);
+    WSP_GGML_ASSERT(op->src[1]->type == WSP_GGML_TYPE_F32);
+    WSP_GGML_ASSERT(op->src[2]->type == WSP_GGML_TYPE_I32);
+    WSP_GGML_ASSERT(op->type         == WSP_GGML_TYPE_F32);
+
+    WSP_GGML_ASSERT(wsp_ggml_is_contiguous_rows(op->src[0]));
+
+    wsp_ggml_metal_kargs_add_id args = {
+        /*.ne0  =*/ ne0,
+        /*.ne1  =*/ ne1,
+        /*.nb01 =*/ nb01,
+        /*.nb02 =*/ nb02,
+        /*.nb11 =*/ nb11,
+        /*.nb21 =*/ nb21,
+    };
+
+    wsp_ggml_metal_pipeline_t pipeline = wsp_ggml_metal_library_get_pipeline_base(lib, WSP_GGML_OP_ADD_ID);
+
+    wsp_ggml_metal_encoder_set_pipeline(enc, pipeline);
+    wsp_ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[0]), 1);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[1]), 2);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[2]), 3);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op),         4);
+
+    const int nth = std::min(wsp_ggml_metal_pipeline_max_theads_per_threadgroup(pipeline), ne00);
+
+    wsp_ggml_metal_encoder_dispatch_threadgroups(enc, ne01, ne02, 1, nth, 1, 1);
+
+    return 1;
+}
+
+bool wsp_ggml_metal_op_flash_attn_ext_use_vec(const wsp_ggml_tensor * op) {
+    assert(op->op == WSP_GGML_OP_FLASH_ATTN_EXT);
+
+    const int64_t ne00 = op->src[0]->ne[0]; // head size
+    const int64_t ne01 = op->src[0]->ne[1]; // batch size
+
+    // use vec kernel if the batch size is small and if the head size is supported
+    return (ne01 < 20) && (ne00 % 32 == 0);
+}
+
+size_t wsp_ggml_metal_op_flash_attn_ext_extra_pad(const wsp_ggml_tensor * op) {
+    assert(op->op == WSP_GGML_OP_FLASH_ATTN_EXT);
+
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne2, op->src[2], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb2, op->src[2], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne3, op->src[3], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb3, op->src[3], nb);
+
+    size_t res = 0;
+
+    const bool has_mask = op->src[3] != nullptr;
+
+    if (wsp_ggml_metal_op_flash_attn_ext_use_vec(op)) {
+        const bool has_kvpad = ne11 % OP_FLASH_ATTN_EXT_VEC_NCPSG != 0;
+
+        if (has_kvpad) {
+            res += OP_FLASH_ATTN_EXT_VEC_NCPSG*(
+                nb11*ne12*ne13 +
+                nb21*ne22*ne23 +
+                (has_mask ? wsp_ggml_type_size(WSP_GGML_TYPE_F16)*ne31*ne32*ne33 : 0));
+        }
+    } else {
+        const bool has_kvpad = ne11 % OP_FLASH_ATTN_EXT_NCPSG != 0;
+
+        if (has_kvpad) {
+            res += OP_FLASH_ATTN_EXT_NCPSG*(
+                nb11*ne12*ne13 +
+                nb21*ne22*ne23 +
+                (has_mask ? wsp_ggml_type_size(WSP_GGML_TYPE_F16)*ne31*ne32*ne33 : 0));
+        }
+    }
+
+    return res;
+}
+
+size_t wsp_ggml_metal_op_flash_attn_ext_extra_blk(const wsp_ggml_tensor * op) {
+    assert(op->op == WSP_GGML_OP_FLASH_ATTN_EXT);
+
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+  //WSP_GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+  //WSP_GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
+  //WSP_GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
+  //WSP_GGML_TENSOR_LOCALS( int32_t, ne2, op->src[2], ne);
+  //WSP_GGML_TENSOR_LOCALS(uint64_t, nb2, op->src[2], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne3, op->src[3], ne);
+    WSP_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 = wsp_ggml_metal_op_flash_attn_ext_use_vec(op);
+
+    // this optimization is not useful for the vector kernels
+    if (is_vec) {
+        return res;
+    }
+
+    const int nqptg = is_vec ? OP_FLASH_ATTN_EXT_VEC_NQPTG : OP_FLASH_ATTN_EXT_NQPTG;
+    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 += WSP_GGML_PAD(wsp_ggml_type_size(WSP_GGML_TYPE_I8)*ne0*ne1*ne32*ne33, 32);
+
+    return res;
+}
+
+size_t wsp_ggml_metal_op_flash_attn_ext_extra_tmp(const wsp_ggml_tensor * op) {
+    assert(op->op == WSP_GGML_OP_FLASH_ATTN_EXT);
+
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+  //WSP_GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
+  //WSP_GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne2, op->src[2], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb2, op->src[2], nb);
+  //WSP_GGML_TENSOR_LOCALS( int32_t, ne3, op->src[3], ne);
+  //WSP_GGML_TENSOR_LOCALS(uint64_t, nb3, op->src[3], nb);
+
+    size_t res = 0;
+
+    if (wsp_ggml_metal_op_flash_attn_ext_use_vec(op)) {
+        const int64_t nwg = 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
+        res += wsp_ggml_type_size(WSP_GGML_TYPE_F32)*(ne01*ne02*ne03*nwg*(ne20 + 2));
+    }
+
+    return res;
+}
+
+int wsp_ggml_metal_op_flash_attn_ext(wsp_ggml_metal_op_t ctx, int idx) {
+    wsp_ggml_tensor * op = ctx->node(idx);
+
+    wsp_ggml_metal_library_t lib = ctx->lib;
+    wsp_ggml_metal_encoder_t enc = ctx->enc;
+
+    const wsp_ggml_metal_device_props * props_dev = wsp_ggml_metal_device_get_props(ctx->dev);
+
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne2, op->src[2], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb2, op->src[2], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne3, op->src[3], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb3, op->src[3], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
+    WSP_GGML_TENSOR_LOCALS( int32_t, nb,  op,         nb);
+
+    WSP_GGML_ASSERT(ne00 % 4 == 0);
+
+    WSP_GGML_ASSERT(op->src[0]->type == WSP_GGML_TYPE_F32);
+    WSP_GGML_ASSERT(op->src[1]->type == op->src[2]->type);
+
+    //WSP_GGML_ASSERT(wsp_ggml_are_same_shape (src1, src2));
+    WSP_GGML_ASSERT(ne11 == ne21);
+    WSP_GGML_ASSERT(ne12 == ne22);
+
+    WSP_GGML_ASSERT(!op->src[3] || op->src[3]->type == WSP_GGML_TYPE_F16);
+    WSP_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;
+    float logit_softcap;
+
+    memcpy(&scale,         ((const int32_t *) op->op_params) + 0, sizeof(scale));
+    memcpy(&max_bias,      ((const int32_t *) op->op_params) + 1, sizeof(max_bias));
+    memcpy(&logit_softcap, ((const int32_t *) op->op_params) + 2, sizeof(logit_softcap));
+
+    if (logit_softcap != 0.0f) {
+        scale /= logit_softcap;
+    }
+
+    const bool has_mask  = op->src[3] != NULL;
+    const bool has_sinks = op->src[4] != NULL;
+    const bool has_bias  = max_bias != 0.0f;
+    const bool has_scap  = logit_softcap != 0.0f;
+
+    const uint32_t n_head      = op->src[0]->ne[2];
+    const  int32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head));
+
+    const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
+    const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
+
+    WSP_GGML_ASSERT(ne01 < 65536);
+
+    wsp_ggml_metal_buffer_id bid_src0 = wsp_ggml_metal_get_buffer_id(op->src[0]);
+    wsp_ggml_metal_buffer_id bid_src1 = wsp_ggml_metal_get_buffer_id(op->src[1]);
+    wsp_ggml_metal_buffer_id bid_src2 = wsp_ggml_metal_get_buffer_id(op->src[2]);
+    wsp_ggml_metal_buffer_id bid_src3 = has_mask  ? wsp_ggml_metal_get_buffer_id(op->src[3]) : bid_src0;
+    wsp_ggml_metal_buffer_id bid_src4 = has_sinks ? wsp_ggml_metal_get_buffer_id(op->src[4]) : bid_src0;
+
+    wsp_ggml_metal_buffer_id bid_dst = wsp_ggml_metal_get_buffer_id(op);
+
+    wsp_ggml_metal_buffer_id bid_pad = bid_dst;
+    bid_pad.offs += wsp_ggml_nbytes(op);
+
+    wsp_ggml_metal_buffer_id bid_blk = bid_pad;
+    bid_blk.offs += wsp_ggml_metal_op_flash_attn_ext_extra_pad(op);
+
+    wsp_ggml_metal_buffer_id bid_tmp = bid_blk;
+    bid_tmp.offs += wsp_ggml_metal_op_flash_attn_ext_extra_blk(op);
+
+    if (!wsp_ggml_metal_op_flash_attn_ext_use_vec(op)) {
+        // half8x8 kernel
+        const int nqptg = OP_FLASH_ATTN_EXT_NQPTG; // queries per threadgroup
+        const int ncpsg = OP_FLASH_ATTN_EXT_NCPSG; // cache values per simdgroup
+
+        WSP_GGML_ASSERT(nqptg <= 32);
+        WSP_GGML_ASSERT(nqptg  % 8  == 0);
+        WSP_GGML_ASSERT(ncpsg  % 32 == 0);
+
+        bool need_sync = false;
+
+        const bool has_kvpad = ne11 % ncpsg != 0;
+
+        if (has_kvpad) {
+            assert(wsp_ggml_metal_op_flash_attn_ext_extra_pad(op) != 0);
+
+            wsp_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,
+            };
+
+            wsp_ggml_metal_pipeline_t pipeline0 = wsp_ggml_metal_library_get_pipeline_flash_attn_ext_pad(lib, op, has_mask, ncpsg);
+
+            wsp_ggml_metal_encoder_set_pipeline(enc, pipeline0);
+            wsp_ggml_metal_encoder_set_bytes   (enc, &args0, sizeof(args0), 0);
+            wsp_ggml_metal_encoder_set_buffer  (enc, bid_src1, 1);
+            wsp_ggml_metal_encoder_set_buffer  (enc, bid_src2, 2);
+            wsp_ggml_metal_encoder_set_buffer  (enc, bid_src3, 3);
+            wsp_ggml_metal_encoder_set_buffer  (enc, bid_pad,  4);
+
+            assert(ne12 == ne22);
+            assert(ne13 == ne23);
+
+            wsp_ggml_metal_encoder_dispatch_threadgroups(enc, ncpsg, std::max(ne12, ne32), std::max(ne13, ne33), 32, 1, 1);
+
+            need_sync = true;
+        } else {
+            assert(wsp_ggml_metal_op_flash_attn_ext_extra_pad(op) == 0);
+        }
+
+        if (has_mask) {
+            assert(wsp_ggml_metal_op_flash_attn_ext_extra_blk(op) != 0);
+
+            wsp_ggml_metal_kargs_flash_attn_ext_blk args0 = {
+                /*.ne01 =*/ ne01,
+                /*.ne30 =*/ ne30,
+                /*.ne31 =*/ ne31,
+                /*.ne32 =*/ ne32,
+                /*.ne33 =*/ ne33,
+                /*.nb31 =*/ nb31,
+                /*.nb32 =*/ nb32,
+                /*.nb33 =*/ nb33,
+            };
+
+            wsp_ggml_metal_pipeline_t pipeline0 = wsp_ggml_metal_library_get_pipeline_flash_attn_ext_blk(lib, op, nqptg, ncpsg);
+
+            wsp_ggml_metal_encoder_set_pipeline(enc, pipeline0);
+            wsp_ggml_metal_encoder_set_bytes   (enc, &args0, sizeof(args0), 0);
+            wsp_ggml_metal_encoder_set_buffer  (enc, bid_src3, 1);
+            wsp_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);
+
+            wsp_ggml_metal_encoder_dispatch_threadgroups(enc, nblk0, nblk1, ne32*ne33, 32, 1, 1);
+
+            need_sync = true;
+        } else {
+            assert(wsp_ggml_metal_op_flash_attn_ext_extra_blk(op) == 0);
+        }
+
+        if (need_sync) {
+            wsp_ggml_metal_op_concurrency_reset(ctx);
+        }
+
+        const int is_q = wsp_ggml_is_quantized(op->src[1]->type) ? 1 : 0;
+
+        // 2*(2*ncpsg)
+        // ncpsg soft_max values + ncpsg mask values
+        //
+        // 16*32*(nsg)
+        // the shared memory needed for the simdgroups to load the KV cache
+        // each thread loads (dequantizes) 16 head elements, there are 32 threads in th SG
+        //
+#define FATTN_SMEM(nsg) (WSP_GGML_PAD((nqptg*(ne00 + 2*WSP_GGML_PAD(ne20, 64) + 2*(2*ncpsg)) + is_q*(16*32*(nsg)))*(sizeof(float)/2), 16))
+
+        //int64_t nsgmax = 4;
+        //
+        //if (is_q) {
+        //    nsgmax = 2;
+        //    while (true) {
+        //        const size_t smem = FATTN_SMEM(nsgmax);
+        //        if (smem > props_dev->max_theadgroup_memory_size) {
+        //            break;
+        //        }
+        //        nsgmax *= 2;
+        //    }
+        //    nsgmax /= 2;
+        //}
+
+        // 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;
+
+        const size_t smem = FATTN_SMEM(nsg);
+
+        wsp_ggml_metal_kargs_flash_attn_ext args = {
+            /*.ne01          =*/ ne01,
+            /*.ne02          =*/ ne02,
+            /*.ne03          =*/ ne03,
+            /*.nb01          =*/ nb01,
+            /*.nb02          =*/ nb02,
+            /*.nb03          =*/ nb03,
+            /*.ne11          =*/ ne11,
+            /*.ne_12_2       =*/ ne12,
+            /*.ne_12_3       =*/ ne13,
+            /*.ns10          =*/ int32_t(nb11/nb10),
+            /*.nb11          =*/ nb11,
+            /*.nb12          =*/ nb12,
+            /*.nb13          =*/ nb13,
+            /*.ns20          =*/ int32_t(nb21/nb20),
+            /*.nb21          =*/ nb21,
+            /*.nb22          =*/ nb22,
+            /*.nb23          =*/ nb23,
+            /*.ne31          =*/ ne31,
+            /*.ne32          =*/ ne32,
+            /*.ne33          =*/ ne33,
+            /*.nb31          =*/ nb31,
+            /*.nb32          =*/ nb32,
+            /*.nb33          =*/ nb33,
+            /*.ne1           =*/ ne1,
+            /*.ne2           =*/ ne2,
+            /*.ne3           =*/ ne3,
+            /*.scale         =*/ scale,
+            /*.max_bias      =*/ max_bias,
+            /*.m0            =*/ m0,
+            /*.m1            =*/ m1,
+            /*.n_head_log2   =*/ n_head_log2,
+            /*.logit_softcap =*/ logit_softcap,
+        };
+
+        wsp_ggml_metal_pipeline_t pipeline = wsp_ggml_metal_library_get_pipeline_flash_attn_ext(lib, op, has_mask, has_sinks, has_bias, has_scap, has_kvpad, nsg);
+
+        wsp_ggml_metal_encoder_set_pipeline(enc, pipeline);
+        wsp_ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+        wsp_ggml_metal_encoder_set_buffer  (enc, bid_src0, 1);
+        wsp_ggml_metal_encoder_set_buffer  (enc, bid_src1, 2);
+        wsp_ggml_metal_encoder_set_buffer  (enc, bid_src2, 3);
+        wsp_ggml_metal_encoder_set_buffer  (enc, bid_src3, 4);
+        wsp_ggml_metal_encoder_set_buffer  (enc, bid_src4, 5);
+        wsp_ggml_metal_encoder_set_buffer  (enc, bid_pad,  6);
+        wsp_ggml_metal_encoder_set_buffer  (enc, bid_blk,  7);
+        wsp_ggml_metal_encoder_set_buffer  (enc, bid_dst,  8);
+
+        wsp_ggml_metal_encoder_set_threadgroup_memory_size(enc, smem, 0);
+
+        wsp_ggml_metal_encoder_dispatch_threadgroups(enc, (ne01 + nqptg - 1)/nqptg, ne02, ne03, 32, nsg, 1);
+#undef FATTN_SMEM
+    } else {
+        // half4x4 kernel
+        const int nqptg = OP_FLASH_ATTN_EXT_VEC_NQPTG; // queries per threadgroup
+        const int ncpsg = OP_FLASH_ATTN_EXT_VEC_NCPSG; // cache values per simdgroup !! sync with kernel template arguments !!
+        const int nkpsg = 1*ncpsg;
+
+        WSP_GGML_ASSERT(nqptg <= 32);
+        WSP_GGML_ASSERT(nqptg  % 1  == 0);
+        WSP_GGML_ASSERT(ncpsg  % 32 == 0);
+
+        bool need_sync = false;
+
+        const bool has_kvpad = ne11 % ncpsg != 0;
+
+        if (has_kvpad) {
+            assert(wsp_ggml_metal_op_flash_attn_ext_extra_pad(op) != 0);
+
+            wsp_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,
+            };
+
+            wsp_ggml_metal_pipeline_t pipeline0 = wsp_ggml_metal_library_get_pipeline_flash_attn_ext_pad(lib, op, has_mask, ncpsg);
+
+            wsp_ggml_metal_encoder_set_pipeline(enc, pipeline0);
+            wsp_ggml_metal_encoder_set_bytes   (enc, &args0, sizeof(args0), 0);
+            wsp_ggml_metal_encoder_set_buffer  (enc, bid_src1, 1);
+            wsp_ggml_metal_encoder_set_buffer  (enc, bid_src2, 2);
+            wsp_ggml_metal_encoder_set_buffer  (enc, bid_src3, 3);
+            wsp_ggml_metal_encoder_set_buffer  (enc, bid_pad,  4);
+
+            assert(ne12 == ne22);
+            assert(ne13 == ne23);
+
+            wsp_ggml_metal_encoder_dispatch_threadgroups(enc, ncpsg, std::max(ne12, ne32), std::max(ne13, ne33), 32, 1, 1);
+
+            need_sync = true;
+        } else {
+            assert(wsp_ggml_metal_op_flash_attn_ext_extra_pad(op) == 0);
+        }
+
+        if (need_sync) {
+            wsp_ggml_metal_op_concurrency_reset(ctx);
+        }
+
+        // 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
+        //
+        // ne20*(nsg)
+        // each simdgroup has a full f32 head vector in shared mem to accumulate results
+        //
+#define FATTN_SMEM(nsg) (WSP_GGML_PAD((nqptg*(WSP_GGML_PAD(ne00, 128) + 4*ncpsg*(nsg)) + 2*WSP_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)));
+
+        int64_t nsg = 1;
+        while (nsg <= nsgt) {
+            nsg *= 2;
+        }
+        nsg /= 2;
+
+        // workgroups
+        // each workgroup handles nsg*nkpsg cache values
+        int32_t nwg = 1;
+        if (false) {
+            // for small KV caches, we could launch a single workgroup and write the results directly to dst/
+            // however, this does not lead to significant improvement, so disabled
+            nwg = 1;
+            nsg = 4;
+        } else {
+            nwg = 32;
+            nsg = 1;
+            while (2*nwg*nsg*nkpsg < ne11 && nsg < 4) {
+                nsg *= 2;
+            }
+        }
+
+        wsp_ggml_metal_kargs_flash_attn_ext_vec args = {
+            /*.ne01          =*/ ne01,
+            /*.ne02          =*/ ne02,
+            /*.ne03          =*/ ne03,
+            /*.nb01          =*/ nb01,
+            /*.nb02          =*/ nb02,
+            /*.nb03          =*/ nb03,
+            /*.ne11          =*/ ne11,
+            /*.ne_12_2       =*/ ne12,
+            /*.ne_12_3       =*/ ne13,
+            /*.ns10          =*/ int32_t(nb11/nb10),
+            /*.nb11          =*/ nb11,
+            /*.nb12          =*/ nb12,
+            /*.nb13          =*/ nb13,
+            /*.ns20          =*/ int32_t(nb21/nb20),
+            /*.nb21          =*/ nb21,
+            /*.nb22          =*/ nb22,
+            /*.nb23          =*/ nb23,
+            /*.ne31          =*/ ne31,
+            /*.ne32          =*/ ne32,
+            /*.ne33          =*/ ne33,
+            /*.nb31          =*/ nb31,
+            /*.nb32          =*/ nb32,
+            /*.nb33          =*/ nb33,
+            /*.ne1           =*/ ne1,
+            /*.ne2           =*/ ne2,
+            /*.ne3           =*/ ne3,
+            /*.scale         =*/ scale,
+            /*.max_bias      =*/ max_bias,
+            /*.m0            =*/ m0,
+            /*.m1            =*/ m1,
+            /*.n_head_log2   =*/ n_head_log2,
+            /*.logit_softcap =*/ logit_softcap,
+        };
+
+        wsp_ggml_metal_pipeline_t pipeline = wsp_ggml_metal_library_get_pipeline_flash_attn_ext_vec(lib, op, has_mask, has_sinks, has_bias, has_scap, has_kvpad, nsg, nwg);
+
+        WSP_GGML_ASSERT(nsg*32 <= wsp_ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));
+
+        wsp_ggml_metal_encoder_set_pipeline(enc, pipeline);
+        wsp_ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+        wsp_ggml_metal_encoder_set_buffer  (enc, bid_src0, 1);
+        wsp_ggml_metal_encoder_set_buffer  (enc, bid_src1, 2);
+        wsp_ggml_metal_encoder_set_buffer  (enc, bid_src2, 3);
+        wsp_ggml_metal_encoder_set_buffer  (enc, bid_src3, 4);
+        wsp_ggml_metal_encoder_set_buffer  (enc, bid_src4, 5);
+
+        const size_t smem = FATTN_SMEM(nsg);
+
+        //printf("smem: %zu, max: %zu, nsg = %d, nsgmax = %d\n", smem, props_dev->max_theadgroup_memory_size, (int) nsg, (int) nsgmax);
+        WSP_GGML_ASSERT(smem <= props_dev->max_theadgroup_memory_size);
+
+        if (nwg == 1) {
+            assert(wsp_ggml_metal_op_flash_attn_ext_extra_tmp(op) == 0);
+
+            // using 1 workgroup -> write the result directly into dst
+            wsp_ggml_metal_encoder_set_buffer(enc, bid_pad, 6);
+            wsp_ggml_metal_encoder_set_buffer(enc, bid_dst, 7);
+
+            wsp_ggml_metal_encoder_set_threadgroup_memory_size(enc, smem, 0);
+
+            wsp_ggml_metal_encoder_dispatch_threadgroups(enc, (ne01 + nqptg - 1)/nqptg, ne02, ne03*nwg, 32, nsg, 1);
+        } else {
+            // sanity checks
+            assert(wsp_ggml_metal_op_flash_attn_ext_extra_tmp(op) != 0);
+
+            WSP_GGML_ASSERT(ne01*ne02*ne03 == ne1*ne2*ne3);
+            WSP_GGML_ASSERT((uint64_t)ne1*ne2*ne3 <= (1u << 31));
+
+            // write the results from each workgroup into a temp buffer
+            wsp_ggml_metal_encoder_set_buffer(enc, bid_pad, 6);
+            wsp_ggml_metal_encoder_set_buffer(enc, bid_tmp, 7);
+
+            wsp_ggml_metal_encoder_set_threadgroup_memory_size(enc, smem, 0);
+            wsp_ggml_metal_encoder_dispatch_threadgroups(enc, (ne01 + nqptg - 1)/nqptg, ne02, ne03*nwg, 32, nsg, 1);
+
+            // sync the 2 kernels
+            wsp_ggml_metal_op_concurrency_reset(ctx);
+
+            // reduce the results from the workgroups
+            {
+                const int32_t nrows = ne1*ne2*ne3;
+
+                wsp_ggml_metal_kargs_flash_attn_ext_vec_reduce args0 = {
+                    nrows,
+                };
+
+                wsp_ggml_metal_pipeline_t pipeline0 = wsp_ggml_metal_library_get_pipeline_flash_attn_ext_vec_reduce(lib, op, ne20, nwg);
+
+                wsp_ggml_metal_encoder_set_pipeline(enc, pipeline0);
+                wsp_ggml_metal_encoder_set_bytes   (enc, &args0, sizeof(args0), 0);
+                wsp_ggml_metal_encoder_set_buffer  (enc, bid_tmp, 1);
+                wsp_ggml_metal_encoder_set_buffer  (enc, bid_dst, 2);
+
+                wsp_ggml_metal_encoder_dispatch_threadgroups(enc, nrows, 1, 1, 32*nwg, 1, 1);
+            }
+        }
+#undef FATTN_SMEM
+    }
+
+    return 1;
+}
+
+int wsp_ggml_metal_op_bin(wsp_ggml_metal_op_t ctx, int idx) {
+    wsp_ggml_tensor * op = ctx->node(idx);
+
+    wsp_ggml_metal_library_t lib = ctx->lib;
+    wsp_ggml_metal_encoder_t enc = ctx->enc;
+
+    const bool use_fusion = ctx->use_fusion;
+
+    const int debug_fusion = ctx->debug_fusion;
+
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
+
+    WSP_GGML_ASSERT(op->src[0]->type == WSP_GGML_TYPE_F32);
+    WSP_GGML_ASSERT(op->src[1]->type == WSP_GGML_TYPE_F32);
+
+    WSP_GGML_ASSERT(wsp_ggml_is_contiguous_rows(op->src[0]));
+    WSP_GGML_ASSERT(wsp_ggml_is_contiguous_rows(op->src[1]));
+
+    bool bcast_row = false;
+
+    wsp_ggml_metal_buffer_id bid_src0 = wsp_ggml_metal_get_buffer_id(op->src[0]);
+    wsp_ggml_metal_buffer_id bid_src1 = wsp_ggml_metal_get_buffer_id(op->src[1]);
+    wsp_ggml_metal_buffer_id bid_dst  = wsp_ggml_metal_get_buffer_id(op);
+
+    wsp_ggml_metal_kargs_bin 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,
+        /*.offs =*/ 0,
+        /*.o1   =*/ { bid_src1.offs },
+    };
+
+    wsp_ggml_op fops[8];
+
+    int n_fuse = 1;
+
+    // c[0] = add(a,    b[0])
+    // c[1] = add(c[0], b[1])
+    // c[2] = add(c[1], b[2])
+    // ...
+    if (use_fusion) {
+        fops[0] = WSP_GGML_OP_ADD;
+        fops[1] = WSP_GGML_OP_ADD;
+        fops[2] = WSP_GGML_OP_ADD;
+        fops[3] = WSP_GGML_OP_ADD;
+        fops[4] = WSP_GGML_OP_ADD;
+        fops[5] = WSP_GGML_OP_ADD;
+        fops[6] = WSP_GGML_OP_ADD;
+        fops[7] = WSP_GGML_OP_ADD;
+
+        // note: in metal, we sometimes encode the graph in parallel so we have to avoid fusing ops
+        //       across splits. idx_end indicates the last node in the current split
+        for (n_fuse = 0; n_fuse <= 6; ++n_fuse) {
+            if (!ctx->can_fuse(idx + n_fuse, fops + n_fuse, 2)) {
+                break;
+            }
+
+            wsp_ggml_tensor * f0 = ctx->node(idx + n_fuse);
+            wsp_ggml_tensor * f1 = ctx->node(idx + n_fuse + 1);
+
+            if (f0 != f1->src[0]) {
+                break;
+            }
+
+            // b[0] === b[1] === ...
+            if (!wsp_ggml_are_same_layout(f0->src[1], f1->src[1])) {
+                break;
+            }
+
+            // only fuse ops if src1 is in the same Metal buffer
+            wsp_ggml_metal_buffer_id bid_fuse = wsp_ggml_metal_get_buffer_id(f1->src[1]);
+            if (bid_fuse.metal != bid_src1.metal) {
+                break;
+            }
+
+            //ctx->fuse_cnt[ops[n_fuse + 1]->op]++;
+
+            args.o1[n_fuse + 1] = bid_fuse.offs;
+        }
+
+        ++n_fuse;
+
+        if (debug_fusion > 1 && n_fuse > 1) {
+            WSP_GGML_LOG_DEBUG("%s: fuse: ADD x %d\n", __func__, n_fuse);
+        }
+    }
+
+    // the offsets of src1 and all fused buffers are relative to the start of the src1 buffer
+    bid_src1.offs = 0;
+
+    wsp_ggml_metal_pipeline_t pipeline = nullptr;
+
+    if (wsp_ggml_nelements(op->src[1]) == ne10 && wsp_ggml_is_contiguous(op->src[1]) && ne00 % 4 == 0 && ne10 % 4 == 0) {
+        WSP_GGML_ASSERT(wsp_ggml_is_contiguous(op->src[0]));
+
+        // src1 is a row
+        WSP_GGML_ASSERT(ne11 == 1);
+
+        pipeline = wsp_ggml_metal_library_get_pipeline_bin(lib, op->op, n_fuse, true);
+
+        bcast_row = true;
+    } else {
+        pipeline = wsp_ggml_metal_library_get_pipeline_bin(lib, op->op, n_fuse, false);
+    }
+
+    if (n_fuse > 1) {
+        bid_dst = wsp_ggml_metal_get_buffer_id(ctx->node(idx + n_fuse - 1));
+
+        for (int i = 1; i < n_fuse; ++i) {
+            if (!wsp_ggml_metal_op_concurrency_check(ctx, ctx->node(idx + i))) {
+                wsp_ggml_metal_op_concurrency_reset(ctx);
+
+                break;
+            }
+        }
+    }
+
+    wsp_ggml_metal_encoder_set_pipeline(enc, pipeline);
+    wsp_ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+    wsp_ggml_metal_encoder_set_buffer  (enc, bid_src0, 1);
+    wsp_ggml_metal_encoder_set_buffer  (enc, bid_src1, 2);
+    wsp_ggml_metal_encoder_set_buffer  (enc, bid_dst,  3);
+
+    if (bcast_row) {
+        const int64_t n = wsp_ggml_nelements(op)/4;
+
+        wsp_ggml_metal_encoder_dispatch_threadgroups(enc, n, 1, 1, 1, 1, 1);
+    } else {
+        int nth = 32;
+
+        while (16*nth < ne0 && nth < wsp_ggml_metal_pipeline_max_theads_per_threadgroup(pipeline)) {
+            nth *= 2;
+        }
+
+        wsp_ggml_metal_encoder_dispatch_threadgroups(enc, ne01, ne02, ne03, nth, 1, 1);
+    }
+
+    return n_fuse;
+}
+
+int wsp_ggml_metal_op_l2_norm(wsp_ggml_metal_op_t ctx, int idx) {
+    wsp_ggml_tensor * op = ctx->node(idx);
+
+    wsp_ggml_metal_library_t lib = ctx->lib;
+    wsp_ggml_metal_encoder_t enc = ctx->enc;
+
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
+    WSP_GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+
+    float eps;
+    memcpy(&eps, op->op_params, sizeof(float));
+
+    int nth = 32; // SIMD width
+
+    wsp_ggml_metal_kargs_l2_norm args = {
+        /*.ne00   =*/ ne00,
+        /*.ne00_4 =*/ ne00/4,
+        /*.nb01   =*/ nb01,
+        /*.eps    =*/ eps,
+    };
+
+    wsp_ggml_metal_pipeline_t pipeline = wsp_ggml_metal_library_get_pipeline_l2_norm(lib, op);
+
+    while (nth < ne00/4 && nth < wsp_ggml_metal_pipeline_max_theads_per_threadgroup(pipeline)) {
+        nth *= 2;
+    }
+
+    nth = std::min(nth, wsp_ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));
+    nth = std::min(nth, ne00/4);
+
+    const size_t smem = wsp_ggml_metal_pipeline_get_smem(pipeline);
+
+    const int64_t nrows = wsp_ggml_nrows(op->src[0]);
+
+    wsp_ggml_metal_encoder_set_pipeline(enc, pipeline);
+    wsp_ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[0]), 1);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op),         2);
+
+    wsp_ggml_metal_encoder_set_threadgroup_memory_size(enc, smem, 0);
+
+    wsp_ggml_metal_encoder_dispatch_threadgroups(enc, nrows, 1, 1, nth, 1, 1);
+
+    return 1;
+}
+
+int wsp_ggml_metal_op_group_norm(wsp_ggml_metal_op_t ctx, int idx) {
+    wsp_ggml_tensor * op = ctx->node(idx);
+
+    wsp_ggml_metal_library_t lib = ctx->lib;
+    wsp_ggml_metal_encoder_t enc = ctx->enc;
+
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
+    WSP_GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+
+    const int32_t ngrp = ((const int32_t *) op->op_params)[0];
+
+    float eps;
+    memcpy(&eps, op->op_params + 1, sizeof(float));
+
+    wsp_ggml_metal_kargs_group_norm args = {
+        /*.ne00 =*/ ne00,
+        /*.ne01 =*/ ne01,
+        /*.ne02 =*/ ne02,
+        /*.nb00 =*/ nb00,
+        /*.nb01 =*/ nb01,
+        /*.nb02 =*/ nb02,
+        /*.ngrp =*/ ngrp,
+        /*.eps  =*/ eps,
+    };
+
+    wsp_ggml_metal_pipeline_t pipeline = wsp_ggml_metal_library_get_pipeline_group_norm(lib, op);
+
+    int nth = 32; // SIMD width
+    //while (nth < ne00/4 && nth < wsp_ggml_metal_pipeline_max_theads_per_threadgroup(pipeline)) {
+    //    nth *= 2;
+    //}
+
+    //nth = std::min(nth, wsp_ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));
+    //nth = std::min(nth, ne00/4);
+
+    const size_t smem = wsp_ggml_metal_pipeline_get_smem(pipeline);
+
+    wsp_ggml_metal_encoder_set_pipeline(enc, pipeline);
+    wsp_ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[0]), 1);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op),         2);
+
+    wsp_ggml_metal_encoder_set_threadgroup_memory_size(enc, smem, 0);
+
+    wsp_ggml_metal_encoder_dispatch_threadgroups(enc, ngrp, 1, 1, nth, 1, 1);
+
+    return 1;
+}
+
+int wsp_ggml_metal_op_norm(wsp_ggml_metal_op_t ctx, int idx) {
+    wsp_ggml_tensor * op = ctx->node(idx);
+
+    wsp_ggml_metal_library_t lib = ctx->lib;
+    wsp_ggml_metal_encoder_t enc = ctx->enc;
+
+    const bool use_fusion = ctx->use_fusion;
+
+    const int debug_fusion = ctx->debug_fusion;
+
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
+    WSP_GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+
+    float eps;
+    memcpy(&eps, op->op_params, sizeof(float));
+
+    wsp_ggml_metal_buffer_id bid_src0 = wsp_ggml_metal_get_buffer_id(op->src[0]);
+    wsp_ggml_metal_buffer_id bid_dst  = wsp_ggml_metal_get_buffer_id(op);
+
+    wsp_ggml_metal_kargs_norm args = {
+        /*.ne00   =*/ ne00,
+        /*.ne00_t =*/ ne00 % 4 == 0 ? ne00/4 : ne00,
+        /*.nb1    =*/ nb1,
+        /*.nb2    =*/ nb2,
+        /*.nb3    =*/ nb3,
+        /*.eps    =*/ eps,
+        /*.nef1   =*/ { ne01 },
+        /*.nef2   =*/ { ne02 },
+        /*.nef3   =*/ { ne03 },
+        /*.nbf1   =*/ { nb01 },
+        /*.nbf2   =*/ { nb02 },
+        /*.nbf3   =*/ { nb03 },
+    };
+
+    wsp_ggml_op fops[8];
+
+    int n_fuse = 1;
+
+    wsp_ggml_metal_buffer_id bid_fuse[2] = { bid_src0, bid_src0 };
+
+    // d[0] = norm(a)
+    // d[1] = mul(d[0], b)
+    // d[2] = add(d[1], c)
+    if (use_fusion) {
+        fops[0] = op->op;
+        fops[1] = WSP_GGML_OP_MUL;
+        fops[2] = WSP_GGML_OP_ADD;
+
+        for (n_fuse = 0; n_fuse <= 1; ++n_fuse) {
+            if (!ctx->can_fuse(idx + n_fuse, fops + n_fuse, 2)) {
+                break;
+            }
+
+            wsp_ggml_tensor * f0 = ctx->node(idx + n_fuse);
+            wsp_ggml_tensor * f1 = ctx->node(idx + n_fuse + 1);
+
+            if (f0 != f1->src[0]) {
+                break;
+            }
+
+            if (f1->src[1]->ne[0] != op->ne[0]) {
+                break;
+            }
+
+            if (!wsp_ggml_is_contiguous_rows(f1->src[1])) {
+                break;
+            }
+
+            if (f1->type != WSP_GGML_TYPE_F32) {
+                break;
+            }
+
+            //ctx->fuse_cnt[f1->op]++;
+
+            bid_fuse[n_fuse] = wsp_ggml_metal_get_buffer_id(f1->src[1]);
+
+            args.nef1[n_fuse + 1] = f1->src[1]->ne[1];
+            args.nef2[n_fuse + 1] = f1->src[1]->ne[2];
+            args.nef3[n_fuse + 1] = f1->src[1]->ne[3];
+
+            args.nbf1[n_fuse + 1] = f1->src[1]->nb[1];
+            args.nbf2[n_fuse + 1] = f1->src[1]->nb[2];
+            args.nbf3[n_fuse + 1] = f1->src[1]->nb[3];
+        }
+
+        ++n_fuse;
+
+        if (debug_fusion > 1 && n_fuse > 1) {
+            if (n_fuse == 2) {
+                WSP_GGML_LOG_DEBUG("%s: fuse: %s + MUL\n", __func__, wsp_ggml_op_name(op->op));
+            }
+            if (n_fuse == 3) {
+                WSP_GGML_LOG_DEBUG("%s: fuse: %s + MUL + ADD\n", __func__, wsp_ggml_op_name(op->op));
+            }
+        }
+    }
+
+    if (n_fuse > 1) {
+        bid_dst = wsp_ggml_metal_get_buffer_id(ctx->node(idx + n_fuse - 1));
+
+        for (int i = 1; i < n_fuse; ++i) {
+            if (!wsp_ggml_metal_op_concurrency_check(ctx, ctx->node(idx + i))) {
+                wsp_ggml_metal_op_concurrency_reset(ctx);
+
+                break;
+            }
+        }
+    }
+
+    wsp_ggml_metal_pipeline_t pipeline = wsp_ggml_metal_library_get_pipeline_norm(lib, op, n_fuse);
+
+    int nth = 32; // SIMD width
+
+    while (nth < args.ne00_t && nth < wsp_ggml_metal_pipeline_max_theads_per_threadgroup(pipeline)) {
+        nth *= 2;
+    }
+
+    nth = std::min(nth, wsp_ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));
+    nth = std::min(nth, args.ne00_t);
+
+    const size_t smem = wsp_ggml_metal_pipeline_get_smem(pipeline);
+
+    wsp_ggml_metal_encoder_set_pipeline(enc, pipeline);
+    wsp_ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+    wsp_ggml_metal_encoder_set_buffer  (enc, bid_src0,    1);
+    wsp_ggml_metal_encoder_set_buffer  (enc, bid_fuse[0], 2);
+    wsp_ggml_metal_encoder_set_buffer  (enc, bid_fuse[1], 3);
+    wsp_ggml_metal_encoder_set_buffer  (enc, bid_dst,     4);
+
+    wsp_ggml_metal_encoder_set_threadgroup_memory_size(enc, smem, 0);
+
+    wsp_ggml_metal_encoder_dispatch_threadgroups(enc, ne01, ne02, ne03, nth, 1, 1);
+
+    return n_fuse;
+}
+
+int wsp_ggml_metal_op_rope(wsp_ggml_metal_op_t ctx, int idx) {
+    wsp_ggml_tensor * op = ctx->node(idx);
+
+    wsp_ggml_metal_library_t lib = ctx->lib;
+    wsp_ggml_metal_encoder_t enc = ctx->enc;
+
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
+    WSP_GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+
+    // make sure we have one or more position id(ne10) per token(ne02)
+    WSP_GGML_ASSERT(ne10 % ne02 == 0);
+    WSP_GGML_ASSERT(ne10 >= ne02);
+
+    const int nth = std::min(1024, ne00);
+
+    const int n_past     = ((const int32_t *) op->op_params)[0];
+    const int n_dims     = ((const int32_t *) op->op_params)[1];
+  //const int mode       = ((const int32_t *) op->op_params)[2];
+    // skip 3, n_ctx, used in GLM RoPE, unimplemented in metal
+    const int n_ctx_orig = ((const int32_t *) op->op_params)[4];
+
+    float freq_base;
+    float freq_scale;
+    float ext_factor;
+    float attn_factor;
+    float beta_fast;
+    float beta_slow;
+
+    memcpy(&freq_base,   (const int32_t *) op->op_params +  5, sizeof(float));
+    memcpy(&freq_scale,  (const int32_t *) op->op_params +  6, sizeof(float));
+    memcpy(&ext_factor,  (const int32_t *) op->op_params +  7, sizeof(float));
+    memcpy(&attn_factor, (const int32_t *) op->op_params +  8, sizeof(float));
+    memcpy(&beta_fast,   (const int32_t *) op->op_params +  9, sizeof(float));
+    memcpy(&beta_slow,   (const int32_t *) op->op_params + 10, sizeof(float));
+
+    // mrope
+    const int sect_0 = ((const int32_t *) op->op_params)[11];
+    const int sect_1 = ((const int32_t *) op->op_params)[12];
+    const int sect_2 = ((const int32_t *) op->op_params)[13];
+    const int sect_3 = ((const int32_t *) op->op_params)[14];
+
+    wsp_ggml_metal_kargs_rope 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,
+        /*.n_past      =*/ n_past,
+        /*.n_dims      =*/ n_dims,
+        /*.n_ctx_orig  =*/ n_ctx_orig,
+        /*.freq_base   =*/ freq_base,
+        /*.freq_scale  =*/ freq_scale,
+        /*.ext_factor  =*/ ext_factor,
+        /*.attn_factor =*/ attn_factor,
+        /*.beta_fast   =*/ beta_fast,
+        /*.beta_slow   =*/ beta_slow,
+        /* sect_0      =*/ sect_0,
+        /* sect_1      =*/ sect_1,
+        /* sect_2      =*/ sect_2,
+        /* sect_3      =*/ sect_3,
+        /* src2        =*/ op->src[2] != nullptr,
+    };
+
+    wsp_ggml_metal_pipeline_t pipeline = wsp_ggml_metal_library_get_pipeline_rope(lib, op);
+
+    wsp_ggml_metal_encoder_set_pipeline(enc, pipeline);
+    wsp_ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[0]), 1);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[1]), 2);
+    if (op->src[2]) {
+        wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[2]), 3);
+    } else {
+        wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[0]), 3);
+    }
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op),         4);
+
+    wsp_ggml_metal_encoder_dispatch_threadgroups(enc, ne01, ne02, ne03, nth, 1, 1);
+
+    return 1;
+}
+
+int wsp_ggml_metal_op_im2col(wsp_ggml_metal_op_t ctx, int idx) {
+    wsp_ggml_tensor * op = ctx->node(idx);
+
+    wsp_ggml_metal_library_t lib = ctx->lib;
+    wsp_ggml_metal_encoder_t enc = ctx->enc;
+
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
+    WSP_GGML_TENSOR_LOCALS(uint32_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];
+    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];
+
+    const bool is_2D = ((const int32_t *)(op->op_params))[6] == 1;
+
+    const int32_t N  = op->src[1]->ne[is_2D ? 3 : 2];
+    const int32_t IC = op->src[1]->ne[is_2D ? 2 : 1];
+    const int32_t IH = is_2D ? op->src[1]->ne[1] : 1;
+    const int32_t IW =         op->src[1]->ne[0];
+
+    const int32_t KH = is_2D ? op->src[0]->ne[1] : 1;
+    const int32_t KW =         op->src[0]->ne[0];
+
+    const int32_t OH = is_2D ? op->ne[2] : 1;
+    const int32_t OW =         op->ne[1];
+
+    const int32_t CHW = IC * KH * KW;
+
+    const uint64_t ofs0 = op->src[1]->nb[is_2D ? 3 : 2] / 4;
+    const uint64_t ofs1 = op->src[1]->nb[is_2D ? 2 : 1] / 4;
+
+    wsp_ggml_metal_kargs_im2col args = {
+        /*.ofs0 =*/ ofs0,
+        /*.ofs1 =*/ ofs1,
+        /*.IW   =*/ IW,
+        /*.IH   =*/ IH,
+        /*.CHW  =*/ CHW,
+        /*.s0   =*/ s0,
+        /*.s1   =*/ s1,
+        /*.p0   =*/ p0,
+        /*.p1   =*/ p1,
+        /*.d0   =*/ d0,
+        /*.d1   =*/ d1,
+        /*.N    =*/ N,
+        /*.KH   =*/ KH,
+        /*.KW   =*/ KW,
+        /*.KHW  =*/ KH * KW,
+    };
+
+    wsp_ggml_metal_pipeline_t pipeline = wsp_ggml_metal_library_get_pipeline_im2col(lib, op);
+
+    WSP_GGML_ASSERT(KH*KW <= wsp_ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));
+
+    const uint64_t ntptg0 = std::min(wsp_ggml_metal_pipeline_max_theads_per_threadgroup(pipeline)/(KH*KW), N);
+
+    wsp_ggml_metal_encoder_set_pipeline(enc, pipeline);
+    wsp_ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[1]), 1);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op),         2);
+
+    wsp_ggml_metal_encoder_dispatch_threadgroups(enc, IC, OH, OW, ntptg0, KH, KW);
+
+    return 1;
+}
+
+int wsp_ggml_metal_op_conv_transpose_1d(wsp_ggml_metal_op_t ctx, int idx) {
+    wsp_ggml_tensor * op = ctx->node(idx);
+
+    wsp_ggml_metal_library_t lib = ctx->lib;
+    wsp_ggml_metal_encoder_t enc = ctx->enc;
+
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
+    WSP_GGML_TENSOR_LOCALS(uint32_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];
+
+    wsp_ggml_metal_kargs_conv_transpose_1d args = {
+        /*.IC  =*/ IC,
+        /*.IL  =*/ IL,
+        /*.K   =*/ K,
+        /*.s0  =*/ s0,
+        /*.nb0 =*/ nb0,
+        /*.nb1 =*/ nb1,
+    };
+
+    wsp_ggml_metal_pipeline_t pipeline = wsp_ggml_metal_library_get_pipeline_conv_transpose_1d(lib, op);
+
+    wsp_ggml_metal_encoder_set_pipeline(enc, pipeline);
+    wsp_ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[0]), 1);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[1]), 2);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op),         3);
+
+    wsp_ggml_metal_encoder_dispatch_threadgroups(enc, OL, OC, 1, 1, 1, 1);
+
+    return 1;
+}
+
+int wsp_ggml_metal_op_conv_transpose_2d(wsp_ggml_metal_op_t ctx, int idx) {
+    wsp_ggml_tensor * op = ctx->node(idx);
+
+    wsp_ggml_metal_library_t lib = ctx->lib;
+    wsp_ggml_metal_encoder_t enc = ctx->enc;
+
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
+    WSP_GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+
+    const int32_t s0 = ((const int32_t *)(op->op_params))[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 KH = op->src[0]->ne[1];
+    const int32_t KW = op->src[0]->ne[0];
+
+    const int32_t OW = op->ne[0];
+    const int32_t OH = op->ne[1];
+    const int32_t OC = op->ne[2];
+
+    wsp_ggml_metal_kargs_conv_transpose_2d args = {
+        /*.IC  =*/ IC,
+        /*.IH  =*/ IH,
+        /*.IW  =*/ IW,
+        /*.KH  =*/ KH,
+        /*.KW  =*/ KW,
+        /*.OC  =*/ OC,
+        /*.s0  =*/ s0,
+        /*.nb0 =*/ nb0,
+        /*.nb1 =*/ nb1,
+        /*.nb2 =*/ nb2,
+    };
+
+    wsp_ggml_metal_pipeline_t pipeline = wsp_ggml_metal_library_get_pipeline_conv_transpose_2d(lib, op);
+
+    wsp_ggml_metal_encoder_set_pipeline(enc, pipeline);
+    wsp_ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[0]), 1);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[1]), 2);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op),         3);
+
+    // Metal requires buffer size to be multiple of 16 bytes
+    const size_t smem = WSP_GGML_PAD(KW * KH * sizeof(float), 16);
+    wsp_ggml_metal_encoder_set_threadgroup_memory_size(enc, smem, 0);
+
+    wsp_ggml_metal_encoder_dispatch_threadgroups(enc, OW, OH, OC, KW, KH, 1);
+
+    return 1;
+}
+
+int wsp_ggml_metal_op_upscale(wsp_ggml_metal_op_t ctx, int idx) {
+    wsp_ggml_tensor * op = ctx->node(idx);
+
+    wsp_ggml_metal_library_t lib = ctx->lib;
+    wsp_ggml_metal_encoder_t enc = ctx->enc;
+
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
+    WSP_GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+
+    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];
+
+    wsp_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
+    };
+
+    wsp_ggml_metal_pipeline_t pipeline = wsp_ggml_metal_library_get_pipeline_upscale(lib, op);
+
+    const int nth = std::min(wsp_ggml_metal_pipeline_max_theads_per_threadgroup(pipeline), ne0);
+
+    wsp_ggml_metal_encoder_set_pipeline(enc, pipeline);
+    wsp_ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[0]), 1);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op),         2);
+
+    wsp_ggml_metal_encoder_dispatch_threadgroups(enc, ne1, ne2, ne3, nth, 1, 1);
+
+    return 1;
+}
+
+int wsp_ggml_metal_op_pad(wsp_ggml_metal_op_t ctx, int idx) {
+    wsp_ggml_tensor * op = ctx->node(idx);
+
+    wsp_ggml_metal_library_t lib = ctx->lib;
+    wsp_ggml_metal_encoder_t enc = ctx->enc;
+
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
+    WSP_GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+
+    wsp_ggml_metal_kargs_pad 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
+    };
+
+    wsp_ggml_metal_pipeline_t pipeline = wsp_ggml_metal_library_get_pipeline_pad(lib, op);
+
+    const int nth = std::min(1024, ne0);
+
+    wsp_ggml_metal_encoder_set_pipeline(enc, pipeline);
+    wsp_ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[0]), 1);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op),         2);
+
+    wsp_ggml_metal_encoder_dispatch_threadgroups(enc, ne1, ne2, ne3, nth, 1, 1);
+
+    return 1;
+}
+
+int wsp_ggml_metal_op_pad_reflect_1d(wsp_ggml_metal_op_t ctx, int idx) {
+    wsp_ggml_tensor * op = ctx->node(idx);
+
+    wsp_ggml_metal_library_t lib = ctx->lib;
+    wsp_ggml_metal_encoder_t enc = ctx->enc;
+
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
+    WSP_GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+
+    wsp_ggml_metal_kargs_pad_reflect_1d 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,
+        /*.p0 =*/ ((const int32_t *)(op->op_params))[0],
+        /*.p1 =*/ ((const int32_t *)(op->op_params))[1]
+    };
+
+    wsp_ggml_metal_pipeline_t pipeline = wsp_ggml_metal_library_get_pipeline_pad_reflect_1d(lib, op);
+
+    const int nth = std::min(1024, ne0);
+
+    wsp_ggml_metal_encoder_set_pipeline(enc, pipeline);
+    wsp_ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[0]), 1);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op),         2);
+
+    wsp_ggml_metal_encoder_dispatch_threadgroups(enc, ne1, ne2, ne3, nth, 1, 1);
+
+    return 1;
+}
+
+int wsp_ggml_metal_op_arange(wsp_ggml_metal_op_t ctx, int idx) {
+    wsp_ggml_tensor * op = ctx->node(idx);
+
+    wsp_ggml_metal_library_t lib = ctx->lib;
+    wsp_ggml_metal_encoder_t enc = ctx->enc;
+
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
+    WSP_GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+
+    float start;
+    float step;
+
+    memcpy(&start, ((const int32_t *) op->op_params) + 0, sizeof(float));
+    memcpy(&step,  ((const int32_t *) op->op_params) + 2, sizeof(float));
+
+    wsp_ggml_metal_kargs_arange args = {
+        /*.ne0   =*/ ne0,
+        /*.start =*/ start,
+        /*.step  =*/ step
+    };
+
+    const int nth = std::min(1024, ne0);
+
+    wsp_ggml_metal_pipeline_t pipeline = wsp_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)];
+
+    wsp_ggml_metal_encoder_set_pipeline(enc, pipeline);
+    wsp_ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op), 1);
+
+    wsp_ggml_metal_encoder_dispatch_threadgroups(enc, 1, 1, 1, nth, 1, 1);
+
+    return 1;
+}
+
+int wsp_ggml_metal_op_timestep_embedding(wsp_ggml_metal_op_t ctx, int idx) {
+    wsp_ggml_tensor * op = ctx->node(idx);
+
+    wsp_ggml_metal_library_t lib = ctx->lib;
+    wsp_ggml_metal_encoder_t enc = ctx->enc;
+
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
+    WSP_GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+
+    const int dim        = op->op_params[0];
+    const int max_period = op->op_params[1];
+
+    wsp_ggml_metal_kargs_timestep_embedding args = {
+        /*.nb1 =*/ nb1,
+        /*.dim =*/ dim,
+        /*.max_period =*/ max_period,
+    };
+
+    wsp_ggml_metal_pipeline_t pipeline = wsp_ggml_metal_library_get_pipeline_timestep_embedding(lib, op);
+
+    const int nth = std::max(1, std::min(1024, dim/2));
+
+    wsp_ggml_metal_encoder_set_pipeline(enc, pipeline);
+    wsp_ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[0]), 1);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op),         2);
+
+    wsp_ggml_metal_encoder_dispatch_threadgroups(enc, ne00, 1, 1, nth, 1, 1);
+
+    return 1;
+}
+
+int wsp_ggml_metal_op_argmax(wsp_ggml_metal_op_t ctx, int idx) {
+    wsp_ggml_tensor * op = ctx->node(idx);
+
+    wsp_ggml_metal_library_t lib = ctx->lib;
+    wsp_ggml_metal_encoder_t enc = ctx->enc;
+
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
+    WSP_GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+
+    wsp_ggml_metal_kargs_argmax args = {
+        /*.ne00 = */ ne00,
+        /*.nb01 = */ nb01,
+    };
+
+    wsp_ggml_metal_pipeline_t pipeline = wsp_ggml_metal_library_get_pipeline_argmax(lib, op);
+
+    const int64_t nrows = wsp_ggml_nrows(op->src[0]);
+
+    int nth = 32; // SIMD width
+    while (nth < ne00 && nth*ne01*ne02*ne03 < 256) {
+        nth *= 2;
+    }
+
+    const size_t smem = wsp_ggml_metal_pipeline_get_smem(pipeline);
+
+    wsp_ggml_metal_encoder_set_pipeline(enc, pipeline);
+    wsp_ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[0]), 1);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op),         2);
+
+    wsp_ggml_metal_encoder_set_threadgroup_memory_size(enc, smem, 0);
+
+    wsp_ggml_metal_encoder_dispatch_threadgroups(enc, nrows, 1, 1, nth, 1, 1);
+
+    return 1;
+}
+
+int wsp_ggml_metal_op_argsort(wsp_ggml_metal_op_t ctx, int idx) {
+    wsp_ggml_tensor * op = ctx->node(idx);
+
+    wsp_ggml_metal_library_t lib = ctx->lib;
+    wsp_ggml_metal_encoder_t enc = ctx->enc;
+
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
+    WSP_GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+
+    // bitonic sort requires the number of elements to be power of 2
+    int64_t ne00_padded = 1;
+    while (ne00_padded < ne00) {
+        ne00_padded *= 2;
+    }
+
+    wsp_ggml_metal_pipeline_t pipeline = wsp_ggml_metal_library_get_pipeline_argsort(lib, op);
+
+    const int64_t nrows = wsp_ggml_nrows(op->src[0]);
+
+    // 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 = WSP_GGML_PAD(ne00_padded*sizeof(int32_t), 16);
+
+    wsp_ggml_metal_kargs_argsort args = {
+        /*.ncols =*/ ne00,
+        /*.ncols_pad =*/ ne00_padded
+    };
+
+    wsp_ggml_metal_encoder_set_pipeline(enc, pipeline);
+    wsp_ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[0]), 1);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op),         2);
+
+    wsp_ggml_metal_encoder_set_threadgroup_memory_size(enc, smem, 0);
+
+    wsp_ggml_metal_encoder_dispatch_threadgroups(enc, 1, nrows, 1, ne00_padded, 1, 1);
+
+    return 1;
+}
+
+int wsp_ggml_metal_op_leaky_relu(wsp_ggml_metal_op_t ctx, int idx) {
+    wsp_ggml_tensor * op = ctx->node(idx);
+
+    wsp_ggml_metal_library_t lib = ctx->lib;
+    wsp_ggml_metal_encoder_t enc = ctx->enc;
+
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
+    WSP_GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+
+    float slope;
+    memcpy(&slope, op->op_params, sizeof(float));
+
+    wsp_ggml_metal_kargs_leaky_relu args = {
+        /*.slope =*/ slope
+    };
+
+    wsp_ggml_metal_pipeline_t pipeline = wsp_ggml_metal_library_get_pipeline_unary(lib, op);
+
+    int64_t n = wsp_ggml_nelements(op);
+
+    if (n % 4 == 0) {
+        n /= 4;
+    }
+
+    wsp_ggml_metal_encoder_set_pipeline(enc, pipeline);
+    wsp_ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[0]), 1);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op),         2);
+
+    wsp_ggml_metal_encoder_dispatch_threadgroups(enc, n, 1, 1, 1, 1, 1);
+
+    return 1;
+}
+
+int wsp_ggml_metal_op_opt_step_adamw(wsp_ggml_metal_op_t ctx, int idx) {
+    wsp_ggml_tensor * op = ctx->node(idx);
+
+    wsp_ggml_metal_library_t lib = ctx->lib;
+    wsp_ggml_metal_encoder_t enc = ctx->enc;
+
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
+    WSP_GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+
+    wsp_ggml_metal_pipeline_t pipeline = wsp_ggml_metal_library_get_pipeline_opt_step_adamw(lib, op);
+
+    const int64_t np = wsp_ggml_nelements(op->src[0]);
+    wsp_ggml_metal_kargs_opt_step_adamw args = {
+        /*.np =*/ np,
+    };
+
+    int ida = 0;
+
+    wsp_ggml_metal_encoder_set_pipeline(enc, pipeline);
+    wsp_ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), ida++);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[0]), ida++);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[1]), ida++);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[2]), ida++);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[3]), ida++);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[4]), ida++);
+
+    const int nth = std::min(wsp_ggml_metal_pipeline_max_theads_per_threadgroup(pipeline), ne0);
+    const int64_t n = (np + nth - 1) / nth;
+
+    wsp_ggml_metal_encoder_dispatch_threadgroups(enc, n, 1, 1, nth, 1, 1);
+
+    return 1;
+}
+
+int wsp_ggml_metal_op_opt_step_sgd(wsp_ggml_metal_op_t ctx, int idx) {
+    wsp_ggml_tensor * op = ctx->node(idx);
+
+    wsp_ggml_metal_library_t lib = ctx->lib;
+    wsp_ggml_metal_encoder_t enc = ctx->enc;
+
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    WSP_GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    WSP_GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
+    WSP_GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+
+    wsp_ggml_metal_pipeline_t pipeline = wsp_ggml_metal_library_get_pipeline_opt_step_sgd(lib, op);
+
+    const int64_t np = wsp_ggml_nelements(op->src[0]);
+    wsp_ggml_metal_kargs_opt_step_sgd args = {
+        /*.np =*/ np,
+    };
+
+    int ida = 0;
+
+    wsp_ggml_metal_encoder_set_pipeline(enc, pipeline);
+    wsp_ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), ida++);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[0]), ida++);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[1]), ida++);
+    wsp_ggml_metal_encoder_set_buffer  (enc, wsp_ggml_metal_get_buffer_id(op->src[2]), ida++);
+
+    const int nth = std::min(wsp_ggml_metal_pipeline_max_theads_per_threadgroup(pipeline), ne0);
+    const int64_t n = (np + nth - 1) / nth;
+
+    wsp_ggml_metal_encoder_dispatch_threadgroups(enc, n, 1, 1, nth, 1, 1);
+
+    return 1;
+}