ggml : sync resolve (skip) (#0)

ggml/CMakeLists.txt

@@ -225,6 +225,7 @@ set(GGML_PUBLIC_HEADERS
     include/ggml-cann.h
     include/ggml-cuda.h
     include/ggml-kompute.h
+    include/ggml-opt.h
     include/ggml-metal.h
     include/ggml-rpc.h
     include/ggml-sycl.h
@@ -241,7 +242,7 @@ install(TARGETS ggml-base LIBRARY)
 if (GGML_METAL)
     # FIXME: does this need to be installed with GGML_METAL_EMBED_LIBRARY?
     install(
-        FILES ggml/src/ggml-metal/ggml-metal.metal
+        FILES src/ggml-metal/ggml-metal.metal
         PERMISSIONS
             OWNER_READ
             OWNER_WRITE

ggml/src/ggml-amx.cpp

@@ -1,436 +0,0 @@
-#include "ggml-amx.h"
-#include "ggml-amx/common.h"
-#include "ggml-amx/mmq.h"
-#include "ggml-backend-impl.h"
-#include "ggml-impl.h"
-
-#if defined(__gnu_linux__)
-#include <sys/syscall.h>
-#include <unistd.h>
-#endif
-
-#include <cstdlib>
-#include <cstring>
-#include <memory>
-
-#if defined(__AMX_INT8__)
-
-// AMX buffer interface
-static void ggml_backend_amx_buffer_free_buffer(ggml_backend_buffer_t buffer) {
-    free(buffer->context);
-}
-
-static void * ggml_backend_amx_buffer_get_base(ggml_backend_buffer_t buffer) {
-    return (void *)(buffer->context);
-}
-
-static void ggml_backend_amx_buffer_memset_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) {
-    memset((char *)tensor->data + offset, value, size);
-
-    GGML_UNUSED(buffer);
-}
-
-static void ggml_backend_amx_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
-    if (qtype_has_amx_kernels(tensor->type)) {
-        ggml_backend_amx_convert_weight(tensor, data, offset, size);
-    } else {
-        memcpy((char *)tensor->data + offset, data, size);
-    }
-
-    GGML_UNUSED(buffer);
-}
-
-static void ggml_backend_amx_buffer_get_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
-    GGML_ASSERT(!qtype_has_amx_kernels(tensor->type));
-    memcpy(data, (const char *)tensor->data + offset, size);
-
-    GGML_UNUSED(buffer);
-}
-
-static bool ggml_backend_amx_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst) {
-    if (ggml_backend_buffer_is_host(src->buffer)) {
-        if (qtype_has_amx_kernels(src->type)) {
-            ggml_backend_amx_convert_weight(dst, src->data, 0, ggml_backend_amx_get_alloc_size(dst));
-        } else {
-            memcpy(dst->data, src->data, ggml_nbytes(src));
-        }
-        return true;
-    }
-    return false;
-
-    GGML_UNUSED(buffer);
-}
-
-static void ggml_backend_amx_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
-    memset(buffer->context, value, buffer->size);
-}
-
-static ggml_backend_buffer_i ggml_backend_amx_buffer_interface = {
-    /* .free_buffer     = */ ggml_backend_amx_buffer_free_buffer,
-    /* .get_base        = */ ggml_backend_amx_buffer_get_base,
-    /* .init_tensor     = */ NULL, // no initialization required
-    /* .memset_tensor   = */ ggml_backend_amx_buffer_memset_tensor,
-    /* .set_tensor      = */ ggml_backend_amx_buffer_set_tensor,
-    /* .get_tensor      = */ ggml_backend_amx_buffer_get_tensor,
-    /* .cpy_tensor      = */ ggml_backend_amx_buffer_cpy_tensor,
-    /* .clear           = */ ggml_backend_amx_buffer_clear,
-    /* .reset           = */ NULL,
-};
-
-static const char * ggml_backend_amx_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
-    return "AMX";
-
-    GGML_UNUSED(buft);
-}
-
-static ggml_backend_buffer_t ggml_backend_amx_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
-    void * data = aligned_alloc(TENSOR_ALIGNMENT, size);
-    if (data == NULL) {
-        fprintf(stderr, "%s: failed to allocate buffer of size %zu\n", __func__, size);
-        return NULL;
-    }
-
-    return ggml_backend_buffer_init(buft, ggml_backend_amx_buffer_interface, data, size);
-}
-
-static size_t ggml_backend_amx_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
-    return TENSOR_ALIGNMENT;
-
-    GGML_UNUSED(buft);
-}
-
-static size_t ggml_backend_amx_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor* tensor) {
-    return ggml_backend_amx_get_alloc_size(tensor);
-
-    GGML_UNUSED(buft);
-}
-
-static bool ggml_backend_amx_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
-    return false;
-
-    GGML_UNUSED(buft);
-}
-
-ggml_backend_buffer_type_t ggml_backend_amx_buffer_type() {
-    static struct ggml_backend_buffer_type ggml_backend_buffer_type_amx = {
-        /* .iface = */ {
-            /* .get_name         = */ ggml_backend_amx_buffer_type_get_name,
-            /* .alloc_buffer     = */ ggml_backend_amx_buffer_type_alloc_buffer,
-            /* .get_alignment    = */ ggml_backend_amx_buffer_type_get_alignment,
-            /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
-            /* .get_alloc_size   = */ ggml_backend_amx_buffer_type_get_alloc_size,
-            /* .is_host          = */ ggml_backend_amx_buffer_type_is_host,
-        },
-        /* .device  = */ ggml_backend_reg_dev_get(ggml_backend_amx_reg(), 0),
-        /* .context = */ NULL,
-    };
-
-    return &ggml_backend_buffer_type_amx;
-}
-
-// backend interface
-
-static const char * ggml_backend_amx_name(ggml_backend_t backend) {
-    return "AMX";
-
-    GGML_UNUSED(backend);
-}
-
-static void ggml_backend_amx_free(ggml_backend_t backend) {
-    ggml_backend_amx_context * ctx = (ggml_backend_amx_context *)backend->context;
-    delete ctx;
-    delete backend;
-}
-
-static enum ggml_status ggml_backend_amx_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
-    ggml_backend_amx_context * ctx = (ggml_backend_amx_context *)backend->context;
-
-    for (int i = 0; i < cgraph->n_nodes; i++) {
-        struct ggml_tensor * node = cgraph->nodes[i];
-
-        switch (node->op) {
-        case GGML_OP_MUL_MAT:
-            ggml_backend_amx_mul_mat(ctx, node);
-            break;
-
-        case GGML_OP_NONE:
-        case GGML_OP_RESHAPE:
-        case GGML_OP_VIEW:
-        case GGML_OP_PERMUTE:
-        case GGML_OP_TRANSPOSE:
-            break;
-
-        default:
-            fprintf(stderr, "%s: unsupported op %s\n", __func__, ggml_op_desc(node));
-            GGML_ASSERT(false);
-        }
-    }
-
-    return GGML_STATUS_SUCCESS;
-
-    GGML_UNUSED(backend);
-}
-
-static struct ggml_backend_i ggml_backend_amx_i = {
-    /* .get_name                = */ ggml_backend_amx_name,
-    /* .free                    = */ ggml_backend_amx_free,
-    /* .set_tensor_async        = */ NULL,
-    /* .get_tensor_async        = */ NULL,
-    /* .cpy_tensor_async        = */ NULL,
-    /* .synchronize             = */ NULL,
-    /* .graph_plan_create       = */ NULL,
-    /* .graph_plan_free         = */ NULL,
-    /* .graph_plan_update       = */ NULL,
-    /* .graph_plan_compute      = */ NULL,
-    /* .graph_compute           = */ ggml_backend_amx_graph_compute,
-    /* .event_record            = */ NULL,
-    /* .event_wait              = */ NULL,
-};
-
-static ggml_guid_t ggml_backend_amx_guid() {
-    static ggml_guid guid = { 0x13, 0xb8, 0xa4, 0xc4, 0xba, 0xfe, 0x51, 0x67, 0x87, 0x44, 0x55, 0x15, 0xb2, 0x35, 0x62, 0x3e };
-    return &guid;
-}
-
-#define ARCH_GET_XCOMP_PERM     0x1022
-#define ARCH_REQ_XCOMP_PERM     0x1023
-#define XFEATURE_XTILECFG       17
-#define XFEATURE_XTILEDATA      18
-
-static bool ggml_amx_init() {
-#if defined(__gnu_linux__)
-    if (syscall(SYS_arch_prctl, ARCH_REQ_XCOMP_PERM, XFEATURE_XTILEDATA)) {
-        fprintf(stderr, "AMX is not ready to be used!\n");
-        return false;
-    }
-    return true;
-#elif defined(_WIN32)
-    return true;
-#endif
-}
-
-ggml_backend_t ggml_backend_amx_init() {
-
-    // invoke a Linux system call to request access to AMX features
-    ggml_amx_init();
-
-    // backend context
-    ggml_backend_amx_context * ctx = new ggml_backend_amx_context;
-
-    // ggml amx backend
-    ggml_backend_t backend = new ggml_backend {
-        /* .guid      = */ ggml_backend_amx_guid(),
-        /* .interface = */ ggml_backend_amx_i,
-        /* .device    = */ ggml_backend_reg_dev_get(ggml_backend_amx_reg(), 0),
-        /* .context   = */ ctx,
-    };
-
-    return backend;
-}
-
-bool ggml_backend_is_amx(ggml_backend_t backend) {
-    return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_amx_guid());
-}
-
-void ggml_backend_amx_set_n_threads(ggml_backend_t backend_amx, int n_threads) {
-    GGML_ASSERT(ggml_backend_is_amx(backend_amx));
-
-    ggml_backend_amx_context * ctx = (ggml_backend_amx_context *)backend_amx->context;
-    ctx->n_threads = n_threads;
-}
-
-// device interface
-
-static const char * ggml_backend_amx_device_get_name(ggml_backend_dev_t dev) {
-    return "AMX";
-
-    GGML_UNUSED(dev);
-}
-
-static const char * ggml_backend_amx_device_get_description(ggml_backend_dev_t dev) {
-    return "Intel Advanced Matrix Extensions";
-
-    GGML_UNUSED(dev);
-}
-
-static void ggml_backend_amx_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
-    // TODO
-    *free = 0;
-    *total = 0;
-
-    GGML_UNUSED(dev);
-}
-
-static enum ggml_backend_dev_type ggml_backend_amx_device_get_type(ggml_backend_dev_t dev) {
-    return GGML_BACKEND_DEVICE_TYPE_ACCEL;
-
-    GGML_UNUSED(dev);
-}
-
-static void ggml_backend_amx_device_get_props(ggml_backend_dev_t dev, struct ggml_backend_dev_props * props) {
-    props->name        = ggml_backend_amx_device_get_name(dev);
-    props->description = ggml_backend_amx_device_get_description(dev);
-    props->type        = ggml_backend_amx_device_get_type(dev);
-    ggml_backend_amx_device_get_memory(dev, &props->memory_free, &props->memory_total);
-
-    // `buffer_from_host_ptr` is intended to be used in mmap, when memory layout unchanged
-    props->caps = {
-        /* .async                 = */ false,
-        /* .host_buffer           = */ false,
-        /* .buffer_from_host_ptr  = */ false,
-        /* .events                = */ false,
-    };
-}
-
-static ggml_backend_t ggml_backend_amx_device_init(ggml_backend_dev_t dev, const char * params) {
-    return ggml_backend_amx_init();
-
-    GGML_UNUSED(dev);
-    GGML_UNUSED(params);
-}
-
-static ggml_backend_buffer_type_t ggml_backend_amx_device_get_buffer_type(ggml_backend_dev_t dev) {
-    return ggml_backend_amx_buffer_type();
-
-    GGML_UNUSED(dev);
-}
-
-static bool ggml_backend_amx_device_supports_op(ggml_backend_dev_t dev, const struct ggml_tensor * op) {
-
-    // handle only 2d gemm for now
-    auto is_contiguous_2d = [](const struct ggml_tensor * t) {
-        return ggml_is_contiguous(t) && t->ne[3] == 1 && t->ne[2] == 1;
-    };
-
-    switch (op->op) {
-        case GGML_OP_NONE:
-        case GGML_OP_RESHAPE:
-        case GGML_OP_VIEW:
-        case GGML_OP_PERMUTE:
-        case GGML_OP_TRANSPOSE:
-            return true;
-
-        case GGML_OP_MUL_MAT: {
-            const struct ggml_tensor * src0 = op->src[0];
-            const struct ggml_tensor * src1 = op->src[1];
-
-            const enum ggml_type type = src0->type;
-            const int64_t ne0 = op->ne[0];
-
-            bool is_training = src0->grad || src1->grad;
-
-            // amx kernels enables for Q4_0, Q4_1, Q8_0, F16
-            // Q4_K, Q5_K, Q6_K, IQ4_XS enabled for QK_K = 256
-            bool has_amx_kernels = qtype_has_amx_kernels(type) || (type == GGML_TYPE_F16);
-
-            bool can_use_amx =
-                is_contiguous_2d(src0) &&       // src0 must be contiguous
-                is_contiguous_2d(src1) &&       // src1 must be contiguous
-                !is_training &&                 // inference only
-                src1->type == GGML_TYPE_F32 &&  // src1 must be float32
-                has_amx_kernels &&              // with amx kernel impls
-                ne0 % (TILE_N * 2) == 0;        // out_features is 32x
-
-            return can_use_amx;
-        }
-        default:
-            return false;
-    }
-
-    GGML_UNUSED(dev);
-}
-
-static bool ggml_backend_amx_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
-    return buft->iface.get_name == ggml_backend_amx_buffer_type_get_name;
-
-    GGML_UNUSED(dev);
-}
-
-static const struct ggml_backend_device_i ggml_backend_amx_device_i = {
-    /* .get_name             = */ ggml_backend_amx_device_get_name,
-    /* .get_description      = */ ggml_backend_amx_device_get_description,
-    /* .get_memory           = */ ggml_backend_amx_device_get_memory,
-    /* .get_type             = */ ggml_backend_amx_device_get_type,
-    /* .get_props            = */ ggml_backend_amx_device_get_props,
-    /* .init_backend         = */ ggml_backend_amx_device_init,
-    /* .get_buffer_type      = */ ggml_backend_amx_device_get_buffer_type,
-    /* .get_host_buffer_type = */ NULL,
-    /* .buffer_from_host_ptr = */ NULL,
-    /* .supports_op          = */ ggml_backend_amx_device_supports_op,
-    /* .supports_buft        = */ ggml_backend_amx_device_supports_buft,
-    /* .offload_op           = */ NULL,
-    /* .event_new            = */ NULL,
-    /* .event_free           = */ NULL,
-    /* .event_synchronize    = */ NULL,
-};
-
-// backend reg interface
-
-static const char * ggml_backend_amx_reg_get_name(ggml_backend_reg_t reg) {
-    return "AMX";
-
-    GGML_UNUSED(reg);
-}
-
-static size_t ggml_backend_amx_reg_get_device_count(ggml_backend_reg_t reg) {
-    return 1;
-
-    GGML_UNUSED(reg);
-}
-
-static ggml_backend_dev_t ggml_backend_amx_reg_get_device(ggml_backend_reg_t reg, size_t index) {
-    GGML_ASSERT(index == 0);
-
-    static ggml_backend_device ggml_backend_amx_device = {
-        /* .iface   = */ ggml_backend_amx_device_i,
-        /* .reg     = */ reg,
-        /* .context = */ nullptr,
-    };
-
-    return &ggml_backend_amx_device;
-
-    GGML_UNUSED(reg);
-    GGML_UNUSED(index);
-}
-
-static void * ggml_backend_amx_get_proc_address(ggml_backend_reg_t reg, const char * name) {
-    if (std::strcmp(name, "ggml_backend_set_n_threads") == 0) {
-        return (void *)ggml_backend_amx_set_n_threads;
-    }
-    return NULL;
-
-    GGML_UNUSED(reg);
-    GGML_UNUSED(name);
-}
-
-static const struct ggml_backend_reg_i ggml_backend_amx_reg_i = {
-    /* .get_name         = */ ggml_backend_amx_reg_get_name,
-    /* .get_device_count = */ ggml_backend_amx_reg_get_device_count,
-    /* .get_device       = */ ggml_backend_amx_reg_get_device,
-    /* .get_proc_address = */ ggml_backend_amx_get_proc_address,
-};
-
-ggml_backend_reg_t ggml_backend_amx_reg(void) {
-    static struct ggml_backend_reg ggml_backend_amx_reg = {
-        /* .iface   = */ ggml_backend_amx_reg_i,
-        /* .context = */ NULL,
-    };
-
-    return &ggml_backend_amx_reg;
-}
-
-#else // if defined(__AMX_INT8__)
-
-ggml_backend_t ggml_backend_amx_init(void) {
-    fprintf(stderr, "GGML is not compiled with AMX support!\n");
-    return ggml_backend_t{};
-}
-
-void ggml_backend_amx_set_n_threads(ggml_backend_t backend_amx, int n_threads) {
-    fprintf(stderr, "GGML is not compiled with AMX support!\n");
-
-    GGML_UNUSED(backend_amx);
-    GGML_UNUSED(n_threads);
-}
-
-#endif

ggml/src/ggml-backend.cpp

@@ -525,197 +525,6 @@ void * ggml_backend_reg_get_proc_address(ggml_backend_reg_t reg, const char * na
     return reg->iface.get_proc_address(reg, name);
 }
 
-// Backend registry
-
-#ifdef GGML_USE_CUDA
-#include "ggml-cuda.h"
-#endif
-
-#ifdef GGML_USE_METAL
-#include "ggml-metal.h"
-#endif
-
-#ifdef GGML_USE_SYCL
-#include "ggml-sycl.h"
-#endif
-
-#ifdef GGML_USE_VULKAN
-#include "ggml-vulkan.h"
-#endif
-
-#ifdef GGML_USE_BLAS
-#include "ggml-blas.h"
-#endif
-
-#ifdef GGML_USE_RPC
-#include "ggml-rpc.h"
-#endif
-
-#ifndef __AMX_INT8__
-#undef GGML_USE_AMX
-#endif
-
-#ifdef GGML_USE_AMX
-#  include "ggml-amx.h"
-#endif
-
-#ifdef GGML_USE_CANN
-#include "ggml-cann.h"
-#endif
-
-#ifdef GGML_USE_KOMPUTE
-#include "ggml-kompute.h"
-#endif
-
-#include "ggml-cpu.h"
-
-struct ggml_backend_registry {
-    std::vector<ggml_backend_reg_t> backends;
-    std::vector<ggml_backend_dev_t> devices;
-
-    ggml_backend_registry() {
-#ifdef GGML_USE_CUDA
-        register_backend(ggml_backend_cuda_reg());
-#endif
-#ifdef GGML_USE_METAL
-        register_backend(ggml_backend_metal_reg());
-#endif
-#ifdef GGML_USE_SYCL
-        register_backend(ggml_backend_sycl_reg());
-#endif
-#ifdef GGML_USE_VULKAN
-        register_backend(ggml_backend_vk_reg());
-#endif
-#ifdef GGML_USE_CANN
-        register_backend(ggml_backend_cann_reg());
-#endif
-#ifdef GGML_USE_BLAS
-        register_backend(ggml_backend_blas_reg());
-#endif
-#ifdef GGML_USE_RPC
-        register_backend(ggml_backend_rpc_reg());
-#endif
-#ifdef GGML_USE_AMX
-        register_backend(ggml_backend_amx_reg());
-#endif
-#ifdef GGML_USE_KOMPUTE
-        register_backend(ggml_backend_kompute_reg());
-#endif
-
-        register_backend(ggml_backend_cpu_reg());
-    }
-
-    void register_backend(ggml_backend_reg_t reg) {
-#ifndef NDEBUG
-        GGML_LOG_DEBUG("%s: registered backend %s (%zu devices)\n",
-            __func__, ggml_backend_reg_name(reg), ggml_backend_reg_dev_count(reg));
-#endif
-        backends.push_back(reg);
-        for (size_t i = 0; i < ggml_backend_reg_dev_count(reg); i++) {
-            register_device(ggml_backend_reg_dev_get(reg, i));
-        }
-    }
-
-    void register_device(ggml_backend_dev_t device) {
-#ifndef NDEBUG
-        GGML_LOG_DEBUG("%s: registered device %s (%s)\n", __func__, ggml_backend_dev_name(device), ggml_backend_dev_description(device));
-#endif
-        devices.push_back(device);
-    }
-};
-
-static ggml_backend_registry & get_reg() {
-    static ggml_backend_registry reg;
-    return reg;
-}
-
-// Internal API
-void ggml_backend_register(ggml_backend_reg_t reg) {
-    get_reg().register_backend(reg);
-}
-
-void ggml_backend_device_register(ggml_backend_dev_t device) {
-    get_reg().register_device(device);
-}
-
-// Backend (reg) enumeration
-size_t ggml_backend_reg_count() {
-    return get_reg().backends.size();
-}
-
-ggml_backend_reg_t ggml_backend_reg_get(size_t index) {
-    GGML_ASSERT(index < ggml_backend_reg_count());
-    return get_reg().backends[index];
-}
-
-ggml_backend_reg_t ggml_backend_reg_by_name(const char * name) {
-    for (size_t i = 0; i < ggml_backend_reg_count(); i++) {
-        ggml_backend_reg_t reg = ggml_backend_reg_get(i);
-        if (strcmp(ggml_backend_reg_name(reg), name) == 0) {
-            return reg;
-        }
-    }
-    return NULL;
-}
-
-// Device enumeration
-size_t ggml_backend_dev_count() {
-    return get_reg().devices.size();
-}
-
-ggml_backend_dev_t ggml_backend_dev_get(size_t index) {
-    GGML_ASSERT(index < ggml_backend_dev_count());
-    return get_reg().devices[index];
-}
-
-ggml_backend_dev_t ggml_backend_dev_by_name(const char * name) {
-    for (size_t i = 0; i < ggml_backend_dev_count(); i++) {
-        ggml_backend_dev_t dev = ggml_backend_dev_get(i);
-        if (strcmp(ggml_backend_dev_name(dev), name) == 0) {
-            return dev;
-        }
-    }
-    return NULL;
-}
-
-ggml_backend_dev_t ggml_backend_dev_by_type(enum ggml_backend_dev_type type) {
-    for (size_t i = 0; i < ggml_backend_dev_count(); i++) {
-        ggml_backend_dev_t dev = ggml_backend_dev_get(i);
-        if (ggml_backend_dev_type(dev) == type) {
-            return dev;
-        }
-    }
-    return NULL;
-}
-
-// Convenience functions
-ggml_backend_t ggml_backend_init_by_name(const char * name, const char * params) {
-    ggml_backend_dev_t dev = ggml_backend_dev_by_name(name);
-    if (!dev) {
-        return NULL;
-    }
-    return ggml_backend_dev_init(dev, params);
-}
-
-ggml_backend_t ggml_backend_init_by_type(enum ggml_backend_dev_type type, const char * params) {
-    ggml_backend_dev_t dev = ggml_backend_dev_by_type(type);
-    if (!dev) {
-        return NULL;
-    }
-    return ggml_backend_dev_init(dev, params);
-}
-
-ggml_backend_t ggml_backend_init_best(void) {
-    ggml_backend_dev_t dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_GPU);
-    if (!dev) {
-        dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
-    }
-    if (!dev) {
-        return NULL;
-    }
-    return ggml_backend_dev_init(dev, NULL);
-}
-
 // multi-buffer buffer
 
 struct ggml_backend_multi_buffer_context {
@@ -1641,7 +1450,7 @@ ggml_backend_sched_t ggml_backend_sched_new(
         bool parallel) {
     GGML_ASSERT(n_backends > 0);
     GGML_ASSERT(n_backends <= GGML_SCHED_MAX_BACKENDS);
-    GGML_ASSERT(ggml_backend_is_cpu(backends[n_backends - 1])); // last backend must be CPU
+    GGML_ASSERT(ggml_backend_dev_type(ggml_backend_get_device(backends[n_backends - 1])) == GGML_BACKEND_DEVICE_TYPE_CPU);
 
     struct ggml_backend_sched * sched = (ggml_backend_sched *) calloc(1, sizeof(struct ggml_backend_sched));
 
@@ -2038,17 +1847,6 @@ bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t
     return true;
 }
 
-
-
-#include "ggml-backend.h"
-#include "ggml-backend-impl.h"
-#include "ggml-cpu.h"
-#include "ggml-impl.h"
-#include <cctype>
-#include <string>
-
-// ggml-backend interface
-
 // CPU backend - buffer
 
 static void * ggml_backend_cpu_buffer_get_base(ggml_backend_buffer_t buffer) {
@@ -2122,7 +1920,9 @@ static const struct ggml_backend_buffer_i ggml_backend_cpu_buffer_from_ptr_i = {
     /* .reset           = */ NULL,
 };
 
-// CPU backend - buffer type
+// CPU backend buffer type
+
+// this buffer type is defined here to make it available to all backends
 
 static const char * ggml_backend_cpu_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
     return "CPU";
@@ -2163,7 +1963,7 @@ ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void) {
             /* .get_alloc_size   = */ NULL, // defaults to ggml_nbytes
             /* .is_host          = */ ggml_backend_cpu_buffer_type_is_host,
         },
-        /* .device  = */ ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
+        /* .device  = */ NULL, // FIXME ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
         /* .context = */ NULL,
     };
 
@@ -2186,479 +1986,14 @@ static ggml_backend_buffer_type_t ggml_backend_cpu_buffer_from_ptr_type(void) {
             /* .get_alloc_size   = */ NULL, // defaults to ggml_nbytes
             /* .is_host          = */ ggml_backend_cpu_buffer_type_is_host,
         },
-        /* .device  = */ ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
+        /* .device  = */ NULL, // FIXME ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
         /* .context = */ NULL,
     };
 
     return &ggml_backend_cpu_buffer_type;
 }
 
-#ifdef GGML_USE_CPU_HBM
-
-// buffer type HBM
-
-#include <hbwmalloc.h>
-
-static const char * ggml_backend_cpu_hbm_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
-    return "CPU_HBM";
-
-    GGML_UNUSED(buft);
-}
-
-static void ggml_backend_cpu_hbm_buffer_free_buffer(ggml_backend_buffer_t buffer) {
-    hbw_free(buffer->context);
-}
-
-static ggml_backend_buffer_t ggml_backend_cpu_hbm_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
-    void * ptr;
-    int result = hbw_posix_memalign(&ptr, ggml_backend_cpu_buffer_type_get_alignment(buft), size);
-    if (result != 0) {
-        GGML_LOG_ERROR("failed to allocate HBM buffer of size %zu\n", size);
-        return NULL;
-    }
-
-    ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(ptr, size);
-    buffer->buft = buft;
-    buffer->iface.free_buffer = ggml_backend_cpu_hbm_buffer_free_buffer;
-
-    return buffer;
-}
-
-ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void) {
-    static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type_hbm = {
-        /* .iface    = */ {
-            /* .get_name         = */ ggml_backend_cpu_hbm_buffer_type_get_name,
-            /* .alloc_buffer     = */ ggml_backend_cpu_hbm_buffer_type_alloc_buffer,
-            /* .get_alignment    = */ ggml_backend_cpu_buffer_type_get_alignment,
-            /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
-            /* .get_alloc_size   = */ NULL, // defaults to ggml_nbytes
-            /* .is_host          = */ ggml_backend_cpu_buffer_type_is_host,
-        },
-        /* .context  = */ NULL,
-    };
-
-    return &ggml_backend_cpu_buffer_type_hbm;
-}
-#endif
-
-static ggml_backend_buffer_type_t * ggml_backend_cpu_get_extra_bufts(ggml_backend_dev_t device) {
-    static ggml_backend_buffer_type_t bufts[] = {
-#ifdef GGML_USE_CPU_HBM
-        ggml_backend_cpu_hbm_buffer_type(),
-#endif
-        NULL
-    };
-
-    return bufts;
-
-    GGML_UNUSED(device);
-}
-
-// CPU backend - backend (stream)
-
-struct ggml_backend_cpu_context {
-    int                 n_threads;
-    ggml_threadpool_t   threadpool;
-
-    uint8_t *           work_data;
-    size_t              work_size;
-
-    ggml_abort_callback abort_callback;
-    void *              abort_callback_data;
-};
-
-static const char * ggml_backend_cpu_get_name(ggml_backend_t backend) {
-    return "CPU";
-
-    GGML_UNUSED(backend);
-}
-
-static void ggml_backend_cpu_free(ggml_backend_t backend) {
-    struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
-    delete[] cpu_ctx->work_data;
-    delete cpu_ctx;
-    delete backend;
-}
-
-struct ggml_backend_plan_cpu {
-    struct ggml_cplan cplan;
-    struct ggml_cgraph cgraph;
-};
-
-static ggml_backend_graph_plan_t ggml_backend_cpu_graph_plan_create(ggml_backend_t backend, const struct ggml_cgraph * cgraph) {
-    struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
-
-    struct ggml_backend_plan_cpu * cpu_plan = new ggml_backend_plan_cpu;
-
-    cpu_plan->cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads, cpu_ctx->threadpool);
-    cpu_plan->cgraph = *cgraph; // FIXME: deep copy
-
-    if (cpu_plan->cplan.work_size > 0) {
-        cpu_plan->cplan.work_data = new uint8_t[cpu_plan->cplan.work_size];
-        if (cpu_plan->cplan.work_data == NULL) {
-            delete cpu_plan;
-            return NULL;
-        }
-    }
-
-    cpu_plan->cplan.abort_callback      = cpu_ctx->abort_callback;
-    cpu_plan->cplan.abort_callback_data = cpu_ctx->abort_callback_data;
-
-    return cpu_plan;
-}
-
-static void ggml_backend_cpu_graph_plan_free(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
-    struct ggml_backend_plan_cpu * cpu_plan = (struct ggml_backend_plan_cpu *)plan;
-
-    delete[] cpu_plan->cplan.work_data;
-    delete cpu_plan;
-
-    GGML_UNUSED(backend);
-}
-
-static enum ggml_status ggml_backend_cpu_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
-    struct ggml_backend_plan_cpu * cpu_plan = (struct ggml_backend_plan_cpu *)plan;
-
-    return ggml_graph_compute(&cpu_plan->cgraph, &cpu_plan->cplan);
-
-    GGML_UNUSED(backend);
-}
-
-static enum ggml_status ggml_backend_cpu_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
-    struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
-
-    struct ggml_cplan cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads, cpu_ctx->threadpool);
-
-    if (cpu_ctx->work_size < cplan.work_size) {
-        delete[] cpu_ctx->work_data;
-        cpu_ctx->work_data = new uint8_t[cplan.work_size];
-        if (cpu_ctx->work_data == NULL) {
-            cpu_ctx->work_size = 0;
-            return GGML_STATUS_ALLOC_FAILED;
-        }
-        cpu_ctx->work_size = cplan.work_size;
-    }
-    cplan.work_data = (uint8_t *)cpu_ctx->work_data;
-
-    cplan.abort_callback      = cpu_ctx->abort_callback;
-    cplan.abort_callback_data = cpu_ctx->abort_callback_data;
-
-    return ggml_graph_compute(cgraph, &cplan);
-}
-
-static const struct ggml_backend_i ggml_backend_cpu_i = {
-    /* .get_name                = */ ggml_backend_cpu_get_name,
-    /* .free                    = */ ggml_backend_cpu_free,
-    /* .set_tensor_async        = */ NULL,
-    /* .get_tensor_async        = */ NULL,
-    /* .cpy_tensor_async        = */ NULL,
-    /* .synchronize             = */ NULL,
-    /* .graph_plan_create       = */ ggml_backend_cpu_graph_plan_create,
-    /* .graph_plan_free         = */ ggml_backend_cpu_graph_plan_free,
-    /* .graph_plan_update       = */ NULL,
-    /* .graph_plan_compute      = */ ggml_backend_cpu_graph_plan_compute,
-    /* .graph_compute           = */ ggml_backend_cpu_graph_compute,
-    /* .event_record            = */ NULL,
-    /* .event_wait              = */ NULL,
-};
-
-static ggml_guid_t ggml_backend_cpu_guid(void) {
-    static ggml_guid guid = { 0xaa, 0x67, 0xc7, 0x43, 0x96, 0xe6, 0xa3, 0x8a, 0xe3, 0xaf, 0xea, 0x92, 0x36, 0xbc, 0xfc, 0x89 };
-    return &guid;
-}
-
-ggml_backend_t ggml_backend_cpu_init(void) {
-    // initialize CPU backend now to avoid slowing the first graph computation
-    ggml_cpu_init();
-
-    struct ggml_backend_cpu_context * ctx = new ggml_backend_cpu_context;
-    if (ctx == NULL) {
-        return NULL;
-    }
-
-    ctx->n_threads           = GGML_DEFAULT_N_THREADS;
-    ctx->threadpool          = NULL;
-    ctx->work_data           = NULL;
-    ctx->work_size           = 0;
-    ctx->abort_callback      = NULL;
-    ctx->abort_callback_data = NULL;
-
-    ggml_backend_t cpu_backend = new ggml_backend {
-        /* .guid      = */ ggml_backend_cpu_guid(),
-        /* .interface = */ ggml_backend_cpu_i,
-        /* .device    = */ ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
-        /* .context   = */ ctx,
-    };
-
-    if (cpu_backend == NULL) {
-        delete ctx;
-        return NULL;
-    }
-
-    return cpu_backend;
-}
-
-bool ggml_backend_is_cpu(ggml_backend_t backend) {
-    return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_cpu_guid());
-}
-
-void ggml_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads) {
-    GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
-
-    struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
-    ctx->n_threads = n_threads;
-}
-
-void ggml_backend_cpu_set_threadpool(ggml_backend_t backend_cpu, ggml_threadpool_t threadpool) {
-    GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
-
-    struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
-
-    if (ctx->threadpool && ctx->threadpool != threadpool) {
-        // already had a different threadpool, pause/suspend it before switching
-        ggml_threadpool_pause(ctx->threadpool);
-    }
-    ctx->threadpool = threadpool;
-}
-
-void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data) {
-    GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
-
-    struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
-    ctx->abort_callback = abort_callback;
-    ctx->abort_callback_data = abort_callback_data;
-}
-
 ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(void * ptr, size_t size) {
     GGML_ASSERT((uintptr_t)ptr % TENSOR_ALIGNMENT == 0 && "buffer pointer must be aligned");
     return ggml_backend_buffer_init(ggml_backend_cpu_buffer_from_ptr_type(), ggml_backend_cpu_buffer_from_ptr_i, ptr, size);
 }
-
-// CPU backend - device
-
-struct ggml_backend_cpu_device_context {
-    std::string description = "CPU";
-
-    ggml_backend_cpu_device_context() {
-#ifdef __APPLE__
-        size_t len = 0;
-        if (!sysctlbyname("machdep.cpu.brand_string", NULL, &len, NULL, 0)) {
-            description.resize(len);
-            sysctlbyname("machdep.cpu.brand_string", &description[0], &len, NULL, 0); // NOLINT
-        }
-#elif defined(__linux__)
-        FILE * f = fopen("/proc/cpuinfo", "r");
-        if (f) {
-            char buf[1024];
-            while (fgets(buf, sizeof(buf), f)) {
-                if (strncmp(buf, "model name", 10) == 0) {
-                    char * p = strchr(buf, ':');
-                    if (p) {
-                        p++;
-                        while (std::isspace(*p)) {
-                            p++;
-                        }
-                        while (std::isspace(p[strlen(p) - 1])) {
-                            p[strlen(p) - 1] = '\0';
-                        }
-                        description = p;
-                        break;
-                    }
-                }
-            }
-            fclose(f);
-        }
-#elif defined(_WIN32)
-        HKEY hKey;
-        if (RegOpenKeyEx(HKEY_LOCAL_MACHINE,
-                        TEXT("HARDWARE\\DESCRIPTION\\System\\CentralProcessor\\0"),
-                        0,
-                        KEY_READ,
-                        &hKey) == ERROR_SUCCESS) {
-            DWORD cpu_brand_size = 0;
-            if (RegQueryValueExA(hKey,
-                                TEXT("ProcessorNameString"),
-                                NULL,
-                                NULL,
-                                NULL,
-                                &cpu_brand_size) == ERROR_SUCCESS) {
-                description.resize(cpu_brand_size);
-                if (RegQueryValueExA(hKey,
-                                    TEXT("ProcessorNameString"),
-                                    NULL,
-                                    NULL,
-                                    (LPBYTE)&description[0], // NOLINT
-                                    &cpu_brand_size) == ERROR_SUCCESS) {
-                    if (description.find('\0') != std::string::npos) {
-                        description.resize(description.find('\0'));
-                    }
-                }
-            }
-            RegCloseKey(hKey);
-        }
-#endif
-    }
-};
-
-static const char * ggml_backend_cpu_device_get_name(ggml_backend_dev_t dev) {
-    return "CPU";
-
-    GGML_UNUSED(dev);
-}
-
-static const char * ggml_backend_cpu_device_get_description(ggml_backend_dev_t dev) {
-    struct ggml_backend_cpu_device_context * ctx = (struct ggml_backend_cpu_device_context *)dev->context;
-
-    return ctx->description.c_str();
-}
-
-static void ggml_backend_cpu_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
-    // TODO
-    *free = 0;
-    *total = 0;
-
-    GGML_UNUSED(dev);
-}
-
-static enum ggml_backend_dev_type ggml_backend_cpu_device_get_type(ggml_backend_dev_t dev) {
-    return GGML_BACKEND_DEVICE_TYPE_CPU;
-
-    GGML_UNUSED(dev);
-}
-
-static void ggml_backend_cpu_device_get_props(ggml_backend_dev_t dev, struct ggml_backend_dev_props * props) {
-    props->name        = ggml_backend_cpu_device_get_name(dev);
-    props->description = ggml_backend_cpu_device_get_description(dev);
-    props->type        = ggml_backend_cpu_device_get_type(dev);
-    ggml_backend_cpu_device_get_memory(dev, &props->memory_free, &props->memory_total);
-    props->caps = {
-        /* .async                 = */ false,
-        /* .host_buffer           = */ false,
-        /* .buffer_from_host_ptr  = */ true,
-        /* .events                = */ false,
-    };
-}
-
-static ggml_backend_t ggml_backend_cpu_device_init_backend(ggml_backend_dev_t dev, const char * params) {
-    return ggml_backend_cpu_init();
-
-    GGML_UNUSED(dev);
-    GGML_UNUSED(params);
-}
-
-static ggml_backend_buffer_type_t ggml_backend_cpu_device_get_buffer_type(ggml_backend_dev_t dev) {
-    return ggml_backend_cpu_buffer_type();
-
-    GGML_UNUSED(dev);
-}
-
-static ggml_backend_buffer_t ggml_backend_cpu_device_buffer_from_host_ptr(ggml_backend_dev_t dev, void * ptr, size_t size, size_t max_tensor_size) {
-    return ggml_backend_cpu_buffer_from_ptr(ptr, size);
-
-    GGML_UNUSED(dev);
-    GGML_UNUSED(max_tensor_size);
-}
-
-static bool ggml_backend_cpu_device_supports_op(ggml_backend_dev_t dev, const struct ggml_tensor * op) {
-    switch (op->op) {
-        case GGML_OP_CPY:
-            return
-                op->type != GGML_TYPE_IQ2_XXS &&
-                op->type != GGML_TYPE_IQ2_XS  &&
-                op->type != GGML_TYPE_IQ1_S   &&
-                op->type != GGML_TYPE_IQ1_M; // missing type_traits.from_float
-        case GGML_OP_MUL_MAT:
-            //return op->src[1]->type == GGML_TYPE_F32; // TMP: workaround until sync with latest ggml
-            return op->src[1]->type == GGML_TYPE_F32 || op->src[1]->type == ggml_get_type_traits_cpu(op->src[0]->type)->vec_dot_type;
-        case GGML_OP_ROPE_BACK:
-            return op->src[2] == NULL && (op->op_params[2] & 4) == 0;
-        case GGML_OP_IM2COL_BACK:
-            return op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32;
-        case GGML_OP_OUT_PROD:
-            return (op->src[0]->type == GGML_TYPE_F32 || ggml_is_quantized(op->src[0]->type)) && op->src[1]->type == GGML_TYPE_F32;
-        default:
-            return true;
-    }
-
-    GGML_UNUSED(dev);
-}
-
-static bool ggml_backend_cpu_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
-    return ggml_backend_buft_is_host(buft);
-
-    GGML_UNUSED(dev);
-}
-
-static const struct ggml_backend_device_i ggml_backend_cpu_device_i = {
-    /* .get_name             = */ ggml_backend_cpu_device_get_name,
-    /* .get_description      = */ ggml_backend_cpu_device_get_description,
-    /* .get_memory           = */ ggml_backend_cpu_device_get_memory,
-    /* .get_type             = */ ggml_backend_cpu_device_get_type,
-    /* .get_props            = */ ggml_backend_cpu_device_get_props,
-    /* .init_backend         = */ ggml_backend_cpu_device_init_backend,
-    /* .get_buffer_type      = */ ggml_backend_cpu_device_get_buffer_type,
-    /* .get_host_buffer_type = */ NULL,
-    /* .buffer_from_host_ptr = */ ggml_backend_cpu_device_buffer_from_host_ptr,
-    /* .supports_op          = */ ggml_backend_cpu_device_supports_op,
-    /* .supports_buft        = */ ggml_backend_cpu_device_supports_buft,
-    /* .offload_op           = */ NULL,
-    /* .event_new            = */ NULL,
-    /* .event_free           = */ NULL,
-    /* .event_synchronize    = */ NULL,
-};
-
-// CPU backend - backend (reg)
-
-static const char * ggml_backend_cpu_reg_get_name(ggml_backend_reg_t reg) {
-    return "CPU";
-
-    GGML_UNUSED(reg);
-}
-
-static size_t ggml_backend_cpu_reg_get_device_count(ggml_backend_reg_t reg) {
-    return 1;
-
-    GGML_UNUSED(reg);
-}
-
-static ggml_backend_dev_t ggml_backend_cpu_reg_get_device(ggml_backend_reg_t reg, size_t index) {
-    GGML_ASSERT(index == 0);
-
-    static ggml_backend_cpu_device_context ctx;
-    static ggml_backend_device ggml_backend_cpu_device = {
-        /* .iface   = */ ggml_backend_cpu_device_i,
-        /* .reg     = */ reg,
-        /* .context = */ &ctx,
-    };
-
-    return &ggml_backend_cpu_device;
-}
-
-static void * ggml_backend_cpu_get_proc_address(ggml_backend_reg_t reg, const char * name) {
-    if (strcmp(name, "ggml_backend_set_n_threads") == 0) {
-        return (void *)ggml_backend_cpu_set_n_threads;
-    }
-    if (strcmp(name, "ggml_backend_dev_get_extra_bufts") == 0) {
-        return (void *)ggml_backend_cpu_get_extra_bufts;
-    }
-
-    return NULL;
-
-    GGML_UNUSED(reg);
-}
-
-static const struct ggml_backend_reg_i ggml_backend_cpu_reg_i = {
-    /* .get_name         = */ ggml_backend_cpu_reg_get_name,
-    /* .get_device_count = */ ggml_backend_cpu_reg_get_device_count,
-    /* .get_device       = */ ggml_backend_cpu_reg_get_device,
-    /* .get_proc_address = */ ggml_backend_cpu_get_proc_address,
-};
-
-ggml_backend_reg_t ggml_backend_cpu_reg(void) {
-    static struct ggml_backend_reg ggml_backend_cpu_reg = {
-        /* .iface   = */ ggml_backend_cpu_reg_i,
-        /* .context = */ NULL,
-    };
-
-    return &ggml_backend_cpu_reg;
-}

ggml/src/ggml-blas.cpp

@@ -1,514 +0,0 @@
-#include "ggml-impl.h"
-#include "ggml-blas.h"
-#include "ggml-backend-impl.h"
-
-#include <future>
-#include <vector>
-#include <cstring>
-
-#if defined(GGML_USE_ACCELERATE)
-#   include <Accelerate/Accelerate.h>
-#elif defined(GGML_BLAS_USE_MKL)
-#   include <mkl.h>
-#elif defined(GGML_BLAS_USE_BLIS)
-#   include <blis.h>
-#elif defined(GGML_BLAS_USE_NVPL)
-#   include <nvpl_blas.h>
-#else
-#   include <cblas.h>
-#endif
-
-struct ggml_backend_blas_context {
-    int n_threads = GGML_DEFAULT_N_THREADS;
-    std::unique_ptr<char[]> work_data;
-    size_t work_size = 0;
-#ifndef GGML_USE_OPENMP
-    std::vector<std::future<void>> tasks;
-#endif
-};
-
-static void ggml_backend_blas_mul_mat(ggml_backend_blas_context * ctx, struct ggml_tensor * dst) {
-    const struct ggml_tensor * src0 = dst->src[0];
-    const struct ggml_tensor * src1 = dst->src[1];
-
-    GGML_TENSOR_BINARY_OP_LOCALS
-
-    const enum ggml_type type = src0->type;
-
-    GGML_ASSERT(ne0 == ne01);
-    GGML_ASSERT(ne1 == ne11);
-    GGML_ASSERT(ne2 == ne12);
-    GGML_ASSERT(ne3 == ne13);
-
-    // we don't support permuted src0 or src1
-    GGML_ASSERT(nb00 == ggml_type_size(type));
-    GGML_ASSERT(nb10 == ggml_type_size(src1->type));
-
-    // dst cannot be transposed or permuted
-    GGML_ASSERT(nb0 == sizeof(float));
-    GGML_ASSERT(nb0 <= nb1);
-    GGML_ASSERT(nb1 <= nb2);
-    GGML_ASSERT(nb2 <= nb3);
-
-    // broadcast factors
-    const int64_t r2 = ne12/ne02;
-    const int64_t r3 = ne13/ne03;
-
-    const int64_t ne_plane      = ne01*ne00;
-    const size_t  desired_wsize = type == GGML_TYPE_F32 ? 0 : ne03*ne02*ne_plane*sizeof(float);
-
-    if (ctx->work_size < desired_wsize) {
-        ctx->work_data.reset(new char[desired_wsize]);
-        ctx->work_size = desired_wsize;
-    }
-    void * wdata = ctx->work_data.get();
-
-    // convert src0 to float
-    if (type != GGML_TYPE_F32) {
-        const auto * type_traits = ggml_get_type_traits(type);
-        ggml_to_float_t const to_float = type_traits->to_float;
-
-        for (int64_t i03 = 0; i03 < ne03; i03++) {
-            for (int64_t i02 = 0; i02 < ne02; i02++) {
-                const void  *       x      = (char *)  src0->data + i02*nb02          + i03*nb03;
-                      float * const wplane = (float *) wdata      + i02*ne_plane      + i03*ne02*ne_plane;
-
-                const int min_cols_per_thread = 4096;
-                const int min_rows_per_thread = std::max((int)(min_cols_per_thread/ne00), 1);
-                const int n_threads = std::max(std::min(ctx->n_threads, (int)(ne01/min_rows_per_thread)), 1);
-
-#ifdef GGML_USE_OPENMP
-                #pragma omp parallel for num_threads(n_threads)
-                for (int64_t i01 = 0; i01 < ne01; i01++) {
-                    to_float((const char *) x + i01*nb01, wplane + i01*ne00, ne00);
-                }
-#else
-                for (int i = 1; i < n_threads; i++) {
-                    const int64_t start =       i*ne01/n_threads;
-                    const int64_t end   = (i + 1)*ne01/n_threads;
-                    if (start < end) {
-                        ctx->tasks.push_back(std::async(std::launch::async, [=]() {
-                            for (int64_t i01 = start; i01 < end; i01++) {
-                                to_float((const char *) x + i01*nb01, wplane + i01*ne00, ne00);
-                            }
-                        }));
-                    }
-                }
-                {
-                    // reuse the current thread for the first task
-                    const int64_t start = 0;
-                    const int64_t end   = ne01/n_threads;
-                    for (int64_t i01 = start; i01 < end; i01++) {
-                        to_float((const char *) x + i01*nb01, wplane + i01*ne00, ne00);
-                    }
-                }
-#endif
-            }
-        }
-
-#ifndef GGML_USE_OPENMP
-        // wait for all tasks to finish
-        for (auto & task : ctx->tasks) {
-            task.get();
-        }
-        ctx->tasks.clear();
-#endif
-    }
-
-#if defined(OPENBLAS_VERSION)
-    openblas_set_num_threads(ctx->n_threads);
-#endif
-
-#if defined(GGML_BLAS_USE_BLIS)
-    bli_thread_set_num_threads(ctx->n_threads);
-#endif
-
-#if defined(GGML_BLAS_USE_NVPL)
-    nvpl_blas_set_num_threads(ctx->n_threads);
-#endif
-
-    for (int64_t i13 = 0; i13 < ne13; i13++) {
-        for (int64_t i12 = 0; i12 < ne12; i12++) {
-            const int64_t i03 = i13/r3;
-            const int64_t i02 = i12/r2;
-
-            const float * x = (float *) ((char *) src0->data + i02*nb02 + i03*nb03);
-            const float * y = (float *) ((char *) src1->data + i12*nb12 + i13*nb13);
-                  float * d = (float *) ((char *)  dst->data + i12*nb2  + i13*nb3);
-
-            if (type != GGML_TYPE_F32) {
-                x = (float *) wdata + i02*ne_plane + i03*ne02*ne_plane;
-            }
-
-            cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasTrans,
-                        ne1, ne01, ne10,
-                        1.0f,   y, ne10,
-                                x, ne00,
-                        0.0f,   d, ne01);
-        }
-    }
-}
-
-static void ggml_backend_blas_out_prod(ggml_backend_blas_context * ctx, struct ggml_tensor * dst) {
-    const struct ggml_tensor * src0 = dst->src[0];
-    const struct ggml_tensor * src1 = dst->src[1];
-
-    GGML_TENSOR_BINARY_OP_LOCALS
-
-    GGML_ASSERT(ne0  == ne00);
-    GGML_ASSERT(ne1  == ne10);
-    GGML_ASSERT(ne2  == ne02);
-    GGML_ASSERT(ne02 == ne12);
-    GGML_ASSERT(ne3  == ne13);
-    GGML_ASSERT(ne03 == ne13);
-
-    // we don't support permuted src0 or src1
-    GGML_ASSERT(nb00 == sizeof(float));
-
-    // dst cannot be transposed or permuted
-    GGML_ASSERT(nb0 == sizeof(float));
-    // GGML_ASSERT(nb0 <= nb1);
-    // GGML_ASSERT(nb1 <= nb2);
-    // GGML_ASSERT(nb2 <= nb3);
-
-    // Arguments to ggml_compute_forward_out_prod (expressed as major,minor)
-    // src0: (k,n)
-    // src1: (k,m)
-    // dst:  (m,n)
-    //
-    // Arguments to sgemm (see https://github.com/Reference-LAPACK/lapack/blob/master/BLAS/SRC/sgemm.f)
-    // Also expressed as (major,minor)
-    // a: (m,k): so src1 transposed
-    // b: (k,n): so src0
-    // c: (m,n)
-    //
-    // However, if ggml_is_transposed(src1) is true, then
-    // src1->data already contains a transposed version, so sgemm mustn't
-    // transpose it further.
-
-    int n = src0->ne[0];
-    int k = src0->ne[1];
-    int m = src1->ne[0];
-
-    CBLAS_TRANSPOSE transposeA;
-    int lda;
-
-    if (!ggml_is_transposed(src1)) {
-        transposeA = CblasTrans;
-        lda = m;
-    } else {
-        transposeA = CblasNoTrans;
-        lda = k;
-    }
-
-    float * a = (float *) ((char *) src1->data);
-    float * b = (float *) ((char *) src0->data);
-    float * c = (float *) ((char *) dst->data);
-
-    cblas_sgemm(CblasRowMajor, transposeA, CblasNoTrans, m, n, k, 1.0, a, lda, b, n, 0.0, c, n);
-
-    GGML_UNUSED(ctx);
-}
-
-// backend interface
-
-static const char * ggml_backend_blas_get_name(ggml_backend_t backend) {
-    return "BLAS";
-
-    GGML_UNUSED(backend);
-}
-
-static void ggml_backend_blas_free(ggml_backend_t backend) {
-    ggml_backend_blas_context * ctx = (ggml_backend_blas_context *)backend->context;
-    delete ctx;
-    delete backend;
-}
-
-static enum ggml_status ggml_backend_blas_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
-    ggml_backend_blas_context * ctx = (ggml_backend_blas_context *)backend->context;
-
-    for (int i = 0; i < cgraph->n_nodes; i++) {
-        struct ggml_tensor * node = cgraph->nodes[i];
-
-        switch (node->op) {
-            case GGML_OP_MUL_MAT:
-                ggml_backend_blas_mul_mat(ctx, node);
-                break;
-
-            case GGML_OP_OUT_PROD:
-                ggml_backend_blas_out_prod(ctx, node);
-                break;
-
-            case GGML_OP_NONE:
-            case GGML_OP_RESHAPE:
-            case GGML_OP_VIEW:
-            case GGML_OP_PERMUTE:
-            case GGML_OP_TRANSPOSE:
-                break;
-
-            default:
-                GGML_ABORT("%s: unsupported op %s\n", __func__, ggml_op_desc(node));
-        }
-    }
-
-    return GGML_STATUS_SUCCESS;
-
-    GGML_UNUSED(backend);
-}
-
-static struct ggml_backend_i blas_backend_i = {
-    /* .get_name                = */ ggml_backend_blas_get_name,
-    /* .free                    = */ ggml_backend_blas_free,
-    /* .set_tensor_async        = */ NULL,
-    /* .get_tensor_async        = */ NULL,
-    /* .cpy_tensor_async        = */ NULL,
-    /* .synchronize             = */ NULL,
-    /* .graph_plan_create       = */ NULL,
-    /* .graph_plan_free         = */ NULL,
-    /* .graph_plan_update       = */ NULL,
-    /* .graph_plan_compute      = */ NULL,
-    /* .graph_compute           = */ ggml_backend_blas_graph_compute,
-    /* .event_record            = */ NULL,
-    /* .event_wait              = */ NULL,
-};
-
-static ggml_guid_t ggml_backend_blas_guid(void) {
-    static ggml_guid guid = { 0x12, 0xa8, 0xae, 0xf4, 0xc0, 0x1e, 0x61, 0x97, 0x8f, 0xeb, 0x33, 0x04, 0xa1, 0x33, 0x51, 0x2d };
-    return &guid;
-}
-
-ggml_backend_t ggml_backend_blas_init(void) {
-    ggml_backend_blas_context * ctx = new ggml_backend_blas_context;
-
-    ggml_backend_t backend = new ggml_backend {
-        /* .guid      = */ ggml_backend_blas_guid(),
-        /* .interface = */ blas_backend_i,
-        /* .device    = */ ggml_backend_reg_dev_get(ggml_backend_blas_reg(), 0),
-        /* .context   = */ ctx,
-    };
-
-#if defined(OPENBLAS_VERSION) && defined(GGML_USE_OPENMP)
-    if (openblas_get_parallel() != OPENBLAS_OPENMP) {
-        GGML_LOG_DEBUG("%s: warning: ggml is using OpenMP, but OpenBLAS was compiled without OpenMP support\n", __func__);
-    }
-#endif
-
-#if defined(BLIS_ENABLE_CBLAS) && defined(GGML_USE_OPENMP) && !defined(BLIS_ENABLE_OPENMP)
-    GGML_LOG_DEBUG("%s: warning: ggml is using OpenMP, but BLIS was compiled without OpenMP support\n", __func__);
-#endif
-
-    return backend;
-}
-
-bool ggml_backend_is_blas(ggml_backend_t backend) {
-    return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_blas_guid());
-}
-
-void ggml_backend_blas_set_n_threads(ggml_backend_t backend_blas, int n_threads) {
-    GGML_ASSERT(ggml_backend_is_blas(backend_blas));
-
-    ggml_backend_blas_context * ctx = (ggml_backend_blas_context *)backend_blas->context;
-    ctx->n_threads = n_threads;
-}
-
-// device interface
-
-static const char * ggml_backend_blas_device_get_name(ggml_backend_dev_t dev) {
-    return "BLAS";
-
-    GGML_UNUSED(dev);
-}
-
-static const char * ggml_backend_blas_device_get_description(ggml_backend_dev_t dev) {
-    #if defined(GGML_USE_ACCELERATE)
-        return "Accelerate";
-    #elif defined(GGML_BLAS_USE_MKL)
-        return "MKL";
-    #elif defined(GGML_BLAS_USE_BLIS)
-        return "BLIS";
-    #elif defined(GGML_BLAS_USE_NVPL)
-        return "NVPL";
-    #elif defined(OPENBLAS_VERSION)
-        return "OpenBLAS";
-    #else
-        return "BLAS";
-    #endif
-
-    GGML_UNUSED(dev);
-}
-
-static void ggml_backend_blas_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
-    // TODO
-    *free = 0;
-    *total = 0;
-
-    GGML_UNUSED(dev);
-}
-
-static enum ggml_backend_dev_type ggml_backend_blas_device_get_type(ggml_backend_dev_t dev) {
-    return GGML_BACKEND_DEVICE_TYPE_ACCEL;
-
-    GGML_UNUSED(dev);
-}
-
-static void ggml_backend_blas_device_get_props(ggml_backend_dev_t dev, struct ggml_backend_dev_props * props) {
-    props->name        = ggml_backend_blas_device_get_name(dev);
-    props->description = ggml_backend_blas_device_get_description(dev);
-    props->type        = ggml_backend_blas_device_get_type(dev);
-    ggml_backend_blas_device_get_memory(dev, &props->memory_free, &props->memory_total);
-    props->caps = {
-        /* .async                 = */ false,
-        /* .host_buffer           = */ false,
-        /* .buffer_from_host_ptr  = */ true,
-        /* .events                = */ false,
-    };
-}
-
-static ggml_backend_t ggml_backend_blas_device_init_backend(ggml_backend_dev_t dev, const char * params) {
-    return ggml_backend_blas_init();
-
-    GGML_UNUSED(dev);
-    GGML_UNUSED(params);
-}
-
-static ggml_backend_buffer_type_t ggml_backend_blas_device_get_buffer_type(ggml_backend_dev_t dev) {
-    return ggml_backend_cpu_buffer_type();
-
-    GGML_UNUSED(dev);
-}
-
-static ggml_backend_buffer_t ggml_backend_blas_device_buffer_from_host_ptr(ggml_backend_dev_t dev, void * ptr, size_t size, size_t max_tensor_size) {
-    return ggml_backend_cpu_buffer_from_ptr(ptr, size);
-
-    GGML_UNUSED(dev);
-    GGML_UNUSED(max_tensor_size);
-}
-
-static bool ggml_backend_blas_device_supports_op(ggml_backend_dev_t dev, const struct ggml_tensor * op) {
-    const struct ggml_tensor * src0 = op->src[0];
-    const struct ggml_tensor * src1 = op->src[1];
-
-    switch (op->op) {
-        case GGML_OP_NONE:
-        case GGML_OP_RESHAPE:
-        case GGML_OP_VIEW:
-        case GGML_OP_PERMUTE:
-        case GGML_OP_TRANSPOSE:
-            return true;
-
-        case GGML_OP_MUL_MAT:
-        {
-            // BLAS usually is only faster for large matrices
-            const struct ggml_tensor * src0 = op->src[0];
-            const struct ggml_tensor * src1 = op->src[1];
-
-            const int64_t ne10 = src1->ne[0];
-
-            const int64_t ne0 = op->ne[0];
-            const int64_t ne1 = op->ne[1];
-
-            // TODO: find the optimal value
-            const int64_t min_batch = 32;
-
-            return ggml_is_contiguous(src0) &&
-                   ggml_is_contiguous(src1) &&
-                   src1->type == GGML_TYPE_F32 &&
-                   (ne0 >= min_batch && ne1 >= min_batch && ne10 >= min_batch) &&
-                   (src0->type == GGML_TYPE_F32 || ggml_get_type_traits(src0->type)->to_float != NULL);
-        }
-
-        case GGML_OP_OUT_PROD:
-            return op->src[0]->type == GGML_TYPE_F32 &&
-                   op->src[1]->type == GGML_TYPE_F32 &&
-                   ggml_is_matrix(src0) &&
-                   ggml_is_matrix(src1) &&
-                   ggml_is_contiguous(src0) &&
-                   (ggml_is_contiguous(src1) || ggml_is_transposed(src1)) &&
-                   (src0->type == GGML_TYPE_F32 || ggml_get_type_traits(src0->type)->to_float != NULL);
-
-        default:
-            return false;
-
-    }
-
-    GGML_UNUSED(dev);
-}
-
-static bool ggml_backend_blas_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
-    return ggml_backend_buft_is_host(buft);
-
-    GGML_UNUSED(dev);
-}
-
-static const struct ggml_backend_device_i ggml_backend_blas_device_i = {
-    /* .get_name             = */ ggml_backend_blas_device_get_name,
-    /* .get_description      = */ ggml_backend_blas_device_get_description,
-    /* .get_memory           = */ ggml_backend_blas_device_get_memory,
-    /* .get_type             = */ ggml_backend_blas_device_get_type,
-    /* .get_props            = */ ggml_backend_blas_device_get_props,
-    /* .init_backend         = */ ggml_backend_blas_device_init_backend,
-    /* .get_buffer_type      = */ ggml_backend_blas_device_get_buffer_type,
-    /* .get_host_buffer_type = */ NULL,
-    /* .buffer_from_host_ptr = */ ggml_backend_blas_device_buffer_from_host_ptr,
-    /* .supports_op          = */ ggml_backend_blas_device_supports_op,
-    /* .supports_buft        = */ ggml_backend_blas_device_supports_buft,
-    /* .offload_op           = */ NULL,
-    /* .event_new            = */ NULL,
-    /* .event_free           = */ NULL,
-    /* .event_synchronize    = */ NULL,
-};
-
-// backend reg interface
-
-static const char * ggml_backend_blas_reg_get_name(ggml_backend_reg_t reg) {
-    return "BLAS";
-
-    GGML_UNUSED(reg);
-}
-
-static size_t ggml_backend_blas_reg_get_device_count(ggml_backend_reg_t reg) {
-    return 1;
-
-    GGML_UNUSED(reg);
-}
-
-static ggml_backend_dev_t ggml_backend_blas_reg_get_device(ggml_backend_reg_t reg, size_t index) {
-    GGML_ASSERT(index == 0);
-
-    static ggml_backend_device ggml_backend_blas_device = {
-        /* .iface   = */ ggml_backend_blas_device_i,
-        /* .reg     = */ reg,
-        /* .context = */ nullptr,
-    };
-
-    return &ggml_backend_blas_device;
-
-    GGML_UNUSED(reg);
-    GGML_UNUSED(index);
-}
-
-static void * ggml_backend_blas_get_proc_address(ggml_backend_reg_t reg, const char * name) {
-    if (std::strcmp(name, "ggml_backend_set_n_threads") == 0) {
-        return (void *)ggml_backend_blas_set_n_threads;
-    }
-    return NULL;
-
-    GGML_UNUSED(reg);
-    GGML_UNUSED(name);
-}
-
-static const struct ggml_backend_reg_i ggml_backend_blas_reg_i = {
-    /* .get_name         = */ ggml_backend_blas_reg_get_name,
-    /* .get_device_count = */ ggml_backend_blas_reg_get_device_count,
-    /* .get_device       = */ ggml_backend_blas_reg_get_device,
-    /* .get_proc_address = */ ggml_backend_blas_get_proc_address,
-};
-
-ggml_backend_reg_t ggml_backend_blas_reg(void) {
-    static struct ggml_backend_reg ggml_backend_blas_reg = {
-        /* .iface   = */ ggml_backend_blas_reg_i,
-        /* .context = */ NULL,
-    };
-
-    return &ggml_backend_blas_reg;
-}

ggml/src/ggml-cann.cpp

@@ -1,2128 +0,0 @@
-/*
- * Copyright (c) 2023-2024 The ggml authors
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to
- * deal in the Software without restriction, including without limitation the
- * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
- * sell copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in
- * all copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
- * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
- * IN THE SOFTWARE.
- */
-
-#include "ggml-cann.h"
-
-#include <acl/acl.h>
-#include <stdarg.h>
-
-#include <cmath>
-#include <cstdio>
-#include <cstring>
-#include <mutex>
-
-#include "ggml-impl.h"
-#include "ggml-backend-impl.h"
-#include "ggml-cann/aclnn_ops.h"
-#include "ggml-cann/common.h"
-
-#define GGML_COMMON_DECL_C
-
-#include "ggml-common.h"
-
-#define GGML_CANN_NAME "CANN"
-
-/**
- * @brief Handles CANN errors by printing an error message and aborting.
- *
- * @param stmt The statement that caused the error.
- * @param func The function in which the error occurred.
- * @param file The file in which the error occurred.
- * @param line The line number where the error occurred.
- * @param msg The error message.
- */
-[[noreturn]] void ggml_cann_error(const char* stmt, const char* func,
-                                  const char* file, int line, const char* msg) {
-    int32_t id = -1;
-    aclrtGetDevice(&id);
-
-    GGML_LOG_ERROR("CANN error: %s\n", msg);
-    GGML_LOG_ERROR("  current device: %d, in function %s at %s:%d\n", id, func,
-            file, line);
-    GGML_LOG_ERROR("  %s\n", stmt);
-    // abort with GGML_ASSERT to get a stack trace
-    GGML_ABORT("CANN error");
-}
-
-/**
- * @brief Sets the device to be used by CANN.
- *
- * @param device The device ID to set.
- */
-void ggml_cann_set_device(const int32_t device) {
-    // TODO: uncomment these lines after empty context has fixed.
-    // int current_device;
-    // ACL_CHECK(aclrtGetDevice(&current_device));
-
-    // if (device == current_device) {
-    //   return;
-    // }
-    ACL_CHECK(aclrtSetDevice(device));
-}
-
-/**
- * @brief Retrieves the current device ID.
- *
- * @return The current device ID.
- */
-int32_t ggml_cann_get_device() {
-    int32_t id;
-    ACL_CHECK(aclrtGetDevice(&id));
-    return id;
-}
-
-/**
- * @brief Initialize the CANN device information.
- *
- * This function initializes the CANN device information by obtaining the
- * device count and setting the memory allocation granularity for each device.
- *
- * @return A structure containing the device information.
- */
-static ggml_cann_device_info ggml_cann_init() {
-    ggml_cann_device_info info = {};
-
-    aclError err = aclrtGetDeviceCount((uint32_t*)&info.device_count);
-
-    if (err != ACL_SUCCESS) {
-        GGML_LOG_ERROR("%s: failed to initialize CANN: %s\n",
-                __func__, aclGetRecentErrMsg());
-        return info;
-    }
-
-    GGML_ASSERT(info.device_count <= GGML_CANN_MAX_DEVICES);
-
-    for (int id = 0; id < info.device_count; ++id) {
-        aclrtPhysicalMemProp prop = {};
-        prop.handleType = ACL_MEM_HANDLE_TYPE_NONE;
-        prop.allocationType = ACL_MEM_ALLOCATION_TYPE_PINNED;
-        prop.memAttr = ACL_HBM_MEM_HUGE;
-        prop.location.type = ACL_MEM_LOCATION_TYPE_DEVICE;
-        prop.location.id = id;
-        prop.reserve = 0;
-        ACL_CHECK(aclrtMemGetAllocationGranularity(
-            &prop, ACL_RT_MEM_ALLOC_GRANULARITY_RECOMMENDED,
-            &info.devices[id].vmm_granularity));
-    }
-
-    // TODO: add more device info later.
-    return info;
-}
-
-/**
- * @brief Retrieve the CANN device information.
- *
- * This function returns a reference to a structure containing the CANN device
- * information. The device information is initialized once and reused on
- * subsequent calls.
- *
- * @return A reference to the structure containing the device information.
- */
-const ggml_cann_device_info& ggml_cann_info() {
-    static ggml_cann_device_info info = ggml_cann_init();
-    return info;
-}
-
-//#define DEBUG_CANN_MALLOC
-/**
- * @brief A pool of CANN buffers(legacy).
- *
- * This class manages a pool of CANN buffers for a specific device.
- */
-struct ggml_cann_pool_leg : public ggml_cann_pool {
-    /**
-     * @brief The maximum number of buffers in the pool.
-     */
-    static const int MAX_BUFFERS = 256;
-
-    /**
-     * @brief The device ID associated with this buffer pool.
-     */
-    int device;
-
-    /**
-     * @brief Structure representing a CANN buffer.
-     */
-    struct ggml_cann_buffer {
-        void* ptr = nullptr;  ///< Pointer to the buffer memory.
-        size_t size = 0;      ///< Size of the buffer.
-    };
-
-    /**
-     * @brief Array of CANN buffers in the pool.
-     */
-    ggml_cann_buffer buffer_pool[MAX_BUFFERS] = {};
-
-    /**
-     * @brief Total size of all buffers in the pool.
-     */
-    size_t pool_size = 0;
-
-    /**
-     * @brief Constructor to initialize the buffer pool for a specific device.
-     *
-     * @param device The device ID to associate with this buffer pool.
-     */
-    explicit ggml_cann_pool_leg(int device) : device(device) {}
-
-    /**
-     * @brief Destructor to free all buffers in the pool.
-     */
-    ~ggml_cann_pool_leg() {
-        ggml_cann_set_device(device);
-        for (int i = 0; i < MAX_BUFFERS; ++i) {
-            ggml_cann_buffer& b = buffer_pool[i];
-            if (b.ptr != nullptr) {
-                ACL_CHECK(aclrtFree(b.ptr));
-                pool_size -= b.size;
-            }
-        }
-        GGML_ASSERT(pool_size == 0);
-    }
-
-    /**
-     * @brief Allocate a buffer of the given size.
-     *
-     * @param size The size of the buffer to allocate.
-     * @param actual_size A pointer to a variable to receive the actual size of
-     * the allocated buffer.
-     * @return A pointer to the allocated buffer.
-     */
-    void* alloc(size_t size, size_t* actual_size) override {
-#ifdef DEBUG_CANN_MALLOC
-        int nnz = 0;
-        size_t max_size = 0;
-#endif
-        size_t best_diff = 1ull << 36;
-        int ibest = -1;
-        for (int i = 0; i < MAX_BUFFERS; ++i) {
-            ggml_cann_buffer& b = buffer_pool[i];
-            if (b.ptr != nullptr) {
-#ifdef DEBUG_CANN_MALLOC
-                ++nnz;
-                if (b.size > max_size) max_size = b.size;
-#endif
-                if (b.size >= size) {
-                    size_t diff = b.size - size;
-                    if (diff < best_diff) {
-                        best_diff = diff;
-                        ibest = i;
-                        if (!best_diff) {
-                            void* ptr = b.ptr;
-                            *actual_size = b.size;
-                            b.ptr = nullptr;
-                            b.size = 0;
-                            return ptr;
-                        }
-                    }
-                }
-            }
-        }
-        if (ibest >= 0) {
-            ggml_cann_buffer& b = buffer_pool[ibest];
-            void* ptr = b.ptr;
-            *actual_size = b.size;
-            b.ptr = nullptr;
-            b.size = 0;
-            return ptr;
-        }
-        void* ptr;
-        size_t look_ahead_size = (size_t)(1.05 * size);
-        look_ahead_size = 256 * ((look_ahead_size + 255) / 256);
-        ggml_cann_set_device(device);
-        ACL_CHECK(
-            aclrtMalloc(&ptr, look_ahead_size, ACL_MEM_MALLOC_HUGE_FIRST));
-        *actual_size = look_ahead_size;
-        pool_size += look_ahead_size;
-#ifdef DEBUG_CANN_MALLOC
-        GGML_LOG_INFO(
-            "%s[%d]: %d buffers, max_size = %u MB, pool_size = %u MB, "
-            "requested %u MB\n",
-            __func__, device, nnz, (uint32_t)(max_size / 1024 / 1024),
-            (uint32_t)(pool_size / 1024 / 1024),
-            (uint32_t)(size / 1024 / 1024));
-#endif
-        return ptr;
-    }
-
-    /**
-     * @brief Free a buffer and return it to the pool.
-     *
-     * @param ptr Pointer to the buffer to free.
-     * @param size Size of the buffer to free.
-     */
-    void free(void* ptr, size_t size) override {
-        for (int i = 0; i < MAX_BUFFERS; ++i) {
-            ggml_cann_buffer& b = buffer_pool[i];
-            if (b.ptr == nullptr) {
-                b.ptr = ptr;
-                b.size = size;
-                return;
-            }
-        }
-        // memory should always buffered. these memory may still needed by
-        // tasks in stream.
-        // TODO, fix me.
-        GGML_ABORT("Cann buffer pool full, increase MAX_CANN_BUFFERS\n");
-    }
-};
-
-/**
- * @brief A pool of CANN buffers with virtual memory.
- *
- * This class manages a pool of CANN buffers with virtual memory for a specific
- * device.
- */
-struct ggml_cann_pool_vmm : public ggml_cann_pool {
-    /**
-     * @brief The maximum size of the virtual memory pool (32 GB).
-     */
-    static const size_t CANN_POOL_VMM_MAX_SIZE = 1ull << 35;  // 32 GB
-
-    /**
-     * @brief The device ID associated with this buffer pool.
-     */
-    int device;
-
-    /**
-     * @brief Pointer to the start of the virtual memory pool.
-     */
-    void* pool_addr = 0;
-
-    /**
-     * @brief Amount of virtual memory used in the pool.
-     */
-    size_t pool_used = 0;
-
-    /**
-     * @brief Total size of the virtual memory pool.
-     */
-    size_t pool_size = 0;
-
-    /**
-     * @brief Allocation granularity for the virtual memory pool.
-     */
-    size_t granularity;
-
-    /**
-     * @brief Handles for the physical memory allocated.
-     */
-    std::vector<aclrtDrvMemHandle> handles;
-
-    /**
-     * @brief Offsets for the mapped memory regions.
-     */
-    std::vector<void*> map_offsets;
-
-    /**
-     * @brief Constructor to initialize the buffer pool with virtual memory for
-     * a specific device.
-     *
-     * @param device The device ID to associate with this buffer pool.
-     */
-    explicit ggml_cann_pool_vmm(int device)
-        : device(device),
-          granularity(ggml_cann_info().devices[device].vmm_granularity) {}
-
-    /**
-     * @brief Destructor to free all buffers in the virtual memory pool.
-     */
-    ~ggml_cann_pool_vmm() {
-        if (pool_addr != 0) {
-            for (auto& offset : map_offsets) {
-                ACL_CHECK(aclrtUnmapMem(offset));
-            }
-            for (auto& handle : handles) {
-                ACL_CHECK(aclrtFreePhysical(handle));
-            }
-            ACL_CHECK(aclrtReleaseMemAddress(pool_addr));
-        }
-    }
-
-    /**
-     * @brief Allocate a buffer of the given size in the virtual memory pool.
-     *
-     * @param size The size of the buffer to allocate.
-     * @param actual_size A pointer to a variable to receive the actual size of
-     * the allocated buffer.
-     * @return A pointer to the allocated buffer.
-     */
-    void* alloc(size_t size, size_t* actual_size) override {
-        // round up the allocation size to the alignment to ensure that all
-        // allocations are aligned for all data types
-        const size_t alignment = 128;
-        size = alignment * ((size + alignment - 1) / alignment);
-
-        size_t avail = pool_size - pool_used;
-
-        if (size > avail) {
-            // round up to the next multiple of the granularity
-            size_t reserve_size = size - avail;
-            reserve_size =
-                granularity * ((reserve_size + granularity - 1) / granularity);
-
-            GGML_ASSERT(pool_size + reserve_size <= CANN_POOL_VMM_MAX_SIZE);
-
-            // allocate more physical memory
-            aclrtPhysicalMemProp prop = {};
-            prop.handleType = ACL_MEM_HANDLE_TYPE_NONE;
-            prop.allocationType = ACL_MEM_ALLOCATION_TYPE_PINNED;
-            prop.memAttr = ACL_HBM_MEM_HUGE;
-            prop.location.type = ACL_MEM_LOCATION_TYPE_DEVICE;
-            prop.location.id = device;
-            prop.reserve = 0;
-            aclrtDrvMemHandle handle;
-            ACL_CHECK(aclrtMallocPhysical(&handle, reserve_size, &prop, 0));
-
-            // reserve virtual address space (if not already reserved)
-            if (pool_addr == 0) {
-                ACL_CHECK(aclrtReserveMemAddress(
-                    &pool_addr, CANN_POOL_VMM_MAX_SIZE, 0, NULL, 1));
-            }
-
-            // map at the end of the pool
-            ACL_CHECK(aclrtMapMem((char*)pool_addr + pool_size, reserve_size, 0,
-                                  handle, 0));
-
-            handles.push_back(handle);
-            map_offsets.push_back((char*)pool_addr + pool_size);
-
-            // add to the pool
-            pool_size += reserve_size;
-
-            // GGML_LOG_INFO("cann pool[%d]: size increased to %llu MB (
-            // reserved %llu MB)\n",
-            //       device, (unsigned long long) (pool_size/1024/1024),
-            //       (unsigned long long) (reserve_size/1024/1024));
-        }
-
-        GGML_ASSERT(pool_addr != 0);
-
-        void* ptr = (void*)((char*)pool_addr + pool_used);
-        *actual_size = size;
-        pool_used += size;
-
-#ifdef DEBUG_CANN_MALLOC
-        GGML_LOG_INFO("cann pool[%d]: allocated %llu bytes at %llx\n", device,
-               (unsigned long long)size, (unsigned long long)ptr);
-#endif
-        return ptr;
-    }
-
-    /**
-     * @brief Free a buffer and return it to the virtual memory pool.
-     *
-     * @param ptr Pointer to the buffer to free.
-     * @param size Size of the buffer to free.
-     */
-    void free(void* ptr, size_t size) override {
-#ifdef DEBUG_CANN_MALLOC
-        GGML_LOG_INFO("cann pool[%d]: freed %llu bytes at %llx\n", device,
-               (unsigned long long)size, (unsigned long long)ptr);
-#endif
-
-        pool_used -= size;
-
-        // all deallocations must be in reverse order of the allocations
-        GGML_ASSERT(ptr == (void*)((char*)pool_addr + pool_used));
-    }
-};
-
-/**
- * @brief Create a new CANN pool for a specific device.
- *
- * Factory method to create a new CANN pool object based on the device type.
- *
- * @param device The device ID for which to create the pool.
- * @return A unique pointer to the created CANN pool.
- */
-std::unique_ptr<ggml_cann_pool> ggml_backend_cann_context::new_pool_for_device(
-    int device) {
-    // return std::unique_ptr<ggml_cann_pool>(new ggml_cann_pool_leg(device));
-    return std::unique_ptr<ggml_cann_pool>(new ggml_cann_pool_vmm(device));
-}
-
-// cann buffer
-/**
- * @brief Context for managing a CANN buffer associated with a specific device.
- *
- * This structure holds information about a CANN buffer, including the device
- * ID, device pointer, and a name derived from GGML_CANN_NAME and the device ID.
- */
-struct ggml_backend_cann_buffer_context {
-    int32_t device;  ///< The device ID associated with this buffer context.
-    void* dev_ptr =
-        nullptr;  ///< Pointer to the device memory allocated for the buffer.
-
-    /**
-     * @brief Constructor to initialize the CANN buffer context.
-     *
-     * @param device The device ID associated with this buffer context.
-     * @param dev_ptr Pointer to the device memory allocated for the buffer.
-     */
-    ggml_backend_cann_buffer_context(int32_t device, void* dev_ptr)
-        : device(device),
-          dev_ptr(dev_ptr) {}
-
-    /**
-     * @brief Destructor to free the device memory allocated for the buffer.
-     */
-    ~ggml_backend_cann_buffer_context() { ACL_CHECK(aclrtFree(dev_ptr)); }
-};
-
-/**
- * @brief Check if a buffer is a CANN buffer.
- *
- * This function checks if a given buffer is a CANN buffer by comparing its
- * `get_name` function pointer to `ggml_backend_cann_buffer_get_name`.
- *
- * @param buffer The buffer to check.
- * @return true if the buffer is a CANN buffer, false otherwise.
- */
-static bool ggml_backend_buft_is_cann(ggml_backend_buffer_type_t buft);
-static bool ggml_backend_buffer_is_cann(
-    ggml_backend_buffer_t buffer) {
-    return ggml_backend_buft_is_cann(buffer->buft);
-}
-
-/**
- * @brief Free resources associated with a CANN buffer.
- *
- * This function frees the resources associated with a CANN buffer, including
- * its context.
- *
- * @param buffer The CANN buffer to free.
- */
-static void ggml_backend_cann_buffer_free_buffer(
-    ggml_backend_buffer_t buffer) {
-    ggml_backend_cann_buffer_context* ctx =
-        (ggml_backend_cann_buffer_context*)buffer->context;
-    delete ctx;
-}
-
-/**
- * @brief Retrieve the base pointer of a CANN buffer.
- *
- * This function returns the base pointer of a CANN buffer, which points to the
- * device memory allocated for the buffer.
- *
- * @param buffer The CANN buffer whose base pointer is to be retrieved.
- * @return A pointer to the base of the device memory allocated for the buffer.
- */
-static void* ggml_backend_cann_buffer_get_base(
-    ggml_backend_buffer_t buffer) {
-    ggml_backend_cann_buffer_context* ctx =
-        (ggml_backend_cann_buffer_context*)buffer->context;
-    return ctx->dev_ptr;
-}
-
-/**
- * @brief Transform quantized Q4.0 tensor data into a format suitable for CANN
- * processing.
- *
- * This function transforms quantized Q4.0 tensor data into a format suitable
- * for CANN processing. It extracts quantization values and scales from the
- * source data and prepares them in a format expected by CANN operations.
- *
- * @param tensor Pointer to the tensor information.
- * @param src Pointer to the source data in Q4.0 format.
- * @param dst Pointer to the destination buffer where transformed data will be
- * stored.
- */
-static void ggml_backend_cann_transform_q4_0(ggml_tensor* tensor,
-                                             const void* src,
-                                             void* dst) {
-
-    int64_t n_elems = ggml_nelements(tensor);
-    int64_t groups = n_elems / QK4_0;
-    size_t quant_bytes = n_elems * sizeof(uint8_t) / 2;
-
-    uint8_t* quant_offset = (uint8_t*)dst;
-    uint16_t* scale_offset = (uint16_t*)((char*)dst + quant_bytes);
-
-    for (int i = 0; i < groups; i++) {
-        const block_q4_0* group =
-            (const block_q4_0*)((const char*)src + i * sizeof(block_q4_0));
-        *scale_offset = group->d;
-        scale_offset++;
-
-        // 0-15
-        for (int j = 0; j < QK4_0 / 2; j += 2) {
-            (*quant_offset) = (group->qs[j] & 0x0F);
-            (*quant_offset) |= ((group->qs[j + 1] << 4));
-            quant_offset++;
-        }
-
-        // 16-31
-        for (int j = 0; j < QK4_0 / 2; j += 2) {
-            (*quant_offset) = (group->qs[j] >> 4);
-            (*quant_offset) |= (group->qs[j + 1] & 0xF0);
-            quant_offset++;
-        }
-    }
-
-    // put (uint4b_t -8) into int4b_t
-    for (quant_offset = (uint8_t*)dst;
-         quant_offset < (uint8_t*)dst + quant_bytes; quant_offset++) {
-        (*quant_offset) ^= 0x88;
-    }
-}
-
-/**
- * @brief Transform CANN processed data back into quantized Q4.0 format.
- *
- * This function transforms CANN processed data back into quantized Q4.0 format.
- * It reverses the transformation performed by
- * ggml_backend_cann_transform_q4_0(), converting the data back into its
- * original quantized form.
- *
- * @param tensor Pointer to the tensor information.
- * @param src Pointer to the source buffer containing transformed data.
- * @param dst Pointer to the destination buffer where the Q4.0 formatted data
- * will be stored.
- */
-static void ggml_backend_cann_transform_back_q4_0(
-    const ggml_tensor* tensor, void* src, void* dst) {
-
-    int64_t n_elems = ggml_nelements(tensor);
-    int64_t groups = n_elems / QK4_0;
-    size_t quant_bytes = n_elems * sizeof(uint8_t) / 2;
-
-    uint8_t* quant_offset = (uint8_t*)src;
-    uint16_t* scale_offset = (uint16_t*)((char*)src + quant_bytes);
-
-    for (; quant_offset < (uint8_t*)src + quant_bytes; quant_offset++) {
-        (*quant_offset) ^= 0x88;
-    }
-    quant_offset = (uint8_t*)src;
-
-    for (int i = 0; i < groups; i++) {
-        block_q4_0* group = (block_q4_0*)((char*)dst + i * sizeof(block_q4_0));
-        group->d = *scale_offset;
-        scale_offset++;
-
-        // 0-15
-        for (int j = 0; j < QK4_0 / 2; j += 2) {
-            group->qs[j] = ((*quant_offset) & 0x0F);
-            group->qs[j + 1] = ((*quant_offset) >> 4);
-            quant_offset++;
-        }
-
-        // 16-31
-        for (int j = 0; j < QK4_0 / 2; j += 2) {
-            group->qs[j] |= ((*quant_offset) << 4);
-            group->qs[j + 1] |= ((*quant_offset) & 0xF0);
-            quant_offset++;
-        }
-    }
-}
-
-/**
- * @brief Transform quantized Q8.0 tensor data into a format suitable for CANN
- * processing.
- *
- * This function transforms quantized Q8.0 tensor data into a format suitable
- * for CANN processing. It extracts quantization values and scales from the
- * source data and prepares them in a format expected by CANN operations.
- *
- * @param tensor Pointer to the tensor information.
- * @param src Pointer to the source data in Q8.0 format.
- * @param dst Pointer to the destination buffer where transformed data will be
- * stored.
- */
-static void ggml_backend_cann_transform_q8_0(ggml_tensor* tensor,
-                                             const void* src,
-                                             void* dst) {
-    int64_t n_elems = ggml_nelements(tensor);
-    int64_t groups = n_elems / QK8_0;
-    size_t quant_bytes = n_elems * sizeof(uint8_t);
-
-    uint8_t* quant_offset = (uint8_t*)dst;
-    uint16_t* scale_offset = (uint16_t*)((char*)dst + quant_bytes);
-
-    for (int i = 0; i < groups; i++) {
-        const block_q8_0* group =
-            (const block_q8_0*)((const char*)src + i * sizeof(block_q8_0));
-        *scale_offset = group->d;
-        scale_offset++;
-        size_t group_quant_size = QK8_0 * sizeof(uint8_t);
-        memcpy(quant_offset, group->qs, group_quant_size);
-        quant_offset += group_quant_size;
-    }
-}
-
-/**
- * @brief Transform CANN processed data back into quantized Q8.0 format.
- *
- * This function transforms CANN processed data back into quantized Q8.0 format.
- * It reverses the transformation performed by
- * ggml_backend_cann_transform_q8_0(), converting the data back into its
- * original quantized form.
- *
- * @param tensor Pointer to the tensor information.
- * @param src Pointer to the source buffer containing transformed data.
- * @param dst Pointer to the destination buffer where the Q8.0 formatted data
- * will be stored.
- */
-static void ggml_backend_cann_transform_back_q8_0(
-    const ggml_tensor* tensor, const void* src, void* dst) {
-    int64_t n_elems = ggml_nelements(tensor);
-    int64_t groups = n_elems / QK8_0;
-    size_t quant_bytes = n_elems * sizeof(uint8_t);
-
-    const uint8_t* quant_offset = (const uint8_t*)src;
-    const uint16_t* scale_offset =
-        (const uint16_t*)((const char*)src + quant_bytes);
-
-    for (int i = 0; i < groups; i++) {
-        block_q8_0* group = (block_q8_0*)((char*)dst + i * sizeof(block_q8_0));
-        group->d = *scale_offset;
-        scale_offset++;
-        size_t group_quant_size = QK8_0 * sizeof(uint8_t);
-        memcpy(group->qs, quant_offset, group_quant_size);
-        quant_offset += group_quant_size;
-    }
-}
-
-/**
- * @brief Transform tensor data based on its type for CANN processing.
- *
- * This function transforms tensor data based on its quantization type for CANN
- * processing. It dispatches the transformation based on the tensor's type to
- * specialized functions handling Q4.0 and Q8.0 formats.
- *
- * @param tensor Pointer to the tensor information.
- * @param src Pointer to the source data to be transformed.
- * @param dst Pointer to the destination buffer where transformed data will be
- * stored.
- */
-static void ggml_backend_cann_transform(ggml_tensor* tensor,
-                                        const void* src, void* dst) {
-    switch (tensor->type) {
-        case GGML_TYPE_Q4_0:
-            ggml_backend_cann_transform_q4_0(tensor, src, dst);
-            break;
-        case GGML_TYPE_Q8_0:
-            ggml_backend_cann_transform_q8_0(tensor, src, dst);
-            break;
-        default:
-            break;
-    }
-}
-
-/**
- * @brief Transform CANN processed data back into tensor data based on its type.
- *
- * This function transforms CANN processed data back into tensor data based on
- * its quantization type for Q4.0 and Q8.0 formats. It dispatches the
- * transformation based on the tensor's type to specialized functions.
- *
- * @param tensor Pointer to the tensor information.
- * @param src Pointer to the source data containing CANN processed data.
- * @param dst Pointer to the destination buffer where transformed tensor data
- * will be stored.
- */
-static void ggml_backend_cann_transform_back(
-    const ggml_tensor* tensor, void* src, void* dst) {
-    switch (tensor->type) {
-        case GGML_TYPE_Q4_0:
-            ggml_backend_cann_transform_back_q4_0(tensor, src, dst);
-            break;
-        case GGML_TYPE_Q8_0:
-            ggml_backend_cann_transform_back_q8_0(tensor, src, dst);
-            break;
-        default:
-            break;
-    }
-}
-
-/**
- * @brief Check if transformation is needed for a given tensor type.
- *
- * This function checks if transformation is needed for a given tensor type
- * to prepare data for CANN processing.
- *
- * @param type The tensor type to check.
- * @return true if transformation is needed, false otherwise.
- */
-static bool need_transform(ggml_type type) {
-    switch (type) {
-        case GGML_TYPE_Q4_0:
-        case GGML_TYPE_Q8_0:
-            return true;
-        default:
-            return false;
-    }
-}
-
-/**
- * @brief Initialize a tensor using data from a CANN buffer.
- *
- * This function initializes a tensor using data from a CANN buffer.
- * It handles special cases such as views and quantization.
- *
- * @param buffer The CANN buffer from which to initialize the tensor.
- * @param tensor Pointer to the tensor to be initialized.
- */
-static void ggml_backend_cann_buffer_init_tensor(
-    ggml_backend_buffer_t buffer, ggml_tensor* tensor) {
-    if (tensor->view_src != NULL && tensor->view_offs == 0) {
-        GGML_ASSERT(tensor->view_src->buffer->buft == buffer->buft);
-        return;
-    }
-
-    // TODO: can backend doesn't support quantized yet. Just leave the code
-    // here.
-    if (ggml_is_quantized(tensor->type)) {
-        // Initialize padding to 0 to avoid possible NaN values
-        size_t original_size = ggml_nbytes(tensor);
-        size_t padded_size =
-            ggml_backend_buft_get_alloc_size(buffer->buft, tensor);
-
-        if (padded_size > original_size && tensor->view_src == nullptr) {
-            size_t memset_size = padded_size - original_size;
-            ACL_CHECK(aclrtMemset((char*)tensor->data + original_size,
-                                  memset_size, 0, memset_size));
-        }
-    }
-}
-
-// TODO: need handle tensor which has paddings.
-/**
- * @brief Set tensor data in a CANN buffer.
- *
- * This function sets tensor data in a CANN buffer, handling transformations
- * if needed based on the tensor's type.
- *
- * @param buffer The CANN buffer where the tensor data will be set.
- * @param tensor Pointer to the tensor whose data will be set.
- * @param data Pointer to the source data to be copied into the tensor.
- * @param offset Offset in the source data from where to start copying.
- * @param size Size of the data to be copied, in bytes.
- */
-static void ggml_backend_cann_buffer_set_tensor(
-    ggml_backend_buffer_t buffer, ggml_tensor *tensor, const void *data,
-    size_t offset, size_t size) {
-    ggml_backend_cann_buffer_context *ctx =
-        (ggml_backend_cann_buffer_context *)buffer->context;
-
-    ggml_cann_set_device(ctx->device);
-    // TODO: refer to cann(#6017), it use thread's default stream.
-    // For acl, synchronous functions use this default stream.
-    // Why aclrtSynchronizeDevice?
-
-    if (!need_transform(tensor->type)) {
-        ACL_CHECK(aclrtMemcpy((char *)tensor->data + offset, size, data, size,
-                              ACL_MEMCPY_HOST_TO_DEVICE));
-    } else {
-        void *transform_buffer = malloc(size);
-        ggml_backend_cann_transform(tensor, data, transform_buffer);
-
-        ACL_CHECK(aclrtMemcpy((char *)tensor->data + offset, size,
-                              transform_buffer, size,
-                              ACL_MEMCPY_HOST_TO_DEVICE));
-        free(transform_buffer);
-    }
-}
-
-/**
- * @brief Get tensor data from a CANN buffer.
- *
- * This function retrieves tensor data from a CANN buffer, handling
- * transformations if needed based on the tensor's type.
- *
- * @param buffer The CANN buffer from which to retrieve tensor data.
- * @param tensor Pointer to the tensor whose data will be retrieved.
- * @param data Pointer to the destination buffer where the tensor data will be
- * copied.
- * @param offset Offset in the destination buffer where to start copying.
- * @param size Size of the data to be copied, in bytes.
- */
-static void ggml_backend_cann_buffer_get_tensor(
-    ggml_backend_buffer_t buffer, const ggml_tensor* tensor, void* data,
-    size_t offset, size_t size) {
-    ggml_backend_cann_buffer_context* ctx =
-        (ggml_backend_cann_buffer_context*)buffer->context;
-
-    ggml_cann_set_device(ctx->device);
-
-    if (!need_transform(tensor->type)) {
-        ACL_CHECK(aclrtMemcpy(data, size, (char*)tensor->data + offset, size,
-                              ACL_MEMCPY_DEVICE_TO_HOST));
-    } else {
-        void* transform_buffer = malloc(size);
-        ACL_CHECK(aclrtMemcpy(transform_buffer, size,
-                              (char*)tensor->data + offset, size,
-                              ACL_MEMCPY_DEVICE_TO_HOST));
-        ggml_backend_cann_transform_back(tensor, transform_buffer, data);
-        free(transform_buffer);
-    }
-}
-
-/**
- * @brief Copy tensor data between CANN buffers if possible.
- *
- * This function copies tensor data between CANN buffers if the source and
- * destination buffers are CANN buffers and they meet the necessary conditions
- * (same device or devices can access each other).
- *
- * @param buffer The destination CANN buffer where the tensor data will be
- * copied.
- * @param src Pointer to the source tensor whose data will be copied.
- * @param dst Pointer to the destination tensor where the data will be copied.
- * @return true if the copy operation succeeded, false otherwise.
- */
-static bool ggml_backend_cann_buffer_cpy_tensor(
-    ggml_backend_buffer_t buffer, const ggml_tensor* src, ggml_tensor* dst) {
-    if (ggml_backend_buffer_is_cann(src->buffer)) {
-        ggml_backend_cann_buffer_context* src_ctx =
-            (ggml_backend_cann_buffer_context*)src->buffer->context;
-        ggml_backend_cann_buffer_context* dst_ctx =
-            (ggml_backend_cann_buffer_context*)buffer->context;
-
-        size_t memcpy_size = ggml_nbytes(src);
-        // Same device.
-        if (src_ctx->device == dst_ctx->device) {
-            ACL_CHECK(aclrtMemcpy((char*)dst->data, memcpy_size,
-                                  (const char*)src->data, memcpy_size,
-                                  ACL_MEMCPY_DEVICE_TO_DEVICE));
-            return true;
-        } else {
-            // Different device but can access by peer.
-            int32_t canAccessPeer = 0;
-            ACL_CHECK(aclrtDeviceCanAccessPeer(&canAccessPeer, src_ctx->device,
-                                               dst_ctx->device));
-            if (canAccessPeer) {
-                ggml_cann_set_device(src_ctx->device);
-                ACL_CHECK(aclrtDeviceEnablePeerAccess(dst_ctx->device, 0));
-                ACL_CHECK(aclrtMemcpy((char*)dst->data, memcpy_size,
-                                      (const char*)src->data, memcpy_size,
-                                      ACL_MEMCPY_DEVICE_TO_DEVICE));
-                return true;
-            }
-        }
-    }
-    return false;
-}
-
-/**
- * @brief Clear a CANN buffer by setting all its memory to a specified value.
- *
- * This function clears a CANN buffer by setting all its memory to a specified
- * value.
- *
- * @param buffer The CANN buffer to be cleared.
- * @param value The value to which each byte in the buffer will be set.
- */
-static void ggml_backend_cann_buffer_clear(
-    ggml_backend_buffer_t buffer, uint8_t value) {
-    ggml_backend_cann_buffer_context* ctx =
-        (ggml_backend_cann_buffer_context*)buffer->context;
-
-    ggml_cann_set_device(ctx->device);
-    ACL_CHECK(aclrtMemset(ctx->dev_ptr, buffer->size, value, buffer->size));
-}
-
-/**
- * @brief Interface for a CANN buffer in the backend.
- *
- * This structure defines function pointers to operations that can be performed
- * on a CANN buffer within the backend.
- */
-static const ggml_backend_buffer_i ggml_backend_cann_buffer_interface = {
-    /* .free_buffer     = */ ggml_backend_cann_buffer_free_buffer,
-    /* .get_base        = */ ggml_backend_cann_buffer_get_base,
-    /* .init_tensor     = */ ggml_backend_cann_buffer_init_tensor,
-    /* .memset_tensor   = */ NULL,
-    /* .set_tensor      = */ ggml_backend_cann_buffer_set_tensor,
-    /* .get_tensor      = */ ggml_backend_cann_buffer_get_tensor,
-    /* .cpy_tensor      = */ ggml_backend_cann_buffer_cpy_tensor,
-    /* .clear           = */ ggml_backend_cann_buffer_clear,
-    /* .reset           = */ NULL,
-};
-
-// cann buffer type
-/**
- * @brief Structure representing context information for a specific backend
- * buffer type.
- */
-struct ggml_backend_cann_buffer_type_context {
-    int32_t
-        device; /**< Device identifier associated with the buffer context. */
-    std::string name; /**< Name associated with the buffer context. */
-};
-
-/**
- * @brief Retrieves the name associated with a CANN buffer type.
- *
- * This function returns the descriptive name associated with the specified
- * CANN buffer type context.
- *
- * @param buft Pointer to the buffer type context.
- * @return Const pointer to the C-style string containing the name.
- */
-static const char* ggml_backend_cann_buffer_type_name(
-    ggml_backend_buffer_type_t buft) {
-    ggml_backend_cann_buffer_type_context* buft_ctx =
-        (ggml_backend_cann_buffer_type_context*)buft->context;
-
-    return buft_ctx->name.c_str();
-}
-
-/**
- * @brief Allocates a new CANN buffer of the specified type and size.
- *
- * This function allocates a new CANN buffer on the specified device with the
- * given size.
- *
- * @param buft Pointer to the buffer type context.
- * @param size Size in bytes of the buffer to allocate.
- * @return Pointer to the allocated buffer, or nullptr if allocation fails.
- */
-static ggml_backend_buffer_t
-ggml_backend_cann_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft,
-                                           size_t size) {
-    ggml_backend_cann_buffer_type_context* buft_ctx =
-        (ggml_backend_cann_buffer_type_context*)buft->context;
-
-    ggml_cann_set_device(buft_ctx->device);
-
-    size = std::max(size, (size_t)1);
-
-    void* dev_ptr;
-    aclError err = aclrtMalloc(&dev_ptr, size, ACL_MEM_MALLOC_HUGE_FIRST);
-    if (err != ACL_SUCCESS) {
-        GGML_LOG_ERROR(
-            "%s: allocating %.2f MiB on device %d: aclrtMalloc failed: %s\n",
-            __func__, size / 1024.0 / 1024.0, buft_ctx->device,
-            aclGetRecentErrMsg());
-        return nullptr;
-    }
-
-    ggml_backend_cann_buffer_context* ctx =
-        new ggml_backend_cann_buffer_context(buft_ctx->device, dev_ptr);
-
-    return ggml_backend_buffer_init(buft, ggml_backend_cann_buffer_interface,
-                                    ctx, size);
-}
-
-/**
- * @brief Retrieves the memory alignment requirement for CANN buffers of this
- * type.
- *
- * This function returns the alignment requirement in bytes for memory allocated
- * by the CANN buffer type.
- *
- * @param buft Pointer to the buffer type context (unused in this
- * implementation).
- * @return The alignment requirement in bytes (fixed at 128 bytes for CANN
- * buffers).
- */
-static size_t ggml_backend_cann_buffer_type_get_alignment(
-    ggml_backend_buffer_type_t buft) {
-    return 128;
-
-    GGML_UNUSED(buft);
-}
-
-/**
- * @brief Calculates the allocation size required for a tensor in a CANN buffer.
- *
- * Computes the total allocation size needed for storing the tensor's data in a
- * CANN buffer, considering any necessary padding or adjustments for quantized
- * types.
- *
- * @param buft Pointer to the buffer type context (unused in this
- * implementation).
- * @param tensor Pointer to the tensor for which the allocation size is
- * calculated.
- * @return The total allocation size in bytes required for the tensor in the
- * CANN buffer.
- */
-static size_t ggml_backend_cann_buffer_type_get_alloc_size(
-    ggml_backend_buffer_type_t buft, const ggml_tensor* tensor) {
-    size_t size = ggml_nbytes(tensor);
-    int64_t ne0 = tensor->ne[0];
-
-    // last line must bigger than 32, because every single op deal at
-    // least 32 bytes.
-    // TODO: quantized type?
-    // int64_t line_size = ne0 * ggml_element_size(tensor);
-    // int64_t line_size_align_32 = (line_size + 31) & ~31;
-    // size += (line_size_align_32 - line_size);
-
-    // TODO: not support quantized yet.
-    // TODO: consider un-continue tensor.
-    if (ggml_is_quantized(tensor->type)) {
-        if (ne0 % MATRIX_ROW_PADDING != 0) {
-            size += ggml_row_size(
-                tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING);
-        }
-    }
-
-    return size;
-
-    GGML_UNUSED(buft);
-}
-
-static bool ggml_backend_cann_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
-    return false;
-
-    GGML_UNUSED(buft);
-}
-
-/**
- * @brief Interface for managing CANN buffer types in the GGML backend.
- *
- * Provides function pointers for allocating, querying properties, and managing
- * memory for CANN buffer types in the GGML backend.
- */
-static const ggml_backend_buffer_type_i ggml_backend_cann_buffer_type_interface = {
-    /* .get_name         = */ ggml_backend_cann_buffer_type_name,
-    /* .alloc_buffer     = */ ggml_backend_cann_buffer_type_alloc_buffer,
-    /* .get_alignment    = */ ggml_backend_cann_buffer_type_get_alignment,
-    /* .get_max_size     = */ NULL,  // defaults to SIZE_MAX
-    /* .get_alloc_size   = */ ggml_backend_cann_buffer_type_get_alloc_size,
-    /* .is_host          = */ ggml_backend_cann_buffer_type_is_host,
-};
-
-/**
- * @brief Retrieves the CANN buffer type for a specified device.
- *
- * This function initializes and returns the buffer type interface associated
- * with the given device. It ensures thread-safe access using a mutex.
- *
- * @param device The device index for which to retrieve the buffer type.
- * @return A pointer to the buffer type interface for the specified device, or
- * nullptr if the device index is out of range.
- */
-ggml_backend_buffer_type_t
-ggml_backend_cann_buffer_type(int32_t device) {
-    static std::mutex mutex;
-    std::lock_guard<std::mutex> lock(mutex);
-
-    if (device >= ggml_backend_cann_get_device_count()) {
-        return nullptr;
-    }
-
-    static ggml_backend_buffer_type
-        ggml_backend_cann_buffer_types[GGML_CANN_MAX_DEVICES];
-
-    static bool ggml_backend_cann_buffer_type_initialized = false;
-
-    if (!ggml_backend_cann_buffer_type_initialized) {
-        for (int32_t i = 0; i < GGML_CANN_MAX_DEVICES; i++) {
-            ggml_backend_cann_buffer_types[i] = {
-                /* .iface    = */ ggml_backend_cann_buffer_type_interface,
-                /* .device    = */ ggml_backend_reg_dev_get(ggml_backend_cann_reg(), device),
-                /* .context  = */
-                 new ggml_backend_cann_buffer_type_context{
-                    i, "CANN" + std::to_string(i)},
-            };
-        }
-        ggml_backend_cann_buffer_type_initialized = true;
-    }
-
-    return &ggml_backend_cann_buffer_types[device];
-}
-
-/**
- * @brief Retrieves the name associated with a CANN host buffer type.
- *
- * This function returns the descriptive name associated with the specified
- * CANN host buffer type context.
- *
- * @param buft Pointer to the host buffer type context.
- * @return Const pointer to the C-style string containing the name.
- */
-static const char * ggml_backend_cann_host_buffer_type_name(ggml_backend_buffer_type_t buft) {
-    return "CANN_Host";
-
-    GGML_UNUSED(buft);
-}
-
-/**
- * @brief Retrieves the name associated with a CANN host buffer.
- *
- * This function returns the descriptive name associated with the specified
- * CANN host buffer context.
- *
- * @param buft Pointer to the host buffer context.
- * @return Const pointer to the C-style string containing the name.
- */
-static const char * ggml_backend_cann_host_buffer_name(ggml_backend_buffer_t buffer) {
-    return "CANN_Host";
-
-    GGML_UNUSED(buffer);
-}
-
-/**
- * @brief Free resources associated with a CANN host buffer.
- *
- * This function frees the resources associated with a CANN host buffer, including
- * its context.
- *
- * @param buffer The CANN host buffer to free.
- */
-static void ggml_backend_cann_host_buffer_free(ggml_backend_buffer_t buffer) {
-    ACL_CHECK(aclrtFreeHost(buffer->context));
-}
-
-/**
- * @brief Allocates a new CANN host buffer of the specified size.
- *
- * This function allocates a new CANN host buffer with the given size.
- * @param size Size in bytes of the host buffer to allocate.
- * @return Pointer to the allocated host buffer, or nullptr if allocation fails.
- */
-static void * ggml_cann_host_malloc(size_t size) {
-    if (getenv("GGML_CANN_NO_PINNED") != nullptr) {
-        return nullptr;
-    }
-
-    void * hostPtr = nullptr;
-    aclError err = aclrtMallocHost((void **) &hostPtr, size);
-    if (err != ACL_SUCCESS) {
-
-        GGML_LOG_WARN("%s: failed to allocate %.2f MiB of pinned memory: %s\n", __func__,
-                           size / 1024.0 / 1024.0, aclGetRecentErrMsg());
-        return nullptr;
-    }
-    return hostPtr;
-}
-
-/**
- * @brief Allocates a new CANN host buffer of the specified type and size.
- *
- * @param buft Pointer to the host buffer type context.
- * @param size Size in bytes of the host buffer to allocate.
- * @return Pointer to the allocated host buffer, or CPU buffer pointer if allocation fails.
- */
-static ggml_backend_buffer_t ggml_backend_cann_host_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
-    void * hostPtr = ggml_cann_host_malloc(size);
-
-    if (hostPtr == nullptr) {
-        // fallback to cpu buffer
-        return ggml_backend_buft_alloc_buffer(ggml_backend_cpu_buffer_type(), size);
-    }
-
-    ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(hostPtr, size);
-    buffer->buft = buft;
-    buffer->iface.free_buffer = ggml_backend_cann_host_buffer_free;
-
-    return buffer;
-}
-
-/**
- * @brief Interface for managing CANN host buffer types in the GGML backend.
- *
- * Provides function pointers for allocating, querying properties, and managing
- * memory for CANN buffer types in the GGML backend.
- */
-ggml_backend_buffer_type_t ggml_backend_cann_host_buffer_type() {
-    static struct ggml_backend_buffer_type ggml_backend_cann_buffer_type_host = {
-        /* .iface    = */ {
-            /* .get_name         = */ ggml_backend_cann_host_buffer_type_name,
-            /* .alloc_buffer     = */ ggml_backend_cann_host_buffer_type_alloc_buffer,
-            /* .get_alignment    = */ ggml_backend_cpu_buffer_type()->iface.get_alignment,
-            /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
-            /* .get_alloc_size   = */ ggml_backend_cpu_buffer_type()->iface.get_alloc_size,
-            /* .is_host          = */ ggml_backend_cpu_buffer_type()->iface.is_host,
-        },
-        /* .device   = */ ggml_backend_reg_dev_get(ggml_backend_cann_reg(), 0),
-        /* .context  = */ nullptr,
-    };
-
-    return &ggml_backend_cann_buffer_type_host;
-}
-
-/**
- * @brief Computes the forward operation for a given tensor using CANN
- * operations.
- *
- * This function selects the appropriate CANN operation based on the type of
- * operation specified in the tensor and performs the computation.
- *
- * @param ctx The CANN context containing necessary resources and
- * configurations.
- * @param dst The destination tensor where the result of the computation will be
- * stored.
- * @return true if the computation was successful; false otherwise.
- */
-static bool ggml_cann_compute_forward(ggml_backend_cann_context& ctx,
-                                      struct ggml_tensor* dst) {
-    switch (dst->op) {
-        case GGML_OP_REPEAT:
-            ggml_cann_repeat(ctx, dst);
-            break;
-        case GGML_OP_GET_ROWS:
-            ggml_cann_get_rows(ctx, dst);
-            break;
-        case GGML_OP_DUP:
-            ggml_cann_dup(ctx, dst);
-            break;
-        case GGML_OP_ADD:
-            ggml_cann_add(ctx, dst);
-            break;
-        case GGML_OP_ACC:
-            ggml_cann_acc(ctx, dst);
-            break;
-        case GGML_OP_MUL:
-            ggml_cann_mul_div<aclnnMulGetWorkspaceSize, aclnnMul>(ctx, dst);
-            break;
-        case GGML_OP_DIV:
-            ggml_cann_mul_div<aclnnDivGetWorkspaceSize, aclnnDiv>(ctx, dst);
-            break;
-        case GGML_OP_UNARY:
-            switch (ggml_get_unary_op(dst)) {
-                case GGML_UNARY_OP_GELU:
-                    ggml_cann_activation<aclnnGeluGetWorkspaceSize, aclnnGelu>(
-                        ctx, dst);
-                    break;
-                case GGML_UNARY_OP_SILU:
-                    ggml_cann_activation<aclnnSiluGetWorkspaceSize, aclnnSilu>(
-                        ctx, dst);
-                    break;
-                // TODO: Use faster gelu??
-                case GGML_UNARY_OP_GELU_QUICK:
-                    ggml_cann_activation<aclnnGeluGetWorkspaceSize, aclnnGelu>(
-                        ctx, dst);
-                    break;
-                case GGML_UNARY_OP_TANH:
-                    ggml_cann_activation<aclnnTanhGetWorkspaceSize, aclnnTanh>(
-                        ctx, dst);
-                    break;
-                case GGML_UNARY_OP_RELU:
-                    ggml_cann_activation<aclnnReluGetWorkspaceSize, aclnnRelu>(
-                        ctx, dst);
-                    break;
-                case GGML_UNARY_OP_HARDSIGMOID:
-                    ggml_cann_activation<aclnnHardsigmoidGetWorkspaceSize,
-                                         aclnnHardsigmoid>(ctx, dst);
-                    break;
-                case GGML_UNARY_OP_HARDSWISH:
-                    ggml_cann_activation<aclnnHardswishGetWorkspaceSize,
-                                         aclnnHardswish>(ctx, dst);
-                    break;
-                default:
-                    return false;
-            }
-            break;
-        case GGML_OP_NORM:
-            ggml_cann_norm(ctx, dst);
-            break;
-        case GGML_OP_GROUP_NORM:
-            ggml_cann_group_norm(ctx, dst);
-            break;
-        case GGML_OP_CONCAT:
-            ggml_cann_concat(ctx, dst);
-            break;
-        case GGML_OP_UPSCALE:
-            ggml_cann_upsample_nearest2d(ctx, dst);
-            break;
-        case GGML_OP_PAD:
-            ggml_cann_pad(ctx, dst);
-            break;
-        case GGML_OP_ARANGE:
-            ggml_cann_arange(ctx, dst);
-            break;
-        case GGML_OP_TIMESTEP_EMBEDDING:
-            ggml_cann_timestep_embedding(ctx, dst);
-            break;
-        case GGML_OP_LEAKY_RELU:
-            ggml_cann_leaky_relu(ctx, dst);
-            break;
-        case GGML_OP_RMS_NORM:
-            ggml_cann_rms_norm(ctx, dst);
-            break;
-        case GGML_OP_MUL_MAT:
-            ggml_cann_mul_mat(ctx, dst);
-            break;
-        case GGML_OP_MUL_MAT_ID:
-            return false;
-        case GGML_OP_SCALE:
-            ggml_cann_scale(ctx, dst);
-            break;
-        case GGML_OP_SQR:
-            ggml_cann_sqr(ctx, dst);
-            break;
-        case GGML_OP_CLAMP:
-            ggml_cann_clamp(ctx, dst);
-            break;
-        case GGML_OP_CPY:
-            ggml_cann_cpy(ctx, dst);
-            break;
-        case GGML_OP_CONT:
-            ggml_cann_dup(ctx, dst);
-            break;
-        case GGML_OP_NONE:
-        case GGML_OP_RESHAPE:
-        case GGML_OP_VIEW:
-        case GGML_OP_PERMUTE:
-        case GGML_OP_TRANSPOSE:
-            break;
-        case GGML_OP_DIAG_MASK_INF:
-            ggml_cann_diag_mask(ctx, dst, -INFINITY);
-            break;
-        case GGML_OP_SOFT_MAX:
-            ggml_cann_softmax(ctx, dst);
-            break;
-        case GGML_OP_ROPE:
-            ggml_cann_rope(ctx, dst);
-            break;
-        case GGML_OP_IM2COL:
-            ggml_cann_im2col(ctx, dst);
-            break;
-        case GGML_OP_POOL_2D:
-            ggml_cann_pool2d(ctx, dst);
-            break;
-        case GGML_OP_SUM_ROWS:
-            ggml_cann_sum_rows(ctx, dst);
-            break;
-        case GGML_OP_ARGSORT:
-            ggml_cann_argsort(ctx, dst);
-            break;
-        default:
-            return false;
-    }
-
-    return true;
-}
-
-// backend
-/**
- * @brief Retrieves the name associated with the CANN backend.
- *
- * This function returns the name assigned to the CANN backend, which is stored
- * in the context of the provided backend structure.
- *
- * @param backend Pointer to the CANN backend structure.
- * @return A pointer to a constant string representing the backend name.
- */
-static const char* ggml_backend_cann_name(ggml_backend_t backend) {
-    ggml_backend_cann_context* cann_ctx =
-        (ggml_backend_cann_context*)backend->context;
-
-    return cann_ctx->name.c_str();
-}
-
-/**
- * @brief Frees resources associated with the CANN backend.
- *
- * This function releases resources associated with the CANN backend context
- * and resets the device associated with the backend to its initial state.
- *
- * @param backend Pointer to the CANN backend structure to be freed.
- */
-static void ggml_backend_cann_free(ggml_backend_t backend) {
-    ggml_backend_cann_context* cann_ctx =
-        (ggml_backend_cann_context*)backend->context;
-    ACL_CHECK(aclrtSynchronizeDevice());
-    ACL_CHECK(aclrtResetDevice(cann_ctx->device));
-
-    // finalize when last backend freed.
-    if (cann_ctx->device == ggml_backend_cann_get_device_count() - 1) {
-        ACL_CHECK(aclFinalize());
-    }
-
-    delete cann_ctx;
-    delete backend;
-}
-
-/**
- * @brief Sets tensor data asynchronously in the CANN backend.
- *
- * This function asynchronously sets tensor data in the CANN backend. Depending
- * on the tensor type, it may perform data transformations before copying data
- * to the device.
- *
- * @param backend Pointer to the CANN backend structure.
- * @param tensor Pointer to the tensor structure to set data for.
- * @param data Pointer to the host data to copy to the tensor.
- * @param offset Offset in bytes within the host data.
- * @param size Size of the data to copy in bytes.
- */
-static void ggml_backend_cann_set_tensor_async(ggml_backend_t backend,
-                                               ggml_tensor *tensor,
-                                               const void *data,
-                                               size_t offset,
-                                               size_t size) {
-    ggml_backend_cann_context *cann_ctx =
-        (ggml_backend_cann_context *)backend->context;
-
-    if (!need_transform(tensor->type)) {
-        ACL_CHECK(aclrtMemcpyAsync((char *)tensor->data + offset, size, data,
-                                   size, ACL_MEMCPY_HOST_TO_DEVICE,
-                                   cann_ctx->stream()));
-    } else {
-        void *transform_buffer = malloc(size);
-        ggml_backend_cann_transform(tensor, data, transform_buffer);
-
-        ACL_CHECK(aclrtMemcpyAsync(
-            (char *)tensor->data + offset, size, transform_buffer, size,
-            ACL_MEMCPY_HOST_TO_DEVICE, cann_ctx->stream()));
-        ACL_CHECK(aclrtSynchronizeStream(cann_ctx->stream()));
-        free(transform_buffer);
-    }
-}
-
-static void ggml_backend_cann_get_tensor_async(
-    ggml_backend_t backend, const ggml_tensor *tensor, void *data,
-    size_t offset, size_t size) {
-    ggml_backend_cann_context *cann_ctx =
-        (ggml_backend_cann_context *)backend->context;
-    ggml_backend_buffer_t buf =
-        tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
-
-    GGML_ASSERT(buf->buft == ggml_backend_cann_buffer_type(cann_ctx->device) &&
-                "unsupported buffer type");
-
-    if (!need_transform(tensor->type)) {
-        ACL_CHECK(aclrtMemcpyAsync(data, size, (char *)tensor->data + offset,
-                                   size, ACL_MEMCPY_DEVICE_TO_HOST,
-                                   cann_ctx->stream()));
-    } else {
-        void *transform_buffer = malloc(size);
-        ACL_CHECK(aclrtMemcpyAsync(
-            transform_buffer, size, (char *)tensor->data + offset, size,
-            ACL_MEMCPY_DEVICE_TO_HOST, cann_ctx->stream()));
-        ACL_CHECK(aclrtSynchronizeStream(cann_ctx->stream()));
-        ggml_backend_cann_transform_back(tensor, transform_buffer, data);
-        free(transform_buffer);
-    }
-}
-
-/**
- * @brief Asynchronously copies tensor data between CANN backends.
- *
- * This function copies tensor data asynchronously between two CANN backends. It
- * checks if both tensors reside in CANN buffers and whether the devices support
- * peer-to-peer access for direct copying. If not, it returns false.
- *
- * @param backend_src Pointer to the source CANN backend structure.
- * @param backend_dst Pointer to the destination CANN backend structure.
- * @param src Pointer to the source tensor to copy data from.
- * @param dst Pointer to the destination tensor to copy data to.
- * @return true if the copy operation succeeds, false otherwise.
- */
-static bool ggml_backend_cann_cpy_tensor_async(
-    ggml_backend_t backend_src, ggml_backend_t backend_dst,
-    const ggml_tensor* src, ggml_tensor* dst) {
-    GGML_ASSERT(ggml_backend_is_cann(backend_src) ||
-                ggml_backend_is_cann(backend_dst));
-
-    if (!ggml_backend_buffer_is_cann(src->buffer) ||
-        !ggml_backend_buffer_is_cann(dst->buffer)) {
-        return false;
-    }
-
-    ggml_backend_buffer_t buf_src =
-        src->view_src ? src->view_src->buffer : src->buffer;
-    ggml_backend_buffer_t buf_dst =
-        dst->view_src ? dst->view_src->buffer : dst->buffer;
-
-    ggml_backend_cann_context* cann_ctx_src =
-        (ggml_backend_cann_context*)backend_src->context;
-    ggml_backend_cann_context* cann_ctx_dst =
-        (ggml_backend_cann_context*)backend_dst->context;
-
-    size_t copy_size = ggml_nbytes(dst);
-    if (backend_src != backend_dst) {
-        ggml_backend_cann_buffer_context* buf_ctx_src =
-            (ggml_backend_cann_buffer_context*)buf_src->context;
-        ggml_backend_cann_buffer_context* buf_ctx_dst =
-            (ggml_backend_cann_buffer_context*)buf_dst->context;
-
-        GGML_ASSERT(cann_ctx_src->device == buf_ctx_src->device);
-        GGML_ASSERT(cann_ctx_dst->device == buf_ctx_dst->device);
-
-        int32_t canAccessPeer = 0;
-        ACL_CHECK(aclrtDeviceCanAccessPeer(&canAccessPeer, cann_ctx_src->device,
-                                           cann_ctx_dst->device));
-        if (!canAccessPeer) {
-            return false;
-        }
-
-        // need open both directions for memcpyasync between devices.
-        ggml_cann_set_device(cann_ctx_dst->device);
-        ACL_CHECK(aclrtDeviceEnablePeerAccess(cann_ctx_src->device, 0));
-        ggml_cann_set_device(cann_ctx_src->device);
-        ACL_CHECK(aclrtDeviceEnablePeerAccess(cann_ctx_dst->device, 0));
-
-        ACL_CHECK(aclrtMemcpyAsync(dst->data, copy_size, src->data, copy_size,
-                                   ACL_MEMCPY_DEVICE_TO_DEVICE,
-                                   cann_ctx_src->stream()));
-
-        //TODO: workaround for Event didn`t work here.
-        aclrtSynchronizeStream(cann_ctx_src->stream());
-    } else {
-        // src and dst are on the same backend
-        ACL_CHECK(aclrtMemcpyAsync(dst->data, copy_size, src->data, copy_size,
-                                   ACL_MEMCPY_DEVICE_TO_DEVICE,
-                                   cann_ctx_dst->stream()));
-    }
-
-    return true;
-}
-
-/**
- * @brief Synchronizes a CANN backend.
- *
- * This function synchronizes the specified CANN backend by waiting for all
- * operations in its associated stream to complete.
- *
- * @param backend Pointer to the CANN backend structure to synchronize.
- */
-static void ggml_backend_cann_synchronize(ggml_backend_t backend) {
-    ggml_backend_cann_context* cann_ctx =
-        (ggml_backend_cann_context*)backend->context;
-
-    ggml_cann_set_device(cann_ctx->device);
-
-    ACL_CHECK(aclrtSynchronizeStream(cann_ctx->stream()));
-}
-
-/**
- * @brief Computes a computational graph using a CANN backend.
- *
- * This function computes the operations defined in the computational graph
- * using the specified CANN backend.
- *
- * @param backend Pointer to the CANN backend structure to use for computation.
- * @param cgraph Pointer to the computational graph structure containing nodes
- *               representing operations to be computed.
- * @return enum ggml_status Returns GGML_STATUS_SUCCESS if computation
- *         completes successfully, otherwise an appropriate error status.
- */
-static enum ggml_status ggml_backend_cann_graph_compute(
-    ggml_backend_t backend, ggml_cgraph* cgraph) {
-    ggml_backend_cann_context* cann_ctx =
-        (ggml_backend_cann_context*)backend->context;
-
-    ggml_cann_set_device(cann_ctx->device);
-
-    for (int i = 0; i < cgraph->n_nodes; i++) {
-        ggml_tensor* node = cgraph->nodes[i];
-
-        if (ggml_is_empty(node) || node->op == GGML_OP_NONE) {
-            continue;
-        }
-
-        bool ok = ggml_cann_compute_forward(*cann_ctx, node);
-
-        if (!ok) {
-            GGML_LOG_ERROR("%s: error: op not supported %s (%s)\n", __func__,
-                    node->name, ggml_op_name(node->op));
-        }
-        GGML_ASSERT(ok);
-    }
-
-    return GGML_STATUS_SUCCESS;
-}
-
-/**
- * @brief Checks if the CANN backend supports a specific operation.
- *
- * This function checks whether the specified operation is supported by the
- * CANN backend.
- *
- * @param backend Pointer to the CANN backend structure to check support for
- *                the operation.
- * @param op Pointer to the tensor representing the operation to check.
- * @return bool Returns true if the operation is supported by the backend,
- *              otherwise false.
- */
-static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
-                                                    const ggml_tensor* op) {
-    switch (op->op) {
-        case GGML_OP_UNARY:
-            switch (ggml_get_unary_op(op)) {
-                case GGML_UNARY_OP_GELU:
-                case GGML_UNARY_OP_SILU:
-                case GGML_UNARY_OP_RELU:
-                case GGML_UNARY_OP_HARDSIGMOID:
-                case GGML_UNARY_OP_HARDSWISH:
-                case GGML_UNARY_OP_GELU_QUICK:
-                case GGML_UNARY_OP_TANH:
-                    return true;
-                default:
-                    return false;
-            }
-        case GGML_OP_MUL_MAT: {
-            switch (op->src[0]->type) {
-                case GGML_TYPE_F16:
-                case GGML_TYPE_F32:
-                case GGML_TYPE_Q8_0:
-                    // TODO: fix me
-                    // Current groupsize should not be greater than k-1 in
-                    // aclnnWeightQuantBatchMatmulV2GetWorkspaceSize().
-                case GGML_TYPE_Q4_0:
-                    return true;
-                default:
-                    return false;
-            }
-        }
-        case GGML_OP_MUL_MAT_ID:
-            return false;
-        // embedding
-        case GGML_OP_GET_ROWS: {
-            switch (op->src[0]->type) {
-                case GGML_TYPE_F32:
-                case GGML_TYPE_F16:
-                case GGML_TYPE_Q4_0:
-                case GGML_TYPE_Q8_0:
-                    return true;
-                default:
-                    return false;
-            }
-        } break;
-        case GGML_OP_CPY: {
-            switch (op->type) {
-                case GGML_TYPE_F32:
-                case GGML_TYPE_F16:
-                case GGML_TYPE_Q8_0:
-                case GGML_TYPE_Q4_0:
-                    return true;
-                default:
-                    return false;
-            }
-        }
-        case GGML_OP_DUP:
-        case GGML_OP_REPEAT:
-        case GGML_OP_CONCAT:
-        case GGML_OP_NONE:
-        case GGML_OP_RESHAPE:
-        case GGML_OP_VIEW:
-        case GGML_OP_PERMUTE:
-        case GGML_OP_TRANSPOSE:
-        case GGML_OP_NORM:
-        case GGML_OP_ADD:
-        case GGML_OP_MUL:
-        case GGML_OP_DIV:
-        case GGML_OP_RMS_NORM:
-        case GGML_OP_SCALE:
-        case GGML_OP_SQR:
-        case GGML_OP_CLAMP:
-        case GGML_OP_CONT:
-        case GGML_OP_DIAG_MASK_INF:
-        case GGML_OP_SOFT_MAX:
-        case GGML_OP_ROPE:
-        case GGML_OP_IM2COL:
-        case GGML_OP_POOL_2D:
-        case GGML_OP_SUM_ROWS:
-        case GGML_OP_ARGSORT:
-        case GGML_OP_ACC:
-        case GGML_OP_GROUP_NORM:
-        case GGML_OP_UPSCALE:
-        case GGML_OP_PAD:
-        case GGML_OP_ARANGE:
-        case GGML_OP_TIMESTEP_EMBEDDING:
-        case GGML_OP_LEAKY_RELU:
-            return true;
-        default:
-            return false;
-    }
-
-    GGML_UNUSED(dev);
-}
-
-/**
- * @brief Checks if the backend buffer type is associated with the CANN backend.
- *
- * This function checks whether the provided backend buffer type is associated
- * with the CANN backend based on the comparison of its name retrieval function
- * pointer.
- *
- * @param buft Pointer to the backend buffer type to check.
- * @return bool Returns true if the buffer type is associated with the CANN
- * backend, otherwise false.
- */
-static bool ggml_backend_buft_is_cann(ggml_backend_buffer_type_t buft) {
-    return buft->iface.get_name == ggml_backend_cann_buffer_type_name;
-}
-
-/**
- * @brief Determines if a tensor operation should be offloaded to the CANN
- * backend.
- *
- * This function checks if a given tensor operation should be offloaded to the
- * CANN backend based on the operation type and the size of the tensor. It
- * returns true if the second dimension (ne[1]) of the tensor is greater than or
- * equal to the minimum batch size and the operation is not GGML_OP_GET_ROWS.
- *
- * @param backend Pointer to the CANN backend.
- * @param op Pointer to the tensor operation to check.
- * @return bool Returns true if the operation should be offloaded, otherwise
- * false.
- */
-static bool ggml_backend_cann_offload_op(ggml_backend_dev_t dev,
-                                                   const ggml_tensor* op) {
-    const int min_batch_size = 32;
-    GGML_UNUSED(dev);
-
-    return op->ne[1] >= min_batch_size && op->op != GGML_OP_GET_ROWS;
-}
-
-/**
- * @brief Records an event on the CANN backend stream.
- *
- * This function records the given event on the ACL runtime stream associated
- * with the backend context.
- *
- * @param event Pointer to the event structure to be recorded.
- */
-static void ggml_backend_cann_event_record(ggml_backend_t backend, ggml_backend_event_t event) {
-    ggml_backend_cann_context* cann_ctx =
-        (ggml_backend_cann_context*)backend->context;
-    ACL_CHECK(aclrtRecordEvent((aclrtEvent)event->context, cann_ctx->stream()));
-}
-
-/**
- * @brief Waits for a recorded event to complete on the CANN backend stream.
- *
- * This function makes the given backend wait for the event to complete on its
- * ACL runtime stream.
- *
- * @param backend Pointer to the backend structure.
- * @param event Pointer to the event structure that the backend needs to wait
- * for.
- */
-static void ggml_backend_cann_event_wait(ggml_backend_t backend,
-                                         ggml_backend_event_t event) {
-    ggml_backend_cann_context* cann_ctx =
-        (ggml_backend_cann_context*)backend->context;
-    if (ggml_backend_is_cann(backend)) {
-        ACL_CHECK(aclrtStreamWaitEvent(cann_ctx->stream(),
-                                       (aclrtEvent)event->context));
-    } else {
-        GGML_ABORT("fatal error");
-    }
-}
-
-/**
- * @brief Structure defining the interface for the CANN backend.
- *
- * This structure contains function pointers for various operations
- * supported by the CANN backend, including name retrieval, memory
- * management, tensor operations, synchronization, and event handling.
- */
-static const ggml_backend_i ggml_backend_cann_interface = {
-    /* .get_name                = */ ggml_backend_cann_name,
-    /* .free                    = */ ggml_backend_cann_free,
-    /* .set_tensor_async        = */ ggml_backend_cann_set_tensor_async,
-    /* .get_tensor_async        = */ ggml_backend_cann_get_tensor_async,
-    /* .cpy_tensor_async        = */ ggml_backend_cann_cpy_tensor_async,
-    /* .synchronize             = */ ggml_backend_cann_synchronize,
-    /* .graph_plan_create       = */ NULL,
-    /* .graph_plan_free         = */ NULL,
-    /* .graph_plan_update       = */ NULL,
-    /* .graph_plan_compute      = */ NULL,
-    /* .graph_compute           = */ ggml_backend_cann_graph_compute,
-    /* .event_record            = */ ggml_backend_cann_event_record,
-    /* .event_wait              = */ ggml_backend_cann_event_wait,
-};
-
-/**
- * @brief Return the hardcoded GUID for the CANN backend.
- *
- * This function returns a static GUID which uniquely identifies the CANN
- * backend.
- *
- * @return A pointer to the static GUID.
- */
-static ggml_guid_t ggml_backend_cann_guid() {
-    static ggml_guid guid = {0xa1, 0x94, 0xaf, 0xac, 0xbd, 0x4f, 0x47, 0x34,
-                             0xbe, 0x1a, 0x9e, 0x71, 0x1f, 0x9e, 0xed, 0x64};
-    return &guid;
-}
-
-// backend device
-struct ggml_backend_cann_device_context {
-    int device;
-    std::string name;
-    std::string description;
-};
-
-static const char * ggml_backend_cann_device_get_name(ggml_backend_dev_t dev) {
-    ggml_backend_cann_device_context * ctx = (ggml_backend_cann_device_context *)dev->context;
-    return ctx->name.c_str();
-}
-
-static const char* ggml_backend_cann_device_get_description(ggml_backend_dev_t dev) {
-    ggml_backend_cann_device_context * ctx = (ggml_backend_cann_device_context *)dev->context;
-    return ctx->description.c_str();
-}
-
-static void ggml_backend_cann_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
-    ggml_backend_cann_device_context * ctx = (ggml_backend_cann_device_context *)dev->context;
-    ggml_backend_cann_get_device_memory(ctx->device, free, total);
-}
-
-static enum ggml_backend_dev_type ggml_backend_cann_device_get_type(ggml_backend_dev_t dev) {
-    GGML_UNUSED(dev);
-    return GGML_BACKEND_DEVICE_TYPE_GPU;
-}
-
-static void ggml_backend_cann_device_get_props(ggml_backend_dev_t dev, ggml_backend_dev_props * props) {
-    props->name        = ggml_backend_cann_device_get_name(dev);
-    props->description = ggml_backend_cann_device_get_description(dev);
-    props->type        = ggml_backend_cann_device_get_type(dev);
-    ggml_backend_cann_device_get_memory(dev, &props->memory_free, &props->memory_total);
-
-    bool host_buffer = getenv("GGML_CANN_NO_PINNED") == nullptr;
-
-    props->caps = {
-        /* .async                 = */ false,
-        /* .host_buffer           = */ host_buffer,
-        /* .buffer_from_host_ptr  = */ false,
-        /* .events                = */ true,
-    };
-}
-
-static ggml_backend_t ggml_backend_cann_device_init(ggml_backend_dev_t dev, const char * params) {
-    GGML_UNUSED(params);
-    ggml_backend_cann_device_context * ctx = (ggml_backend_cann_device_context *)dev->context;
-    return ggml_backend_cann_init(ctx->device);
-}
-
-/**
- * @brief Checks if the CANN backend supports a specific backend buffer type.
- *
- * This function determines whether the CANN backend supports the given backend
- * buffer type by comparing the device context of the backend and buffer type.
- * It returns true if the devices are same between the backend context and
- * buffer type context.
- *
- * @param backend Pointer to the CANN backend.
- * @param buft Pointer to the backend buffer type to check.
- * @return bool Returns true if the CANN backend supports the buffer type,
- *              otherwise false.
- */
-static bool ggml_backend_cann_supports_buft(
-    ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
-    if (ggml_backend_buft_is_cann(buft)) {
-        ggml_backend_cann_device_context * dev_ctx = (ggml_backend_cann_device_context *)dev->context;
-        ggml_backend_cann_buffer_type_context * buft_ctx =
-                        (ggml_backend_cann_buffer_type_context *)buft->context;
-        return buft_ctx->device == dev_ctx->device;
-    }
-    return false;
-}
-
-static ggml_backend_buffer_type_t ggml_backend_cann_device_get_buffer_type(ggml_backend_dev_t dev) {
-    ggml_backend_cann_device_context * ctx = (ggml_backend_cann_device_context *)dev->context;
-    return ggml_backend_cann_buffer_type(ctx->device);
-}
-
-static ggml_backend_buffer_type_t ggml_backend_cann_device_get_host_buffer_type(ggml_backend_dev_t dev) {
-    GGML_UNUSED(dev);
-    return ggml_backend_cann_host_buffer_type();
-}
-
-/**
- * @brief Creates a new event for the CANN backend device.
- *
- * This function initializes a new event for the CANN backend by setting the
- * device and creating an ACL runtime event. The created event is then wrapped
- * in a ggml_backend_event structure and returned.
- *
- * @param backend Pointer to the CANN backend.
- * @return ggml_backend_event_t Returns a pointer to the new event structure.
- */
-static ggml_backend_event_t ggml_backend_cann_device_event_new(
-    ggml_backend_dev_t dev) {
-    ggml_backend_cann_device_context * dev_ctx = (ggml_backend_cann_device_context *)dev->context;
-
-    ggml_cann_set_device(dev_ctx->device);
-
-    aclrtEvent event;
-    ACL_CHECK(aclrtCreateEvent(&event));
-
-    return new ggml_backend_event{
-        /* .device = */ ggml_backend_reg_dev_get(ggml_backend_cann_reg(), dev_ctx->device),
-        /* .context = */ event,
-    };
-}
-
-/**
- * @brief Frees a CANN backend event.
- *
- * This function destroys the ACL runtime event associated with the given CANN
- * backend event and then deletes the event structure itself.
- *
- * @param event Pointer to the event structure to be freed.
- */
-static void ggml_backend_cann_device_event_free(ggml_backend_dev_t dev, ggml_backend_event_t event) {
-    ACL_CHECK(aclrtDestroyEvent((aclrtEvent)event->context));
-
-    delete event;
-    GGML_UNUSED(dev);
-}
-
-/**
- * @brief Synchronizes the given event on the CANN backend.
- *
- * This function waits for the specified event to complete on the ACL runtime.
- *
- * @param event Pointer to the event structure to be synchronized.
- */
-static void ggml_backend_cann_device_event_synchronize(ggml_backend_dev_t dev, ggml_backend_event_t event) {
-    ACL_CHECK(aclrtSynchronizeEvent((aclrtEvent)event->context));
-
-    GGML_UNUSED(dev);
-}
-
-static const ggml_backend_device_i ggml_backend_cann_device_interface = {
-    /* .get_name                = */ ggml_backend_cann_device_get_name,
-    /* .get_description         = */ ggml_backend_cann_device_get_description,
-    /* .get_memory              = */ ggml_backend_cann_device_get_memory,
-    /* .get_type                = */ ggml_backend_cann_device_get_type,
-    /* .get_props               = */ ggml_backend_cann_device_get_props,
-    /* .init_backend            = */ ggml_backend_cann_device_init,    // called for every card
-    /* .get_buffer_type         = */ ggml_backend_cann_device_get_buffer_type,
-    /* .get_host_buffer_type    = */ ggml_backend_cann_device_get_host_buffer_type,
-    /* .buffer_from_host_ptr    = */ NULL, // not supported for CANN
-    /* .supports_op             = */ ggml_backend_cann_supports_op,
-    /* .supports_buft           = */ ggml_backend_cann_supports_buft,
-    /* .offload_op              = */ ggml_backend_cann_offload_op,
-    /* .event_new               = */ ggml_backend_cann_device_event_new,
-    /* .event_free              = */ ggml_backend_cann_device_event_free,
-    /* .event_synchronize       = */ ggml_backend_cann_device_event_synchronize,
-};
-
-
-// backend reg
-struct ggml_backend_cann_reg_context {
-    std::vector<ggml_backend_dev_t> devices;
-};
-
-static const char * ggml_backend_cann_reg_get_name(ggml_backend_reg_t reg) {
-    GGML_UNUSED(reg);
-    return GGML_CANN_NAME;
-}
-
-static size_t ggml_backend_cann_reg_get_device_count(ggml_backend_reg_t reg) {
-    ggml_backend_cann_reg_context * ctx = (ggml_backend_cann_reg_context *)reg->context;
-    return ctx->devices.size();
-}
-
-static ggml_backend_dev_t ggml_backend_cann_reg_get_device(ggml_backend_reg_t reg, size_t index) {
-    ggml_backend_cann_reg_context * ctx = (ggml_backend_cann_reg_context *)reg->context;
-    GGML_ASSERT(index < ctx->devices.size());
-    return ctx->devices[index];
-}
-
-static void * ggml_backend_cann_reg_get_proc_address(ggml_backend_reg_t reg, const char * name) {
-    GGML_UNUSED(reg);
-    GGML_UNUSED(name);
-    // reserved for future use
-    return nullptr;
-}
-
-static const ggml_backend_reg_i ggml_backend_cann_reg_interface = {
-    /* .get_name          = */ ggml_backend_cann_reg_get_name,
-    /* .get_device_count  = */ ggml_backend_cann_reg_get_device_count,
-    /* .get_device_get    = */ ggml_backend_cann_reg_get_device,
-    /* .get_proc_address  = */ ggml_backend_cann_reg_get_proc_address,
-};
-
-// backend registry, called only once for cann backend
-ggml_backend_reg_t ggml_backend_cann_reg() {
-    static ggml_backend_reg reg;
-    static bool initialized = false;
-
-    {
-        static std::mutex mutex;
-        std::lock_guard<std::mutex> lock(mutex);
-        if (!initialized) {
-            aclInit(nullptr);
-            ggml_backend_cann_reg_context * ctx = new ggml_backend_cann_reg_context;
-
-            for (int i = 0; i < ggml_cann_info().device_count; i++) {
-                ggml_backend_cann_device_context* dev_ctx = new ggml_backend_cann_device_context();
-                dev_ctx->description = aclrtGetSocName();
-                dev_ctx->device = i;
-                dev_ctx->name = GGML_CANN_NAME + std::to_string(i);
-                ggml_cann_set_device(i);
-                ggml_backend_dev_t dev = new ggml_backend_device {
-                    /* .interface = */ ggml_backend_cann_device_interface,
-                    /* .reg       = */ &reg,
-                    /* .context   = */ dev_ctx
-                };
-                ctx->devices.push_back(dev);
-            }
-
-            reg = ggml_backend_reg {
-                /* .interface = */ ggml_backend_cann_reg_interface,
-                /* .context   = */ ctx
-            };
-        }
-
-        initialized = true;
-    }
-
-    return &reg;
-}
-
-ggml_backend_t ggml_backend_cann_init(int32_t device) {
-    aclInit(nullptr);
-    if (device < 0 || device >= ggml_backend_cann_get_device_count()) {
-        GGML_LOG_ERROR("%s: error: invalid device %d\n", __func__, device);
-        return nullptr;
-    }
-
-    ggml_backend_cann_context* ctx = new ggml_backend_cann_context(device);
-    if (ctx == nullptr) {
-        GGML_LOG_ERROR("%s: error: failed to allocate context\n", __func__);
-        return nullptr;
-    }
-    ggml_cann_set_device(ctx->device);
-    ggml_backend_t cann_backend =
-        new ggml_backend{/* .guid      = */ ggml_backend_cann_guid(),
-                         /* .interface = */ ggml_backend_cann_interface,
-                         /* .device    = */ ggml_backend_reg_dev_get(ggml_backend_cann_reg(), device),
-                         /* .context   = */ ctx};
-
-    return cann_backend;
-}
-
-bool ggml_backend_is_cann(ggml_backend_t backend) {
-    return backend != NULL &&
-           ggml_guid_matches(backend->guid, ggml_backend_cann_guid());
-}
-
-int32_t ggml_backend_cann_get_device_count() {
-    return ggml_cann_info().device_count;
-}
-
-void ggml_backend_cann_get_device_description(
-    int32_t device, char* description, size_t description_size) {
-    ggml_cann_set_device(device);
-    const char* soc_name = aclrtGetSocName();
-    snprintf(description, description_size, "%s", soc_name);
-}
-
-void ggml_backend_cann_get_device_memory(int32_t device, size_t* free,
-                                         size_t* total) {
-    ggml_cann_set_device(device);
-    ACL_CHECK(aclrtGetMemInfo(ACL_HBM_MEM, free, total));
-}

ggml/src/ggml-cpu-impl.h

@@ -1,614 +0,0 @@
-#pragma once
-
-// GGML CPU internal header
-
-#include "ggml.h"
-#include "ggml-impl.h"
-#include <stdlib.h> // load `stdlib.h` before other headers to work around MinGW bug: https://sourceforge.net/p/mingw-w64/bugs/192/
-//#include <stddef.h>
-#include <stdbool.h>
-#include <string.h> // memcpy
-#include <math.h>   // fabsf
-
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-#if defined(_MSC_VER)
-
-#define m512bh(p) p
-#define m512i(p) p
-
-#else
-
-#define m512bh(p) (__m512bh)(p)
-#define m512i(p) (__m512i)(p)
-
-#endif
-
-/**
- * Converts brain16 to float32.
- *
- * The bfloat16 floating point format has the following structure:
- *
- *       ┌sign
- *       │
- *       │   ┌exponent
- *       │   │
- *       │   │      ┌mantissa
- *       │   │      │
- *       │┌──┴───┐┌─┴───┐
- *     0b0000000000000000 brain16
- *
- * Since bf16 has the same number of exponent bits as a 32bit float,
- * encoding and decoding numbers becomes relatively straightforward.
- *
- *       ┌sign
- *       │
- *       │   ┌exponent
- *       │   │
- *       │   │      ┌mantissa
- *       │   │      │
- *       │┌──┴───┐┌─┴───────────────────┐
- *     0b00000000000000000000000000000000 IEEE binary32
- *
- * For comparison, the standard fp16 format has fewer exponent bits.
- *
- *       ┌sign
- *       │
- *       │  ┌exponent
- *       │  │
- *       │  │    ┌mantissa
- *       │  │    │
- *       │┌─┴─┐┌─┴──────┐
- *     0b0000000000000000 IEEE binary16
- *
- * @see IEEE 754-2008
- */
-static inline float ggml_compute_bf16_to_fp32(ggml_bf16_t h) {
-    union {
-        float f;
-        uint32_t i;
-    } u;
-    u.i = (uint32_t)h.bits << 16;
-    return u.f;
-}
-
-/**
- * Converts float32 to brain16.
- *
- * This is binary identical with Google Brain float conversion.
- * Floats shall round to nearest even, and NANs shall be quiet.
- * Subnormals aren't flushed to zero, except perhaps when used.
- * This code should vectorize nicely if using modern compilers.
- */
-static inline ggml_bf16_t ggml_compute_fp32_to_bf16(float s) {
-    ggml_bf16_t h;
-    union {
-        float f;
-        uint32_t i;
-    } u;
-    u.f = s;
-    if ((u.i & 0x7fffffff) > 0x7f800000) { /* nan */
-        h.bits = (u.i >> 16) | 64; /* force to quiet */
-        return h;
-    }
-    h.bits = (u.i + (0x7fff + ((u.i >> 16) & 1))) >> 16;
-    return h;
-}
-
-#define GGML_FP32_TO_BF16(x) ggml_compute_fp32_to_bf16(x)
-#define GGML_BF16_TO_FP32(x) ggml_compute_bf16_to_fp32(x)
-
-// __FMA__ and __F16C__ are not defined in MSVC, however they are implied with AVX2/AVX512
-#if defined(_MSC_VER) && (defined(__AVX2__) || defined(__AVX512F__))
-#ifndef __FMA__
-#define __FMA__
-#endif
-#ifndef __F16C__
-#define __F16C__
-#endif
-#endif
-
-// __SSE3__ and __SSSE3__ are not defined in MSVC, but SSE3/SSSE3 are present when AVX/AVX2/AVX512 are available
-#if defined(_MSC_VER) && (defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__))
-#ifndef __SSE3__
-#define __SSE3__
-#endif
-#ifndef __SSSE3__
-#define __SSSE3__
-#endif
-#endif
-
-#if defined(__ARM_FEATURE_SVE)
-#include <arm_sve.h>
-#include <sys/prctl.h>
-#endif
-
-// 16-bit float
-// on Arm, we use __fp16
-// on x86, we use uint16_t
-#if defined(__ARM_NEON)
-
-// if YCM cannot find <arm_neon.h>, make a symbolic link to it, for example:
-//
-//   $ ln -sfn /Library/Developer/CommandLineTools/usr/lib/clang/13.1.6/include/arm_neon.h ./src/
-//
-#include <arm_neon.h>
-
-#ifdef _MSC_VER
-
-typedef uint16_t ggml_fp16_internal_t;
-
-#define ggml_vld1q_u32(w,x,y,z) { ((w) + ((uint64_t)(x) << 32)), ((y) + ((uint64_t)(z) << 32)) }
-
-#else
-
-typedef __fp16 ggml_fp16_internal_t;
-
-#define ggml_vld1q_u32(w,x,y,z) { (w), (x), (y), (z) }
-
-#endif // _MSC_VER
-
-#if !defined(__aarch64__)
-
-// 32-bit ARM compatibility
-
-// vaddlvq_s16
-// vpaddq_s16
-// vpaddq_s32
-// vaddvq_s32
-// vaddvq_f32
-// vmaxvq_f32
-// vcvtnq_s32_f32
-// vzip1_u8
-// vzip2_u8
-
-inline static int32_t vaddlvq_s16(int16x8_t v) {
-    int32x4_t v0 = vreinterpretq_s32_s64(vpaddlq_s32(vpaddlq_s16(v)));
-    return vgetq_lane_s32(v0, 0) + vgetq_lane_s32(v0, 2);
-}
-
-inline static int16x8_t vpaddq_s16(int16x8_t a, int16x8_t b) {
-    int16x4_t a0 = vpadd_s16(vget_low_s16(a), vget_high_s16(a));
-    int16x4_t b0 = vpadd_s16(vget_low_s16(b), vget_high_s16(b));
-    return vcombine_s16(a0, b0);
-}
-
-inline static int32x4_t vpaddq_s32(int32x4_t a, int32x4_t b) {
-    int32x2_t a0 = vpadd_s32(vget_low_s32(a), vget_high_s32(a));
-    int32x2_t b0 = vpadd_s32(vget_low_s32(b), vget_high_s32(b));
-    return vcombine_s32(a0, b0);
-}
-
-inline static int32_t vaddvq_s32(int32x4_t v) {
-    return vgetq_lane_s32(v, 0) + vgetq_lane_s32(v, 1) + vgetq_lane_s32(v, 2) + vgetq_lane_s32(v, 3);
-}
-
-inline static float vaddvq_f32(float32x4_t v) {
-    return vgetq_lane_f32(v, 0) + vgetq_lane_f32(v, 1) + vgetq_lane_f32(v, 2) + vgetq_lane_f32(v, 3);
-}
-
-inline static float vmaxvq_f32(float32x4_t v) {
-    return
-        MAX(MAX(vgetq_lane_f32(v, 0), vgetq_lane_f32(v, 1)),
-            MAX(vgetq_lane_f32(v, 2), vgetq_lane_f32(v, 3)));
-}
-
-inline static int32x4_t vcvtnq_s32_f32(float32x4_t v) {
-    int32x4_t res;
-
-    res[0] = roundf(vgetq_lane_f32(v, 0));
-    res[1] = roundf(vgetq_lane_f32(v, 1));
-    res[2] = roundf(vgetq_lane_f32(v, 2));
-    res[3] = roundf(vgetq_lane_f32(v, 3));
-
-    return res;
-}
-
-inline static uint8x8_t vzip1_u8(uint8x8_t a, uint8x8_t b) {
-    uint8x8_t res;
-
-    res[0] = a[0]; res[1] = b[0];
-    res[2] = a[1]; res[3] = b[1];
-    res[4] = a[2]; res[5] = b[2];
-    res[6] = a[3]; res[7] = b[3];
-
-    return res;
-}
-
-inline static uint8x8_t vzip2_u8(uint8x8_t a, uint8x8_t b) {
-    uint8x8_t res;
-
-    res[0] = a[4]; res[1] = b[4];
-    res[2] = a[5]; res[3] = b[5];
-    res[4] = a[6]; res[5] = b[6];
-    res[6] = a[7]; res[7] = b[7];
-
-    return res;
-}
-
-// vld1q_s16_x2
-// vld1q_u8_x2
-// vld1q_u8_x4
-// vld1q_s8_x2
-// vld1q_s8_x4
-// TODO: double-check these work correctly
-
-typedef struct ggml_int16x8x2_t {
-    int16x8_t val[2];
-} ggml_int16x8x2_t;
-
-inline static ggml_int16x8x2_t ggml_vld1q_s16_x2(const int16_t * ptr) {
-    ggml_int16x8x2_t res;
-
-    res.val[0] = vld1q_s16(ptr + 0);
-    res.val[1] = vld1q_s16(ptr + 8);
-
-    return res;
-}
-
-typedef struct ggml_uint8x16x2_t {
-    uint8x16_t val[2];
-} ggml_uint8x16x2_t;
-
-inline static ggml_uint8x16x2_t ggml_vld1q_u8_x2(const uint8_t * ptr) {
-    ggml_uint8x16x2_t res;
-
-    res.val[0] = vld1q_u8(ptr + 0);
-    res.val[1] = vld1q_u8(ptr + 16);
-
-    return res;
-}
-
-typedef struct ggml_uint8x16x4_t {
-    uint8x16_t val[4];
-} ggml_uint8x16x4_t;
-
-inline static ggml_uint8x16x4_t ggml_vld1q_u8_x4(const uint8_t * ptr) {
-    ggml_uint8x16x4_t res;
-
-    res.val[0] = vld1q_u8(ptr + 0);
-    res.val[1] = vld1q_u8(ptr + 16);
-    res.val[2] = vld1q_u8(ptr + 32);
-    res.val[3] = vld1q_u8(ptr + 48);
-
-    return res;
-}
-
-typedef struct ggml_int8x16x2_t {
-    int8x16_t val[2];
-} ggml_int8x16x2_t;
-
-inline static ggml_int8x16x2_t ggml_vld1q_s8_x2(const int8_t * ptr) {
-    ggml_int8x16x2_t res;
-
-    res.val[0] = vld1q_s8(ptr + 0);
-    res.val[1] = vld1q_s8(ptr + 16);
-
-    return res;
-}
-
-typedef struct ggml_int8x16x4_t {
-    int8x16_t val[4];
-} ggml_int8x16x4_t;
-
-inline static ggml_int8x16x4_t ggml_vld1q_s8_x4(const int8_t * ptr) {
-    ggml_int8x16x4_t res;
-
-    res.val[0] = vld1q_s8(ptr + 0);
-    res.val[1] = vld1q_s8(ptr + 16);
-    res.val[2] = vld1q_s8(ptr + 32);
-    res.val[3] = vld1q_s8(ptr + 48);
-
-    return res;
-}
-
-// NOTE: not tested
-inline static int8x16_t ggml_vqtbl1q_s8(int8x16_t a, uint8x16_t b) {
-    int8x16_t res;
-
-    res[ 0] = a[b[ 0]];
-    res[ 1] = a[b[ 1]];
-    res[ 2] = a[b[ 2]];
-    res[ 3] = a[b[ 3]];
-    res[ 4] = a[b[ 4]];
-    res[ 5] = a[b[ 5]];
-    res[ 6] = a[b[ 6]];
-    res[ 7] = a[b[ 7]];
-    res[ 8] = a[b[ 8]];
-    res[ 9] = a[b[ 9]];
-    res[10] = a[b[10]];
-    res[11] = a[b[11]];
-    res[12] = a[b[12]];
-    res[13] = a[b[13]];
-    res[14] = a[b[14]];
-    res[15] = a[b[15]];
-
-    return res;
-}
-
-// NOTE: not tested
-inline static uint8x16_t ggml_vqtbl1q_u8(uint8x16_t a, uint8x16_t b) {
-    uint8x16_t res;
-
-    res[ 0] = a[b[ 0]];
-    res[ 1] = a[b[ 1]];
-    res[ 2] = a[b[ 2]];
-    res[ 3] = a[b[ 3]];
-    res[ 4] = a[b[ 4]];
-    res[ 5] = a[b[ 5]];
-    res[ 6] = a[b[ 6]];
-    res[ 7] = a[b[ 7]];
-    res[ 8] = a[b[ 8]];
-    res[ 9] = a[b[ 9]];
-    res[10] = a[b[10]];
-    res[11] = a[b[11]];
-    res[12] = a[b[12]];
-    res[13] = a[b[13]];
-    res[14] = a[b[14]];
-    res[15] = a[b[15]];
-
-    return res;
-}
-
-#else
-
-#define ggml_int16x8x2_t  int16x8x2_t
-#define ggml_uint8x16x2_t uint8x16x2_t
-#define ggml_uint8x16x4_t uint8x16x4_t
-#define ggml_int8x16x2_t  int8x16x2_t
-#define ggml_int8x16x4_t  int8x16x4_t
-
-#define ggml_vld1q_s16_x2 vld1q_s16_x2
-#define ggml_vld1q_u8_x2  vld1q_u8_x2
-#define ggml_vld1q_u8_x4  vld1q_u8_x4
-#define ggml_vld1q_s8_x2  vld1q_s8_x2
-#define ggml_vld1q_s8_x4  vld1q_s8_x4
-#define ggml_vqtbl1q_s8   vqtbl1q_s8
-#define ggml_vqtbl1q_u8   vqtbl1q_u8
-
-#endif // !defined(__aarch64__)
-
-#if !defined(__ARM_FEATURE_DOTPROD)
-
-inline static int32x4_t ggml_vdotq_s32(int32x4_t acc, int8x16_t a, int8x16_t b) {
-    const int16x8_t p0 = vmull_s8(vget_low_s8 (a), vget_low_s8 (b));
-    const int16x8_t p1 = vmull_s8(vget_high_s8(a), vget_high_s8(b));
-
-    return vaddq_s32(acc, vaddq_s32(vpaddlq_s16(p0), vpaddlq_s16(p1)));
-}
-
-#else
-
-#define ggml_vdotq_s32(a, b, c) vdotq_s32(a, b, c)
-
-#endif // !defined(__ARM_FEATURE_DOTPROD)
-
-#endif // defined(__ARM_NEON)
-
-#if defined(__ARM_NEON) && !defined(_MSC_VER)
-
-#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
-#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
-
-#define GGML_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
-
-static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
-    ggml_fp16_internal_t tmp;
-    memcpy(&tmp, &h, sizeof(ggml_fp16_t));
-    return (float)tmp;
-}
-
-static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
-    ggml_fp16_t res;
-    ggml_fp16_internal_t tmp = f;
-    memcpy(&res, &tmp, sizeof(ggml_fp16_t));
-    return res;
-}
-
-#else
-
-#ifdef __wasm_simd128__
-#include <wasm_simd128.h>
-#else
-#ifdef __POWER9_VECTOR__
-#include <altivec.h>
-#undef bool
-#define bool _Bool
-#else
-#if defined(_MSC_VER) || defined(__MINGW32__)
-#include <intrin.h>
-#else
-#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__) || defined(__SSE3__) || defined(__SSE__)
-#if !defined(__riscv)
-#include <immintrin.h>
-#endif
-#endif
-#endif
-#endif
-#endif
-
-#ifdef __riscv_v_intrinsic
-#include <riscv_vector.h>
-#endif
-
-#if defined(__loongarch64)
-#if defined(__loongarch_asx)
-#include <lasxintrin.h>
-#endif
-#if defined(__loongarch_sx)
-#include <lsxintrin.h>
-#endif
-#endif
-
-#if defined(__loongarch_asx)
-
-typedef union {
-    int32_t i;
-    float f;
-} ft_union;
-
-/* float type data load instructions */
-static __m128 __lsx_vreplfr2vr_s(float val) {
-    ft_union fi_tmpval = {.f = val};
-    return (__m128)__lsx_vreplgr2vr_w(fi_tmpval.i);
-}
-
-static __m256 __lasx_xvreplfr2vr_s(float val) {
-    ft_union fi_tmpval = {.f = val};
-    return (__m256)__lasx_xvreplgr2vr_w(fi_tmpval.i);
-}
-#endif
-
-#ifdef __F16C__
-
-#ifdef _MSC_VER
-#define GGML_COMPUTE_FP16_TO_FP32(x) _mm_cvtss_f32(_mm_cvtph_ps(_mm_cvtsi32_si128(x)))
-#define GGML_COMPUTE_FP32_TO_FP16(x) _mm_extract_epi16(_mm_cvtps_ph(_mm_set_ss(x), 0), 0)
-#else
-#define GGML_COMPUTE_FP16_TO_FP32(x) _cvtsh_ss(x)
-#define GGML_COMPUTE_FP32_TO_FP16(x) _cvtss_sh(x, 0)
-#endif
-
-#elif defined(__POWER9_VECTOR__)
-
-#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
-#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
-/* the inline asm below is about 12% faster than the lookup method */
-#define GGML_FP16_TO_FP32(x) GGML_COMPUTE_FP16_TO_FP32(x)
-#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
-
-static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
-    register float f;
-    register double d;
-    __asm__(
-        "mtfprd %0,%2\n"
-        "xscvhpdp %0,%0\n"
-        "frsp %1,%0\n" :
-        /* temp */ "=d"(d),
-        /* out */  "=f"(f):
-        /* in */   "r"(h));
-    return f;
-}
-
-static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
-    register double d;
-    register ggml_fp16_t r;
-    __asm__( /* xscvdphp can work on double or single precision */
-        "xscvdphp %0,%2\n"
-        "mffprd %1,%0\n" :
-        /* temp */ "=d"(d),
-        /* out */  "=r"(r):
-        /* in */   "f"(f));
-    return r;
-}
-
-#else
-
-// FP16 <-> FP32
-// ref: https://github.com/Maratyszcza/FP16
-
-static inline float fp32_from_bits(uint32_t w) {
-    union {
-        uint32_t as_bits;
-        float as_value;
-    } fp32;
-    fp32.as_bits = w;
-    return fp32.as_value;
-}
-
-static inline uint32_t fp32_to_bits(float f) {
-    union {
-        float as_value;
-        uint32_t as_bits;
-    } fp32;
-    fp32.as_value = f;
-    return fp32.as_bits;
-}
-
-static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
-    const uint32_t w = (uint32_t) h << 16;
-    const uint32_t sign = w & UINT32_C(0x80000000);
-    const uint32_t two_w = w + w;
-
-    const uint32_t exp_offset = UINT32_C(0xE0) << 23;
-#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)
-    const float exp_scale = 0x1.0p-112f;
-#else
-    const float exp_scale = fp32_from_bits(UINT32_C(0x7800000));
-#endif
-    const float normalized_value = fp32_from_bits((two_w >> 4) + exp_offset) * exp_scale;
-
-    const uint32_t magic_mask = UINT32_C(126) << 23;
-    const float magic_bias = 0.5f;
-    const float denormalized_value = fp32_from_bits((two_w >> 17) | magic_mask) - magic_bias;
-
-    const uint32_t denormalized_cutoff = UINT32_C(1) << 27;
-    const uint32_t result = sign |
-        (two_w < denormalized_cutoff ? fp32_to_bits(denormalized_value) : fp32_to_bits(normalized_value));
-    return fp32_from_bits(result);
-}
-
-static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
-#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)
-    const float scale_to_inf = 0x1.0p+112f;
-    const float scale_to_zero = 0x1.0p-110f;
-#else
-    const float scale_to_inf = fp32_from_bits(UINT32_C(0x77800000));
-    const float scale_to_zero = fp32_from_bits(UINT32_C(0x08800000));
-#endif
-    float base = (fabsf(f) * scale_to_inf) * scale_to_zero;
-
-    const uint32_t w = fp32_to_bits(f);
-    const uint32_t shl1_w = w + w;
-    const uint32_t sign = w & UINT32_C(0x80000000);
-    uint32_t bias = shl1_w & UINT32_C(0xFF000000);
-    if (bias < UINT32_C(0x71000000)) {
-        bias = UINT32_C(0x71000000);
-    }
-
-    base = fp32_from_bits((bias >> 1) + UINT32_C(0x07800000)) + base;
-    const uint32_t bits = fp32_to_bits(base);
-    const uint32_t exp_bits = (bits >> 13) & UINT32_C(0x00007C00);
-    const uint32_t mantissa_bits = bits & UINT32_C(0x00000FFF);
-    const uint32_t nonsign = exp_bits + mantissa_bits;
-    return (sign >> 16) | (shl1_w > UINT32_C(0xFF000000) ? UINT16_C(0x7E00) : nonsign);
-}
-
-#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
-#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
-
-#endif // __F16C__
-
-#endif // defined(__ARM_NEON) && (!defined(__MSC_VER)
-
-#ifdef __ARM_FEATURE_SVE
-#include <arm_sve.h>
-#endif // __ARM_FEATURE_SVE
-
-// precomputed f32 table for f16 (256 KB)
-// defined in ggml.c, initialized in ggml_init()
-extern float ggml_table_f32_f16[1 << 16];
-
-// On ARM NEON, it's quicker to directly convert x -> x instead of calling into ggml_lookup_fp16_to_fp32,
-// so we define GGML_FP16_TO_FP32 and GGML_FP32_TO_FP16 elsewhere for NEON.
-// This is also true for POWER9.
-#if !defined(GGML_FP16_TO_FP32)
-inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
-    uint16_t s;
-    memcpy(&s, &f, sizeof(uint16_t));
-    return ggml_table_f32_f16[s];
-}
-
-#define GGML_FP16_TO_FP32(x) ggml_lookup_fp16_to_fp32(x)
-#endif
-
-#if !defined(GGML_FP32_TO_FP16)
-#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
-#endif
-
-#ifdef __cplusplus
-}
-#endif

ggml/src/ggml-cuda/CMakeLists.txt

@@ -0,0 +1,155 @@
+cmake_minimum_required(VERSION 3.18)  # for CMAKE_CUDA_ARCHITECTURES
+
+find_package(CUDAToolkit)
+
+if (CUDAToolkit_FOUND)
+    message(STATUS "CUDA Toolkit found")
+
+    if (NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
+        # native == GPUs available at build time
+        # 52     == Maxwell, lowest CUDA 12 standard
+        # 60     == P100, FP16 CUDA intrinsics
+        # 61     == Pascal, __dp4a instruction (per-byte integer dot product)
+        # 70     == V100, FP16 tensor cores
+        # 75     == Turing, int8 tensor cores
+        if (GGML_NATIVE AND CUDAToolkit_VERSION VERSION_GREATER_EQUAL "11.6" AND CMAKE_VERSION VERSION_GREATER_EQUAL "3.24")
+            set(CMAKE_CUDA_ARCHITECTURES "native")
+        elseif(GGML_CUDA_F16 OR GGML_CUDA_DMMV_F16)
+            set(CMAKE_CUDA_ARCHITECTURES "60;61;70;75")
+        else()
+            set(CMAKE_CUDA_ARCHITECTURES "52;61;70;75")
+        endif()
+    endif()
+    message(STATUS "Using CUDA architectures: ${CMAKE_CUDA_ARCHITECTURES}")
+
+    enable_language(CUDA)
+
+    file(GLOB   GGML_HEADERS_CUDA "*.cuh")
+    list(APPEND GGML_HEADERS_CUDA "../../include/ggml-cuda.h")
+
+    file(GLOB   GGML_SOURCES_CUDA "*.cu")
+    file(GLOB   SRCS "template-instances/fattn-wmma*.cu")
+    list(APPEND GGML_SOURCES_CUDA ${SRCS})
+    file(GLOB   SRCS "template-instances/mmq*.cu")
+    list(APPEND GGML_SOURCES_CUDA ${SRCS})
+
+    if (GGML_CUDA_FA_ALL_QUANTS)
+        file(GLOB   SRCS "template-instances/fattn-vec*.cu")
+        list(APPEND GGML_SOURCES_CUDA ${SRCS})
+        add_compile_definitions(GGML_CUDA_FA_ALL_QUANTS)
+    else()
+        file(GLOB   SRCS "template-instances/fattn-vec*q4_0-q4_0.cu")
+        list(APPEND GGML_SOURCES_CUDA ${SRCS})
+        file(GLOB   SRCS "template-instances/fattn-vec*q8_0-q8_0.cu")
+        list(APPEND GGML_SOURCES_CUDA ${SRCS})
+        file(GLOB   SRCS "template-instances/fattn-vec*f16-f16.cu")
+        list(APPEND GGML_SOURCES_CUDA ${SRCS})
+    endif()
+
+    add_library(ggml-cuda
+                ${GGML_HEADERS_CUDA}
+                ${GGML_SOURCES_CUDA}
+                )
+
+    target_link_libraries(ggml-cuda PRIVATE ggml-base)
+    target_include_directories(ggml-cuda PRIVATE . ..)
+
+    add_compile_definitions(GGML_CUDA_PEER_MAX_BATCH_SIZE=${GGML_CUDA_PEER_MAX_BATCH_SIZE})
+
+    if (GGML_CUDA_GRAPHS)
+        add_compile_definitions(GGML_CUDA_USE_GRAPHS)
+    endif()
+
+    if (GGML_CUDA_FORCE_MMQ)
+        add_compile_definitions(GGML_CUDA_FORCE_MMQ)
+    endif()
+
+    if (GGML_CUDA_FORCE_CUBLAS)
+        add_compile_definitions(GGML_CUDA_FORCE_CUBLAS)
+    endif()
+
+    if (GGML_CUDA_NO_VMM)
+        add_compile_definitions(GGML_CUDA_NO_VMM)
+    endif()
+
+    if (GGML_CUDA_F16 OR GGML_CUDA_DMMV_F16)
+        add_compile_definitions(GGML_CUDA_F16)
+    endif()
+
+    if (GGML_CUDA_NO_PEER_COPY)
+        add_compile_definitions(GGML_CUDA_NO_PEER_COPY)
+    endif()
+
+    if (GGML_STATIC)
+        if (WIN32)
+            # As of 12.3.1 CUDA Toolkit for Windows does not offer a static cublas library
+            target_link_libraries(ggml-cuda PRIVATE CUDA::cudart_static CUDA::cublas CUDA::cublasLt)
+        else ()
+            target_link_libraries(ggml-cuda PRIVATE  CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)
+        endif()
+    else()
+        target_link_libraries(ggml-cuda PRIVATE CUDA::cudart CUDA::cublas CUDA::cublasLt)
+    endif()
+
+    if (GGML_CUDA_NO_VMM)
+        # No VMM requested, no need to link directly with the cuda driver lib (libcuda.so)
+    else()
+        target_link_libraries(ggml-cuda PRIVATE CUDA::cuda_driver)
+    endif()
+
+    set(CUDA_CXX_FLAGS "")
+
+    set(CUDA_FLAGS -use_fast_math)
+
+    if (GGML_FATAL_WARNINGS)
+        list(APPEND CUDA_FLAGS -Werror all-warnings)
+    endif()
+
+    if (GGML_ALL_WARNINGS AND NOT MSVC)
+        set(NVCC_CMD ${CMAKE_CUDA_COMPILER} .c)
+        if (NOT CMAKE_CUDA_HOST_COMPILER STREQUAL "")
+            list(APPEND NVCC_CMD -ccbin ${CMAKE_CUDA_HOST_COMPILER})
+        endif()
+
+        execute_process(
+            COMMAND ${NVCC_CMD} -Xcompiler --version
+            OUTPUT_VARIABLE CUDA_CCFULLVER
+            ERROR_QUIET
+        )
+
+        if (NOT CUDA_CCFULLVER MATCHES clang)
+            set(CUDA_CCID "GNU")
+            execute_process(
+                COMMAND ${NVCC_CMD} -Xcompiler "-dumpfullversion -dumpversion"
+                OUTPUT_VARIABLE CUDA_CCVER
+                ERROR_QUIET
+            )
+        else()
+            if (CUDA_CCFULLVER MATCHES Apple)
+                set(CUDA_CCID "AppleClang")
+            else()
+                set(CUDA_CCID "Clang")
+            endif()
+            string(REGEX REPLACE "^.* version ([0-9.]*).*$" "\\1" CUDA_CCVER ${CUDA_CCFULLVER})
+        endif()
+
+        message("-- CUDA host compiler is ${CUDA_CCID} ${CUDA_CCVER}")
+
+        get_flags(${CUDA_CCID} ${CUDA_CCVER})
+        list(APPEND CUDA_CXX_FLAGS ${CXX_FLAGS} ${GF_CXX_FLAGS})  # This is passed to -Xcompiler later
+    endif()
+
+    if (NOT MSVC)
+        list(APPEND CUDA_CXX_FLAGS -Wno-pedantic)
+    endif()
+
+    list(JOIN   CUDA_CXX_FLAGS " " CUDA_CXX_FLAGS_JOINED)  # pass host compiler flags as a single argument
+
+    if (NOT CUDA_CXX_FLAGS_JOINED STREQUAL "")
+        list(APPEND CUDA_FLAGS -Xcompiler ${CUDA_CXX_FLAGS_JOINED})
+    endif()
+
+    target_compile_options(ggml-cuda PRIVATE "$<$<COMPILE_LANGUAGE:CUDA>:${CUDA_FLAGS}>")
+else()
+    message(FATAL_ERROR "CUDA Toolkit not found")
+endif()

ggml/src/ggml-kompute.cpp

@@ -1,2184 +0,0 @@
-#include "ggml-impl.h"
-#include "ggml-backend.h"
-#include "ggml-backend-impl.h"
-#include "ggml-kompute.h"
-
-// These are generated at build time by cmake custom command
-#include "shaderop_scale.h"
-#include "shaderop_scale_8.h"
-#include "shaderop_add.h"
-#include "shaderop_addrow.h"
-#include "shaderop_mul.h"
-#include "shaderop_silu.h"
-#include "shaderop_relu.h"
-#include "shaderop_gelu.h"
-#include "shaderop_softmax.h"
-#include "shaderop_norm.h"
-#include "shaderop_rmsnorm.h"
-#include "shaderop_diagmask.h"
-#include "shaderop_mul_mat_f16.h"
-#include "shaderop_mul_mat_q8_0.h"
-#include "shaderop_mul_mat_q4_0.h"
-#include "shaderop_mul_mat_q4_1.h"
-#include "shaderop_mul_mat_q4_k.h"
-#include "shaderop_mul_mat_q6_k.h"
-#include "shaderop_mul_mat_mat_f32.h"
-#include "shaderop_getrows_f32.h"
-#include "shaderop_getrows_f16.h"
-#include "shaderop_getrows_q4_0.h"
-#include "shaderop_getrows_q4_1.h"
-#include "shaderop_getrows_q6_k.h"
-#include "shaderop_rope_f16.h"
-#include "shaderop_rope_f32.h"
-#include "shaderop_cpy_f16_f16.h"
-#include "shaderop_cpy_f16_f32.h"
-#include "shaderop_cpy_f32_f16.h"
-#include "shaderop_cpy_f32_f32.h"
-
-#include <algorithm>
-#include <array>
-#include <cassert>
-#include <cstdint>
-#include <cstdio>
-#include <cstring>
-#include <iostream>
-#include <memory>
-#include <mutex>
-#include <stdexcept>
-#include <string>
-#include <unordered_map>
-#include <utility>
-#include <vector>
-
-#include <kompute/Kompute.hpp>
-#include <vulkan/vulkan.hpp>
-
-#ifdef __linux__
-#include <cstdlib> // for setenv
-#endif
-
-#define QK4_0 32
-#define QR4_0 2
-#define QK4_1 32
-#define QK_NL 16
-
-typedef ggml_fp16_t half;
-
-static std::string ggml_kompute_format_name(int device) {
-    return "Kompute" + std::to_string(device);
-}
-
-struct ggml_kompute_context {
-    int device;
-    std::string name;
-    std::shared_ptr<vk::DescriptorPool> pool;
-
-    ggml_kompute_context(int device)
-        : device(device), name(ggml_kompute_format_name(device)) {}
-};
-
-// FIXME: It would be good to consolidate the kompute manager and the kompute context into one object
-// and consolidate the init functions and simplify object lifetime management. As it currently stands,
-// we *have* to have the kompute manager no matter what for device discovery, but the kompute context
-// is only created when a device is set and vulkan is explicitly turned on.
-static ggml_kompute_context *s_kompute_context = nullptr;
-
-class kompute_manager {
-    kp::Manager *s_mgr = nullptr;
-
-public:
-    kp::Manager *operator()() {
-        if (s_mgr && !s_mgr->hasInstance()) {
-            destroy();
-        }
-        if (!s_mgr) {
-            s_mgr = new kp::Manager;
-        }
-        return s_mgr;
-    }
-
-    void destroy() {
-        delete s_mgr;
-        s_mgr = nullptr;
-    }
-};
-
-static kompute_manager komputeManager;
-
-struct ggml_vk_memory {
-    void *data = nullptr;
-    size_t size = 0;
-    vk::DeviceMemory *primaryMemory = nullptr;
-    vk::Buffer *primaryBuffer = nullptr;
-    vk::DeviceMemory *stagingMemory = nullptr;
-    vk::Buffer *stagingBuffer = nullptr;
-};
-
-#ifdef __linux__
-__attribute__((constructor))
-static void enable_sam() {
-    setenv("RADV_PERFTEST", "sam", false);
-}
-#endif
-
-static bool ggml_vk_checkPhysicalDeviceFeatures(vk::PhysicalDevice physical_device) {
-    vk::PhysicalDeviceFeatures availableFeatures;
-    physical_device.getFeatures(&availableFeatures);
-
-    if (!availableFeatures.shaderInt16)
-        return false;
-
-    vk::PhysicalDeviceVulkan11Features availableFeatures11;
-    vk::PhysicalDeviceVulkan12Features availableFeatures12;
-
-    availableFeatures11.pNext = &availableFeatures12;
-    availableFeatures12.pNext = nullptr;
-
-    vk::PhysicalDeviceFeatures2 features2;
-    features2.pNext = &availableFeatures11;
-
-    physical_device.getFeatures2(&features2);
-
-    if (!availableFeatures11.uniformAndStorageBuffer16BitAccess ||
-        !availableFeatures11.storageBuffer16BitAccess) {
-        return false;
-    }
-
-    if (!availableFeatures12.storageBuffer8BitAccess ||
-        !availableFeatures12.uniformAndStorageBuffer8BitAccess ||
-        !availableFeatures12.shaderFloat16 ||
-        !availableFeatures12.shaderInt8) {
-        return false;
-    }
-
-    return true;
-}
-
-static const char * ggml_vk_getVendorName(uint32_t vendorID) {
-    switch (vendorID) {
-        case 0x10DE:
-            return "nvidia";
-        case 0x1002:
-            return "amd";
-        case 0x8086:
-            return "intel";
-        default:
-            return "unknown";
-    }
-}
-
-static std::vector<ggml_vk_device> ggml_vk_available_devices_internal(size_t memoryRequired) {
-    std::vector<ggml_vk_device> results;
-    if (!komputeManager()->hasVulkan() || !komputeManager()->hasInstance())
-        return results;
-
-    std::vector<vk::PhysicalDevice> physical_devices;
-    try {
-        physical_devices = komputeManager()->listDevices();
-    } catch (vk::SystemError & err) {
-        std::cerr << __func__ << ": ignoring Vulkan exception: " << err.what() << "\n";
-        return results;
-    }
-
-    uint32_t deviceCount = physical_devices.size();
-    if (deviceCount == 0)
-        return results;
-
-    std::unordered_map<std::string, size_t> count_by_name;
-
-    for (uint32_t i = 0; i < deviceCount; i++) {
-        const auto & physical_device = physical_devices[i];
-
-        VkPhysicalDeviceProperties dev_props = physical_device.getProperties();
-        VkPhysicalDeviceMemoryProperties memoryProperties = physical_device.getMemoryProperties();
-        const uint32_t major = VK_VERSION_MAJOR(dev_props.apiVersion);
-        const uint32_t minor = VK_VERSION_MINOR(dev_props.apiVersion);
-        if (major < 1 || minor < 2)
-            continue;
-
-        if (!ggml_vk_checkPhysicalDeviceFeatures(physical_device))
-            continue;
-
-        size_t heapSize = 0;
-        for (uint32_t j = 0; j < memoryProperties.memoryHeapCount; ++j) {
-            VkMemoryHeap heap = memoryProperties.memoryHeaps[j];
-            if (heap.flags & VK_MEMORY_HEAP_DEVICE_LOCAL_BIT) {
-                heapSize = heap.size;
-                break;
-            }
-        }
-
-        if (heapSize < memoryRequired)
-            continue;
-
-        auto ext_props = physical_device.enumerateDeviceExtensionProperties();
-        bool has_maintenance4 = false;
-
-        // Check if maintenance4 is supported
-        for (const auto & properties : ext_props) {
-            if (strcmp("VK_KHR_maintenance4", properties.extensionName) == 0) {
-                has_maintenance4 = true;
-            }
-        }
-
-        vk::PhysicalDeviceSubgroupProperties subgroup_props;
-        vk::PhysicalDeviceProperties2 dev_props2;
-        vk::PhysicalDeviceMaintenance3Properties dev_props3;
-        vk::PhysicalDeviceMaintenance4Properties dev_props4;
-        dev_props2.pNext = &dev_props3;
-        dev_props3.pNext = &subgroup_props;
-        if (has_maintenance4) {
-            subgroup_props.pNext = &dev_props4;
-        }
-        physical_device.getProperties2(&dev_props2);
-
-        if (subgroup_props.subgroupSize < 32)
-            continue;
-
-        ggml_vk_device d;
-        d.index = i;
-        d.type = dev_props.deviceType;
-        d.heapSize = heapSize;
-        d.vendor = strdup(ggml_vk_getVendorName(dev_props.vendorID));
-        d.subgroupSize = subgroup_props.subgroupSize;
-        d.bufferAlignment = dev_props.limits.minStorageBufferOffsetAlignment;
-
-        if (has_maintenance4) {
-            d.maxAlloc = std::min(dev_props3.maxMemoryAllocationSize, dev_props4.maxBufferSize);
-        } else {
-            d.maxAlloc = dev_props3.maxMemoryAllocationSize;
-        }
-
-        std::string name(dev_props.deviceName);
-        size_t n_idx = ++count_by_name[name];
-        if (n_idx > 1) {
-            name += " (" + std::to_string(n_idx) + ")";
-        }
-        d.name = strdup(name.c_str());
-
-        results.push_back(d);
-    }
-
-    std::stable_sort(results.begin(), results.end(),
-        [](const ggml_vk_device& lhs, const ggml_vk_device& rhs) -> bool {
-            if (lhs.type != rhs.type) {
-                if (lhs.type == VK_PHYSICAL_DEVICE_TYPE_DISCRETE_GPU) return true;
-                if (rhs.type == VK_PHYSICAL_DEVICE_TYPE_DISCRETE_GPU) return false;
-
-                if (lhs.type == VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU) return true;
-                if (rhs.type == VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU) return false;
-            }
-            return lhs.heapSize < rhs.heapSize;
-        }
-    );
-
-    return results;
-}
-
-static std::vector<ggml_vk_device>& ggml_vk_available_devices() {
-    static std::vector<ggml_vk_device> devices = ggml_vk_available_devices_internal(0);
-    return devices;
-}
-
-static void ggml_vk_filterByVendor(std::vector<ggml_vk_device>& devices, const std::string& targetVendor) {
-    devices.erase(
-        std::remove_if(devices.begin(), devices.end(),
-            [&targetVendor](const ggml_vk_device& device) {
-                return device.vendor != targetVendor;
-            }),
-        devices.end()
-    );
-}
-
-static void ggml_vk_filterByName(std::vector<ggml_vk_device>& devices, const std::string& targetName) {
-    devices.erase(
-        std::remove_if(devices.begin(), devices.end(),
-            [&targetName](const ggml_vk_device& device) {
-                return device.name != targetName;
-            }),
-        devices.end()
-    );
-}
-
-static bool ggml_vk_get_device(ggml_vk_device * device, size_t memoryRequired, const std::string & name) {
-    if (name.empty())
-        return false;
-
-    auto devices = ggml_vk_available_devices_internal(memoryRequired);
-    if (name == "amd" || name == "nvidia" || name == "intel") {
-        ggml_vk_filterByVendor(devices, name);
-    } else if (name != "gpu") {
-        ggml_vk_filterByName(devices, name);
-    }
-
-    if (devices.empty())
-        return false;
-
-    *device = devices.front();
-    return true;
-}
-
-bool ggml_vk_get_device(ggml_vk_device * device, size_t memoryRequired, const char * name) {
-    return ggml_vk_get_device(device, memoryRequired, std::string(name));
-}
-
-bool ggml_vk_has_vulkan() {
-    return komputeManager()->hasVulkan();
-}
-
-bool ggml_vk_has_device() {
-    return komputeManager()->hasDevice();
-}
-
-ggml_vk_device ggml_vk_current_device() {
-    if (!komputeManager()->hasDevice())
-        return ggml_vk_device();
-
-    auto devices = ggml_vk_available_devices();
-    ggml_vk_filterByName(devices, komputeManager()->physicalDevice()->getProperties().deviceName.data());
-    GGML_ASSERT(!devices.empty());
-    return devices.front();
-}
-
-static
-void ggml_vk_allocate_descriptor_pool(struct ggml_kompute_context * ctx, size_t size) {
-    std::vector<vk::DescriptorPoolSize> descriptorPoolSizes = {
-        vk::DescriptorPoolSize(
-          vk::DescriptorType::eStorageBuffer,
-          3 * size // Descriptor count is number of possible tensors to pass into an algorithm
-          )
-    };
-
-    vk::DescriptorPoolCreateInfo descriptorPoolInfo(
-      vk::DescriptorPoolCreateFlags(),
-      size, // Max sets
-      static_cast<uint32_t>(descriptorPoolSizes.size()),
-      descriptorPoolSizes.data());
-
-    ctx->pool = std::make_shared<vk::DescriptorPool>();
-    vk::Result r = komputeManager()->device()->createDescriptorPool(
-      &descriptorPoolInfo, nullptr, ctx->pool.get());
-    if (r != vk::Result::eSuccess)
-        std::cerr << "Error allocating descriptor pool" << vk::to_string(r);
-}
-
-static
-void ggml_vk_free_descriptor_pool(struct ggml_kompute_context * ctx) {
-    if (ctx->pool) {
-        komputeManager()->device()->destroy(
-          *ctx->pool,
-          (vk::Optional<const vk::AllocationCallbacks>)nullptr);
-        ctx->pool = nullptr;
-    }
-}
-
-static
-vk::Buffer *ggml_vk_allocate_buffer(size_t size) {
-    vk::BufferCreateInfo bufferCreateInfo;
-    bufferCreateInfo.size = size;
-    bufferCreateInfo.usage = vk::BufferUsageFlagBits::eStorageBuffer |
-                             vk::BufferUsageFlagBits::eTransferSrc |
-                             vk::BufferUsageFlagBits::eTransferDst;
-    bufferCreateInfo.sharingMode = vk::SharingMode::eExclusive;
-
-    vk::Buffer *vkBuffer = new vk::Buffer;
-    vk::Result r = komputeManager()->device()->createBuffer(&bufferCreateInfo, nullptr, vkBuffer);
-    if (r != vk::Result::eSuccess)
-        std::cerr << "Error allocating buffer " << vk::to_string(r) << std::endl;
-    return vkBuffer;
-}
-
-static
-vk::DeviceMemory *ggml_vk_allocate(size_t size, vk::MemoryPropertyFlags flags, vk::MemoryRequirements requirements, bool *isHostVisible) {
-
-    uint32_t memoryTypeIndex = -1;
-    bool memoryTypeIndexFound = false;
-    vk::PhysicalDeviceMemoryProperties memoryProperties = komputeManager()->physicalDevice()->getMemoryProperties();
-    for (uint32_t i = 0; i < memoryProperties.memoryTypeCount; i++) {
-        const vk::MemoryType &memoryType = memoryProperties.memoryTypes[i];
-        const vk::MemoryHeap &memoryHeap = memoryProperties.memoryHeaps[memoryType.heapIndex];
-        if (memoryHeap.size < size) {
-            continue;
-        }
-
-        if (requirements.memoryTypeBits & (1 << i)) {
-            if (((memoryProperties.memoryTypes[i]).propertyFlags &
-                 flags) == flags) {
-                memoryTypeIndex = i;
-                memoryTypeIndexFound = true;
-                if (isHostVisible && (memoryProperties.memoryTypes[i].propertyFlags & vk::MemoryPropertyFlagBits::eHostVisible)) {
-                    *isHostVisible = true;
-                }
-                break;
-            }
-        }
-    }
-    if (!memoryTypeIndexFound) {
-        throw std::runtime_error(
-          "Memory type index for buffer creation not found");
-    }
-
-    vk::MemoryAllocateInfo allocInfo;
-    allocInfo.allocationSize = size;
-    allocInfo.memoryTypeIndex = memoryTypeIndex;
-    vk::DeviceMemory *vkDeviceMemory =  new vk::DeviceMemory;
-    vk::Result r = komputeManager()->device()->allocateMemory(&allocInfo, nullptr, vkDeviceMemory);
-    if (r != vk::Result::eSuccess) {
-        std::cerr << "Error allocating memory " << vk::to_string(r) << std::endl;
-        throw std::runtime_error("Error allocating vulkan memory.");
-    }
-    return vkDeviceMemory;
-}
-
-static size_t ggml_vk_aligned_offset(ggml_backend_buffer_t buffer, size_t offset) {
-    size_t minStorageBufferOffsetAlignment = ggml_backend_buffer_get_alignment(buffer);
-
-    // If offset is already aligned, return it directly
-    if (offset % minStorageBufferOffsetAlignment == 0) {
-        return offset;
-    }
-
-    // Otherwise, return the largest multiple of minStorageBufferOffsetAlignment less than offset
-    return (offset / minStorageBufferOffsetAlignment) * minStorageBufferOffsetAlignment;
-}
-
-static ggml_vk_memory ggml_vk_allocate(size_t size) {
-    ggml_vk_memory memory;
-    bool isHostVisible = false;
-    {
-        memory.primaryBuffer = ggml_vk_allocate_buffer(size);
-        vk::MemoryRequirements memoryRequirements = komputeManager()->device()->getBufferMemoryRequirements(*memory.primaryBuffer);
-        vk::MemoryPropertyFlags memoryPropertyFlags = vk::MemoryPropertyFlagBits::eDeviceLocal;
-        memory.primaryMemory = ggml_vk_allocate(size, memoryPropertyFlags, memoryRequirements, &isHostVisible);
-        komputeManager()->device()->bindBufferMemory(*memory.primaryBuffer, *memory.primaryMemory, 0);
-        if (isHostVisible) {
-            vk::Result r = komputeManager()->device()->mapMemory(*memory.primaryMemory, 0, size, vk::MemoryMapFlags(), &memory.data);
-            if (r != vk::Result::eSuccess)
-                std::cerr << "Error mapping memory" << vk::to_string(r);
-        }
-    }
-
-    if (!isHostVisible) {
-        memory.stagingBuffer = ggml_vk_allocate_buffer(size);
-        vk::MemoryRequirements memoryRequirements = komputeManager()->device()->getBufferMemoryRequirements(*memory.stagingBuffer);
-        vk::MemoryPropertyFlags memoryPropertyFlags = vk::MemoryPropertyFlagBits::eHostVisible |
-                                                      vk::MemoryPropertyFlagBits::eHostCoherent |
-                                                      vk::MemoryPropertyFlagBits::eHostCached;
-        memory.stagingMemory = ggml_vk_allocate(size, memoryPropertyFlags, memoryRequirements, &isHostVisible);
-        komputeManager()->device()->bindBufferMemory(*memory.stagingBuffer, *memory.stagingMemory, 0);
-        vk::Result r = komputeManager()->device()->mapMemory(*memory.stagingMemory, 0, size, vk::MemoryMapFlags(), &memory.data);
-        if (r != vk::Result::eSuccess)
-            std::cerr << "Error mapping memory" << vk::to_string(r);
-    }
-
-    memory.size = size;
-    return memory;
-}
-
-static void ggml_vk_free_memory(ggml_vk_memory &memory)
-{
-    komputeManager()->device()->destroy(
-      *memory.primaryBuffer,
-      (vk::Optional<const vk::AllocationCallbacks>)nullptr);
-    if (memory.stagingBuffer) {
-        komputeManager()->device()->destroy(
-          *memory.stagingBuffer,
-          (vk::Optional<const vk::AllocationCallbacks>)nullptr);
-    }
-    komputeManager()->device()->freeMemory(
-      *memory.primaryMemory,
-      (vk::Optional<const vk::AllocationCallbacks>)nullptr);
-    if (memory.stagingMemory) {
-        komputeManager()->device()->freeMemory(
-          *memory.stagingMemory,
-          (vk::Optional<const vk::AllocationCallbacks>)nullptr);
-    }
-}
-
-static const char * ggml_backend_kompute_buffer_type_get_name(ggml_backend_buffer_type_t buft);
-
-static
-ggml_vk_memory * ggml_vk_find_tensor(const struct ggml_tensor * t, uint64_t & offset) {
-    ggml_backend_buffer_t buffer = t->view_src ? t->view_src->buffer : t->buffer;
-
-    // compatibility with ggml-backend
-    GGML_ASSERT(buffer && buffer->buft->iface.get_name == ggml_backend_kompute_buffer_type_get_name);
-
-    ggml_vk_memory * buf_ctx = static_cast<ggml_vk_memory *>(buffer->context);
-
-    const intptr_t ioffs = intptr_t(t->data) - intptr_t(buf_ctx->data);
-
-    GGML_ASSERT(ioffs >= 0 && ioffs + int64_t(ggml_nbytes(t)) <= int64_t(buffer->size));
-
-    offset = uint64_t(ioffs);
-    return buf_ctx;
-}
-
-static
-const std::shared_ptr<kp::Tensor> ggml_vk_get_tensor(const struct ggml_tensor * t, uint32_t * alignedOffset = nullptr) {
-    uint64_t originalOffset = 0;
-    auto * res = ggml_vk_find_tensor(t, originalOffset);
-    if (!res) {
-        static std::shared_ptr<kp::Tensor> nullTensor = nullptr;
-        return nullTensor;
-    }
-
-    // Create a tensor whose memory will be composed of our buffers at the correct offset
-    const size_t nelements = ggml_nelements(t);
-    size_t nbytes = ggml_nbytes(t);
-
-    size_t vulkanOffset = ggml_vk_aligned_offset(t->buffer, originalOffset);
-    if (alignedOffset) {
-        *alignedOffset = originalOffset - vulkanOffset;
-        nbytes += *alignedOffset;
-    }
-
-    return komputeManager()->tensor(
-        t->data,
-        nelements,
-        nbytes, kp::Tensor::TensorDataTypes::eFloat,
-        res->primaryMemory, res->primaryBuffer,
-        res->stagingMemory, res->stagingBuffer,
-        vulkanOffset);
-}
-
-static std::vector<uint32_t> getSpirvShader(const unsigned char* rawData, size_t size) {
-    if (size % sizeof(uint32_t) != 0) {
-        throw std::runtime_error("Invalid size: must be divisible by sizeof(uint32_t)");
-    }
-
-    const uint32_t* data_ptr = reinterpret_cast<const uint32_t*>(rawData);
-    size_t count = size / sizeof(uint32_t);
-    return std::vector<uint32_t>(data_ptr, data_ptr + count);
-}
-
-inline static
-uint32_t safe_divide(uint32_t a, uint32_t b) {
-    if (b <= 1) {
-        return a;
-    }
-    if ((a % b) != 0) {
-        fprintf(stderr, "((%u %% %u) == %u) != 0\n", a, b, a % b);
-        GGML_ABORT("safe_divide result would've had remainder");
-    }
-    return a / b;
-}
-
-static void ggml_vk_add(
-    kp::Sequence& seq,
-    const std::shared_ptr<kp::Tensor>& inA,
-    const std::shared_ptr<kp::Tensor>& inB,
-    const std::shared_ptr<kp::Tensor>& out,
-    uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
-    int32_t ne00, int32_t ne01, int32_t ne02, int32_t ne03,
-    int32_t nb00, int32_t nb01, int32_t nb02, int32_t nb03,
-    int32_t ne10, int32_t ne11, int32_t ne12, int32_t ne13,
-    int32_t nb10, int32_t nb11, int32_t nb12, int32_t nb13,
-    int32_t ne0,
-    int32_t nb0,  int32_t nb1,  int32_t nb2,  int32_t nb3
-) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_add_comp_spv,
-        kp::shader_data::op_add_comp_spv_len);
-
-    struct PushConstants {
-        uint32_t inAOff, inBOff, outOff;
-        int32_t ne00;
-        int32_t nb00, nb01, nb02, nb03;
-        int32_t ne10, ne11, ne12, ne13;
-        int32_t nb10, nb11, nb12, nb13;
-        int32_t ne0;
-        int32_t nb0, nb1, nb2, nb3;
-    } const pushConsts {
-        safe_divide(inAOff, 4), safe_divide(inBOff, 4), safe_divide(outOff, 4),
-        ne00,
-        nb00, nb01, nb02, nb03,
-        ne10, ne11, ne12, ne13,
-        nb10, nb11, nb12, nb13,
-        ne0,
-        nb0, nb1, nb2, nb3
-    };
-
-    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
-    if (!komputeManager()->hasAlgorithm(__func__)) {
-        s_algo = komputeManager()->algorithm<float, PushConstants>(__func__, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {unsigned(ne01), unsigned(ne02), unsigned(ne03)}, {}, {pushConsts});
-    } else {
-        s_algo = komputeManager()->getAlgorithm(__func__);
-        s_algo->setTensors({inA, inB, out});
-        s_algo->setWorkgroup({unsigned(ne01), unsigned(ne02), unsigned(ne03)});
-        s_algo->setPushConstants<PushConstants>({pushConsts});
-        s_algo->updateDescriptors(s_kompute_context->pool.get());
-    }
-    seq.record<kp::OpAlgoDispatch>(s_algo);
-}
-
-static void ggml_vk_addrow(kp::Sequence& seq,
-                 const std::shared_ptr<kp::Tensor>& inA,
-                 const std::shared_ptr<kp::Tensor>& inB,
-                 const std::shared_ptr<kp::Tensor>& out,
-                 uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
-                 uint32_t size, uint32_t row = 0) {
-
-    const static auto spirv = getSpirvShader(kp::shader_data::op_addrow_comp_spv,
-        kp::shader_data::op_addrow_comp_spv_len);
-
-    struct PushConstants {
-        uint32_t inAOff, inBOff, outOff;
-        uint32_t row;
-    } const pushConsts {
-        safe_divide(inAOff, 4), safe_divide(inBOff, 4), safe_divide(outOff, 4),
-        row
-    };
-
-    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
-    if (!komputeManager()->hasAlgorithm(__func__))
-        s_algo = komputeManager()->algorithm<float, PushConstants>(__func__, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {size}, {}, {pushConsts});
-    else {
-        s_algo = komputeManager()->getAlgorithm(__func__);
-        s_algo->setTensors({inA, inB, out});
-        s_algo->setWorkgroup({size});
-        s_algo->setPushConstants<PushConstants>({pushConsts});
-        s_algo->updateDescriptors(s_kompute_context->pool.get());
-    }
-    seq.record<kp::OpAlgoDispatch>(s_algo);
-}
-
-static void ggml_vk_mul(
-    kp::Sequence& seq,
-    const std::shared_ptr<kp::Tensor>& inA,
-    const std::shared_ptr<kp::Tensor>& inB,
-    const std::shared_ptr<kp::Tensor>& out,
-    uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
-    int32_t ne00, int32_t ne01, int32_t ne02, int32_t ne03,
-    int32_t nb00, int32_t nb01, int32_t nb02, int32_t nb03,
-    int32_t ne10, int32_t ne11, int32_t ne12, int32_t ne13,
-    int32_t nb10, int32_t nb11, int32_t nb12, int32_t nb13,
-    int32_t ne0,
-    int32_t nb0,  int32_t nb1,  int32_t nb2,  int32_t nb3
-) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_mul_comp_spv,
-        kp::shader_data::op_mul_comp_spv_len);
-
-    struct PushConstants {
-        uint32_t inAOff, inBOff, outOff;
-        int32_t ne00;
-        int32_t nb00, nb01, nb02, nb03;
-        int32_t ne10, ne11, ne12, ne13;
-        int32_t nb10, nb11, nb12, nb13;
-        int32_t ne0;
-        int32_t nb0, nb1, nb2, nb3;
-    } const pushConsts {
-        safe_divide(inAOff, 4), safe_divide(inBOff, 4), safe_divide(outOff, 4),
-        ne00,
-        nb00, nb01, nb02, nb03,
-        ne10, ne11, ne12, ne13,
-        nb10, nb11, nb12, nb13,
-        ne0,
-        nb0, nb1, nb2, nb3
-    };
-
-    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
-    if (!komputeManager()->hasAlgorithm(__func__)) {
-        s_algo = komputeManager()->algorithm<float, PushConstants>(__func__, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {unsigned(ne01), unsigned(ne02), unsigned(ne03)}, {}, {pushConsts});
-    } else {
-        s_algo = komputeManager()->getAlgorithm(__func__);
-        s_algo->setTensors({inA, inB, out});
-        s_algo->setWorkgroup({unsigned(ne01), unsigned(ne02), unsigned(ne03)});
-        s_algo->setPushConstants<PushConstants>({pushConsts});
-        s_algo->updateDescriptors(s_kompute_context->pool.get());
-    }
-    seq.record<kp::OpAlgoDispatch>(s_algo);
-}
-
-static void ggml_vk_scale(kp::Sequence& seq,
-                   const std::shared_ptr<kp::Tensor>& in,
-                   const std::shared_ptr<kp::Tensor>& out,
-                   uint32_t inOff, uint32_t outOff,
-                   uint32_t size, float scale) {
-    const static auto spirv_1 = getSpirvShader(
-        kp::shader_data::op_scale_comp_spv, kp::shader_data::op_scale_comp_spv_len
-    );
-    const static auto spirv_8 = getSpirvShader(
-        kp::shader_data::op_scale_8_comp_spv, kp::shader_data::op_scale_8_comp_spv_len
-    );
-
-    struct PushConstants {
-        uint32_t inOff, outOff;
-        float scale;
-    } const pushConsts {
-        safe_divide(inOff, 4), safe_divide(outOff, 4),
-        scale
-    };
-
-    const auto * spirv = &spirv_1;
-    std::string name(__func__);
-    if (size % 8 == 0) {
-        size /= 8;
-        name += "_8";
-        spirv = &spirv_8;
-    }
-
-    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
-    if (!komputeManager()->hasAlgorithm(name)) {
-        s_algo = komputeManager()->algorithm<float, PushConstants>(name, s_kompute_context->pool.get(), {in, out}, *spirv, {size}, {}, {pushConsts});
-    } else {
-        s_algo = komputeManager()->getAlgorithm(name);
-        s_algo->setTensors({in, out});
-        s_algo->setWorkgroup({size});
-        s_algo->setPushConstants<PushConstants>({pushConsts});
-        s_algo->updateDescriptors(s_kompute_context->pool.get());
-    }
-    seq.record<kp::OpAlgoDispatch>(s_algo);
-}
-
-static void ggml_vk_xxlu(
-    const std::vector<uint32_t>& spirv, const char * suffix, kp::Sequence& seq,
-    const std::shared_ptr<kp::Tensor>& in,
-    const std::shared_ptr<kp::Tensor>& out,
-    uint32_t inOff, uint32_t outOff,
-    uint32_t size
-) {
-    struct PushConstants {
-        uint32_t inOff, outOff;
-    } const pushConsts {
-        safe_divide(inOff, 4), safe_divide(outOff, 4),
-    };
-
-    auto name = std::string(__func__) + "_" + suffix;
-    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
-    if (!komputeManager()->hasAlgorithm(name)) {
-        s_algo = komputeManager()->algorithm<float, PushConstants>(name, s_kompute_context->pool.get(), {in, out}, spirv, {size}, {}, {pushConsts});
-    } else {
-        s_algo = komputeManager()->getAlgorithm(name);
-        s_algo->setTensors({in, out});
-        s_algo->setWorkgroup({size});
-        s_algo->setPushConstants<PushConstants>({pushConsts});
-        s_algo->updateDescriptors(s_kompute_context->pool.get());
-    }
-    seq.record<kp::OpAlgoDispatch>(s_algo);
-}
-
-template <typename... Args>
-static void ggml_vk_silu(Args&&... args) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_silu_comp_spv,
-        kp::shader_data::op_silu_comp_spv_len);
-
-    ggml_vk_xxlu(spirv, "silu", std::forward<Args>(args)...);
-}
-
-template <typename... Args>
-static void ggml_vk_relu(Args&&... args) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_relu_comp_spv,
-        kp::shader_data::op_relu_comp_spv_len);
-
-    ggml_vk_xxlu(spirv, "relu", std::forward<Args>(args)...);
-}
-
-template <typename... Args>
-static void ggml_vk_gelu(Args&&... args) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_gelu_comp_spv,
-        kp::shader_data::op_gelu_comp_spv_len);
-
-    ggml_vk_xxlu(spirv, "gelu", std::forward<Args>(args)...);
-}
-
-static void ggml_vk_soft_max(
-    kp::Sequence& seq,
-    const std::shared_ptr<kp::Tensor>& inA,
-    const std::shared_ptr<kp::Tensor>& inB,
-    const std::shared_ptr<kp::Tensor>& out,
-    uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
-    int32_t ne00, int32_t ne01, int32_t ne02, uint32_t ne03,
-    float scale
-) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_softmax_comp_spv,
-        kp::shader_data::op_softmax_comp_spv_len);
-
-    struct PushConstants {
-        uint32_t inAOff, inBOff, outOff;
-        int32_t ne00, ne01, ne02;
-        float scale;
-        int32_t mask;
-    } pushConsts {
-        safe_divide(inAOff, 4), safe_divide(inBOff, 4), safe_divide(outOff, 4),
-        ne00, ne01, ne02,
-        scale,
-        bool(inB)
-    };
-
-    auto & inB_ = inB ? inB : inA;
-
-    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
-    if (!komputeManager()->hasAlgorithm(__func__)) {
-        // FIXME: The softmax kernel needs to be fixed to use the subgroupsize which can vary by device
-        const uint32_t local_x = 32;
-        s_algo = komputeManager()->algorithm<uint32_t, PushConstants>(__func__, s_kompute_context->pool.get(), {inA, inB_, out}, spirv, {unsigned(ne01), unsigned(ne02), unsigned(ne03)}, {local_x}, {pushConsts});
-    } else {
-        s_algo = komputeManager()->getAlgorithm(__func__);
-        s_algo->setTensors({inA, inB_, out});
-        s_algo->setWorkgroup({unsigned(ne01), unsigned(ne02), unsigned(ne03)});
-        s_algo->setPushConstants<PushConstants>({pushConsts});
-        s_algo->updateDescriptors(s_kompute_context->pool.get());
-    }
-    seq.record<kp::OpAlgoDispatch>(s_algo);
-}
-
-static void ggml_vk_norm_(
-    const std::vector<uint32_t>& spirv, const char * suffix, kp::Sequence& seq,
-    const std::shared_ptr<kp::Tensor>& in,
-    const std::shared_ptr<kp::Tensor>& out,
-    uint32_t inOff, uint32_t outOff,
-    int32_t ne00, int32_t nb01,
-    int32_t nrows, float epsilon
-) {
-    GGML_ASSERT(nb01%sizeof(float) == 0);
-    GGML_ASSERT(ne00%sizeof(float) == 0);
-
-    struct PushConstants {
-        uint32_t inOff, outOff;
-        uint32_t ne00, nb01;
-        float eps;
-    } pushConsts {
-        safe_divide(inOff, 4), safe_divide(outOff, 4),
-        (uint32_t)ne00, (uint32_t)nb01, epsilon
-    };
-
-    auto name = std::string(__func__) + "_" + suffix;
-    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
-    if (!komputeManager()->hasAlgorithm(name)) {
-        s_algo = komputeManager()->algorithm<float, PushConstants>(name, s_kompute_context->pool.get(), {in, out}, spirv, {(uint32_t)nrows}, {}, {pushConsts});
-    } else {
-        s_algo = komputeManager()->getAlgorithm(name);
-        s_algo->setTensors({in, out});
-        s_algo->setWorkgroup({(uint32_t)nrows});
-        s_algo->setPushConstants<PushConstants>({pushConsts});
-        s_algo->updateDescriptors(s_kompute_context->pool.get());
-    }
-    seq.record<kp::OpAlgoDispatch>(s_algo);
-}
-
-template <typename... Args>
-static void ggml_vk_norm(Args&&... args) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_norm_comp_spv,
-        kp::shader_data::op_norm_comp_spv_len);
-
-    ggml_vk_norm_(spirv, "norm", std::forward<Args>(args)...);
-}
-
-template <typename... Args>
-static void ggml_vk_rms_norm(Args&&... args) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_rmsnorm_comp_spv,
-        kp::shader_data::op_rmsnorm_comp_spv_len);
-
-    ggml_vk_norm_(spirv, "rms", std::forward<Args>(args)...);
-}
-
-static void ggml_vk_diag_mask_inf(kp::Sequence& seq,
-                           const std::shared_ptr<kp::Tensor>& in,
-                           const std::shared_ptr<kp::Tensor>& out,
-                           uint32_t inOff, uint32_t outOff,
-                           uint32_t n_past,
-                           int32_t ne00, int32_t ne01, int32_t ne02) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_diagmask_comp_spv,
-        kp::shader_data::op_diagmask_comp_spv_len);
-
-    struct PushConstants {
-        uint32_t inOff, outOff;
-        uint32_t n_past;
-        int32_t ne00, ne01;
-    } pushConsts {
-        safe_divide(inOff, 4), safe_divide(outOff, 4),
-        n_past,
-        ne00, ne01
-    };
-
-    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
-    if (!komputeManager()->hasAlgorithm(__func__))
-        s_algo = komputeManager()->algorithm<float, PushConstants>(__func__, s_kompute_context->pool.get(), {in, out}, spirv, {unsigned(ne00), unsigned(ne01), unsigned(ne02)}, {}, {pushConsts});
-    else {
-        s_algo = komputeManager()->getAlgorithm(__func__);
-        s_algo->setTensors({in, out});
-        s_algo->setWorkgroup({unsigned(ne00), unsigned(ne01), unsigned(ne02)});
-        s_algo->setPushConstants<PushConstants>({pushConsts});
-        s_algo->updateDescriptors(s_kompute_context->pool.get());
-    }
-    seq.record<kp::OpAlgoDispatch>(s_algo);
-}
-
-static void ggml_vk_mul_mat_f16(
-    kp::Sequence& seq,
-    const std::shared_ptr<kp::Tensor>& inA,
-    const std::shared_ptr<kp::Tensor>& inB,
-    const std::shared_ptr<kp::Tensor>& out,
-    uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
-    int32_t ne00, int32_t ne01, int32_t ne02,
-    uint32_t nb00, uint32_t nb01, uint32_t nb02,
-    int32_t ne10, int32_t ne11, int32_t ne12, int32_t ne13,
-    uint32_t nb10, uint32_t nb11, uint32_t nb12,
-    int32_t ne0, int32_t ne1,
-    uint32_t r2, uint32_t r3
-) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_mul_mat_f16_comp_spv,
-        kp::shader_data::op_mul_mat_f16_comp_spv_len);
-
-    struct PushConstants {
-        uint32_t inAOff, inBOff, outOff;
-        int32_t ne00, ne01, ne02;
-        uint32_t nb00, nb01, nb02;
-        int32_t ne10, ne11, ne12;
-        uint32_t nb10, nb11, nb12;
-        int32_t ne0, ne1;
-        uint32_t r2, r3;
-    } pushConsts {
-        safe_divide(inAOff, 2), safe_divide(inBOff, 4), safe_divide(outOff, 4),
-        ne00, ne01, ne02,
-        nb00, nb01, nb02,
-        ne10, ne11, ne12,
-        nb10, nb11, nb12,
-        ne0, ne1,
-        r2, r3
-    };
-
-    const unsigned ny = unsigned((ne11 + 4 - 1)/4);
-
-    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
-    if (!komputeManager()->hasAlgorithm(__func__)) {
-        const uint32_t local_x = ggml_vk_current_device().subgroupSize * 2;
-        s_algo = komputeManager()->algorithm<uint32_t, PushConstants>(__func__, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {unsigned(ne01), ny, unsigned(ne12*ne13)}, {local_x}, {pushConsts});
-    } else {
-        s_algo = komputeManager()->getAlgorithm(__func__);
-        s_algo->setTensors({inA, inB, out});
-        s_algo->setWorkgroup({unsigned(ne01), ny, unsigned(ne12*ne13)});
-        s_algo->setPushConstants<PushConstants>({pushConsts});
-        s_algo->updateDescriptors(s_kompute_context->pool.get());
-    }
-    seq.record<kp::OpAlgoDispatch>(s_algo);
-}
-
-static void ggml_vk_mul_mat_mat_f32(kp::Sequence& seq,
-                         const std::shared_ptr<kp::Tensor>& inA,
-                         const std::shared_ptr<kp::Tensor>& inB,
-                         const std::shared_ptr<kp::Tensor>& out,
-                         uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
-                         int32_t ne00, int32_t ne01, int32_t ne02,
-                         uint32_t nb01, uint32_t nb02,
-                         int32_t ne11, int32_t ne12,
-                         uint32_t nb11, uint32_t nb12,
-                         uint32_t nb1, uint32_t nb2) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_mul_mat_mat_f32_comp_spv,
-        kp::shader_data::op_mul_mat_mat_f32_comp_spv_len);
-
-    struct PushConstants {
-        uint32_t inAOff, inBOff, outOff;
-        int32_t ne00, ne01, ne02, ne11, ne12;
-        uint32_t nb01, nb02;
-        uint32_t nb11, nb12;
-        uint32_t nb1, nb2;
-    } pushConsts {
-        safe_divide(inAOff, 4), safe_divide(inBOff, 4), safe_divide(outOff, 4),
-        ne00, ne01, ne02, ne11, ne12,
-        nb01, nb02, nb11, nb12,
-        nb1, nb2
-    };
-
-    const uint32_t local_x = ggml_vk_current_device().subgroupSize;
-    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
-    if (!komputeManager()->hasAlgorithm(__func__)) {
-        s_algo = komputeManager()->algorithm<uint32_t, PushConstants>(__func__, s_kompute_context->pool.get(),
-        {inA, inB, out}, spirv,
-        {unsigned(ne01),
-         unsigned(ne11),
-         unsigned(std::max(ne12, ne02))
-         },
-        {local_x},
-        {pushConsts});
-    } else {
-        s_algo = komputeManager()->getAlgorithm(__func__);
-        s_algo->setTensors({inA, inB, out});
-        s_algo->setWorkgroup({unsigned(ne01),
-                              unsigned(ne11),
-                              unsigned(std::max(ne12, ne02)),
-                              });
-        s_algo->setPushConstants<PushConstants>({pushConsts});
-        s_algo->updateDescriptors(s_kompute_context->pool.get());
-    }
-    seq.record<kp::OpAlgoDispatch>(s_algo);
-}
-
-static void ggml_vk_mul_mat_impl(
-    const std::vector<uint32_t>& spirv, const char * suffix, uint32_t block_size, kp::Sequence& seq,
-    const std::shared_ptr<kp::Tensor>& inA,
-    const std::shared_ptr<kp::Tensor>& inB,
-    const std::shared_ptr<kp::Tensor>& out,
-    uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
-    int32_t ne00, int32_t ne01, int32_t ne02,
-    int32_t ne10, int32_t ne11, int32_t ne12, int32_t ne13,
-    int32_t ne0, int32_t ne1,
-    uint32_t r2, uint32_t r3
-) {
-    struct PushConstants {
-        uint32_t inAOff, inBOff, outOff;
-        int32_t ne00, ne01, ne02;
-        int32_t ne10, ne12;
-        int32_t ne0, ne1;
-        uint32_t r2, r3;
-    } pushConsts {
-        safe_divide(inAOff, block_size), safe_divide(inBOff, 4), safe_divide(outOff, 4),
-        ne00, ne01, ne02,
-        ne10, ne12,
-        ne0, ne1,
-        r2, r3
-    };
-
-    auto name = std::string(__func__) + "_" + suffix;
-    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
-    if (!komputeManager()->hasAlgorithm(name)) {
-        const uint32_t local_x = ggml_vk_current_device().subgroupSize * 2;
-        s_algo = komputeManager()->algorithm<uint32_t, PushConstants>(name, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {unsigned((ne01 + 7)/8), unsigned(ne11), unsigned(ne12*ne13)}, {local_x}, {pushConsts});
-    } else {
-        s_algo = komputeManager()->getAlgorithm(name);
-        s_algo->setTensors({inA, inB, out});
-        s_algo->setWorkgroup({unsigned((ne01 + 7)/8), unsigned(ne11), unsigned(ne12*ne13)});
-        s_algo->setPushConstants<PushConstants>({pushConsts});
-        s_algo->updateDescriptors(s_kompute_context->pool.get());
-    }
-    seq.record<kp::OpAlgoDispatch>(s_algo);
-}
-
-template <typename... Args>
-static void ggml_vk_mul_mat_q4_0(Args&&... args) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_mul_mat_q4_0_comp_spv,
-        kp::shader_data::op_mul_mat_q4_0_comp_spv_len);
-
-    ggml_vk_mul_mat_impl(spirv, "q4_0", 1/*We access blocks unaligned*/, std::forward<Args>(args)...);
-}
-
-template <typename... Args>
-static void ggml_vk_mul_mat_q4_1(Args&&... args) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_mul_mat_q4_1_comp_spv,
-        kp::shader_data::op_mul_mat_q4_1_comp_spv_len);
-
-    ggml_vk_mul_mat_impl(spirv, "q4_1", 1/*We access blocks unaligned*/, std::forward<Args>(args)...);
-}
-
-template <typename... Args>
-static void ggml_vk_mul_mat_q8_0(Args&&... args) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_mul_mat_q8_0_comp_spv,
-        kp::shader_data::op_mul_mat_q8_0_comp_spv_len);
-
-    ggml_vk_mul_mat_impl(spirv, "q8_0", 1/*We access blocks unaligned*/, std::forward<Args>(args)...);
-}
-
-static void ggml_vk_mul_mat_q4_k(
-    kp::Sequence& seq,
-    const std::shared_ptr<kp::Tensor>& inA,
-    const std::shared_ptr<kp::Tensor>& inB,
-    const std::shared_ptr<kp::Tensor>& out,
-    uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
-    int32_t ne00, int32_t ne01, int32_t ne02, int32_t ne10,
-    int32_t ne11, int32_t ne12, int32_t ne13, int32_t ne0,
-    int32_t ne1, int32_t r2, int32_t r3
-) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_mul_mat_q4_k_comp_spv,
-        kp::shader_data::op_mul_mat_q4_k_comp_spv_len);
-
-    struct PushConstants {
-        uint32_t inAOff, inBOff, outOff;
-        int32_t ne00, ne10, ne0, ne1, ne01, ne02, ne12, r2, r3;
-    } pushConsts {
-        0, 0, 0,
-        ne00, ne10, ne0, ne1, ne01, ne02, ne12, r2, r3
-    };
-
-    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
-    if (!komputeManager()->hasAlgorithm(__func__)) {
-        s_algo = komputeManager()->algorithm<uint32_t, PushConstants>(__func__, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {unsigned((ne01 + 3)/4), unsigned(ne11), unsigned(ne12) * unsigned(ne13)}, {}, {pushConsts});
-    } else {
-        s_algo = komputeManager()->getAlgorithm(__func__);
-        s_algo->setTensors({inA, inB, out});
-        s_algo->setWorkgroup({unsigned((ne01 + 3)/4), unsigned(ne11), unsigned(ne12) * unsigned(ne13)});
-        s_algo->setPushConstants<PushConstants>({pushConsts});
-        s_algo->updateDescriptors(s_kompute_context->pool.get());
-    }
-    seq.record<kp::OpAlgoDispatch>(s_algo);
-}
-
-static void ggml_vk_mul_mat_q6_k(
-    kp::Sequence& seq,
-    const std::shared_ptr<kp::Tensor>& inA,
-    const std::shared_ptr<kp::Tensor>& inB,
-    const std::shared_ptr<kp::Tensor>& out,
-    uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
-    int32_t ne00, int32_t ne10, int32_t ne0, int32_t ne1,
-    int32_t ne01, int32_t ne11, int32_t ne12, int32_t ne02
-) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_mul_mat_q6_k_comp_spv,
-        kp::shader_data::op_mul_mat_q6_k_comp_spv_len);
-
-    struct PushConstants {
-        uint32_t inAOff, inBOff, outOff;
-        int32_t ne00, ne10, ne0, ne1, ne01, gqa;
-    } pushConsts {
-        inAOff, safe_divide(inBOff, 4), safe_divide(outOff, 4),
-        ne00, ne10, ne0, ne1, ne01, ne12/ne02
-    };
-
-    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
-    if (!komputeManager()->hasAlgorithm(__func__)) {
-        const uint32_t local_x = ggml_vk_current_device().subgroupSize * 2;
-        s_algo = komputeManager()->algorithm<uint32_t, PushConstants>(__func__, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {unsigned((ne01 + 1)/2), unsigned(ne11), unsigned(ne12)}, {local_x}, {pushConsts});
-    } else {
-        s_algo = komputeManager()->getAlgorithm(__func__);
-        s_algo->setTensors({inA, inB, out});
-        s_algo->setWorkgroup({unsigned((ne01 + 1)/2), unsigned(ne11), unsigned(ne12)});
-        s_algo->setPushConstants<PushConstants>({pushConsts});
-        s_algo->updateDescriptors(s_kompute_context->pool.get());
-    }
-    seq.record<kp::OpAlgoDispatch>(s_algo);
-}
-
-static void ggml_vk_get_rows(
-    const std::vector<uint32_t>& spirv,
-    const char * suffix,
-    unsigned element_size, unsigned qk,
-    kp::Sequence& seq,
-    const std::shared_ptr<kp::Tensor>& inA,
-    const std::shared_ptr<kp::Tensor>& inB,
-    const std::shared_ptr<kp::Tensor>& out,
-    uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
-    int32_t ne00, int32_t nb01, int32_t nb1,
-    uint32_t size
-) {
-    GGML_ASSERT(nb01%element_size == 0);
-    GGML_ASSERT(nb1%sizeof(float) == 0);
-    if (qk) GGML_ASSERT(ne00%qk == 0);
-
-    struct PushConstants {
-        uint32_t inAOff, inBOff, outOff;
-        int32_t ne00, nb01, nb1;
-    } pushConsts {
-        safe_divide(inAOff, element_size), safe_divide(inBOff, 4), safe_divide(outOff, 4),
-        ne00, nb01, nb1
-    };
-
-    auto name = std::string(__func__) + "_" + suffix;
-    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
-    if (!komputeManager()->hasAlgorithm(name)) {
-        s_algo = komputeManager()->algorithm<float, PushConstants>(name, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {size}, {}, {pushConsts});
-    } else {
-        s_algo = komputeManager()->getAlgorithm(name);
-        s_algo->setTensors({inA, inB, out});
-        s_algo->setWorkgroup({size});
-        s_algo->setPushConstants<PushConstants>({pushConsts});
-        s_algo->updateDescriptors(s_kompute_context->pool.get());
-    }
-    seq.record<kp::OpAlgoDispatch>(s_algo);
-}
-
-template <typename... Args>
-static void ggml_vk_get_rows_f32(Args&&... args) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_getrows_f32_comp_spv,
-        kp::shader_data::op_getrows_f32_comp_spv_len);
-
-    ggml_vk_get_rows(spirv, "f32", sizeof(float), 0, std::forward<Args>(args)...);
-}
-
-template <typename... Args>
-static void ggml_vk_get_rows_f16(Args&&... args) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_getrows_f16_comp_spv,
-        kp::shader_data::op_getrows_f16_comp_spv_len);
-
-    ggml_vk_get_rows(spirv, "f16", sizeof(half), 0, std::forward<Args>(args)...);
-}
-
-template <typename... Args>
-static void ggml_vk_get_rows_q4_0(Args&&... args) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_getrows_q4_0_comp_spv,
-        kp::shader_data::op_getrows_q4_0_comp_spv_len);
-
-    ggml_vk_get_rows(spirv, "q4_0", 1/*We access blocks unaligned*/, QK4_0, std::forward<Args>(args)...);
-}
-
-template <typename... Args>
-static void ggml_vk_get_rows_q4_1(Args&&... args) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_getrows_q4_1_comp_spv,
-        kp::shader_data::op_getrows_q4_1_comp_spv_len);
-
-    ggml_vk_get_rows(spirv, "q4_1", 1/*We access blocks unaligned*/, QK4_1, std::forward<Args>(args)...);
-}
-
-template <typename... Args>
-static void ggml_vk_get_rows_q6_k(Args&&... args) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_getrows_q6_k_comp_spv,
-        kp::shader_data::op_getrows_q6_k_comp_spv_len);
-    ggml_vk_get_rows(spirv, "q6_k", 1/*We access blocks unaligned*/, QK_NL, std::forward<Args>(args)...);
-}
-
-static void ggml_vk_rope(
-    kp::Sequence& seq,
-    const std::shared_ptr<kp::Tensor>& inA,
-    const std::shared_ptr<kp::Tensor>& inB,
-    const std::shared_ptr<kp::Tensor>& out,
-    uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
-    ggml_type src0t, int32_t n_dims, int32_t mode, int32_t n_ctx_orig,
-    float freq_base, float freq_scale, float ext_factor, float attn_factor, float beta_fast, float beta_slow,
-    int32_t ne01, int32_t ne02, int32_t ne03,
-    uint32_t nb00, uint32_t nb01, uint32_t nb02, uint32_t nb03,
-    int32_t ne0,
-    uint32_t nb0, uint32_t nb1, uint32_t nb2, uint32_t nb3
-) {
-    GGML_ASSERT(src0t == GGML_TYPE_F16 || src0t == GGML_TYPE_F32);
-
-    static const auto spirv_f16 = getSpirvShader(
-        kp::shader_data::op_rope_f16_comp_spv, kp::shader_data::op_rope_f16_comp_spv_len
-    );
-    static const auto spirv_f32 = getSpirvShader(
-        kp::shader_data::op_rope_f32_comp_spv, kp::shader_data::op_rope_f32_comp_spv_len
-    );
-
-    int type_size = src0t == GGML_TYPE_F16 ? 2 : 4;
-
-    GGML_ASSERT(nb03 % type_size == 0);
-    GGML_ASSERT(nb02 % type_size == 0);
-    GGML_ASSERT(nb01 % type_size == 0);
-    GGML_ASSERT(nb00 % type_size == 0);
-    GGML_ASSERT(nb3  % type_size == 0);
-    GGML_ASSERT(nb2  % type_size == 0);
-    GGML_ASSERT(nb1  % type_size == 0);
-    GGML_ASSERT(nb0  % type_size == 0);
-
-    struct PushConstants {
-        uint32_t inAOff, inBOff, outOff;
-        int32_t n_dims, mode, n_ctx_orig;
-        float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
-        uint32_t nb00, nb01, nb02, nb03;
-        int32_t ne0;
-        uint32_t nb0, nb1, nb2, nb3;
-    } pushConsts {
-        safe_divide(inAOff, type_size), safe_divide(inBOff, 4), safe_divide(outOff, type_size),
-        n_dims, mode, n_ctx_orig,
-        freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow,
-        nb00, nb01, nb02, nb03,
-        ne0,
-        nb0, nb1, nb2, nb3
-    };
-
-    auto name = std::string(__func__) + (src0t == GGML_TYPE_F16 ? "_f16" : "_f32");
-    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
-    if (!komputeManager()->hasAlgorithm(name)) {
-        s_algo = komputeManager()->algorithm<float, PushConstants>(
-            name, s_kompute_context->pool.get(), {inA, inB, out},
-            src0t == GGML_TYPE_F16 ? spirv_f16 : spirv_f32,
-            {unsigned(ne01), unsigned(ne02), unsigned(ne03)}, {}, {pushConsts}
-        );
-    } else {
-        s_algo = komputeManager()->getAlgorithm(name);
-        s_algo->setTensors({inA, inB, out});
-        s_algo->setWorkgroup({unsigned(ne01), unsigned(ne02), unsigned(ne03)});
-        s_algo->setPushConstants<PushConstants>({pushConsts});
-        s_algo->updateDescriptors(s_kompute_context->pool.get());
-    }
-    seq.record<kp::OpAlgoDispatch>(s_algo);
-}
-
-static void ggml_vk_cpy(
-    const std::vector<uint32_t>& spirv,
-    uint32_t in_element_size, uint32_t out_element_size,
-    kp::Sequence& seq,
-    const std::shared_ptr<kp::Tensor>& in,
-    const std::shared_ptr<kp::Tensor>& out,
-    uint32_t inOff, uint32_t outOff,
-    int32_t ne00, int32_t ne01, int32_t ne02, int32_t ne03,
-    uint32_t nb00, uint32_t nb01, uint32_t nb02, uint32_t nb03,
-    int32_t ne0, int32_t ne1, int32_t ne2,
-    uint32_t nb0, uint32_t nb1, uint32_t nb2, uint32_t nb3
-) {
-    struct PushConstants {
-        uint32_t inOff, outOff;
-        int32_t ne00, ne01, ne02;
-        uint32_t nb00, nb01, nb02, nb03;
-        int32_t ne0, ne1, ne2;
-        uint32_t nb0, nb1, nb2, nb3;
-    } pushConsts {
-        safe_divide(inOff, in_element_size), safe_divide(outOff, out_element_size),
-        ne00, ne01, ne02,
-        nb00, nb01, nb02, nb03,
-        ne0, ne1, ne2,
-        nb0, nb1, nb2, nb3
-    };
-
-    std::string name = std::string(__func__)
-                       + "_i_" + std::to_string(in_element_size)
-                       + "_o_" + std::to_string(out_element_size);
-    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
-    if (!komputeManager()->hasAlgorithm(name))
-        s_algo = komputeManager()->algorithm<float, PushConstants>(name, s_kompute_context->pool.get(), {in, out}, spirv, {unsigned(ne01), unsigned(ne02), unsigned(ne03)}, {}, {pushConsts});
-    else {
-        s_algo = komputeManager()->getAlgorithm(name);
-        s_algo->setTensors({in, out});
-        s_algo->setWorkgroup({unsigned(ne01), unsigned(ne02), unsigned(ne03)});
-        s_algo->setPushConstants<PushConstants>({pushConsts});
-        s_algo->updateDescriptors(s_kompute_context->pool.get());
-    }
-    seq.record<kp::OpAlgoDispatch>(s_algo);
-}
-
-template <typename... Args>
-static void ggml_vk_cpy_f32_f16(Args&&... args) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_cpy_f32_f16_comp_spv,
-        kp::shader_data::op_cpy_f32_f16_comp_spv_len);
-    ggml_vk_cpy(spirv, 4, 2, std::forward<Args>(args)...);
-}
-
-template <typename... Args>
-static void ggml_vk_cpy_f32_f32(Args&&... args) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_cpy_f32_f32_comp_spv,
-        kp::shader_data::op_cpy_f32_f32_comp_spv_len);
-    ggml_vk_cpy(spirv, 4, 4, std::forward<Args>(args)...);
-}
-
-template <typename... Args>
-static void ggml_vk_cpy_f16_f16(Args&&... args) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_cpy_f16_f16_comp_spv,
-        kp::shader_data::op_cpy_f16_f16_comp_spv_len);
-    ggml_vk_cpy(spirv, 2, 2, std::forward<Args>(args)...);
-}
-
-template <typename... Args>
-static void ggml_vk_cpy_f16_f32(Args&&... args) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_cpy_f16_f32_comp_spv,
-        kp::shader_data::op_cpy_f16_f32_comp_spv_len);
-    ggml_vk_cpy(spirv, 2, 4, std::forward<Args>(args)...);
-}
-
-static bool ggml_backend_kompute_device_supports_op(ggml_backend_dev_t dev, const struct ggml_tensor * op) {
-    switch (op->op) {
-        case GGML_OP_UNARY:
-            switch (ggml_get_unary_op(op)) {
-                case GGML_UNARY_OP_RELU:
-                case GGML_UNARY_OP_GELU:
-                case GGML_UNARY_OP_SILU:
-                    return ggml_is_contiguous(op->src[0]);
-                default:
-                    ;
-            }
-            break;
-        case GGML_OP_NONE:
-        case GGML_OP_RESHAPE:
-        case GGML_OP_VIEW:
-        case GGML_OP_TRANSPOSE:
-        case GGML_OP_PERMUTE:
-        case GGML_OP_ADD:
-        case GGML_OP_MUL:
-        case GGML_OP_SCALE:
-        case GGML_OP_SOFT_MAX:
-        case GGML_OP_RMS_NORM:
-        case GGML_OP_NORM:
-        case GGML_OP_ROPE:
-            return true;
-        case GGML_OP_DUP:
-        case GGML_OP_CPY:
-        case GGML_OP_CONT:
-            switch (op->src[0]->type) {
-                case GGML_TYPE_F32:
-                case GGML_TYPE_F16:
-                    break;
-                default:
-                    return false;
-            }
-            switch (op->type) {
-                case GGML_TYPE_F32:
-                case GGML_TYPE_F16:
-                    break;
-                default:
-                    return false;
-            }
-            return true;
-        case GGML_OP_DIAG_MASK_INF:
-            return op->ne[3] == 1;
-        case GGML_OP_GET_ROWS:
-            switch (op->src[0]->type) {
-                case GGML_TYPE_F32:
-                case GGML_TYPE_F16:
-                case GGML_TYPE_Q4_0:
-                case GGML_TYPE_Q4_1:
-                case GGML_TYPE_Q6_K:
-                    return op->ne[2] == 1 && op->ne[3] == 1;
-                default:
-                    ;
-            }
-            return false;
-        case GGML_OP_MUL_MAT:
-            if (op->src[1]->type != GGML_TYPE_F32 || ggml_is_transposed(op->src[0]) || ggml_is_transposed(op->src[1]))
-                return false;
-
-            switch (op->src[0]->type) {
-                case GGML_TYPE_F32:
-                case GGML_TYPE_Q6_K:
-                    return op->ne[3] == 1;
-                case GGML_TYPE_F16:
-                case GGML_TYPE_Q8_0:
-                case GGML_TYPE_Q4_0:
-                case GGML_TYPE_Q4_1:
-                case GGML_TYPE_Q4_K:
-                    return true;
-                default:
-                    ;
-            }
-        default:
-            ;
-    }
-    return false;
-
-    GGML_UNUSED(dev);
-}
-
-static void ggml_vk_graph_compute(struct ggml_kompute_context * ctx, struct ggml_cgraph * gf) {
-    const int n_seq = 8;
-
-    // FIXME: Figure out if we can somehow optimize the size of the pool... right now we're setting
-    // it to the size of the graph, but I think it can be made smaller?
-    ggml_vk_allocate_descriptor_pool(ctx, gf->n_nodes);
-
-    std::vector<std::shared_ptr<kp::Sequence>> sequences(n_seq);
-
-    for (auto& sequence : sequences) {
-        sequence = komputeManager()->sequence();
-    }
-    for (int seq_idx = 0; seq_idx < n_seq; ++seq_idx) {
-        const int n_nodes_per_seq = (gf->n_nodes + n_seq - 1) / n_seq;
-
-        auto& seq = *sequences[seq_idx];
-
-        const int node_start = (seq_idx + 0) * n_nodes_per_seq;
-        const int node_end   = std::min((seq_idx == n_seq - 1) ? gf->n_nodes : (seq_idx + 1) * n_nodes_per_seq, gf->n_nodes);
-
-        bool any_commands_recorded = false;
-
-        for (int i = node_start; i < node_end; ++i) {
-            struct ggml_tensor * src0 = gf->nodes[i]->src[0];
-            struct ggml_tensor * src1 = gf->nodes[i]->src[1];
-            struct ggml_tensor * src2 = gf->nodes[i]->src[2]; GGML_UNUSED(src2);
-            struct ggml_tensor * dst = gf->nodes[i];
-            GGML_ASSERT(dst->data != nullptr);
-
-            if (ggml_is_empty(dst)) {
-                continue;
-            }
-
-            switch (dst->op) {
-                case GGML_OP_NONE:
-                case GGML_OP_RESHAPE:
-                case GGML_OP_VIEW:
-                case GGML_OP_TRANSPOSE:
-                case GGML_OP_PERMUTE:
-                    continue; // noop -> next node
-                default:
-                    break;
-            }
-
-            any_commands_recorded = true;
-
-            const int32_t ne00 = src0 ? src0->ne[0] : 0;
-            const int32_t ne01 = src0 ? src0->ne[1] : 0;
-            const int32_t ne02 = src0 ? src0->ne[2] : 0;
-            const int32_t ne03 = src0 ? src0->ne[3] : 0;
-
-            const uint32_t nb00 = src0 ? src0->nb[0] : 0;
-            const uint32_t nb01 = src0 ? src0->nb[1] : 0;
-            const uint32_t nb02 = src0 ? src0->nb[2] : 0;
-            const uint32_t nb03 = src0 ? src0->nb[3] : 0;
-
-            const int32_t ne10 = src1 ? src1->ne[0] : 0;
-            const int32_t ne11 = src1 ? src1->ne[1] : 0;
-            const int32_t ne12 = src1 ? src1->ne[2] : 0;
-            const int32_t ne13 = src1 ? src1->ne[3] : 0;
-
-            const uint32_t nb10 = src1 ? src1->nb[0] : 0;
-            const uint32_t nb11 = src1 ? src1->nb[1] : 0;
-            const uint32_t nb12 = src1 ? src1->nb[2] : 0;
-            const uint32_t nb13 = src1 ? src1->nb[3] : 0;
-
-            const int32_t ne0 = dst ? dst->ne[0] : 0;
-            const int32_t ne1 = dst ? dst->ne[1] : 0;
-            const int32_t ne2 = dst ? dst->ne[2] : 0;
-//            const int32_t ne3 = dst ? dst->ne[3] : 0;
-
-            const uint32_t nb0 = dst ? dst->nb[0] : 0;
-            const uint32_t nb1 = dst ? dst->nb[1] : 0;
-            const uint32_t nb2 = dst ? dst->nb[2] : 0;
-            const uint32_t nb3 = dst ? dst->nb[3] : 0;
-
-            const enum ggml_type src0t = src0 ? src0->type : GGML_TYPE_COUNT;
-            const enum ggml_type src1t = src1 ? src1->type : GGML_TYPE_COUNT;
-            const enum ggml_type dstt = dst ? dst->type : GGML_TYPE_COUNT;
-
-            const static std::shared_ptr<kp::Tensor> nullTensor = nullptr;
-            uint32_t off_src0 = 0;
-            uint32_t off_src1 = 0;
-            uint32_t off_dst  = 0;
-            const std::shared_ptr<kp::Tensor>& id_src0 = src0 ? ggml_vk_get_tensor(src0, &off_src0) : nullTensor;
-            const std::shared_ptr<kp::Tensor>& id_src1 = src1 ? ggml_vk_get_tensor(src1, &off_src1) : nullTensor;
-            const std::shared_ptr<kp::Tensor>& id_dst  = dst  ? ggml_vk_get_tensor(dst,  &off_dst)  : nullTensor;
-
-            switch (dst->op) {
-                case GGML_OP_ADD:
-                    {
-                        if (ggml_nelements(src1) == ne10 && ggml_is_contiguous(src1) && ne00 % 4 == 0 && ne10 % 4 == 0) {
-                            // src1 is a row
-                            ggml_vk_addrow(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ggml_nelements(dst)/4, ne00);
-                        } else {
-                            ggml_vk_add(
-                                seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,
-                                ne00, ne01, ne02, ne03,
-                                nb00, nb01, nb02, nb03,
-                                ne10, ne11, ne12, ne13,
-                                nb10, nb11, nb12, nb13,
-                                ne0,
-                                nb0, nb1, nb2, nb3
-                            );
-                        }
-                    } break;
-                case GGML_OP_MUL:
-                    {
-                        ggml_vk_mul(
-                            seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,
-                            ne00, ne01, ne02, ne03,
-                            nb00, nb01, nb02, nb03,
-                            ne10, ne11, ne12, ne13,
-                            nb10, nb11, nb12, nb13,
-                            ne0,
-                            nb0, nb1, nb2, nb3
-                        );
-                    } break;
-                case GGML_OP_SCALE:
-                    {
-                        float scale; memcpy(&scale, dst->op_params, sizeof(float));
-
-                        ggml_vk_scale(seq, id_src0, id_dst, off_src0, off_dst, ggml_nelements(dst), scale);
-                    } break;
-                case GGML_OP_UNARY:
-                    {
-                        int64_t n = ggml_nelements(dst);
-                        GGML_ASSERT(n % 4 == 0);
-                        switch (ggml_get_unary_op(gf->nodes[i])) {
-                            case GGML_UNARY_OP_SILU:
-                                {
-                                    ggml_vk_silu(seq, id_src0, id_dst, off_src0, off_dst, n/4);
-                                } break;
-                            case GGML_UNARY_OP_RELU:
-                                {
-                                    ggml_vk_relu(seq, id_src0, id_dst, off_src0, off_dst, n/4);
-                                } break;
-                            case GGML_UNARY_OP_GELU:
-                                {
-                                    GGML_ASSERT(n % 8 == 0);
-                                    ggml_vk_gelu(seq, id_src0, id_dst, off_src0, off_dst, n/8);
-                                } break;
-                            default:
-                                {
-                                    fprintf(stderr, "%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
-                                    GGML_ABORT("fatal error");
-                                }
-                        }
-                    } break;
-                case GGML_OP_SOFT_MAX:
-                    {
-                        float scale;
-                        float max_bias;
-
-                        memcpy(&scale,    (float *)dst->op_params + 0, sizeof(float));
-                        memcpy(&max_bias, (float *)dst->op_params + 1, sizeof(float));
-
-#pragma message("TODO: add ggml_vk_soft_max() F16 src1 support")
-#pragma message("ref:  https://github.com/ggerganov/llama.cpp/pull/5021")
-                        GGML_ASSERT(!src1 || src1t == GGML_TYPE_F32);
-
-#pragma message("TODO: add ALiBi support")
-#pragma message("ref:  https://github.com/ggerganov/llama.cpp/pull/7192")
-                        GGML_ASSERT(max_bias == 0.0f);
-
-                        ggml_vk_soft_max(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, ne01, ne02, ne03, scale);
-                    } break;
-                case GGML_OP_DIAG_MASK_INF:
-                    {
-                        const int n_past = ((int32_t *)(dst->op_params))[0];
-                        ggml_vk_diag_mask_inf(seq, id_src0, id_dst, off_src0, off_dst, n_past, ne00, ne01, ne02);
-                    } break;
-                case GGML_OP_NORM:
-                    {
-                        float eps;
-                        memcpy(&eps, dst->op_params, sizeof(float));
-                        ggml_vk_norm(seq, id_src0, id_dst, off_src0, off_dst, ne00, nb01, ggml_nrows(src0), eps);
-                    } break;
-                case GGML_OP_RMS_NORM:
-                    {
-                        GGML_ASSERT(ne00 % 4 == 0);
-
-                        float eps;
-                        memcpy(&eps, dst->op_params, sizeof(float));
-                        ggml_vk_rms_norm(seq, id_src0, id_dst, off_src0, off_dst, ne00, nb01, ggml_nrows(src0), eps);
-                    } break;
-                case GGML_OP_MUL_MAT:
-                    {
-                        GGML_ASSERT(ne00 == ne10);
-
-                        GGML_ASSERT(ne12 % ne02 == 0);
-                        GGML_ASSERT(ne13 % ne03 == 0);
-
-                        const uint32_t r2 = ne12/ne02;
-                        const uint32_t r3 = ne13/ne03;
-
-                        if (src1t != GGML_TYPE_F32) {
-                            fprintf(stderr, "%s: %s: Unsupported src1 type: %u/%u\n", __func__, ggml_op_name(dst->op), src0t, src1t);
-                            goto not_implemented;
-                        }
-
-                        if (ggml_is_transposed(src0) ||
-                            ggml_is_transposed(src1)) {
-                            fprintf(stderr, "%s: %s: matmul on tranposed tensor not supported: %u/%u\n", __func__, ggml_op_name(dst->op), src0t, src1t);
-                            goto not_implemented;
-                        }
-
-                        switch (src0t) {
-                            case GGML_TYPE_F32:
-                                ggml_vk_mul_mat_mat_f32(
-                                    seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,
-                                    ne00, ne01, ne02, nb01, nb02, ne11, ne12, nb11, nb12, nb1, nb2
-                                );
-                                break;
-                            case GGML_TYPE_F16:
-                                ggml_vk_mul_mat_f16(
-                                    seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,
-                                    ne00, ne01, ne02, nb00, nb01, nb02, ne10, ne11, ne12, ne13, nb10, nb11, nb12,
-                                    ne0, ne1, r2, r3
-                                );
-                                break;
-                            case GGML_TYPE_Q8_0:
-                                ggml_vk_mul_mat_q8_0(
-                                    seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,
-                                    ne00, ne01, ne02, ne10, ne11, ne12, ne13, ne0, ne1, r2, r3
-                                );
-                                break;
-                            case GGML_TYPE_Q4_0:
-                                ggml_vk_mul_mat_q4_0(
-                                    seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,
-                                    ne00, ne01, ne02, ne10, ne11, ne12, ne13, ne0, ne1, r2, r3
-                                );
-                                break;
-                            case GGML_TYPE_Q4_1:
-                                ggml_vk_mul_mat_q4_1(
-                                    seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,
-                                    ne00, ne01, ne02, ne10, ne11, ne12, ne13, ne0, ne1, r2, r3
-                                );
-                                break;
-                            case GGML_TYPE_Q4_K:
-                                ggml_vk_mul_mat_q4_k(
-                                    seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,
-                                    ne00, ne01, ne02, ne10, ne11, ne12, ne13, ne0, ne1, ne12/ne02, ne13/ne03
-                                );
-                                break;
-                            case GGML_TYPE_Q6_K:
-                                ggml_vk_mul_mat_q6_k(
-                                    seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,
-                                    ne00, ne10, ne0, ne1, ne01, ne11, ne12, ne02
-                                );
-                                break;
-                            default: {
-                                fprintf(stderr, "%s: %s: Unsupported quantization: %u/%u\n", __func__, ggml_op_name(dst->op), src0t, src1t);
-                                goto not_implemented;
-                            }
-                        }
-
-                    } break;
-                case GGML_OP_GET_ROWS:
-                    {
-                        if (src0t == GGML_TYPE_F32) {
-                            ggml_vk_get_rows_f32(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, nb01, nb1, ggml_nelements(src1));
-                        } else if (src0t == GGML_TYPE_F16) {
-                            ggml_vk_get_rows_f16(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, nb01, nb1, ggml_nelements(src1));
-                        } else if (src0t == GGML_TYPE_Q4_0) {
-                            ggml_vk_get_rows_q4_0(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, nb01, nb1, ggml_nelements(src1));
-                        } else if (src0t == GGML_TYPE_Q4_1) {
-                            ggml_vk_get_rows_q4_1(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, nb01, nb1, ggml_nelements(src1));
-                        } else if (src0t == GGML_TYPE_Q6_K) {
-                            ggml_vk_get_rows_q6_k(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, nb01, nb1, ggml_nelements(src1));
-                        } else {
-                            fprintf(stderr, "%s: %s: Unsupported quantization: %u\n", __func__, ggml_op_name(dst->op), src0t);
-                            goto not_implemented;
-                        }
-                    } break;
-                case GGML_OP_ROPE:
-                    {
-#pragma message("TODO: implement phi3 frequency factors support")
-#pragma message("      https://github.com/ggerganov/llama.cpp/pull/7225")
-                        GGML_ASSERT(dst->src[2] == nullptr && "phi3 frequency factors not implemented yet");
-
-#pragma message("TODO: update rope NORM mode to match NEOX mode")
-#pragma message("      https://github.com/ggerganov/llama.cpp/pull/7634")
-
-                        GGML_ASSERT(ne10 == ne02);
-                        GGML_ASSERT(src0t == dstt);
-                        // const int n_past = ((int32_t *) dst->op_params)[0];
-                        const int n_dims     = ((int32_t *) dst->op_params)[1];
-                        const int mode       = ((int32_t *) dst->op_params)[2];
-                        // skip 3, n_ctx used in GLM RoPE, unimplemented in Vulkan
-                        const int n_ctx_orig = ((int32_t *) dst->op_params)[4];
-
-                        float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
-                        memcpy(&freq_base,   (int32_t *) dst->op_params +  5, sizeof(float));
-                        memcpy(&freq_scale,  (int32_t *) dst->op_params +  6, sizeof(float));
-                        memcpy(&ext_factor,  (int32_t *) dst->op_params +  7, sizeof(float));
-                        memcpy(&attn_factor, (int32_t *) dst->op_params +  8, sizeof(float));
-                        memcpy(&beta_fast,   (int32_t *) dst->op_params +  9, sizeof(float));
-                        memcpy(&beta_slow,   (int32_t *) dst->op_params + 10, sizeof(float));
-                        ggml_vk_rope(
-                            seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, src0t, n_dims, mode, n_ctx_orig,
-                            freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow,
-                            ne01, ne02, ne03, nb00, nb01, nb02, nb03, ne0, nb0, nb1, nb2, nb3
-                        );
-                    } break;
-                case GGML_OP_DUP:
-                case GGML_OP_CPY:
-                case GGML_OP_CONT:
-                    {
-                        switch (src0t) {
-                            case GGML_TYPE_F32:
-                                {
-                                    switch (dstt) {
-                                        case GGML_TYPE_F16: ggml_vk_cpy_f32_f16(seq, id_src0, id_dst, off_src0, off_dst, ne00, ne01, ne02, ne03, nb00, nb01, nb02, nb03, ne0, ne1, ne2, nb0, nb1, nb2, nb3); break;
-                                        case GGML_TYPE_F32: ggml_vk_cpy_f32_f32(seq, id_src0, id_dst, off_src0, off_dst, ne00, ne01, ne02, ne03, nb00, nb01, nb02, nb03, ne0, ne1, ne2, nb0, nb1, nb2, nb3); break;
-                                        default: goto not_implemented;
-                                    }
-                                } break;
-                            case GGML_TYPE_F16:
-                                {
-                                    switch (dstt) {
-                                        case GGML_TYPE_F16: ggml_vk_cpy_f16_f16(seq, id_src0, id_dst, off_src0, off_dst, ne00, ne01, ne02, ne03, nb00, nb01, nb02, nb03, ne0, ne1, ne2, nb0, nb1, nb2, nb3); break;
-                                        case GGML_TYPE_F32: ggml_vk_cpy_f16_f32(seq, id_src0, id_dst, off_src0, off_dst, ne00, ne01, ne02, ne03, nb00, nb01, nb02, nb03, ne0, ne1, ne2, nb0, nb1, nb2, nb3); break;
-                                    default: goto not_implemented;
-                                } break;
-                            default: goto not_implemented;
-                            }
-                        }
-                    } break;
-                default: goto not_implemented;
-            }
-            continue;
-            not_implemented: {}
-            fprintf(stderr, "%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
-            //GGML_ABORT("fatal error");
-        }
-
-        // Evaluate sequence
-        if (any_commands_recorded) {
-            seq.evalAsync();
-        }
-    }
-
-    // Wait for all sequences to finish
-    for (auto& sequence : sequences) {
-        if (sequence->isRunning())
-            sequence->evalAwait();
-    }
-
-    ggml_vk_free_descriptor_pool(ctx);
-}
-
-template<>
-kp::Tensor::TensorDataTypes
-kp::TensorT<half>::dataType()
-{
-    return TensorDataTypes::eFloat;
-}
-
-template<>
-kp::Tensor::TensorDataTypes
-kp::TensorT<uint8_t>::dataType()
-{
-    return TensorDataTypes::eUnsignedInt;
-}
-
-////////////////////////////////////////////////////////////////////////////////
-
-// backend interface
-
-struct ggml_backend_kompute_buffer_type_context {
-    int         device;
-    int         device_ref = 0;
-    uint64_t    buffer_alignment;
-    uint64_t    max_alloc;
-    std::string name;
-
-    ggml_backend_kompute_buffer_type_context(int device, uint64_t buffer_alignment, uint64_t max_alloc)
-        : device(device), buffer_alignment(buffer_alignment), max_alloc(max_alloc), name(ggml_kompute_format_name(device)) {}
-};
-
-static void ggml_backend_kompute_device_ref(ggml_backend_buffer_type_t buft) {
-    auto * ctx = static_cast<ggml_backend_kompute_buffer_type_context *>(buft->context);
-
-    if (!ctx->device_ref) {
-        komputeManager()->initializeDevice(
-            ctx->device, {}, {
-                "VK_KHR_shader_float16_int8", "VK_KHR_8bit_storage",
-                "VK_KHR_16bit_storage", "VK_KHR_shader_non_semantic_info"
-            }
-        );
-    }
-
-    assert(ggml_vk_has_device());
-    ctx->device_ref++;
-}
-
-static void ggml_backend_kompute_device_unref(ggml_backend_buffer_type_t buft) {
-    auto * ctx = static_cast<ggml_backend_kompute_buffer_type_context *>(buft->context);
-
-    assert(ctx->device_ref > 0);
-
-    ctx->device_ref--;
-
-    if (!ctx->device_ref) {
-        komputeManager.destroy();
-    }
-}
-
-static void ggml_backend_kompute_buffer_free_buffer(ggml_backend_buffer_t buffer) {
-    auto * memory = (ggml_vk_memory *)buffer->context;
-    if (ggml_vk_has_device()) {
-        ggml_vk_free_memory(*memory);
-    }
-    delete memory;
-}
-
-static void * ggml_backend_kompute_buffer_get_base(ggml_backend_buffer_t buffer) {
-    return ((ggml_vk_memory *)buffer->context)->data;
-}
-
-static void ggml_backend_kompute_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
-    GGML_UNUSED(buffer);
-
-    const auto res = ggml_vk_get_tensor(tensor);
-    GGML_ASSERT(res);
-
-    memcpy((char *)tensor->data + offset, data, size);
-
-    komputeManager()->sequence()->eval<kp::OpTensorSyncDevice>({res});
-}
-
-static void ggml_backend_kompute_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
-    GGML_UNUSED(buffer);
-
-    const auto res = ggml_vk_get_tensor(tensor);
-    GGML_ASSERT(res);
-
-    komputeManager()->sequence()->eval<kp::OpTensorSyncLocal>({res});
-
-    memcpy(data, (const char *)tensor->data + offset, size);
-}
-
-static void ggml_backend_kompute_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
-    auto * memory = (ggml_vk_memory *)buffer->context;
-    memset(memory->data, value, buffer->size);
-
-    if (memory->stagingBuffer)
-        komputeManager()->sequence()->eval<kp::OpBufferSyncDevice>(memory->primaryBuffer, memory->stagingBuffer, memory->size);
-}
-
-static ggml_backend_buffer_i ggml_backend_kompute_buffer_i = {
-    /* .free_buffer     = */ ggml_backend_kompute_buffer_free_buffer,
-    /* .get_base        = */ ggml_backend_kompute_buffer_get_base,
-    /* .init_tensor     = */ NULL,
-    /* .memset_tensor   = */ NULL,
-    /* .set_tensor      = */ ggml_backend_kompute_buffer_set_tensor,
-    /* .get_tensor      = */ ggml_backend_kompute_buffer_get_tensor,
-    /* .cpy_tensor      = */ NULL,
-    /* .clear           = */ ggml_backend_kompute_buffer_clear,
-    /* .reset           = */ NULL,
-};
-
-// default buffer type
-
-static const char * ggml_backend_kompute_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
-    auto * ctx = static_cast<ggml_backend_kompute_buffer_type_context *>(buft->context);
-    return ctx->name.c_str();
-}
-
-static ggml_backend_buffer_t ggml_backend_kompute_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
-    ggml_backend_kompute_device_ref(buft);
-    auto * ctx = new ggml_vk_memory(ggml_vk_allocate(size));
-    return ggml_backend_buffer_init(buft, ggml_backend_kompute_buffer_i, ctx, size);
-}
-
-static size_t ggml_backend_kompute_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
-    auto * ctx = static_cast<ggml_backend_kompute_buffer_type_context *>(buft->context);
-    return ctx->buffer_alignment;
-}
-
-static size_t ggml_backend_vk_buffer_type_get_max_size(ggml_backend_buffer_type_t buft) {
-    auto * ctx = static_cast<ggml_backend_kompute_buffer_type_context *>(buft->context);
-    return ctx->max_alloc;
-}
-
-static ggml_backend_buffer_type_i ggml_backend_kompute_buffer_type_interface = {
-    /* .get_name         = */ ggml_backend_kompute_buffer_type_get_name,
-    /* .alloc_buffer     = */ ggml_backend_kompute_buffer_type_alloc_buffer,
-    /* .get_alignment    = */ ggml_backend_kompute_buffer_type_get_alignment,
-    /* .get_max_size     = */ ggml_backend_vk_buffer_type_get_max_size,
-    /* .get_alloc_size   = */ NULL, // defaults to ggml_nbytes
-    /* .is_host          = */ NULL,
-};
-
-ggml_backend_buffer_type_t ggml_backend_kompute_buffer_type(int device) {
-    static std::mutex mutex;
-    std::lock_guard<std::mutex> lock(mutex);
-
-    auto devices = ggml_vk_available_devices();
-    int32_t device_count = (int32_t) devices.size();
-    GGML_ASSERT(device < device_count);
-    GGML_ASSERT(devices.size() <= GGML_KOMPUTE_MAX_DEVICES);
-
-    static ggml_backend_buffer_type
-        ggml_backend_kompute_buffer_types[GGML_KOMPUTE_MAX_DEVICES];
-
-    static bool ggml_backend_kompute_buffer_type_initialized = false;
-
-    if (!ggml_backend_kompute_buffer_type_initialized) {
-        for (int32_t i = 0; i < device_count; i++) {
-            ggml_backend_kompute_buffer_types[i] = {
-                /* .iface    = */ ggml_backend_kompute_buffer_type_interface,
-                /* .device   = */ ggml_backend_reg_dev_get(ggml_backend_kompute_reg(), i),
-                /* .context  = */ new ggml_backend_kompute_buffer_type_context{ i, devices[i].bufferAlignment, devices[i].maxAlloc },
-            };
-        }
-        ggml_backend_kompute_buffer_type_initialized = true;
-    }
-
-    return &ggml_backend_kompute_buffer_types[device];
-}
-
-// backend
-
-static const char * ggml_backend_kompute_name(ggml_backend_t backend) {
-    auto * ctx = static_cast<ggml_kompute_context *>(backend->context);
-    return ctx->name.c_str();
-}
-
-static void ggml_backend_kompute_free(ggml_backend_t backend) {
-    auto * ctx = static_cast<ggml_kompute_context *>(backend->context);
-
-    assert(ctx == s_kompute_context);
-    s_kompute_context = nullptr;
-    if (ctx != nullptr) {
-        delete ctx;
-    }
-
-    delete backend;
-}
-
-static ggml_status ggml_backend_kompute_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
-    auto * ctx = static_cast<ggml_kompute_context *>(backend->context);
-    ggml_vk_graph_compute(ctx, cgraph);
-    return GGML_STATUS_SUCCESS;
-}
-
-static struct ggml_backend_i kompute_backend_i = {
-    /* .get_name                = */ ggml_backend_kompute_name,
-    /* .free                    = */ ggml_backend_kompute_free,
-    /* .set_tensor_async        = */ NULL,
-    /* .get_tensor_async        = */ NULL,
-    /* .cpy_tensor_async        = */ NULL,
-    /* .synchronize             = */ NULL,
-    /* .graph_plan_create       = */ NULL,
-    /* .graph_plan_free         = */ NULL,
-    /* .graph_plan_update       = */ NULL,
-    /* .graph_plan_compute      = */ NULL,
-    /* .graph_compute           = */ ggml_backend_kompute_graph_compute,
-    /* .event_record            = */ NULL,
-    /* .event_wait              = */ NULL,
-};
-
-static ggml_guid_t ggml_backend_kompute_guid() {
-    static ggml_guid guid = { 0x7b, 0x57, 0xdc, 0xaf, 0xde, 0x12, 0x1d, 0x49, 0xfb, 0x35, 0xfa, 0x9b, 0x18, 0x31, 0x1d, 0xca };
-    return &guid;
-}
-
-ggml_backend_t ggml_backend_kompute_init(int device) {
-    GGML_ASSERT(s_kompute_context == nullptr);
-    s_kompute_context = new ggml_kompute_context(device);
-
-    ggml_backend_t kompute_backend = new ggml_backend {
-        /* .guid      = */ ggml_backend_kompute_guid(),
-        /* .interface = */ kompute_backend_i,
-        /* .device    = */ ggml_backend_reg_dev_get(ggml_backend_kompute_reg(), device),
-        /* .context   = */ s_kompute_context,
-    };
-
-    return kompute_backend;
-}
-
-bool ggml_backend_is_kompute(ggml_backend_t backend) {
-    return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_kompute_guid());
-}
-
-static size_t ggml_backend_kompute_get_device_count() {
-    auto devices = ggml_vk_available_devices();
-    return devices.size();
-}
-
-static void ggml_backend_kompute_get_device_description(int device, char * description, size_t description_size) {
-    auto devices = ggml_vk_available_devices();
-    GGML_ASSERT((size_t) device < devices.size());
-    snprintf(description, description_size, "%s", devices[device].name);
-}
-
-static void ggml_backend_kompute_get_device_memory(int device, size_t * free, size_t * total) {
-    auto devices = ggml_vk_available_devices();
-    GGML_ASSERT((size_t) device < devices.size());
-    *total = devices[device].heapSize;
-    *free = devices[device].heapSize;
-}
-
-//////////////////////////
-
-struct ggml_backend_kompute_device_context {
-    int device;
-    std::string name;
-    std::string description;
-};
-
-static const char * ggml_backend_kompute_device_get_name(ggml_backend_dev_t dev) {
-    ggml_backend_kompute_device_context * ctx = (ggml_backend_kompute_device_context *)dev->context;
-    return ctx->name.c_str();
-}
-
-static const char * ggml_backend_kompute_device_get_description(ggml_backend_dev_t dev) {
-    ggml_backend_kompute_device_context * ctx = (ggml_backend_kompute_device_context *)dev->context;
-    return ctx->description.c_str();
-}
-
-static void ggml_backend_kompute_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
-    ggml_backend_kompute_device_context * ctx = (ggml_backend_kompute_device_context *)dev->context;
-    ggml_backend_kompute_get_device_memory(ctx->device, free, total);
-}
-
-static ggml_backend_buffer_type_t ggml_backend_kompute_device_get_buffer_type(ggml_backend_dev_t dev) {
-    ggml_backend_kompute_device_context * ctx = (ggml_backend_kompute_device_context *)dev->context;
-    return ggml_backend_kompute_buffer_type(ctx->device);
-}
-
-static bool ggml_backend_kompute_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
-    if (buft->iface.get_name != ggml_backend_kompute_buffer_type_get_name) {
-        return false;
-    }
-
-    ggml_backend_kompute_device_context * ctx = (ggml_backend_kompute_device_context *)dev->context;
-    ggml_backend_kompute_buffer_type_context * buft_ctx = (ggml_backend_kompute_buffer_type_context *)buft->context;
-
-    return buft_ctx->device == ctx->device;
-}
-
-static enum ggml_backend_dev_type ggml_backend_kompute_device_get_type(ggml_backend_dev_t dev) {
-    GGML_UNUSED(dev);
-    return GGML_BACKEND_DEVICE_TYPE_GPU;
-}
-
-static void ggml_backend_kompute_device_get_props(ggml_backend_dev_t dev, struct ggml_backend_dev_props * props) {
-    props->name        = ggml_backend_kompute_device_get_name(dev);
-    props->description = ggml_backend_kompute_device_get_description(dev);
-    props->type        = ggml_backend_kompute_device_get_type(dev);
-    ggml_backend_kompute_device_get_memory(dev, &props->memory_free, &props->memory_total);
-    props->caps = {
-        /* async                  = */ false,
-        /* host_buffer            = */ false,
-        /* .buffer_from_host_ptr  = */ false,
-        /* events                 = */ false,
-    };
-}
-
-static ggml_backend_t ggml_backend_kompute_device_init(ggml_backend_dev_t dev, const char * params) {
-    GGML_UNUSED(params);
-    ggml_backend_kompute_device_context * ctx = (ggml_backend_kompute_device_context *)dev->context;
-    return ggml_backend_kompute_init(ctx->device);
-}
-
-static bool ggml_backend_kompute_device_offload_op(ggml_backend_dev_t dev, const ggml_tensor * op) {
-    const int min_batch_size = 32;
-
-    return (op->ne[1] >= min_batch_size && op->op != GGML_OP_GET_ROWS) ||
-           (op->ne[2] >= min_batch_size && op->op == GGML_OP_MUL_MAT_ID);
-
-    GGML_UNUSED(dev);
-}
-
-static const struct ggml_backend_device_i ggml_backend_kompute_device_i = {
-    /* .get_name             = */ ggml_backend_kompute_device_get_name,
-    /* .get_description      = */ ggml_backend_kompute_device_get_description,
-    /* .get_memory           = */ ggml_backend_kompute_device_get_memory,
-    /* .get_type             = */ ggml_backend_kompute_device_get_type,
-    /* .get_props            = */ ggml_backend_kompute_device_get_props,
-    /* .init_backend         = */ ggml_backend_kompute_device_init,
-    /* .get_buffer_type      = */ ggml_backend_kompute_device_get_buffer_type,
-    /* .get_host_buffer_type = */ NULL,
-    /* .buffer_from_host_ptr = */ NULL,
-    /* .supports_op          = */ ggml_backend_kompute_device_supports_op,
-    /* .supports_buft        = */ ggml_backend_kompute_device_supports_buft,
-    /* .offload_op           = */ ggml_backend_kompute_device_offload_op,
-    /* .event_new            = */ NULL,
-    /* .event_free           = */ NULL,
-    /* .event_synchronize    = */ NULL,
-};
-
-static const char * ggml_backend_kompute_reg_get_name(ggml_backend_reg_t reg) {
-    GGML_UNUSED(reg);
-    return "Kompute";
-}
-
-static size_t ggml_backend_kompute_reg_get_device_count(ggml_backend_reg_t reg) {
-    GGML_UNUSED(reg);
-    return ggml_backend_kompute_get_device_count();
-}
-
-static ggml_backend_dev_t ggml_backend_kompute_reg_get_device(ggml_backend_reg_t reg, size_t device) {
-    static std::vector<ggml_backend_dev_t> devices;
-
-    static bool initialized = false;
-
-    {
-        static std::mutex mutex;
-        std::lock_guard<std::mutex> lock(mutex);
-        if (!initialized) {
-            for (size_t i = 0; i < ggml_backend_kompute_get_device_count(); i++) {
-                ggml_backend_kompute_device_context * ctx = new ggml_backend_kompute_device_context;
-                char desc[256];
-                ggml_backend_kompute_get_device_description(i, desc, sizeof(desc));
-                ctx->device = i;
-                ctx->name = "Kompute" + std::to_string(i);
-                ctx->description = desc;
-                devices.push_back(new ggml_backend_device {
-                    /* .iface   = */ ggml_backend_kompute_device_i,
-                    /* .reg     = */ reg,
-                    /* .context = */ ctx,
-                });
-            }
-            initialized = true;
-        }
-    }
-
-    GGML_ASSERT(device < devices.size());
-    return devices[device];
-}
-
-static const struct ggml_backend_reg_i ggml_backend_kompute_reg_i = {
-    /* .get_name         = */ ggml_backend_kompute_reg_get_name,
-    /* .get_device_count = */ ggml_backend_kompute_reg_get_device_count,
-    /* .get_device       = */ ggml_backend_kompute_reg_get_device,
-    /* .get_proc_address = */ NULL,
-};
-
-ggml_backend_reg_t ggml_backend_kompute_reg() {
-    static ggml_backend_reg reg = {
-        /* .iface   = */ ggml_backend_kompute_reg_i,
-        /* .context = */ nullptr,
-    };
-
-    return &reg;
-}

ggml/src/ggml-metal/ggml-metal.m

@@ -3651,7 +3651,7 @@ static enum ggml_status ggml_metal_graph_compute(
         dispatch_apply(n_cb, ctx->d_queue, ctx->encode_async);
 
         // wait for completion and check status of each command buffer
-        // needed to detect if the device ran out-of-memory for example (ggml/1881)
+        // needed to detect if the device ran out-of-memory for example (#1881)
         {
             id<MTLCommandBuffer> command_buffer = ctx->command_buffers[n_cb];
             [command_buffer waitUntilCompleted];

ggml/src/ggml-musa/CMakeLists.txt

@@ -0,0 +1,100 @@
+if (NOT EXISTS $ENV{MUSA_PATH})
+    if (NOT EXISTS /opt/musa)
+        set(MUSA_PATH /usr/local/musa)
+    else()
+        set(MUSA_PATH /opt/musa)
+    endif()
+else()
+    set(MUSA_PATH $ENV{MUSA_PATH})
+endif()
+
+set(CMAKE_C_COMPILER "${MUSA_PATH}/bin/clang")
+set(CMAKE_C_EXTENSIONS OFF)
+set(CMAKE_CXX_COMPILER "${MUSA_PATH}/bin/clang++")
+set(CMAKE_CXX_EXTENSIONS OFF)
+
+list(APPEND CMAKE_MODULE_PATH "${MUSA_PATH}/cmake")
+
+find_package(MUSAToolkit)
+
+if (MUSAToolkit_FOUND)
+    message(STATUS "MUSA Toolkit found")
+
+    file(GLOB   GGML_HEADERS_MUSA "../ggml-cuda/*.cuh")
+    list(APPEND GGML_HEADERS_MUSA "../../include/ggml-cuda.h")
+
+    file(GLOB   GGML_SOURCES_MUSA "../ggml-cuda/*.cu")
+    file(GLOB   SRCS "../ggml-cuda/template-instances/fattn-wmma*.cu")
+    list(APPEND GGML_SOURCES_MUSA ${SRCS})
+    file(GLOB   SRCS "../ggml-cuda/template-instances/mmq*.cu")
+    list(APPEND GGML_SOURCES_MUSA ${SRCS})
+
+    if (GGML_CUDA_FA_ALL_QUANTS)
+        file(GLOB   SRCS "../ggml-cuda/template-instances/fattn-vec*.cu")
+        list(APPEND GGML_SOURCES_MUSA ${SRCS})
+        add_compile_definitions(GGML_CUDA_FA_ALL_QUANTS)
+    else()
+        file(GLOB   SRCS "../ggml-cuda/template-instances/fattn-vec*q4_0-q4_0.cu")
+        list(APPEND GGML_SOURCES_MUSA ${SRCS})
+        file(GLOB   SRCS "../ggml-cuda/template-instances/fattn-vec*q8_0-q8_0.cu")
+        list(APPEND GGML_SOURCES_MUSA ${SRCS})
+        file(GLOB   SRCS "../ggml-cuda/template-instances/fattn-vec*f16-f16.cu")
+        list(APPEND GGML_SOURCES_MUSA ${SRCS})
+    endif()
+
+    set_source_files_properties(${GGML_SOURCES_MUSA} PROPERTIES LANGUAGE CXX)
+    foreach(SOURCE ${GGML_SOURCES_MUSA})
+        set_property(SOURCE ${SOURCE} PROPERTY COMPILE_FLAGS "-x musa -mtgpu --cuda-gpu-arch=mp_21 --cuda-gpu-arch=mp_22")
+    endforeach()
+
+    add_library(ggml-musa
+                ${GGML_HEADERS_MUSA}
+                ${GGML_SOURCES_MUSA})
+
+    target_link_libraries(ggml-musa PRIVATE ggml-base)
+    target_include_directories(ggml-musa PRIVATE . ..)
+
+    # TODO: do not use CUDA definitions for MUSA
+    target_compile_definitions(ggml PUBLIC GGML_USE_CUDA)
+
+    add_compile_definitions(GGML_USE_MUSA)
+    add_compile_definitions(GGML_CUDA_PEER_MAX_BATCH_SIZE=${GGML_CUDA_PEER_MAX_BATCH_SIZE})
+
+    if (GGML_CUDA_GRAPHS)
+        add_compile_definitions(GGML_CUDA_USE_GRAPHS)
+    endif()
+
+    if (GGML_CUDA_FORCE_MMQ)
+        add_compile_definitions(GGML_CUDA_FORCE_MMQ)
+    endif()
+
+    if (GGML_CUDA_FORCE_CUBLAS)
+        add_compile_definitions(GGML_CUDA_FORCE_CUBLAS)
+    endif()
+
+    if (GGML_CUDA_NO_VMM)
+        add_compile_definitions(GGML_CUDA_NO_VMM)
+    endif()
+
+    if (GGML_CUDA_F16 OR GGML_CUDA_DMMV_F16)
+        add_compile_definitions(GGML_CUDA_F16)
+    endif()
+
+    if (GGML_CUDA_NO_PEER_COPY)
+        add_compile_definitions(GGML_CUDA_NO_PEER_COPY)
+    endif()
+
+    if (GGML_STATIC)
+        target_link_libraries(ggml-musa PRIVATE MUSA::musart_static MUSA::mublas_static)
+    else()
+        target_link_libraries(ggml-musa PRIVATE MUSA::musart MUSA::mublas)
+    endif()
+
+    if (GGML_CUDA_NO_VMM)
+        # No VMM requested, no need to link directly with the musa driver lib (libmusa.so)
+    else()
+        target_link_libraries(ggml-musa PRIVATE MUSA::musa_driver)
+    endif()
+else()
+    message(FATAL_ERROR "MUSA Toolkit not found")
+endif()

ggml/src/ggml-rpc.cpp

@@ -1,1403 +0,0 @@
-#include "ggml-rpc.h"
-#include "ggml-impl.h"
-#include "ggml-backend-impl.h"
-
-#include <cinttypes>
-#include <string>
-#include <vector>
-#include <memory>
-#include <mutex>
-#include <unordered_map>
-#include <unordered_set>
-#ifdef _WIN32
-#  define WIN32_LEAN_AND_MEAN
-#  ifndef NOMINMAX
-#     define NOMINMAX
-#  endif
-#  include <windows.h>
-#  include <winsock2.h>
-#else
-#  include <arpa/inet.h>
-#  include <sys/socket.h>
-#  include <sys/types.h>
-#  include <netinet/in.h>
-#  include <netinet/tcp.h>
-#  include <netdb.h>
-#  include <unistd.h>
-#endif
-#include <cstring>
-
-#define UNUSED GGML_UNUSED
-
-#define GGML_DEBUG 0
-#if (GGML_DEBUG >= 1)
-#define GGML_PRINT_DEBUG(...) printf(__VA_ARGS__)
-#else
-#define GGML_PRINT_DEBUG(...)
-#endif
-
-#ifdef _WIN32
-typedef SOCKET sockfd_t;
-using ssize_t = __int64;
-#else
-typedef int sockfd_t;
-#endif
-
-// cross-platform socket
-struct socket_t {
-    sockfd_t fd;
-    socket_t(sockfd_t fd) : fd(fd) {}
-    ~socket_t() {
-        GGML_PRINT_DEBUG("[%s] closing socket %d\n", __func__, this->fd);
-#ifdef _WIN32
-        closesocket(this->fd);
-#else
-        close(this->fd);
-#endif
-    }
-};
-
-// all RPC structures must be packed
-#pragma pack(push, 1)
-// ggml_tensor is serialized into rpc_tensor
-struct rpc_tensor {
-    uint64_t id;
-    uint32_t type;
-    uint64_t buffer;
-    uint32_t ne[GGML_MAX_DIMS];
-    uint32_t nb[GGML_MAX_DIMS];
-    uint32_t op;
-    int32_t  op_params[GGML_MAX_OP_PARAMS / sizeof(int32_t)];
-    int32_t  flags;
-    uint64_t src[GGML_MAX_SRC];
-    uint64_t view_src;
-    uint64_t view_offs;
-    uint64_t data;
-    char name[GGML_MAX_NAME];
-
-    char padding[4];
-};
-
-static_assert(sizeof(rpc_tensor) % 8 == 0, "rpc_tensor size must be multiple of 8");
-
-// RPC commands
-enum rpc_cmd {
-    RPC_CMD_ALLOC_BUFFER = 0,
-    RPC_CMD_GET_ALIGNMENT,
-    RPC_CMD_GET_MAX_SIZE,
-    RPC_CMD_BUFFER_GET_BASE,
-    RPC_CMD_FREE_BUFFER,
-    RPC_CMD_BUFFER_CLEAR,
-    RPC_CMD_SET_TENSOR,
-    RPC_CMD_GET_TENSOR,
-    RPC_CMD_COPY_TENSOR,
-    RPC_CMD_GRAPH_COMPUTE,
-    RPC_CMD_GET_DEVICE_MEMORY,
-    RPC_CMD_COUNT,
-};
-
-struct rpc_msg_alloc_buffer_req {
-    uint64_t size;
-};
-
-struct rpc_msg_alloc_buffer_rsp {
-    uint64_t remote_ptr;
-    uint64_t remote_size;
-};
-
-struct rpc_msg_get_alignment_rsp {
-    uint64_t alignment;
-};
-
-struct rpc_msg_get_max_size_rsp {
-    uint64_t max_size;
-};
-
-struct rpc_msg_buffer_get_base_req {
-    uint64_t remote_ptr;
-};
-
-struct rpc_msg_buffer_get_base_rsp {
-    uint64_t base_ptr;
-};
-
-struct rpc_msg_free_buffer_req {
-    uint64_t remote_ptr;
-};
-
-struct rpc_msg_buffer_clear_req {
-    uint64_t remote_ptr;
-    uint8_t value;
-};
-
-struct rpc_msg_get_tensor_req {
-    rpc_tensor tensor;
-    uint64_t offset;
-    uint64_t size;
-};
-
-struct rpc_msg_copy_tensor_req {
-    rpc_tensor src;
-    rpc_tensor dst;
-};
-
-struct rpc_msg_copy_tensor_rsp {
-    uint8_t result;
-};
-
-struct rpc_msg_graph_compute_rsp {
-    uint8_t result;
-};
-
-struct rpc_msg_get_device_memory_rsp {
-    uint64_t free_mem;
-    uint64_t total_mem;
-};
-#pragma pack(pop)
-
-// RPC data structures
-
-static ggml_guid_t ggml_backend_rpc_guid() {
-    static ggml_guid guid = {0x99, 0x68, 0x5b, 0x6c, 0xd2, 0x83, 0x3d, 0x24, 0x25, 0x36, 0x72, 0xe1, 0x5b, 0x0e, 0x14, 0x03};
-    return &guid;
-}
-
-struct ggml_backend_rpc_buffer_type_context {
-    std::string endpoint;
-    std::string name;
-    size_t alignment;
-    size_t max_size;
-};
-
-struct ggml_backend_rpc_context {
-    std::string endpoint;
-    std::string name;
-};
-
-struct ggml_backend_rpc_buffer_context {
-    std::shared_ptr<socket_t> sock;
-    std::unordered_map<ggml_backend_buffer_t, void *> base_cache;
-    uint64_t remote_ptr;
-};
-
-// RPC helper functions
-
-static std::shared_ptr<socket_t> make_socket(sockfd_t fd) {
-#ifdef _WIN32
-    if (fd == INVALID_SOCKET) {
-        return nullptr;
-    }
-#else
-    if (fd < 0) {
-        return nullptr;
-    }
-#endif
-    return std::make_shared<socket_t>(fd);
-}
-
-static bool set_no_delay(sockfd_t sockfd) {
-    int flag = 1;
-    // set TCP_NODELAY to disable Nagle's algorithm
-    int ret = setsockopt(sockfd, IPPROTO_TCP, TCP_NODELAY, (char *)&flag, sizeof(int));
-    return ret == 0;
-}
-
-static bool set_reuse_addr(sockfd_t sockfd) {
-    int flag = 1;
-    int ret = setsockopt(sockfd, SOL_SOCKET, SO_REUSEADDR, (char *)&flag, sizeof(int));
-    return ret == 0;
-}
-
-static std::shared_ptr<socket_t> socket_connect(const char * host, int port) {
-    struct sockaddr_in addr;
-    auto sockfd = socket(AF_INET, SOCK_STREAM, 0);
-    auto sock_ptr = make_socket(sockfd);
-    if (sock_ptr == nullptr) {
-        return nullptr;
-    }
-    if (!set_no_delay(sockfd)) {
-        fprintf(stderr, "Failed to set TCP_NODELAY\n");
-        return nullptr;
-    }
-    addr.sin_family = AF_INET;
-    addr.sin_port = htons(port);
-    struct hostent * server = gethostbyname(host);
-    if (server == NULL) {
-        fprintf(stderr, "Cannot resolve host '%s'\n", host);
-        return nullptr;
-    }
-    memcpy(&addr.sin_addr.s_addr, server->h_addr, server->h_length);
-    if (connect(sock_ptr->fd, (struct sockaddr *)&addr, sizeof(addr)) < 0) {
-        return nullptr;
-    }
-    return sock_ptr;
-}
-
-static std::shared_ptr<socket_t> socket_accept(sockfd_t srv_sockfd) {
-    auto client_socket_fd = accept(srv_sockfd, NULL, NULL);
-    auto client_socket = make_socket(client_socket_fd);
-    if (client_socket == nullptr) {
-        return nullptr;
-    }
-    if (!set_no_delay(client_socket_fd)) {
-        fprintf(stderr, "Failed to set TCP_NODELAY\n");
-        return nullptr;
-    }
-    return client_socket;
-}
-
-static std::shared_ptr<socket_t> create_server_socket(const char * host, int port) {
-    auto sockfd = socket(AF_INET, SOCK_STREAM, 0);
-    auto sock = make_socket(sockfd);
-    if (sock == nullptr) {
-        return nullptr;
-    }
-    if (!set_reuse_addr(sockfd)) {
-        fprintf(stderr, "Failed to set SO_REUSEADDR\n");
-        return nullptr;
-    }
-    if (inet_addr(host) == INADDR_NONE) {
-        fprintf(stderr, "Invalid host address: %s\n", host);
-        return nullptr;
-    }
-    struct sockaddr_in serv_addr;
-    serv_addr.sin_family = AF_INET;
-    serv_addr.sin_addr.s_addr = inet_addr(host);
-    serv_addr.sin_port = htons(port);
-
-    if (bind(sockfd, (struct sockaddr *) &serv_addr, sizeof(serv_addr)) < 0) {
-        return nullptr;
-    }
-    if (listen(sockfd, 1) < 0) {
-        return nullptr;
-    }
-    return sock;
-}
-
-static bool send_data(sockfd_t sockfd, const void * data, size_t size) {
-    size_t bytes_sent = 0;
-    while (bytes_sent < size) {
-        ssize_t n = send(sockfd, (const char *)data + bytes_sent, size - bytes_sent, 0);
-        if (n < 0) {
-            return false;
-        }
-        bytes_sent += n;
-    }
-    return true;
-}
-
-static bool recv_data(sockfd_t sockfd, void * data, size_t size) {
-    size_t bytes_recv = 0;
-    while (bytes_recv < size) {
-        ssize_t n = recv(sockfd, (char *)data + bytes_recv, size - bytes_recv, 0);
-        if (n <= 0) {
-            return false;
-        }
-        bytes_recv += n;
-    }
-    return true;
-}
-
-static bool send_msg(sockfd_t sockfd, const void * msg, size_t msg_size) {
-    if (!send_data(sockfd, &msg_size, sizeof(msg_size))) {
-        return false;
-    }
-    return send_data(sockfd, msg, msg_size);
-}
-
-static bool recv_msg(sockfd_t sockfd, void * msg, size_t msg_size) {
-    uint64_t size;
-    if (!recv_data(sockfd, &size, sizeof(size))) {
-        return false;
-    }
-    if (size != msg_size) {
-        return false;
-    }
-    return recv_data(sockfd, msg, msg_size);
-}
-
-static bool recv_msg(sockfd_t sockfd, std::vector<uint8_t> & input) {
-    uint64_t size;
-    if (!recv_data(sockfd, &size, sizeof(size))) {
-        return false;
-    }
-    try {
-        input.resize(size);
-    } catch (const std::bad_alloc & e) {
-        fprintf(stderr, "Failed to allocate input buffer of size %" PRIu64 "\n", size);
-        return false;
-    }
-    return recv_data(sockfd, input.data(), size);
-}
-
-static bool parse_endpoint(const std::string & endpoint, std::string & host, int & port) {
-    size_t pos = endpoint.find(':');
-    if (pos == std::string::npos) {
-        return false;
-    }
-    host = endpoint.substr(0, pos);
-    port = std::stoi(endpoint.substr(pos + 1));
-    return true;
-}
-
-// RPC request : | rpc_cmd (1 byte) | request_size (8 bytes) | request_data (request_size bytes) |
-// RPC response: | response_size (8 bytes) | response_data (response_size bytes) |
-static bool send_rpc_cmd(const std::shared_ptr<socket_t> & sock, enum rpc_cmd cmd, const void * input, size_t input_size, void * output, size_t output_size) {
-    uint8_t cmd_byte = cmd;
-    if (!send_data(sock->fd, &cmd_byte, sizeof(cmd_byte))) {
-        return false;
-    }
-    if (!send_data(sock->fd, &input_size, sizeof(input_size))) {
-        return false;
-    }
-    if (!send_data(sock->fd, input, input_size)) {
-        return false;
-    }
-    // TODO: currently the output_size is always known, do we need support for commands with variable output size?
-    // even if we do, we can skip sending output_size from the server for commands with known output size
-    uint64_t out_size;
-    if (!recv_data(sock->fd, &out_size, sizeof(out_size))) {
-        return false;
-    }
-    if (out_size != output_size) {
-        return false;
-    }
-    if (!recv_data(sock->fd, output, output_size)) {
-        return false;
-    }
-    return true;
-}
-
-// RPC client-side implementation
-
-static std::shared_ptr<socket_t> get_socket(const std::string & endpoint) {
-    static std::mutex mutex;
-    std::lock_guard<std::mutex> lock(mutex);
-    static std::unordered_map<std::string, std::weak_ptr<socket_t>> sockets;
-    static bool initialized = false;
-
-    auto it = sockets.find(endpoint);
-    if (it != sockets.end()) {
-        if (auto sock = it->second.lock()) {
-            return sock;
-        }
-    }
-    std::string host;
-    int port;
-    if (!parse_endpoint(endpoint, host, port)) {
-        return nullptr;
-    }
-#ifdef _WIN32
-    if (!initialized) {
-        WSADATA wsaData;
-        int res = WSAStartup(MAKEWORD(2, 2), &wsaData);
-        if (res != 0) {
-            return nullptr;
-        }
-        initialized = true;
-    }
-#else
-    UNUSED(initialized);
-#endif
-    auto sock = socket_connect(host.c_str(), port);
-    if (sock == nullptr) {
-        return nullptr;
-    }
-    GGML_PRINT_DEBUG("[%s] connected to %s, sockfd=%d\n", __func__, endpoint.c_str(), sock->fd);
-    sockets[endpoint] = sock;
-    return sock;
-}
-
-static void ggml_backend_rpc_buffer_free_buffer(ggml_backend_buffer_t buffer) {
-    ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context;
-    rpc_msg_free_buffer_req request = {ctx->remote_ptr};
-    bool status = send_rpc_cmd(ctx->sock, RPC_CMD_FREE_BUFFER, &request, sizeof(request), nullptr, 0);
-    GGML_ASSERT(status);
-    delete ctx;
-}
-
-static void * ggml_backend_rpc_buffer_get_base(ggml_backend_buffer_t buffer) {
-    ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context;
-    if (ctx->base_cache.find(buffer) != ctx->base_cache.end()) {
-        return ctx->base_cache[buffer];
-    }
-    rpc_msg_buffer_get_base_req request = {ctx->remote_ptr};
-    rpc_msg_buffer_get_base_rsp response;
-    bool status = send_rpc_cmd(ctx->sock, RPC_CMD_BUFFER_GET_BASE, &request, sizeof(request), &response, sizeof(response));
-    GGML_ASSERT(status);
-    void * base_ptr = reinterpret_cast<void *>(response.base_ptr);
-    ctx->base_cache[buffer] = base_ptr;
-    return base_ptr;
-}
-
-static rpc_tensor serialize_tensor(const ggml_tensor * tensor) {
-    rpc_tensor result;
-    result.id = reinterpret_cast<uint64_t>(tensor);
-    result.type = tensor->type;
-    if (tensor->buffer) {
-        ggml_backend_buffer_t buffer = tensor->buffer;
-        ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context;
-        result.buffer = ctx->remote_ptr;
-    } else {
-        result.buffer = 0;
-    }
-    for (uint32_t i = 0; i < GGML_MAX_DIMS; i++) {
-        result.ne[i] = tensor->ne[i];
-        result.nb[i] = tensor->nb[i];
-    }
-    result.op = tensor->op;
-    for (uint32_t i = 0; i < GGML_MAX_OP_PARAMS / sizeof(int32_t); i++) {
-        result.op_params[i] = tensor->op_params[i];
-    }
-    result.flags = tensor->flags;
-    for (uint32_t i = 0; i < GGML_MAX_SRC; i++) {
-        result.src[i] = reinterpret_cast<uint64_t>(tensor->src[i]);
-    }
-    result.view_src = reinterpret_cast<uint64_t>(tensor->view_src);
-    result.view_offs = tensor->view_offs;
-    result.data = reinterpret_cast<uint64_t>(tensor->data);
-    snprintf(result.name, GGML_MAX_NAME, "%s", tensor->name);
-    return result;
-}
-
-static void ggml_backend_rpc_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) {
-    UNUSED(buffer);
-    if (ggml_is_quantized(tensor->type)) {
-        // TODO: this check is due to MATRIX_ROW_PADDING in CUDA and should be generalized
-        GGML_ASSERT(tensor->ne[0] % 512 == 0 && "unsupported quantized tensor");
-    }
-}
-
-static void ggml_backend_rpc_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
-    ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context;
-    // input serialization format: | rpc_tensor | offset (8 bytes) | data (size bytes) |
-    size_t input_size = sizeof(rpc_tensor) + sizeof(uint64_t) + size;
-    std::vector<uint8_t> input(input_size, 0);
-    rpc_tensor rpc_tensor = serialize_tensor(tensor);
-    memcpy(input.data(), &rpc_tensor, sizeof(rpc_tensor));
-    memcpy(input.data() + sizeof(rpc_tensor), &offset, sizeof(offset));
-    memcpy(input.data() + sizeof(rpc_tensor) + sizeof(offset), data, size);
-    bool status = send_rpc_cmd(ctx->sock, RPC_CMD_SET_TENSOR, input.data(), input.size(), nullptr, 0);
-    GGML_ASSERT(status);
-}
-
-static void ggml_backend_rpc_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
-    ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context;
-    rpc_msg_get_tensor_req request;
-    request.tensor = serialize_tensor(tensor);
-    request.offset = offset;
-    request.size = size;
-    bool status = send_rpc_cmd(ctx->sock, RPC_CMD_GET_TENSOR, &request, sizeof(request), data, size);
-    GGML_ASSERT(status);
-}
-
-static bool ggml_backend_rpc_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * src, ggml_tensor * dst) {
-    // check if src and dst are on the same server
-    ggml_backend_buffer_t src_buffer = src->buffer;
-    ggml_backend_rpc_buffer_context * src_ctx = (ggml_backend_rpc_buffer_context *)src_buffer->context;
-    ggml_backend_buffer_t dst_buffer = dst->buffer;
-    ggml_backend_rpc_buffer_context * dst_ctx = (ggml_backend_rpc_buffer_context *)dst_buffer->context;
-    if (src_ctx->sock != dst_ctx->sock) {
-        return false;
-    }
-    ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context;
-    rpc_msg_copy_tensor_req request;
-    request.src = serialize_tensor(src);
-    request.dst = serialize_tensor(dst);
-    rpc_msg_copy_tensor_rsp response;
-    bool status = send_rpc_cmd(ctx->sock, RPC_CMD_COPY_TENSOR, &request, sizeof(request), &response, sizeof(response));
-    GGML_ASSERT(status);
-    return response.result;
-}
-
-static void ggml_backend_rpc_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
-    ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context;
-    rpc_msg_buffer_clear_req request = {ctx->remote_ptr, value};
-    bool status = send_rpc_cmd(ctx->sock, RPC_CMD_BUFFER_CLEAR, &request, sizeof(request), nullptr, 0);
-    GGML_ASSERT(status);
-}
-
-static ggml_backend_buffer_i ggml_backend_rpc_buffer_interface = {
-    /* .free_buffer     = */ ggml_backend_rpc_buffer_free_buffer,
-    /* .get_base        = */ ggml_backend_rpc_buffer_get_base,
-    /* .init_tensor     = */ ggml_backend_rpc_buffer_init_tensor,
-    /* .memset_tensor   = */ NULL,
-    /* .set_tensor      = */ ggml_backend_rpc_buffer_set_tensor,
-    /* .get_tensor      = */ ggml_backend_rpc_buffer_get_tensor,
-    /* .cpy_tensor      = */ ggml_backend_rpc_buffer_cpy_tensor,
-    /* .clear           = */ ggml_backend_rpc_buffer_clear,
-    /* .reset           = */ NULL,
-};
-
-static const char * ggml_backend_rpc_buffer_type_name(ggml_backend_buffer_type_t buft) {
-    ggml_backend_rpc_buffer_type_context * buft_ctx = (ggml_backend_rpc_buffer_type_context *)buft->context;
-    return buft_ctx->name.c_str();
-}
-
-static ggml_backend_buffer_t ggml_backend_rpc_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
-    ggml_backend_rpc_buffer_type_context * buft_ctx = (ggml_backend_rpc_buffer_type_context *)buft->context;
-    rpc_msg_alloc_buffer_req request = {size};
-    rpc_msg_alloc_buffer_rsp response;
-    auto sock = get_socket(buft_ctx->endpoint);
-    bool status = send_rpc_cmd(sock, RPC_CMD_ALLOC_BUFFER, &request, sizeof(request), &response, sizeof(response));
-    GGML_ASSERT(status);
-    if (response.remote_ptr != 0) {
-        ggml_backend_buffer_t buffer = ggml_backend_buffer_init(buft,
-            ggml_backend_rpc_buffer_interface,
-            new ggml_backend_rpc_buffer_context{sock, {}, response.remote_ptr},
-            response.remote_size);
-        return buffer;
-    } else {
-        return nullptr;
-    }
-}
-
-static size_t get_alignment(const std::shared_ptr<socket_t> & sock) {
-    rpc_msg_get_alignment_rsp response;
-    bool status = send_rpc_cmd(sock, RPC_CMD_GET_ALIGNMENT, nullptr, 0, &response, sizeof(response));
-    GGML_ASSERT(status);
-    return response.alignment;
-}
-
-static size_t ggml_backend_rpc_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
-    ggml_backend_rpc_buffer_type_context * buft_ctx = (ggml_backend_rpc_buffer_type_context *)buft->context;
-    return buft_ctx->alignment;
-}
-
-static size_t get_max_size(const std::shared_ptr<socket_t> & sock) {
-    rpc_msg_get_max_size_rsp response;
-    bool status = send_rpc_cmd(sock, RPC_CMD_GET_MAX_SIZE, nullptr, 0, &response, sizeof(response));
-    GGML_ASSERT(status);
-    return response.max_size;
-}
-
-static size_t ggml_backend_rpc_get_max_size(ggml_backend_buffer_type_t buft) {
-    ggml_backend_rpc_buffer_type_context * buft_ctx = (ggml_backend_rpc_buffer_type_context *)buft->context;
-    return buft_ctx->max_size;
-}
-
-static size_t ggml_backend_rpc_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) {
-    UNUSED(buft);
-    return ggml_nbytes(tensor);
-}
-
-static ggml_backend_buffer_type_i ggml_backend_rpc_buffer_type_interface = {
-    /* .get_name         = */ ggml_backend_rpc_buffer_type_name,
-    /* .alloc_buffer     = */ ggml_backend_rpc_buffer_type_alloc_buffer,
-    /* .get_alignment    = */ ggml_backend_rpc_buffer_type_get_alignment,
-    /* .get_max_size     = */ ggml_backend_rpc_get_max_size,
-    /* .get_alloc_size   = */ ggml_backend_rpc_buffer_type_get_alloc_size,
-    /* .is_host          = */ NULL,
-};
-
-static const char * ggml_backend_rpc_name(ggml_backend_t backend) {
-    ggml_backend_rpc_context * rpc_ctx = (ggml_backend_rpc_context *)backend->context;
-
-    return rpc_ctx->name.c_str();
-}
-
-static void ggml_backend_rpc_free(ggml_backend_t backend) {
-    ggml_backend_rpc_context * rpc_ctx = (ggml_backend_rpc_context *)backend->context;
-    delete rpc_ctx;
-    delete backend;
-}
-
-static void ggml_backend_rpc_synchronize(ggml_backend_t backend) {
-    UNUSED(backend);
-    // this is no-op because we don't have any async operations
-}
-
-static void add_tensor(ggml_tensor * tensor, std::vector<rpc_tensor> & tensors, std::unordered_set<ggml_tensor*> & visited) {
-    if (tensor == nullptr) {
-        return;
-    }
-    if (visited.find(tensor) != visited.end()) {
-        return;
-    }
-    visited.insert(tensor);
-    for (int i = 0; i < GGML_MAX_SRC; i++) {
-        add_tensor(tensor->src[i], tensors, visited);
-    }
-    add_tensor(tensor->view_src, tensors, visited);
-    tensors.push_back(serialize_tensor(tensor));
-}
-
-static void serialize_graph(const ggml_cgraph * cgraph, std::vector<uint8_t> & output) {
-    uint32_t n_nodes = cgraph->n_nodes;
-    std::vector<rpc_tensor> tensors;
-    std::unordered_set<ggml_tensor*> visited;
-    for (uint32_t i = 0; i < n_nodes; i++) {
-        add_tensor(cgraph->nodes[i], tensors, visited);
-    }
-    // serialization format:
-    // | n_nodes (4 bytes) | nodes (n_nodes * sizeof(uint64_t) | n_tensors (4 bytes) | tensors (n_tensors * sizeof(rpc_tensor)) |
-    uint32_t n_tensors = tensors.size();
-    int output_size = sizeof(uint32_t) + n_nodes * sizeof(uint64_t) + sizeof(uint32_t) + n_tensors * sizeof(rpc_tensor);
-    output.resize(output_size, 0);
-    memcpy(output.data(), &n_nodes, sizeof(n_nodes));
-    for (uint32_t i = 0; i < n_nodes; i++) {
-        memcpy(output.data() + sizeof(n_nodes) + i * sizeof(uint64_t), &cgraph->nodes[i], sizeof(uint64_t));
-    }
-    uint32_t * out_ntensors = (uint32_t *)(output.data() + sizeof(n_nodes) + n_nodes * sizeof(uint64_t));
-    *out_ntensors = n_tensors;
-    rpc_tensor * out_tensors = (rpc_tensor *)(output.data() + sizeof(n_nodes) + n_nodes * sizeof(uint64_t) + sizeof(uint32_t));
-    memcpy(out_tensors, tensors.data(), n_tensors * sizeof(rpc_tensor));
-}
-
-static enum ggml_status ggml_backend_rpc_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
-    ggml_backend_rpc_context * rpc_ctx = (ggml_backend_rpc_context *)backend->context;
-    std::vector<uint8_t> input;
-    serialize_graph(cgraph, input);
-    rpc_msg_graph_compute_rsp response;
-    auto sock = get_socket(rpc_ctx->endpoint);
-    bool status = send_rpc_cmd(sock, RPC_CMD_GRAPH_COMPUTE, input.data(), input.size(), &response, sizeof(response));
-    GGML_ASSERT(status);
-    return (enum ggml_status)response.result;
-}
-
-static ggml_backend_i ggml_backend_rpc_interface = {
-    /* .get_name                = */ ggml_backend_rpc_name,
-    /* .free                    = */ ggml_backend_rpc_free,
-    /* .set_tensor_async        = */ NULL,
-    /* .get_tensor_async        = */ NULL,
-    /* .cpy_tensor_async        = */ NULL,
-    /* .synchronize             = */ ggml_backend_rpc_synchronize,
-    /* .graph_plan_create       = */ NULL,
-    /* .graph_plan_free         = */ NULL,
-    /* .graph_plan_update       = */ NULL,
-    /* .graph_plan_compute      = */ NULL,
-    /* .graph_compute           = */ ggml_backend_rpc_graph_compute,
-    /* .event_record            = */ NULL,
-    /* .event_wait              = */ NULL,
-};
-
-GGML_API ggml_backend_buffer_type_t ggml_backend_rpc_buffer_type(const char * endpoint) {
-    static std::mutex mutex;
-    std::lock_guard<std::mutex> lock(mutex);
-    // NOTE: buffer types are allocated and never freed; this is by design
-    static std::unordered_map<std::string, ggml_backend_buffer_type_t> buft_map;
-    auto it = buft_map.find(endpoint);
-    if (it != buft_map.end()) {
-        return it->second;
-    }
-    auto sock = get_socket(endpoint);
-    if (sock == nullptr) {
-        fprintf(stderr, "Failed to connect to %s\n", endpoint);
-        return nullptr;
-    }
-    size_t alignment = get_alignment(sock);
-    size_t max_size = get_max_size(sock);
-    ggml_backend_rpc_buffer_type_context * buft_ctx = new ggml_backend_rpc_buffer_type_context {
-        /* .endpoint  = */ endpoint,
-        /* .name      = */ "RPC[" + std::string(endpoint) + "]",
-        /* .alignment = */ alignment,
-        /* .max_size  = */ max_size
-    };
-
-    ggml_backend_buffer_type_t buft = new ggml_backend_buffer_type {
-        /* .iface   = */ ggml_backend_rpc_buffer_type_interface,
-        /* .device  = */ ggml_backend_rpc_add_device(endpoint),
-        /* .context = */ buft_ctx
-    };
-    buft_map[endpoint] = buft;
-    return buft;
-}
-
-ggml_backend_t ggml_backend_rpc_init(const char * endpoint) {
-    ggml_backend_rpc_context * ctx = new ggml_backend_rpc_context {
-        /* .endpoint  = */ endpoint,
-        /* .name      = */ "RPC[" + std::string(endpoint) + "]",
-    };
-
-    ggml_backend_t backend = new ggml_backend {
-        /* .guid      = */ ggml_backend_rpc_guid(),
-        /* .interface = */ ggml_backend_rpc_interface,
-        /* .device    = */ ggml_backend_rpc_add_device(endpoint),
-        /* .context   = */ ctx
-    };
-    return backend;
-}
-
-GGML_API bool ggml_backend_is_rpc(ggml_backend_t backend) {
-    return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_rpc_guid());
-}
-
-static void get_device_memory(const std::shared_ptr<socket_t> & sock, size_t * free, size_t * total) {
-    rpc_msg_get_device_memory_rsp response;
-    bool status = send_rpc_cmd(sock, RPC_CMD_GET_DEVICE_MEMORY, nullptr, 0, &response, sizeof(response));
-    GGML_ASSERT(status);
-    *free = response.free_mem;
-    *total = response.total_mem;
-}
-
-GGML_API void ggml_backend_rpc_get_device_memory(const char * endpoint, size_t * free, size_t * total) {
-    auto sock = get_socket(endpoint);
-    if (sock == nullptr) {
-        *free = 0;
-        *total = 0;
-        return;
-    }
-    get_device_memory(sock, free, total);
-}
-
-// RPC server-side implementation
-
-class rpc_server {
-public:
-    rpc_server(ggml_backend_t backend) : backend(backend) {}
-    ~rpc_server();
-
-    void alloc_buffer(const rpc_msg_alloc_buffer_req & request, rpc_msg_alloc_buffer_rsp & response);
-    void get_alignment(rpc_msg_get_alignment_rsp & response);
-    void get_max_size(rpc_msg_get_max_size_rsp & response);
-    bool buffer_get_base(const rpc_msg_buffer_get_base_req & request, rpc_msg_buffer_get_base_rsp & response);
-    bool free_buffer(const rpc_msg_free_buffer_req & request);
-    bool buffer_clear(const rpc_msg_buffer_clear_req & request);
-    bool set_tensor(const std::vector<uint8_t> & input);
-    bool get_tensor(const rpc_msg_get_tensor_req & request, std::vector<uint8_t> & response);
-    bool copy_tensor(const rpc_msg_copy_tensor_req & request, rpc_msg_copy_tensor_rsp & response);
-    bool graph_compute(const std::vector<uint8_t> & input, rpc_msg_graph_compute_rsp & response);
-
-private:
-    ggml_tensor * deserialize_tensor(struct ggml_context * ctx, const rpc_tensor * tensor);
-    ggml_tensor * create_node(uint64_t id,
-                              struct ggml_context * ctx,
-                              const std::unordered_map<uint64_t, const rpc_tensor*> & tensor_ptrs,
-                              std::unordered_map<uint64_t, struct ggml_tensor*> & tensor_map);
-
-
-    ggml_backend_t backend;
-    std::unordered_set<ggml_backend_buffer_t> buffers;
-};
-
-void rpc_server::alloc_buffer(const rpc_msg_alloc_buffer_req & request, rpc_msg_alloc_buffer_rsp & response) {
-    ggml_backend_buffer_type_t buft = ggml_backend_get_default_buffer_type(backend);
-    ggml_backend_buffer_t buffer = ggml_backend_buft_alloc_buffer(buft, request.size);
-    response.remote_ptr = 0;
-    response.remote_size = 0;
-    if (buffer != nullptr) {
-        response.remote_ptr = reinterpret_cast<uint64_t>(buffer);
-        response.remote_size = buffer->size;
-        GGML_PRINT_DEBUG("[%s] size: %" PRIu64 " -> remote_ptr: %" PRIx64 ", remote_size: %" PRIu64 "\n", __func__, request.size, response.remote_ptr, response.remote_size);
-        buffers.insert(buffer);
-    } else {
-        GGML_PRINT_DEBUG("[%s] size: %" PRIu64 " -> failed\n", __func__, request.size);
-    }
-}
-
-void rpc_server::get_alignment(rpc_msg_get_alignment_rsp & response) {
-    ggml_backend_buffer_type_t buft = ggml_backend_get_default_buffer_type(backend);
-    size_t alignment = ggml_backend_buft_get_alignment(buft);
-    GGML_PRINT_DEBUG("[%s] alignment: %lu\n", __func__, alignment);
-    response.alignment = alignment;
-}
-
-void rpc_server::get_max_size(rpc_msg_get_max_size_rsp & response) {
-    ggml_backend_buffer_type_t buft = ggml_backend_get_default_buffer_type(backend);
-    size_t max_size = ggml_backend_buft_get_max_size(buft);
-    GGML_PRINT_DEBUG("[%s] max_size: %lu\n", __func__, max_size);
-    response.max_size = max_size;
-}
-
-bool rpc_server::buffer_get_base(const rpc_msg_buffer_get_base_req & request, rpc_msg_buffer_get_base_rsp & response) {
-    GGML_PRINT_DEBUG("[%s] remote_ptr: %" PRIx64 "\n", __func__, request.remote_ptr);
-    ggml_backend_buffer_t buffer = reinterpret_cast<ggml_backend_buffer_t>(request.remote_ptr);
-    if (buffers.find(buffer) == buffers.end()) {
-        GGML_PRINT_DEBUG("[%s] buffer not found\n", __func__);
-        return false;
-    }
-    void * base = ggml_backend_buffer_get_base(buffer);
-    response.base_ptr = reinterpret_cast<uint64_t>(base);
-    return true;
-}
-
-bool rpc_server::free_buffer(const rpc_msg_free_buffer_req & request) {
-    GGML_PRINT_DEBUG("[%s] remote_ptr: %" PRIx64 "\n", __func__, request.remote_ptr);
-    ggml_backend_buffer_t buffer = reinterpret_cast<ggml_backend_buffer_t>(request.remote_ptr);
-    if (buffers.find(buffer) == buffers.end()) {
-        GGML_PRINT_DEBUG("[%s] buffer not found\n", __func__);
-        return false;
-    }
-    ggml_backend_buffer_free(buffer);
-    buffers.erase(buffer);
-    return true;
-}
-
-bool rpc_server::buffer_clear(const rpc_msg_buffer_clear_req & request) {
-    GGML_PRINT_DEBUG("[%s] remote_ptr: %" PRIx64 ", value: %u\n", __func__, request.remote_ptr, request.value);
-    ggml_backend_buffer_t buffer = reinterpret_cast<ggml_backend_buffer_t>(request.remote_ptr);
-    if (buffers.find(buffer) == buffers.end()) {
-        GGML_PRINT_DEBUG("[%s] buffer not found\n", __func__);
-        return false;
-    }
-    ggml_backend_buffer_clear(buffer, request.value);
-    return true;
-}
-
-ggml_tensor * rpc_server::deserialize_tensor(struct ggml_context * ctx, const rpc_tensor * tensor) {
-    ggml_tensor * result = ggml_new_tensor_4d(ctx, (ggml_type) tensor->type,
-        tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3]);
-    for (uint32_t i = 0; i < GGML_MAX_DIMS; i++) {
-        result->nb[i] = tensor->nb[i];
-    }
-    result->buffer = reinterpret_cast<ggml_backend_buffer_t>(tensor->buffer);
-    if (result->buffer && buffers.find(result->buffer) == buffers.end()) {
-        result->buffer = nullptr;
-    }
-
-    if (result->buffer) {
-        // require that the tensor data does not go beyond the buffer end
-        uint64_t tensor_size = (uint64_t) ggml_nbytes(result);
-        uint64_t buffer_start = (uint64_t) ggml_backend_buffer_get_base(result->buffer);
-        uint64_t buffer_size = (uint64_t) ggml_backend_buffer_get_size(result->buffer);
-        GGML_ASSERT(tensor->data + tensor_size >= tensor->data); // check for overflow
-        GGML_ASSERT(tensor->data >= buffer_start && tensor->data + tensor_size <= buffer_start + buffer_size);
-    }
-
-    result->op = (ggml_op) tensor->op;
-    for (uint32_t i = 0; i < GGML_MAX_OP_PARAMS / sizeof(int32_t); i++) {
-        result->op_params[i] = tensor->op_params[i];
-    }
-    result->flags = tensor->flags;
-    result->data = reinterpret_cast<void *>(tensor->data);
-    ggml_set_name(result, tensor->name);
-    return result;
-}
-
-
-bool rpc_server::set_tensor(const std::vector<uint8_t> & input) {
-    // serialization format: | rpc_tensor | offset (8 bytes) | data (size bytes) |
-    if (input.size() < sizeof(rpc_tensor) + sizeof(uint64_t)) {
-        return false;
-    }
-    const rpc_tensor * in_tensor = (const rpc_tensor *)input.data();
-    uint64_t offset;
-    memcpy(&offset, input.data() + sizeof(rpc_tensor), sizeof(offset));
-    const size_t size = input.size() - sizeof(rpc_tensor) - sizeof(offset);
-
-    struct ggml_init_params params {
-        /*.mem_size   =*/ ggml_tensor_overhead(),
-        /*.mem_buffer =*/ NULL,
-        /*.no_alloc   =*/ true,
-    };
-    struct ggml_context * ctx = ggml_init(params);
-    ggml_tensor * tensor = deserialize_tensor(ctx, in_tensor);
-    if (tensor == nullptr) {
-        GGML_PRINT_DEBUG("[%s] error deserializing tensor\n", __func__);
-        ggml_free(ctx);
-        return false;
-    }
-    GGML_PRINT_DEBUG("[%s] buffer: %p, data: %p, offset: %" PRIu64 ", size: %zu\n", __func__, (void*)tensor->buffer, tensor->data, offset, size);
-
-    // sanitize tensor->data
-    {
-        const size_t p0 = (size_t) ggml_backend_buffer_get_base(tensor->buffer);
-        const size_t p1 = p0 + ggml_backend_buffer_get_size(tensor->buffer);
-
-        if (in_tensor->data + offset < p0 || in_tensor->data + offset >= p1 || size > (p1 - in_tensor->data - offset)) {
-            GGML_ABORT("[%s] tensor->data out of bounds\n", __func__);
-        }
-    }
-
-    const void * data = input.data() + sizeof(rpc_tensor) + sizeof(offset);
-    ggml_backend_tensor_set(tensor, data, offset, size);
-    ggml_free(ctx);
-    return true;
-}
-
-bool rpc_server::get_tensor(const rpc_msg_get_tensor_req & request, std::vector<uint8_t> & response) {
-    struct ggml_init_params params {
-        /*.mem_size   =*/ ggml_tensor_overhead(),
-        /*.mem_buffer =*/ NULL,
-        /*.no_alloc   =*/ true,
-    };
-    struct ggml_context * ctx = ggml_init(params);
-    ggml_tensor * tensor = deserialize_tensor(ctx, &request.tensor);
-    if (tensor == nullptr) {
-        GGML_PRINT_DEBUG("[%s] error deserializing tensor\n", __func__);
-        ggml_free(ctx);
-        return false;
-    }
-    GGML_PRINT_DEBUG("[%s] buffer: %p, data: %p, offset: %" PRIu64 ", size: %" PRIu64 "\n", __func__, (void*)tensor->buffer, tensor->data, request.offset, request.size);
-
-    // sanitize tensor->data
-    {
-        const size_t p0 = (size_t) ggml_backend_buffer_get_base(tensor->buffer);
-        const size_t p1 = p0 + ggml_backend_buffer_get_size(tensor->buffer);
-
-        if (request.tensor.data + request.offset < p0 ||
-            request.tensor.data + request.offset >= p1 ||
-            request.size > (p1 - request.tensor.data - request.offset)) {
-                GGML_ABORT("[%s] tensor->data out of bounds\n", __func__);
-        }
-    }
-
-    response.resize(request.size, 0);
-    ggml_backend_tensor_get(tensor, response.data(), request.offset, request.size);
-    ggml_free(ctx);
-    return true;
-}
-
-bool rpc_server::copy_tensor(const rpc_msg_copy_tensor_req & request, rpc_msg_copy_tensor_rsp & response) {
-    struct ggml_init_params params {
-        /*.mem_size   =*/ 2*ggml_tensor_overhead(),
-        /*.mem_buffer =*/ NULL,
-        /*.no_alloc   =*/ true,
-    };
-    struct ggml_context * ctx = ggml_init(params);
-    ggml_tensor * src = deserialize_tensor(ctx, &request.src);
-    ggml_tensor * dst = deserialize_tensor(ctx, &request.dst);
-    if (src == nullptr || dst == nullptr) {
-        GGML_PRINT_DEBUG("[%s] error deserializing tensors\n", __func__);
-        ggml_free(ctx);
-        return false;
-    }
-    GGML_PRINT_DEBUG("[%s] src->buffer: %p, dst->buffer: %p\n", __func__, (void*)src->buffer, (void*)dst->buffer);
-    response.result = ggml_backend_buffer_copy_tensor(src, dst);
-    ggml_free(ctx);
-    return true;
-}
-
-ggml_tensor * rpc_server::create_node(uint64_t id,
-                                      struct ggml_context * ctx,
-                                      const std::unordered_map<uint64_t, const rpc_tensor*> & tensor_ptrs,
-                                      std::unordered_map<uint64_t, struct ggml_tensor*> & tensor_map) {
-    if (id == 0) {
-        return nullptr;
-    }
-    if (tensor_map.find(id) != tensor_map.end()) {
-        return tensor_map[id];
-    }
-    const rpc_tensor * tensor = tensor_ptrs.at(id);
-    struct ggml_tensor * result = deserialize_tensor(ctx, tensor);
-    if (result == nullptr) {
-        return nullptr;
-    }
-    tensor_map[id] = result;
-    for (int i = 0; i < GGML_MAX_SRC; i++) {
-        result->src[i] = create_node(tensor->src[i], ctx, tensor_ptrs, tensor_map);
-    }
-    result->view_src = create_node(tensor->view_src, ctx, tensor_ptrs, tensor_map);
-    result->view_offs = tensor->view_offs;
-    return result;
-}
-
-bool rpc_server::graph_compute(const std::vector<uint8_t> & input, rpc_msg_graph_compute_rsp & response) {
-    // serialization format:
-    // | n_nodes (4 bytes) | nodes (n_nodes * sizeof(uint64_t) | n_tensors (4 bytes) | tensors (n_tensors * sizeof(rpc_tensor)) |
-    if (input.size() < sizeof(uint32_t)) {
-        return false;
-    }
-    uint32_t n_nodes;
-    memcpy(&n_nodes, input.data(), sizeof(n_nodes));
-    if (input.size() < sizeof(uint32_t) + n_nodes*sizeof(uint64_t) + sizeof(uint32_t)) {
-        return false;
-    }
-    const uint64_t * nodes = (const uint64_t *)(input.data() + sizeof(n_nodes));
-    uint32_t n_tensors;
-    memcpy(&n_tensors, input.data() + sizeof(n_nodes) + n_nodes*sizeof(uint64_t), sizeof(n_tensors));
-    if (input.size() < sizeof(uint32_t) + n_nodes*sizeof(uint64_t) + sizeof(uint32_t) + n_tensors*sizeof(rpc_tensor)) {
-        return false;
-    }
-    const rpc_tensor * tensors = (const rpc_tensor *)(input.data() + sizeof(n_nodes) + n_nodes*sizeof(uint64_t) + sizeof(n_tensors));
-    GGML_PRINT_DEBUG("[%s] n_nodes: %u, n_tensors: %u\n", __func__, n_nodes, n_tensors);
-
-    size_t buf_size = ggml_tensor_overhead()*(n_nodes + n_tensors) + ggml_graph_overhead_custom(n_nodes, false);
-    struct ggml_init_params params = {
-        /*.mem_size   =*/ buf_size,
-        /*.mem_buffer =*/ NULL,
-        /*.no_alloc   =*/ true,
-    };
-    struct ggml_context * ctx = ggml_init(params);
-    struct ggml_cgraph * graph = ggml_new_graph_custom(ctx, n_nodes, false);
-    graph->n_nodes = n_nodes;
-    std::unordered_map<uint64_t, const rpc_tensor*> tensor_ptrs;
-    for (uint32_t i = 0; i < n_tensors; i++) {
-        tensor_ptrs[tensors[i].id] = &tensors[i];
-    }
-    std::unordered_map<uint64_t, ggml_tensor*> tensor_map;
-    for (uint32_t i = 0; i < n_nodes; i++) {
-        int64_t id;
-        memcpy(&id, &nodes[i], sizeof(id));
-        graph->nodes[i] = create_node(id, ctx, tensor_ptrs, tensor_map);
-    }
-    ggml_status status = ggml_backend_graph_compute(backend, graph);
-    response.result = status;
-    ggml_free(ctx);
-    return true;
-}
-
-rpc_server::~rpc_server() {
-    for (auto buffer : buffers) {
-        ggml_backend_buffer_free(buffer);
-    }
-}
-
-static void rpc_serve_client(ggml_backend_t backend, sockfd_t sockfd, size_t free_mem, size_t total_mem) {
-    rpc_server server(backend);
-    while (true) {
-        uint8_t cmd;
-        if (!recv_data(sockfd, &cmd, 1)) {
-            break;
-        }
-        if (cmd >= RPC_CMD_COUNT) {
-            // fail fast if the command is invalid
-            fprintf(stderr, "Unknown command: %d\n", cmd);
-            break;
-        }
-        switch (cmd) {
-            case RPC_CMD_ALLOC_BUFFER: {
-                rpc_msg_alloc_buffer_req request;
-                if (!recv_msg(sockfd, &request, sizeof(request))) {
-                    return;
-                }
-                rpc_msg_alloc_buffer_rsp response;
-                server.alloc_buffer(request, response);
-                if (!send_msg(sockfd, &response, sizeof(response))) {
-                    return;
-                }
-                break;
-            }
-            case RPC_CMD_GET_ALIGNMENT: {
-                if (!recv_msg(sockfd, nullptr, 0)) {
-                    return;
-                }
-                rpc_msg_get_alignment_rsp response;
-                server.get_alignment(response);
-                if (!send_msg(sockfd, &response, sizeof(response))) {
-                    return;
-                }
-                break;
-            }
-            case RPC_CMD_GET_MAX_SIZE: {
-                if (!recv_msg(sockfd, nullptr, 0)) {
-                    return;
-                }
-                rpc_msg_get_max_size_rsp response;
-                server.get_max_size(response);
-                if (!send_msg(sockfd, &response, sizeof(response))) {
-                    return;
-                }
-                break;
-            }
-            case RPC_CMD_BUFFER_GET_BASE: {
-                rpc_msg_buffer_get_base_req request;
-                if (!recv_msg(sockfd, &request, sizeof(request))) {
-                    return;
-                }
-                rpc_msg_buffer_get_base_rsp response;
-                if (!server.buffer_get_base(request, response)) {
-                    return;
-                }
-                if (!send_msg(sockfd, &response, sizeof(response))) {
-                    return;
-                }
-                break;
-            }
-            case RPC_CMD_FREE_BUFFER: {
-                rpc_msg_free_buffer_req request;
-                if (!recv_msg(sockfd, &request, sizeof(request))) {
-                    return;
-                }
-                if (!server.free_buffer(request)) {
-                    return;
-                }
-                if (!send_msg(sockfd, nullptr, 0)) {
-                    return;
-                }
-                break;
-            }
-            case RPC_CMD_BUFFER_CLEAR: {
-                rpc_msg_buffer_clear_req request;
-                if (!recv_msg(sockfd, &request, sizeof(request))) {
-                    return;
-                }
-                if (!server.buffer_clear(request)) {
-                    return;
-                }
-                if (!send_msg(sockfd, nullptr, 0)) {
-                    return;
-                }
-                break;
-            }
-            case RPC_CMD_SET_TENSOR: {
-                std::vector<uint8_t> input;
-                if (!recv_msg(sockfd, input)) {
-                    return;
-                }
-                if (!server.set_tensor(input)) {
-                    return;
-                }
-                if (!send_msg(sockfd, nullptr, 0)) {
-                    return;
-                }
-                break;
-            }
-            case RPC_CMD_GET_TENSOR: {
-                rpc_msg_get_tensor_req request;
-                if (!recv_msg(sockfd, &request, sizeof(request))) {
-                    return;
-                }
-                std::vector<uint8_t> response;
-                if (!server.get_tensor(request, response)) {
-                    return;
-                }
-                if (!send_msg(sockfd, response.data(), response.size())) {
-                    return;
-                }
-                break;
-            }
-            case RPC_CMD_COPY_TENSOR: {
-                rpc_msg_copy_tensor_req request;
-                if (!recv_msg(sockfd, &request, sizeof(request))) {
-                    return;
-                }
-                rpc_msg_copy_tensor_rsp response;
-                if (!server.copy_tensor(request, response)) {
-                    return;
-                }
-                if (!send_msg(sockfd, &response, sizeof(response))) {
-                    return;
-                }
-                break;
-            }
-            case RPC_CMD_GRAPH_COMPUTE: {
-                std::vector<uint8_t> input;
-                if (!recv_msg(sockfd, input)) {
-                    return;
-                }
-                rpc_msg_graph_compute_rsp response;
-                if (!server.graph_compute(input, response)) {
-                    return;
-                }
-                if (!send_msg(sockfd, &response, sizeof(response))) {
-                    return;
-                }
-                break;
-            }
-            case RPC_CMD_GET_DEVICE_MEMORY: {
-                if (!recv_msg(sockfd, nullptr, 0)) {
-                    return;
-                }
-                rpc_msg_get_device_memory_rsp response;
-                response.free_mem = free_mem;
-                response.total_mem = total_mem;
-                if (!send_msg(sockfd, &response, sizeof(response))) {
-                    return;
-                }
-                break;
-            }
-            default: {
-                fprintf(stderr, "Unknown command: %d\n", cmd);
-                return;
-            }
-        }
-    }
-}
-
-void ggml_backend_rpc_start_server(ggml_backend_t backend, const char * endpoint, size_t free_mem, size_t total_mem) {
-    std::string host;
-    int port;
-    if (!parse_endpoint(endpoint, host, port)) {
-        return;
-    }
-#ifdef _WIN32
-    {
-        WSADATA wsaData;
-        int res = WSAStartup(MAKEWORD(2, 2), &wsaData);
-        if (res != 0) {
-            fprintf(stderr, "WSAStartup failed: %d\n", res);
-            return;
-        }
-    }
-#endif
-    auto server_socket = create_server_socket(host.c_str(), port);
-    if (server_socket == nullptr) {
-        fprintf(stderr, "Failed to create server socket\n");
-        return;
-    }
-    while (true) {
-        auto client_socket = socket_accept(server_socket->fd);
-        if (client_socket == nullptr) {
-            fprintf(stderr, "Failed to accept client connection\n");
-            return;
-        }
-        printf("Accepted client connection, free_mem=%zu, total_mem=%zu\n", free_mem, total_mem);
-        fflush(stdout);
-        rpc_serve_client(backend, client_socket->fd, free_mem, total_mem);
-        printf("Client connection closed\n");
-        fflush(stdout);
-    }
-#ifdef _WIN32
-    WSACleanup();
-#endif
-}
-
-// device interface
-
-struct ggml_backend_rpc_device_context {
-    std::string endpoint;
-    std::string name;
-};
-
-static const char * ggml_backend_rpc_device_get_name(ggml_backend_dev_t dev) {
-    ggml_backend_rpc_device_context * ctx = (ggml_backend_rpc_device_context *)dev->context;
-
-    return ctx->name.c_str();
-}
-
-static const char * ggml_backend_rpc_device_get_description(ggml_backend_dev_t dev) {
-    ggml_backend_rpc_device_context * ctx = (ggml_backend_rpc_device_context *)dev->context;
-
-    return ctx->name.c_str();
-}
-
-static void ggml_backend_rpc_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
-    ggml_backend_rpc_device_context * ctx = (ggml_backend_rpc_device_context *)dev->context;
-
-    ggml_backend_rpc_get_device_memory(ctx->endpoint.c_str(), free, total);
-
-    UNUSED(dev);
-}
-
-static enum ggml_backend_dev_type ggml_backend_rpc_device_get_type(ggml_backend_dev_t dev) {
-    // TODO: obtain value from the server
-    return GGML_BACKEND_DEVICE_TYPE_GPU;
-
-    UNUSED(dev);
-}
-
-static void ggml_backend_rpc_device_get_props(ggml_backend_dev_t dev, struct ggml_backend_dev_props * props) {
-    props->name        = ggml_backend_rpc_device_get_name(dev);
-    props->description = ggml_backend_rpc_device_get_description(dev);
-    props->type        = ggml_backend_rpc_device_get_type(dev);
-    ggml_backend_rpc_device_get_memory(dev, &props->memory_free, &props->memory_total);
-    props->caps = {
-        /* .async                 = */ false,
-        /* .host_buffer           = */ false,
-        /* .buffer_from_host_ptr  = */ false,
-        /* .events                = */ false,
-    };
-}
-
-static ggml_backend_t ggml_backend_rpc_device_init(ggml_backend_dev_t dev, const char * params) {
-    ggml_backend_rpc_device_context * ctx = (ggml_backend_rpc_device_context *)dev->context;
-
-    return ggml_backend_rpc_init(ctx->endpoint.c_str());
-
-    UNUSED(params);
-}
-
-static ggml_backend_buffer_type_t ggml_backend_rpc_device_get_buffer_type(ggml_backend_dev_t dev) {
-    ggml_backend_rpc_device_context * ctx = (ggml_backend_rpc_device_context *)dev->context;
-
-    return ggml_backend_rpc_buffer_type(ctx->endpoint.c_str());
-
-    UNUSED(dev);
-}
-
-static bool ggml_backend_rpc_device_supports_op(ggml_backend_dev_t dev, const struct ggml_tensor * op) {
-    UNUSED(dev);
-    UNUSED(op);
-    //TODO: call the remote backend and cache the results
-    return true;
-}
-
-static bool ggml_backend_rpc_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
-    if (!buft || buft->iface.get_name != ggml_backend_rpc_buffer_type_name) {
-        return false;
-    }
-    ggml_backend_rpc_buffer_type_context * buft_ctx = (ggml_backend_rpc_buffer_type_context *)buft->context;
-    ggml_backend_rpc_device_context * dev_ctx = (ggml_backend_rpc_device_context *)dev->context;
-    return buft_ctx->endpoint == dev_ctx->endpoint;
-}
-
-static const struct ggml_backend_device_i ggml_backend_rpc_device_i = {
-    /* .get_name             = */ ggml_backend_rpc_device_get_name,
-    /* .get_description      = */ ggml_backend_rpc_device_get_description,
-    /* .get_memory           = */ ggml_backend_rpc_device_get_memory,
-    /* .get_type             = */ ggml_backend_rpc_device_get_type,
-    /* .get_props            = */ ggml_backend_rpc_device_get_props,
-    /* .init_backend         = */ ggml_backend_rpc_device_init,
-    /* .get_buffer_type      = */ ggml_backend_rpc_device_get_buffer_type,
-    /* .get_host_buffer_type = */ NULL,
-    /* .buffer_from_host_ptr = */ NULL,
-    /* .supports_op          = */ ggml_backend_rpc_device_supports_op,
-    /* .supports_buft        = */ ggml_backend_rpc_device_supports_buft,
-    /* .offload_op           = */ NULL,
-    /* .event_new            = */ NULL,
-    /* .event_free           = */ NULL,
-    /* .event_synchronize    = */ NULL,
-};
-
-// backend reg interface
-
-static const char * ggml_backend_rpc_reg_get_name(ggml_backend_reg_t reg) {
-    return "RPC";
-
-    UNUSED(reg);
-}
-
-static size_t ggml_backend_rpc_reg_get_device_count(ggml_backend_reg_t reg) {
-    return 0;
-
-    UNUSED(reg);
-}
-
-static ggml_backend_dev_t ggml_backend_rpc_reg_get_device(ggml_backend_reg_t reg, size_t index) {
-    GGML_ABORT("The RPC backend does not have enumerated devices - use ggml_backend_add_device instead");
-
-    UNUSED(reg);
-    UNUSED(index);
-}
-
-static void * ggml_backend_rpc_get_proc_address(ggml_backend_reg_t reg, const char * name) {
-    if (std::strcmp(name, "ggml_backend_rpc_add_device") == 0) {
-        return (void *)ggml_backend_rpc_add_device;
-    }
-    return NULL;
-
-    UNUSED(reg);
-}
-
-static const struct ggml_backend_reg_i ggml_backend_rpc_reg_i = {
-    /* .get_name         = */ ggml_backend_rpc_reg_get_name,
-    /* .get_device_count = */ ggml_backend_rpc_reg_get_device_count,
-    /* .get_device       = */ ggml_backend_rpc_reg_get_device,
-    /* .get_proc_address = */ ggml_backend_rpc_get_proc_address,
-};
-
-ggml_backend_reg_t ggml_backend_rpc_reg(void) {
-    static struct ggml_backend_reg ggml_backend_rpc_reg = {
-        /* .iface   = */ ggml_backend_rpc_reg_i,
-        /* .context = */ NULL,
-    };
-
-    return &ggml_backend_rpc_reg;
-}
-
-ggml_backend_dev_t ggml_backend_rpc_add_device(const char * endpoint) {
-    static std::unordered_map<std::string, ggml_backend_dev_t> dev_map;
-
-    static std::mutex mutex;
-    std::lock_guard<std::mutex> lock(mutex);
-
-    if (dev_map.find(endpoint) != dev_map.end()) {
-        return dev_map[endpoint];
-    }
-
-    ggml_backend_rpc_device_context * ctx = new ggml_backend_rpc_device_context {
-        /* .endpoint = */ endpoint,
-        /* .name     = */ "RPC[" + std::string(endpoint) + "]",
-    };
-
-    ggml_backend_dev_t dev = new ggml_backend_device {
-        /* .iface   = */ ggml_backend_rpc_device_i,
-        /* .reg     = */ ggml_backend_rpc_reg(),
-        /* .context = */ ctx,
-    };
-
-    dev_map[endpoint] = dev;
-
-    return dev;
-}

ggml/src/kompute-shaders/common.comp

@@ -1,102 +0,0 @@
-#extension GL_EXT_shader_16bit_storage: require
-#extension GL_EXT_shader_8bit_storage: require
-#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
-#extension GL_EXT_shader_explicit_arithmetic_types_int8: require
-#extension GL_EXT_shader_explicit_arithmetic_types_int16: require
-#extension GL_EXT_control_flow_attributes: enable
-#extension GL_KHR_shader_subgroup_arithmetic : require
-#extension GL_EXT_debug_printf : enable
-
-#define QK4_0 32
-#define QK4_1 32
-
-#define GELU_COEF_A 0.044715
-#define SQRT_2_OVER_PI 0.79788456080286535587989211986876
-#define TWOPI_F 6.283185307179586f
-
-#define QK_K 256
-
-#define u8BufToU16(buf, idx) (((uint16_t(buf[idx + 1]) << 8)) | buf[idx])
-#define u8BufToFloat16(buf, idx) uint16BitsToHalf u8BufToU16(buf, idx)
-#define u8BufToU32(buf, idx) (((uint32_t u8BufToU16(buf, idx + 2) << 8 | buf[idx + 1]) << 8) | buf[idx])
-#define u8BufToFloat(buf, idx) uintBitsToFloat u8BufToU32(buf, idx)
-
-#define sizeof_block_q4_0 0x12
-struct block_q4_0 {
-    float16_t d;
-    uint8_t qs[QK4_0 / 2];
-};
-mat4 dequantize_q4_0(const block_q4_0 xb, uint il) {
-    const float d1 = il != 0 ? (xb.d / 16.f) : xb.d;
-    const float d2 = d1 / 256.f;
-    const float md = -8.f * xb.d;
-    const uint16_t mask0 = il != 0 ? uint16_t(0x00F0) : uint16_t(0x000F);
-    const uint16_t mask1 = mask0 << 8;
-
-    mat4 reg;
-    for (int i=0;i<8;i++) {
-        uint16_t b = (uint16_t(xb.qs[2 * i + 1]) << 8) | uint16_t(xb.qs[2 * i]);
-        reg[i/2][2*(i%2)+0] = d1 * (b & mask0) + md;
-        reg[i/2][2*(i%2)+1] = d2 * (b & mask1) + md;
-    }
-    return reg;
-}
-
-#define sizeof_block_q4_1 0x14
-struct block_q4_1 {
-    float16_t d;
-    float16_t m;
-    uint8_t qs[QK4_1 / 2];
-};
-mat4 dequantize_q4_1(const block_q4_1 xb, uint il) {
-    const float d1 = il != 0 ? (xb.d / 16.f) : xb.d;
-    const float d2 = d1 / 256.f;
-    const float  m = xb.m;
-    const uint16_t mask0 = il != 0 ? uint16_t(0x00F0) : uint16_t(0x000F);
-    const uint16_t mask1 = mask0 << 8;
-
-    mat4 reg;
-    for (int i=0;i<8;i++) {
-        uint16_t b = (uint16_t(xb.qs[2 * i + 1]) << 8) | uint16_t(xb.qs[2 * i]);
-        reg[i/2][2*(i%2)+0] = ((b & mask0) * d1) + m;
-        reg[i/2][2*(i%2)+1] = ((b & mask1) * d2) + m;
-    }
-    return reg;
-}
-
-#define sizeof_block_q6_k 210
-struct block_q6_k {
-    uint8_t ql[QK_K/2];      // quants, lower 4 bits
-    uint8_t qh[QK_K/4];      // quants, upper 2 bits
-    int8_t  scales[QK_K/16]; // scales, quantized with 8 bits
-    float16_t d;             // super-block scale
-};
-mat4 dequantize_q6_k(const block_q6_k xb, uint il) {
-    const float16_t d_all = xb.d;
-
-    const uint qlIndex = 64*(il/8) + 32*((il/2)&1) + 16*(il&1);
-    const uint qhIndex = 32*(il/8) + 16*(il&1);
-    float16_t sc = xb.scales[(il%2) + 2 * ((il/2))];
-    il = (il/2) & 3;
-
-    const uint16_t  kmask1 = il>1 ? uint16_t(il>2 ? 192 : 48) : uint16_t(il>0 ? 12 : 3);
-    const uint16_t  kmask2 = il>1 ? uint8_t(0xF0)             : uint8_t(0x0F);
-    const float16_t coef   = il>1 ? float16_t(1.f/16.f)       : float16_t(1.f);
-    const float16_t ml = float16_t(d_all * sc * 32.f);
-    const float16_t dl = float16_t(d_all * sc * coef);
-    mat4 reg;
-    for (int i = 0; i < 16; ++i) {
-        const float16_t q = (il&1) != 0 ? ((xb.ql[qlIndex + i] & kmask2) | ((xb.qh[qhIndex + i] & kmask1) << 2))
-                                        : ((xb.ql[qlIndex + i] & kmask2) | ((xb.qh[qhIndex + i] & kmask1) << 4));
-        reg[i/4][i%4] = dl * q - ml;
-    }
-    return reg;
-}
-
-
-#define QK8_0 32
-// struct block_q8_0 {
-//     float16_t d;         // delta
-//     int8_t    qs[QK8_0]; // quants
-// };
-#define sizeof_block_q8_0 34

ggml/src/kompute-shaders/op_add.comp

@@ -1,58 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-layout(local_size_x = 1024) in;
-
-layout(binding = 0) buffer restrict readonly tensorInA { float inA[]; };
-layout(binding = 1) buffer restrict readonly tensorInB { float inB[]; };
-layout(binding = 2) buffer restrict writeonly tensorOut { float out_[]; };
-
-layout(push_constant) uniform PushConstants {
-    uint inAOff;
-    uint inBOff;
-    uint outOff;
-    int ne00;
-    int nb00;
-    int nb01;
-    int nb02;
-    int nb03;
-    int ne10;
-    int ne11;
-    int ne12;
-    int ne13;
-    int nb10;
-    int nb11;
-    int nb12;
-    int nb13;
-    int ne0;
-    int nb0;
-    int nb1;
-    int nb2;
-    int nb3;
-  //int offs; // TODO: needed for GGML_OP_ACC, see metal code
-} pcs;
-
-// general-purpose kernel for addition of two tensors
-// pros: works for non-contiguous tensors, supports broadcast across dims 1, 2 and 3
-// cons: not very efficient
-void main() {
-    const uint i03 = gl_WorkGroupID.z;
-    const uint i02 = gl_WorkGroupID.y;
-    const uint i01 = gl_WorkGroupID.x;
-
-    const uint i13 = i03 % pcs.ne13;
-    const uint i12 = i02 % pcs.ne12;
-    const uint i11 = i01 % pcs.ne11;
-
-    int offs = 0; // TMP (see above)
-
-    uint src0_off = uint((i03*pcs.nb03 + i02*pcs.nb02 + i01*pcs.nb01 + offs) / 4);
-    uint src1_off = uint((i13*pcs.nb13 + i12*pcs.nb12 + i11*pcs.nb11       ) / 4);
-    uint dst_off  = uint((i03*pcs.nb3  + i02*pcs.nb2  + i01*pcs.nb1  + offs) / 4);
-
-    for (uint i0 = gl_LocalInvocationID.x; i0 < pcs.ne0; i0 += gl_WorkGroupSize.x) {
-        const uint i10 = i0 % pcs.ne10;
-        out_[pcs.outOff + dst_off + i0] = inA[pcs.inAOff + src0_off + i0] + inB[pcs.inBOff + src1_off + i10];
-    }
-}

ggml/src/kompute-shaders/op_addrow.comp

@@ -1,25 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-layout(local_size_x = 1) in;
-
-layout(binding = 0) buffer restrict readonly tensorInA { float inA[]; };
-layout(binding = 1) buffer restrict readonly tensorInB { float inB[]; };
-layout(binding = 2) buffer restrict writeonly tensorOut { float out_[]; };
-
-layout(push_constant) uniform PushConstants {
-    uint inAOff;
-    uint inBOff;
-    uint outOff;
-    uint row;
-} pcs;
-
-void main() {
-    const uint baseIndex = gl_WorkGroupID.x * 4;
-
-    for (uint x = 0; x < 4; x++) {
-        const uint i = baseIndex + x;
-        out_[i + pcs.outOff] = inA[i + pcs.inAOff] + inB[(i % pcs.row) + pcs.inBOff];
-    }
-}

ggml/src/kompute-shaders/op_cpy_f16_f16.comp

@@ -1,52 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-#define IN_TYPE float16_t
-#define IN_TYPE_SIZE 2
-#define OUT_TYPE float16_t
-#define OUT_TYPE_SIZE 2
-
-layout(local_size_x = 1024) in;
-
-layout (binding = 0) readonly buffer tensorIn { IN_TYPE in_[]; };
-layout (binding = 1) writeonly buffer tensorOut { OUT_TYPE out_[]; };
-
-layout (push_constant) uniform parameter {
-    uint inOff;
-    uint outOff;
-    int ne00;
-    int ne01;
-    int ne02;
-    uint nb00;
-    uint nb01;
-    uint nb02;
-    uint nb03;
-    int ne0;
-    int ne1;
-    int ne2;
-    uint nb0;
-    uint nb1;
-    uint nb2;
-    uint nb3;
-} pcs;
-
-void main() {
-    const uint i03 = gl_WorkGroupID.z;
-    const uint i02 = gl_WorkGroupID.y;
-    const uint i01 = gl_WorkGroupID.x;
-
-    const int n = int(i03)*pcs.ne02*pcs.ne01*pcs.ne00 + int(i02)*pcs.ne01*pcs.ne00 + int(i01)*pcs.ne00;
-
-    const int i3 = n / (pcs.ne2*pcs.ne1*pcs.ne0);
-    const int i2 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0) / (pcs.ne1*pcs.ne0);
-    const int i1 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0) / pcs.ne0;
-    const int i0 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0 - i1*pcs.ne0);
-
-    const uint dst_data = (i3*pcs.nb3 + i2*pcs.nb2 + i1*pcs.nb1 + i0*pcs.nb0) / OUT_TYPE_SIZE + pcs.outOff; // Based from out_
-
-    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
-        const uint src = uint((i03*pcs.nb03 + i02*pcs.nb02 + i01*pcs.nb01 + i00*pcs.nb00) / IN_TYPE_SIZE) + pcs.inOff; // Based from in_
-        out_[dst_data+i00] = OUT_TYPE(in_[src]);
-    }
-}

ggml/src/kompute-shaders/op_cpy_f16_f32.comp

@@ -1,52 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-#define IN_TYPE float16_t
-#define IN_TYPE_SIZE 2
-#define OUT_TYPE float
-#define OUT_TYPE_SIZE 4
-
-layout(local_size_x = 1024) in;
-
-layout (binding = 0) readonly buffer tensorIn { IN_TYPE in_[]; };
-layout (binding = 1) writeonly buffer tensorOut { OUT_TYPE out_[]; };
-
-layout (push_constant) uniform parameter {
-    uint inOff;
-    uint outOff;
-    int ne00;
-    int ne01;
-    int ne02;
-    uint nb00;
-    uint nb01;
-    uint nb02;
-    uint nb03;
-    int ne0;
-    int ne1;
-    int ne2;
-    uint nb0;
-    uint nb1;
-    uint nb2;
-    uint nb3;
-} pcs;
-
-void main() {
-    const uint i03 = gl_WorkGroupID.z;
-    const uint i02 = gl_WorkGroupID.y;
-    const uint i01 = gl_WorkGroupID.x;
-
-    const int n = int(i03)*pcs.ne02*pcs.ne01*pcs.ne00 + int(i02)*pcs.ne01*pcs.ne00 + int(i01)*pcs.ne00;
-
-    const int i3 = n / (pcs.ne2*pcs.ne1*pcs.ne0);
-    const int i2 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0) / (pcs.ne1*pcs.ne0);
-    const int i1 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0) / pcs.ne0;
-    const int i0 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0 - i1*pcs.ne0);
-
-    const uint dst_data = (i3*pcs.nb3 + i2*pcs.nb2 + i1*pcs.nb1 + i0*pcs.nb0) / OUT_TYPE_SIZE + pcs.outOff; // Based from out_
-
-    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
-        const uint src = uint((i03*pcs.nb03 + i02*pcs.nb02 + i01*pcs.nb01 + i00*pcs.nb00) / IN_TYPE_SIZE) + pcs.inOff; // Based from in_
-        out_[dst_data+i00] = OUT_TYPE(in_[src]);
-    }
-}

ggml/src/kompute-shaders/op_cpy_f32_f16.comp

@@ -1,52 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-#define IN_TYPE float
-#define IN_TYPE_SIZE 4
-#define OUT_TYPE float16_t
-#define OUT_TYPE_SIZE 2
-
-layout(local_size_x = 1024) in;
-
-layout (binding = 0) readonly buffer tensorIn { IN_TYPE in_[]; };
-layout (binding = 1) writeonly buffer tensorOut { OUT_TYPE out_[]; };
-
-layout (push_constant) uniform parameter {
-    uint inOff;
-    uint outOff;
-    int ne00;
-    int ne01;
-    int ne02;
-    uint nb00;
-    uint nb01;
-    uint nb02;
-    uint nb03;
-    int ne0;
-    int ne1;
-    int ne2;
-    uint nb0;
-    uint nb1;
-    uint nb2;
-    uint nb3;
-} pcs;
-
-void main() {
-    const uint i03 = gl_WorkGroupID.z;
-    const uint i02 = gl_WorkGroupID.y;
-    const uint i01 = gl_WorkGroupID.x;
-
-    const int n = int(i03)*pcs.ne02*pcs.ne01*pcs.ne00 + int(i02)*pcs.ne01*pcs.ne00 + int(i01)*pcs.ne00;
-
-    const int i3 = n / (pcs.ne2*pcs.ne1*pcs.ne0);
-    const int i2 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0) / (pcs.ne1*pcs.ne0);
-    const int i1 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0) / pcs.ne0;
-    const int i0 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0 - i1*pcs.ne0);
-
-    const uint dst_data = (i3*pcs.nb3 + i2*pcs.nb2 + i1*pcs.nb1 + i0*pcs.nb0) / OUT_TYPE_SIZE + pcs.outOff; // Based from out_
-
-    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
-        const uint src = uint((i03*pcs.nb03 + i02*pcs.nb02 + i01*pcs.nb01 + i00*pcs.nb00) / IN_TYPE_SIZE) + pcs.inOff; // Based from in_
-        out_[dst_data+i00] = OUT_TYPE(in_[src]);
-    }
-}

ggml/src/kompute-shaders/op_cpy_f32_f32.comp

@@ -1,52 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-#define IN_TYPE float
-#define IN_TYPE_SIZE 4
-#define OUT_TYPE float
-#define OUT_TYPE_SIZE 4
-
-layout(local_size_x = 1024) in;
-
-layout (binding = 0) readonly buffer tensorIn { IN_TYPE in_[]; };
-layout (binding = 1) writeonly buffer tensorOut { OUT_TYPE out_[]; };
-
-layout (push_constant) uniform parameter {
-    uint inOff;
-    uint outOff;
-    int ne00;
-    int ne01;
-    int ne02;
-    uint nb00;
-    uint nb01;
-    uint nb02;
-    uint nb03;
-    int ne0;
-    int ne1;
-    int ne2;
-    uint nb0;
-    uint nb1;
-    uint nb2;
-    uint nb3;
-} pcs;
-
-void main() {
-    const uint i03 = gl_WorkGroupID.z;
-    const uint i02 = gl_WorkGroupID.y;
-    const uint i01 = gl_WorkGroupID.x;
-
-    const int n = int(i03)*pcs.ne02*pcs.ne01*pcs.ne00 + int(i02)*pcs.ne01*pcs.ne00 + int(i01)*pcs.ne00;
-
-    const int i3 = n / (pcs.ne2*pcs.ne1*pcs.ne0);
-    const int i2 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0) / (pcs.ne1*pcs.ne0);
-    const int i1 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0) / pcs.ne0;
-    const int i0 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0 - i1*pcs.ne0);
-
-    const uint dst_data = (i3*pcs.nb3 + i2*pcs.nb2 + i1*pcs.nb1 + i0*pcs.nb0) / OUT_TYPE_SIZE + pcs.outOff; // Based from out_
-
-    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
-        const uint src = uint((i03*pcs.nb03 + i02*pcs.nb02 + i01*pcs.nb01 + i00*pcs.nb00) / IN_TYPE_SIZE) + pcs.inOff; // Based from in_
-        out_[dst_data+i00] = OUT_TYPE(in_[src]);
-    }
-}

ggml/src/kompute-shaders/op_diagmask.comp

@@ -1,30 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-layout(local_size_x = 1) in;
-
-layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
-layout(binding = 1) buffer restrict writeonly tensorOut { float out_[]; };
-
-layout(push_constant) uniform PushConstants {
-    uint inOff;
-    uint outOff;
-    uint n_past;
-    int ne00;
-    int ne01;
-} pcs;
-
-void main() {
-    const uint i02 = gl_WorkGroupID.z;
-    const uint i01 = gl_WorkGroupID.y;
-    const uint i00 = gl_WorkGroupID.x;
-
-    const uint index = i02*pcs.ne01*pcs.ne00 + i01*pcs.ne00 + i00;
-
-    if (i00 > pcs.n_past + i01) {
-        out_[index + pcs.outOff] = uintBitsToFloat(0xFF800000);
-    } else {
-        out_[index + pcs.outOff] = in_[index + pcs.inOff];
-    }
-}

ggml/src/kompute-shaders/op_gelu.comp

@@ -1,22 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-layout(local_size_x = 1) in;
-
-layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
-layout(binding = 1) buffer restrict writeonly tensorOut { float out_[]; };
-layout(push_constant) uniform PushConstants {
-    uint inOff;
-    uint outOff;
-} pcs;
-
-void main() {
-    const uint baseIndex = gl_WorkGroupID.x * 8;
-
-    for (uint x = 0; x < 8; x++) {
-        const uint i = baseIndex + x;
-        const float y = in_[i + pcs.inOff];
-        out_[i + pcs.outOff] = 0.5*y*(1.0 + tanh(clamp(SQRT_2_OVER_PI*y*(1.0 + GELU_COEF_A*y*y), -15.0, 15.0)));
-    }
-}

ggml/src/kompute-shaders/op_getrows.comp

@@ -1,17 +0,0 @@
-void main() {
-    const uint i = gl_WorkGroupID.x;
-    const int r = inB[i + pcs.inBOff];
-
-    int z = 0;
-    for (uint ind = gl_LocalInvocationID.x; ind < pcs.ne00/16; ind += gl_WorkGroupSize.x) {
-        const uint inIndex = (r * pcs.nb01 + pcs.inAOff) + ind/NL * SIZE_OF_BLOCK;
-        const mat4 result = dequantize_block(inIndex, ind%NL);
-        for (uint j = 0; j < 4; ++j) {
-            for (uint k = 0; k < 4; ++k) {
-                const uint outIndex = i * pcs.nb1/BYTES_FOR_TYPE + pcs.outOff + z;
-                out_[outIndex] = result[j][k];
-                ++z;
-            }
-        }
-    }
-}

ggml/src/kompute-shaders/op_getrows_f16.comp

@@ -1,31 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-layout(local_size_x = 1) in;
-
-layout (binding = 0) readonly buffer tensorInA { float16_t inA[]; };
-layout (binding = 1) readonly buffer tensorInB { int inB[]; };
-layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
-
-layout (push_constant) uniform parameter {
-    uint inAOff;
-    uint inBOff;
-    uint outOff;
-    int ne00;
-    int nb01;
-    int nb1;
-} pcs;
-
-void dequantize_row_f16(uint x /*Based from inA unaligned*/, uint y /*Based from out_*/, int k) {
-    for (int j = 0; j < k; j++) {
-        out_[y + j] = inA[x + j];
-    }
-}
-
-void main() {
-    const uint i = gl_WorkGroupID.x;
-    const int r = inB[i + pcs.inBOff];
-
-    dequantize_row_f16(r*pcs.nb01/2/*bytes for float16*/ + pcs.inAOff, i*pcs.nb1/4 + pcs.outOff, pcs.ne00);
-}

ggml/src/kompute-shaders/op_getrows_f32.comp

@@ -1,31 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-layout(local_size_x = 1) in;
-
-layout (binding = 0) readonly buffer tensorInA { float inA[]; };
-layout (binding = 1) readonly buffer tensorInB { int inB[]; };
-layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
-
-layout (push_constant) uniform parameter {
-    uint inAOff;
-    uint inBOff;
-    uint outOff;
-    int ne00;
-    int nb01;
-    int nb1;
-} pcs;
-
-void dequantize_row_f32(uint x /*Based from inA unaligned*/, uint y /*Based from out_*/, int k) {
-    for (int j = 0; j < k; j++) {
-        out_[y + j] = inA[x + j];
-    }
-}
-
-void main() {
-    const uint i = gl_WorkGroupID.x;
-    const int r = inB[i + pcs.inBOff];
-
-    dequantize_row_f32(r*pcs.nb01/4 + pcs.inAOff, i*pcs.nb1/4 + pcs.outOff, pcs.ne00);
-}

ggml/src/kompute-shaders/op_getrows_q4_0.comp

@@ -1,38 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-#define NL 2
-#define BYTES_FOR_TYPE 4 /*bytes for float*/
-#define SIZE_OF_BLOCK sizeof_block_q4_0
-
-layout(local_size_x = 1) in;
-
-layout (binding = 0) readonly buffer tensorInA { uint8_t inA[]; };
-layout (binding = 1) readonly buffer tensorInB { int inB[]; };
-layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
-
-layout (push_constant) uniform parameter {
-    uint inAOff;
-    uint inBOff;
-    uint outOff;
-    int ne00;
-    int nb01;
-    int nb1;
-} pcs;
-
-block_q4_0 get_unaligned_block_q4_0(uint index) {
-    block_q4_0 fres;
-    fres.d = u8BufToFloat16(inA, index);
-    [[unroll]] for (uint it = 0; it != QK4_0 / 2; it++) {
-        fres.qs[it] = inA[index+2+it];
-    }
-    return fres;
-}
-
-mat4 dequantize_block(uint index, uint il) {
-    const block_q4_0 block = get_unaligned_block_q4_0(index);
-    return dequantize_q4_0(block, il);
-}
-
-#include "op_getrows.comp"

ggml/src/kompute-shaders/op_getrows_q4_1.comp

@@ -1,39 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-#define NL 2
-#define BYTES_FOR_TYPE 4 /*bytes for float*/
-#define SIZE_OF_BLOCK sizeof_block_q4_1
-
-layout(local_size_x = 1) in;
-
-layout (binding = 0) readonly buffer tensorInA { uint8_t inA[]; };
-layout (binding = 1) readonly buffer tensorInB { int inB[]; };
-layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
-
-layout (push_constant) uniform parameter {
-    uint inAOff;
-    uint inBOff;
-    uint outOff;
-    int ne00;
-    int nb01;
-    int nb1;
-} pcs;
-
-block_q4_1 get_unaligned_block_q4_1(uint index) {
-    block_q4_1 fres;
-    fres.d = u8BufToFloat16(inA, index);
-    fres.m = u8BufToFloat16(inA, index+2);
-    [[unroll]] for (uint it = 0; it != QK4_1 / 2; it++) {
-        fres.qs[it] = inA[index+4+it];
-    }
-    return fres;
-}
-
-mat4 dequantize_block(uint index, uint il) {
-    const block_q4_1 block = get_unaligned_block_q4_1(index);
-    return dequantize_q4_1(block, il);
-}
-
-#include "op_getrows.comp"

ggml/src/kompute-shaders/op_getrows_q6_k.comp

@@ -1,44 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-#define NL 16
-#define BYTES_FOR_TYPE 4 /*bytes for float*/
-#define SIZE_OF_BLOCK sizeof_block_q6_k
-
-layout(local_size_x = 1) in;
-
-layout (binding = 0) readonly buffer tensorInA { uint8_t inA[]; };
-layout (binding = 1) readonly buffer tensorInB { int inB[]; };
-layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
-
-layout (push_constant) uniform parameter {
-    uint inAOff;
-    uint inBOff;
-    uint outOff;
-    int ne00;
-    int nb01;
-    int nb1;
-} pcs;
-
-block_q6_k get_unaligned_block_q6_k(uint index) {
-    block_q6_k fres;
-    [[unroll]] for (uint it = 0; it != QK_K / 2; it++) {
-        fres.ql[it] = inA[index + it];
-    }
-    [[unroll]] for (uint it = 0; it != QK_K / 4; it++) {
-        fres.qh[it] = inA[index + QK_K/2 + it];
-    }
-    [[unroll]] for (uint it = 0; it != QK_K / 16; it++) {
-        fres.scales[it] = int8_t(inA[index + QK_K/2 + QK_K/4 + it]);
-    }
-    fres.d = u8BufToFloat16(inA, index + QK_K/2 + QK_K/4 + QK_K/16);
-    return fres;
-}
-
-mat4 dequantize_block(uint index, uint il) {
-    const block_q6_k block = get_unaligned_block_q6_k(index);
-    return dequantize_q6_k(block, il);
-}
-
-#include "op_getrows.comp"

ggml/src/kompute-shaders/op_mul.comp

@@ -1,52 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-layout(local_size_x = 1024) in;
-
-layout(binding = 0) buffer restrict readonly tensorInA { float inA[]; };
-layout(binding = 1) buffer restrict readonly tensorInB { float inB[]; };
-layout(binding = 2) buffer restrict writeonly tensorOut { float out_[]; };
-
-layout(push_constant) uniform PushConstants {
-    uint inAOff;
-    uint inBOff;
-    uint outOff;
-    int ne00;
-    int nb00;
-    int nb01;
-    int nb02;
-    int nb03;
-    int ne10;
-    int ne11;
-    int ne12;
-    int ne13;
-    int nb10;
-    int nb11;
-    int nb12;
-    int nb13;
-    int ne0;
-    int nb0;
-    int nb1;
-    int nb2;
-    int nb3;
-} pcs;
-
-void main() {
-    const uint i03 = gl_WorkGroupID.z;
-    const uint i02 = gl_WorkGroupID.y;
-    const uint i01 = gl_WorkGroupID.x;
-
-    const uint i13 = i03 % pcs.ne13;
-    const uint i12 = i02 % pcs.ne12;
-    const uint i11 = i01 % pcs.ne11;
-
-    uint src0_off = uint((i03*pcs.nb03 + i02*pcs.nb02 + i01*pcs.nb01) / 4);
-    uint src1_off = uint((i13*pcs.nb13 + i12*pcs.nb12 + i11*pcs.nb11) / 4);
-    uint dst_off  = uint((i03*pcs.nb3  + i02*pcs.nb2  + i01*pcs.nb1)  / 4);
-
-    for (uint i0 = gl_LocalInvocationID.x; i0 < pcs.ne0; i0 += gl_WorkGroupSize.x) {
-        const uint i10 = i0 % pcs.ne10;
-        out_[pcs.outOff + dst_off + i0] = inA[pcs.inAOff + src0_off + i0] * inB[pcs.inBOff + src1_off + i10];
-    }
-}

ggml/src/kompute-shaders/op_mul_mat_f16.comp

@@ -1,67 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-#extension GL_KHR_shader_subgroup_arithmetic : require
-
-layout(local_size_x_id = 0) in;
-
-layout (binding = 0) readonly buffer tensorInA { float16_t inA[]; };
-layout (binding = 1) readonly buffer tensorInB { float inB[]; };
-layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
-
-layout (push_constant) uniform parameter {
-    uint inAOff;
-    uint inBOff;
-    uint outOff;
-    int ne00;
-    int ne01;
-    int ne02;
-    uint nb00;
-    uint nb01;
-    uint nb02;
-    int ne10;
-    int ne11;
-    int ne12;
-    uint nb10;
-    uint nb11;
-    uint nb12;
-    int ne0;
-    int ne1;
-    uint r2;
-    uint r3;
-} pcs;
-
-#define N_F16_F32 4
-
-void main() {
-    const uint r0 = gl_WorkGroupID.x;
-    const uint rb = gl_WorkGroupID.y*N_F16_F32;
-    const uint im = gl_WorkGroupID.z;
-
-    const uint i12 = im%pcs.ne12;
-    const uint i13 = im/pcs.ne12;
-
-    const uint offset0 = r0*pcs.nb01 + (i12/pcs.r2)*pcs.nb02 + (i13/pcs.r3)*pcs.nb02*pcs.ne02;
-
-    const uint x = offset0 / 2 + pcs.inAOff; // Based from inA
-
-    for (uint row = 0; row < N_F16_F32; ++row) {
-        uint r1 = rb + row;
-        if (r1 >= pcs.ne11) {
-            break;
-        }
-
-        const uint y = (r1*pcs.nb11 + im*pcs.nb12) / 4 + pcs.inBOff; // Based from inB
-
-        float sumf = 0;
-        for (uint i = gl_SubgroupInvocationID.x; i < pcs.ne00; i += gl_SubgroupSize) {
-            sumf += float(inA[x+i]) * float(inB[y+i]);
-        }
-
-        const float all_sum = subgroupAdd(sumf);
-        if (subgroupElect()) {
-            out_[im*pcs.ne1*pcs.ne0 + r1*pcs.ne0 + r0 + pcs.outOff] = all_sum;
-        }
-    }
-}

ggml/src/kompute-shaders/op_mul_mat_mat_f32.comp

@@ -1,51 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-#extension GL_KHR_shader_subgroup_arithmetic : require
-#extension GL_EXT_debug_printf : enable
-
-// device subgroup size
-layout (local_size_x_id = 0) in;
-
-layout(binding = 0) readonly buffer tensorInA { float inA[]; };
-layout(binding = 1) readonly buffer tensorInB { float inB[]; };
-layout(binding = 2) writeonly buffer tensorOut { float out_[]; };
-
-layout(push_constant) uniform parameter {
-  uint inAOff;
-  uint inBOff;
-  uint outOff;
-  int ne00;
-  int ne01;
-  int ne02;
-  int ne11;
-  int ne12;
-  uint nb01;
-  uint nb02;
-  uint nb11;
-  uint nb12;
-  uint nb1;
-  uint nb2;
-}
-pcs;
-
-
-void main() {
-  uvec3 gid = gl_WorkGroupID;
-
-  uint bc_ab = pcs.ne12 > pcs.ne02 ? gid.z / (pcs.ne12 / pcs.ne02) : gid.z;
-  uint bc_ba = pcs.ne02 > pcs.ne12 ? gid.z / (pcs.ne02 / pcs.ne12) : gid.z;
-
-  const uint x = (gid.x*pcs.nb01 + bc_ab*pcs.nb02) / 4 + pcs.inAOff; // Based from inA
-  const uint y = (gid.y*pcs.nb11 + bc_ba*pcs.nb12) / 4 + pcs.inBOff; // based from inB
-  float sum = 0.0f;
-  for (uint i = gl_SubgroupInvocationID.x; i < pcs.ne00; i += gl_SubgroupSize) {
-      sum += float(inA[x+i]) * float(inB[y+i]);
-  }
-
-  const float all_sum = subgroupAdd(sum);
-  if (subgroupElect()) {
-    out_[gid.z*(pcs.nb2/4) + gid.y*(pcs.nb1/4) + gid.x + pcs.outOff] = all_sum;
-  }
-}

ggml/src/kompute-shaders/op_mul_mat_q4_0.comp

@@ -1,33 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-#define BLOCKS_IN_QUANT QK4_0
-#define SIZE_OF_BLOCK sizeof_block_q4_0
-#define N_ROWS 4
-
-#include "op_mul_mv_q_n_pre.comp"
-
-// The q4_0 version of this function
-float block_q_n_dot_y(uint block_index, uint yb, uint il) {
-    vec2 acc = vec2(0.0, 0.0);
-    const uint index = (block_index) * SIZE_OF_BLOCK + pcs.inAOff;
-    float d = float(u8BufToFloat16(inA, index));
-    float sumy = 0.0f;
-    for (int i = 0; i < BLOCKS_IN_QUANT/4; i+=2) {
-        const uint16_t b = u8BufToU16(inA, index + 2 + il + i);
-
-        const float yl0 = inB[yb + i];
-        const float yl1 = inB[yb + i + 1];
-        const float yl8 = inB[yb + i + BLOCKS_IN_QUANT/2];
-        const float yl9 = inB[yb + i + BLOCKS_IN_QUANT/2 + 1];
-
-        sumy += yl0 + yl1 + yl8 + yl9;
-
-        acc[0] += yl0 * (b & 0x000F) + yl1 / 256.f * (b & 0x0F00);
-        acc[1] += yl8 / 16.f * (b & 0x00F0) + yl9 / 4096.f * (b & 0xF000);
-    }
-    return d * (sumy * -8.f + acc[0] + acc[1]);
-}
-
-#include "op_mul_mv_q_n.comp"

ggml/src/kompute-shaders/op_mul_mat_q4_1.comp

@@ -1,35 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-#define BLOCKS_IN_QUANT QK4_1
-#define SIZE_OF_BLOCK sizeof_block_q4_1
-#define N_ROWS 4
-
-#include "op_mul_mv_q_n_pre.comp"
-
-// The q4_1 version of this function
-float block_q_n_dot_y(uint block_index, uint yb, uint il) {
-    vec2 acc = vec2(0.0, 0.0);
-    const uint index = (block_index) * SIZE_OF_BLOCK + pcs.inAOff;
-    float d = float(u8BufToFloat16(inA, index));
-    float m = float(u8BufToFloat16(inA, index+2));
-
-    float sumy = 0.0f;
-    for (int i = 0; i < BLOCKS_IN_QUANT/4; i+=2) {
-        const uint16_t b = u8BufToU16(inA, index + 4 + il + i);
-
-        const float yl0 = inB[yb + i];
-        const float yl1 = inB[yb + i + 1];
-        const float yl8 = inB[yb + i + BLOCKS_IN_QUANT/2];
-        const float yl9 = inB[yb + i + BLOCKS_IN_QUANT/2 + 1];
-
-        sumy += yl0 + yl1 + yl8 + yl9;
-
-        acc[0] += yl0 * (b & 0x000F) + yl1 / 256.f * (b & 0x0F00);
-        acc[1] += yl8 / 16.f * (b & 0x00F0) + yl9 / 4096.f * (b & 0xF000);
-    }
-    return d * (acc[0] + acc[1]) + sumy * m;
-}
-
-#include "op_mul_mv_q_n.comp"

ggml/src/kompute-shaders/op_mul_mat_q6_k.comp

@@ -1,94 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-#define SIZE_OF_BLOCK sizeof_block_q6_k
-
-layout(local_size_x_id = 0) in;
-layout(local_size_y_id = 1) in;
-layout(local_size_z = 1) in;
-
-layout (binding = 0) readonly buffer tensorInA { uint8_t inA[]; };
-layout (binding = 1) readonly buffer tensorInB { float inB[]; };
-layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
-
-layout (push_constant) uniform parameter {
-    uint inAOff;
-    uint inBOff;
-    uint outOff;
-    int ne00;
-    int ne10;
-    int ne0;
-    int ne1;
-    int ne01;
-    int gqa;
-} pcs;
-
-void main() {
-    const uint8_t kmask1 = uint8_t(0x03);
-    const uint8_t kmask2 = uint8_t(0x0C);
-    const uint8_t kmask3 = uint8_t(0x30);
-    const uint8_t kmask4 = uint8_t(0xC0);
-
-    const uint nb = pcs.ne00/QK_K;
-
-    const uint r0 = gl_WorkGroupID.x;
-    const uint r1 = gl_WorkGroupID.y;
-    const uint r2 = gl_WorkGroupID.z;
-
-    const uint row = (r0 * gl_NumSubgroups + gl_SubgroupID);
-    const uint offset0 = r2/pcs.gqa*(nb*pcs.ne0);
-    const uint x = row * nb + offset0; // Based from inA without base offset
-    const uint yy = r1*pcs.ne10 + r2*pcs.ne00*pcs.ne1+pcs.inBOff; // Based from inB
-
-    float sumf = 0;
-
-    // bits of invocation ID for gl_SubgroupSize=32:
-    //  x   x   x   x   x
-    //  4   3   2   1   0
-    // (     tid     ) ix
-    //  ip (   il    )
-
-    const uint block_stride = gl_SubgroupSize / 16;         // number of blocks each subgroup processes
-    const uint tid  = gl_SubgroupInvocationID/block_stride; // first block_stride groups have tid=0
-    const uint ix   = gl_SubgroupInvocationID%block_stride; // first block is 0..block_stride-1
-    const uint ip   = tid/8;        // first or second half of block (0 or 1)
-    const uint il   = tid%8;        // each half has 8 parts, one per scale
-    const uint n    = 4;            // 4 scales at a time (and 4 sums)
-    const uint l0   = n*il;         // offset into half-block, 0..28
-    const uint is   = 8*ip + l0/16; // 0, 1, 8, 9
-
-    const uint y_offset = 128*ip + l0;
-    const uint q_offset_l = 64*ip + l0;
-    const uint q_offset_h = 32*ip + l0;
-
-    for (uint i = ix; i < nb; i += block_stride) {
-
-        const uint baseIndex = (x + i) * SIZE_OF_BLOCK + pcs.inAOff;
-
-        const uint qlIndex = q_offset_l;
-        const uint q2Index = qlIndex + QK_K/8;
-        const uint qhIndex = q_offset_h;
-        const uint y = yy + i * QK_K + y_offset;
-
-        float sums[4] = {0.0f, 0.0f, 0.0f, 0.0f};
-        for (uint l = 0; l < n; ++l) {
-            const uint8_t currentQ1 = inA[baseIndex + qlIndex + l];
-            const uint8_t currentQ2 = inA[baseIndex + q2Index + l];
-            const uint8_t currentQh = inA[baseIndex + QK_K/2 + qhIndex + l];
-
-            sums[0] += inB[y+l+ 0] * (int8_t((currentQ1 & 0xF) | ((currentQh & kmask1) << 4)) - 32);
-            sums[1] += inB[y+l+32] * (int8_t((currentQ2 & 0xF) | ((currentQh & kmask2) << 2)) - 32);
-            sums[2] += inB[y+l+64] * (int8_t((currentQ1  >> 4) | ((currentQh & kmask3) << 0)) - 32);
-            sums[3] += inB[y+l+96] * (int8_t((currentQ2  >> 4) | ((currentQh & kmask4) >> 2)) - 32);
-        }
-
-        float d = u8BufToFloat16(inA, baseIndex + QK_K/2 + QK_K/4 + QK_K/16);
-        sumf += d * (sums[0] * int8_t(inA[baseIndex + QK_K/2 + QK_K/4 + is]) + sums[1] * int8_t(inA[baseIndex + QK_K/2 + QK_K/4 + 2 + is]) + sums[2] * int8_t(inA[baseIndex + QK_K/2 + QK_K/4 + 4 + is]) + sums[3] * int8_t(inA[baseIndex + QK_K/2 + QK_K/4 + 6 + is]));
-    }
-
-    const float tot = subgroupAdd(sumf);
-    if (subgroupElect()) {
-        out_[r1*pcs.ne0 + r2*pcs.ne0*pcs.ne1 + row + pcs.outOff] = tot;
-    }
-}

ggml/src/kompute-shaders/op_mul_mat_q8_0.comp

@@ -1,73 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-#include "op_mul_mv_q_n_pre.comp"
-
-#define SIZE_OF_D 2
-
-#define N_DST 4 // each SIMD group works on 4 rows
-#define N_SIMDGROUP 2 // number of SIMD groups in a thread group
-#define N_SIMDWIDTH 32 // assuming SIMD group size is 32
-
-#define NB_Q8_0 8
-
-void main() {
-    // NB: hack to make compatible with AMD GPUs that have a subgroup size of 64
-    if (gl_SubgroupInvocationID > 31)
-        return;
-
-    const int nr  = N_DST;
-    const int nsg = N_SIMDGROUP;
-    const int nw  = N_SIMDWIDTH;
-
-    const int nb = pcs.ne00/QK8_0;
-    const uint r0 = gl_WorkGroupID.x;
-    const uint r1 = gl_WorkGroupID.y;
-    const uint im = gl_WorkGroupID.z;
-
-    const uint first_row = (r0 * nsg + gl_SubgroupID) * nr;
-
-    const uint i12 = im%pcs.ne12;
-    const uint i13 = im/pcs.ne12;
-
-    const uint offset0 = first_row * nb + (i12/pcs.r2)*(nb*pcs.ne01) + (i13/pcs.r3)*(nb*pcs.ne01*pcs.ne02);
-
-    const uint x = offset0*sizeof_block_q8_0 + pcs.inAOff; // Based from inA
-    const uint y = r1*pcs.ne10 + im*pcs.ne00*pcs.ne1 + pcs.inBOff; // based from inB
-
-    float yl[NB_Q8_0];
-    float sumf[N_DST]={0.f, 0.f, 0.f, 0.f};
-
-    const uint ix = gl_SubgroupInvocationID.x/4;
-    const uint il = gl_SubgroupInvocationID.x%4;
-
-    uint yb = y + ix * QK8_0 + NB_Q8_0*il;
-
-    // each thread in a SIMD group deals with NB_Q8_0 quants at a time
-    for (uint ib = ix; ib < nb; ib += nw/4) {
-        for (int i = 0; i < NB_Q8_0; ++i) {
-            yl[i] = inB[yb + i];
-        }
-
-        for (int row = 0; row < nr; row++) {
-            const uint block_offset = (ib+row*nb) * sizeof_block_q8_0;
-            float sumq = 0.f;
-            for (int iq = 0; iq < NB_Q8_0; ++iq) {
-                const int8_t qs_iq = int8_t(inA[x + block_offset + SIZE_OF_D + NB_Q8_0*il + iq]);
-                sumq += qs_iq * yl[iq];
-            }
-            const float16_t d = u8BufToFloat16(inA, x + block_offset);
-            sumf[row] += sumq*d;
-        }
-
-        yb += NB_Q8_0 * nw;
-    }
-
-    for (int row = 0; row < nr; ++row) {
-        const float tot = subgroupAdd(sumf[row]);
-        if (subgroupElect() && first_row + row < pcs.ne01) {
-            out_[r1*pcs.ne0 + im*pcs.ne0*pcs.ne1 + first_row + row] = tot;
-        }
-    }
-}

ggml/src/kompute-shaders/op_mul_mv_q_n.comp

@@ -1,48 +0,0 @@
-void main() {
-    // NB: hack to make compatible with AMD GPUs that have a subgroup size of 64
-    if (gl_SubgroupInvocationID > 31)
-        return;
-
-    const uint nb = uint(pcs.ne00/BLOCKS_IN_QUANT);
-
-    const uint r0 = gl_WorkGroupID.x;
-    const uint r1 = gl_WorkGroupID.y;
-    const uint im = gl_WorkGroupID.z;
-
-    const uint first_row = (r0 * gl_NumSubgroups + gl_SubgroupID) * N_ROWS;
-
-    const uint i12 = im%pcs.ne12;
-    const uint i13 = im/pcs.ne12;
-
-    const uint offset0 = first_row * nb + (i12/pcs.r2)*(nb*pcs.ne01) + (i13/pcs.r3)*(nb*pcs.ne01*pcs.ne02);
-
-    const uint x = offset0; // Based from inA without base offset
-    const uint y = r1*uint(pcs.ne10)+im*pcs.ne00*pcs.ne1+pcs.inBOff; // Based from inB
-
-    float sumf[N_ROWS] = {0.0f, 0.0f, 0.0f, 0.0f};
-
-    const uint ix = gl_SubgroupInvocationID/2;
-    const uint il = (BLOCKS_IN_QUANT/4)*(gl_SubgroupInvocationID%2);
-
-    uint yb = y + ix * BLOCKS_IN_QUANT + il;
-
-    //debugPrintfEXT("gl_NumSubgroups=%d, gl_SubgroupID=%d, gl_SubgroupInvocationID=%d, glSubgroupSize=%d, gl_WorkGroupSize.x=%d, gl_WorkGroupSize.y=%d, gl_WorkGroupSize.z=%d\n",
-    //    gl_NumSubgroups, gl_SubgroupID, gl_SubgroupInvocationID, gl_SubgroupSize,
-    //    gl_WorkGroupSize.x, gl_WorkGroupSize.y, gl_WorkGroupSize.z);
-
-    for (uint ib = ix; ib < nb; ib += 16) {
-        for (int row = 0; row < N_ROWS; row++) {
-            const uint block_index = x + ib + row * nb;
-            sumf[row] += block_q_n_dot_y(block_index, yb, il);
-        }
-
-        yb += BLOCKS_IN_QUANT * 16;
-    }
-
-    for (int row = 0; row < N_ROWS; ++row) {
-        const float tot = subgroupAdd(sumf[row]);
-        if (first_row + row < pcs.ne01 && subgroupElect()) {
-            out_[r1*pcs.ne0 + im*pcs.ne0*pcs.ne1 + first_row + row + pcs.outOff] = tot;
-        }
-    }
-}

ggml/src/kompute-shaders/op_mul_mv_q_n_pre.comp

@@ -1,22 +0,0 @@
-layout(local_size_x_id = 0) in;
-layout(local_size_y = 1) in;
-layout(local_size_z = 1) in;
-
-layout (binding = 0) readonly buffer tensorInA { uint8_t inA[]; };
-layout (binding = 1) readonly buffer tensorInB { float inB[]; };
-layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
-
-layout (push_constant) uniform parameter {
-    uint inAOff;
-    uint inBOff;
-    uint outOff;
-    int  ne00;
-    int  ne01;
-    int  ne02;
-    int  ne10;
-    int  ne12;
-    int  ne0;
-    int  ne1;
-    uint r2;
-    uint r3;
-} pcs;

ggml/src/kompute-shaders/op_norm.comp

@@ -1,84 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-layout(local_size_x = 256) in;
-
-layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
-layout(binding = 1) buffer restrict tensorOut { float out_[]; };
-
-layout(push_constant) uniform PushConstants {
-    uint inOff;
-    uint outOff;
-    uint ne00;
-    uint nb01;
-    float eps;
-} pcs;
-
-shared float sum[gl_WorkGroupSize.x];
-
-void main() {
-    const uint x = (gl_WorkGroupID.x*pcs.nb01/4) + pcs.inOff; // Based from in_
-    // MEAN
-    // parallel sum
-    sum[gl_LocalInvocationID.x] = 0.0;
-    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
-        sum[gl_LocalInvocationID.x] += in_[x+i00];
-    }
-
-    // reduce
-    barrier();
-    memoryBarrierShared();
-    [[unroll]] for (uint i = gl_WorkGroupSize.x/2; i > 0; i /= 2) {
-        if (gl_LocalInvocationID.x < i) {
-            sum[gl_LocalInvocationID.x] += sum[gl_LocalInvocationID.x + i];
-        }
-        barrier();
-        memoryBarrierShared();
-    }
-
-    // broadcast
-    if (gl_LocalInvocationID.x == 0) {
-        sum[0] /= float(pcs.ne00);
-    }
-    barrier();
-    memoryBarrierShared();
-    const float mean = sum[0];
-
-    // recenter
-    const uint y = (gl_WorkGroupID.x*pcs.ne00) + pcs.outOff; // Based from out_
-    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
-        out_[y+i00] = in_[x+i00] - mean;
-    }
-
-    // VARIANCE
-    // parallel sum
-    sum[gl_LocalInvocationID.x] = 0.0;
-    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
-        sum[gl_LocalInvocationID.x] += out_[y+i00] * out_[y+i00];
-    }
-
-    // reduce
-    barrier();
-    memoryBarrierShared();
-    [[unroll]] for (uint i = gl_WorkGroupSize.x/2; i > 0; i /= 2) {
-        if (gl_LocalInvocationID.x < i) {
-            sum[gl_LocalInvocationID.x] += sum[gl_LocalInvocationID.x + i];
-        }
-        barrier();
-        memoryBarrierShared();
-    }
-
-    // broadcast
-    if (gl_LocalInvocationID.x == 0) {
-        sum[0] /= float(pcs.ne00);
-    }
-    barrier();
-    memoryBarrierShared();
-    const float variance = sum[0];
-
-    const float scale = 1.0f/sqrt(variance + pcs.eps);
-    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
-        out_[y+i00] *= scale;
-    }
-}

ggml/src/kompute-shaders/op_relu.comp

@@ -1,21 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-layout(local_size_x = 1) in;
-
-layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
-layout(binding = 1) buffer restrict writeonly tensorOut { float out_[]; };
-layout(push_constant) uniform PushConstants {
-    uint inOff;
-    uint outOff;
-} pcs;
-
-void main() {
-    const uint baseIndex = gl_WorkGroupID.x * 4;
-
-    for (uint x = 0; x < 4; x++) {
-        const uint i = baseIndex + x;
-        out_[i + pcs.outOff] = max(0.0, in_[i + pcs.inOff]);
-    }
-}

ggml/src/kompute-shaders/op_rmsnorm.comp

@@ -1,53 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-layout(local_size_x = 512) in;
-
-layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
-layout(binding = 1) buffer restrict tensorOut { float out_[]; };
-
-layout(push_constant) uniform PushConstants {
-    uint inOff;
-    uint outOff;
-    uint ne00;
-    uint nb01;
-    float eps;
-} pcs;
-
-shared float sum[gl_WorkGroupSize.x];
-
-void main() {
-    const uint x = (gl_WorkGroupID.x*pcs.nb01/4) + pcs.inOff; // Based from in_
-
-    // parallel sum
-    sum[gl_LocalInvocationID.x] = 0.0;
-    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
-        sum[gl_LocalInvocationID.x] += in_[x+i00] * in_[x+i00];
-    }
-
-    // reduce
-    barrier();
-    memoryBarrierShared();
-    [[unroll]] for (uint i = gl_WorkGroupSize.x/2; i > 0; i /= 2) {
-        if (gl_LocalInvocationID.x < i) {
-            sum[gl_LocalInvocationID.x] += sum[gl_LocalInvocationID.x + i];
-        }
-        barrier();
-        memoryBarrierShared();
-    }
-
-    // broadcast
-    if (gl_LocalInvocationID.x == 0) {
-        sum[0] /= float(pcs.ne00);
-    }
-    barrier();
-    memoryBarrierShared();
-
-    const float scale = 1.0f/sqrt(sum[0] + pcs.eps);
-
-    const uint y = (gl_WorkGroupID.x*pcs.ne00) + pcs.outOff; // Based from out_
-    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
-        out_[y+i00] = in_[x+i00] * scale;
-    }
-}

ggml/src/kompute-shaders/op_rope_f16.comp

@@ -1,73 +0,0 @@
-#version 450
-
-#include "rope_common.comp"
-
-layout(binding = 0) buffer restrict readonly  tensorInA { float16_t inA[]; };
-layout(binding = 1) buffer restrict readonly  tensorInB { int       inB[]; };
-layout(binding = 2) buffer restrict writeonly tensorOut { float16_t out_[]; };
-
-void main() {
-    const uint i3 = gl_WorkGroupID.z;
-    const uint i2 = gl_WorkGroupID.y;
-    const uint i1 = gl_WorkGroupID.x;
-
-    const bool is_neox = (pcs.mode & 2) != 0;
-
-    float corr_dims[2];
-    rope_yarn_corr_dims(pcs.n_dims, pcs.n_ctx_orig, pcs.freq_base, pcs.beta_fast, pcs.beta_slow, corr_dims);
-
-    const float theta_scale = pow(pcs.freq_base, -2.0/pcs.n_dims);
-
-    const int p = inB[pcs.inBOff + i2];
-
-    float theta = float(p);
-
-    if (!is_neox) {
-        for (uint i0 = 0; i0 < pcs.ne0; i0 += 2) {
-            float cos_theta, sin_theta;
-            rope_yarn(theta, pcs.freq_scale, corr_dims, i0, pcs.ext_factor, pcs.attn_factor, cos_theta, sin_theta);
-
-            theta *= theta_scale;
-
-            const uint src      = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 2) + pcs.inAOff; // Based from in
-            const uint dst_data = uint((i3*pcs.nb3  + i2*pcs.nb2  + i1*pcs.nb1  + i0*pcs.nb0)  / 2) + pcs.outOff; // Based from out_
-
-            const float x0 = float(inA[src]);
-            const float x1 = float(inA[src+1]);
-
-            out_[dst_data]   = float16_t(x0*cos_theta - x1*sin_theta);
-            out_[dst_data+1] = float16_t(x0*sin_theta + x1*cos_theta);
-        }
-    } else {
-        const float inv_ndims = -1.f/pcs.n_dims;
-        for (uint ic = 0; ic < pcs.n_dims; ic += 2) {
-            const uint cur_rot = ic;
-
-            float cos_theta, sin_theta;
-            rope_yarn(theta, pcs.freq_scale, corr_dims, cur_rot, pcs.ext_factor, pcs.attn_factor, cos_theta, sin_theta);
-
-            theta *= theta_scale;
-
-            const uint i0 = ic/2;
-
-            const uint src      = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 2) + pcs.inAOff; // Based from in
-            const uint dst_data = uint((i3*pcs.nb3  + i2*pcs.nb2  + i1*pcs.nb1  + i0*pcs.nb0)  / 2) + pcs.outOff; // Based from out_
-
-            const float x0 = float(inA[src]);
-            const float x1 = float(inA[src+pcs.n_dims/2]);
-
-            out_[dst_data]              = float16_t(x0*cos_theta - x1*sin_theta);
-            out_[dst_data+pcs.n_dims/2] = float16_t(x0*sin_theta + x1*cos_theta);
-        }
-
-        for (uint ic = pcs.n_dims; ic < pcs.ne0; ic += 2) {
-            const uint i0 = ic;
-
-            const uint src      = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 2) + pcs.inAOff; // Based from in
-            const uint dst_data = uint((i3*pcs.nb3  + i2*pcs.nb2  + i1*pcs.nb1  + i0*pcs.nb0)  / 2) + pcs.outOff; // Based from out_
-
-            out_[dst_data + 0] = inA[src + 0];
-            out_[dst_data + 1] = inA[src + 1];
-        }
-    }
-}

ggml/src/kompute-shaders/op_rope_f32.comp

@@ -1,73 +0,0 @@
-#version 450
-
-#include "rope_common.comp"
-
-layout(binding = 0) buffer restrict readonly  tensorInA { float inA[]; };
-layout(binding = 1) buffer restrict readonly  tensorInB { int   inB[]; };
-layout(binding = 2) buffer restrict writeonly tensorOut { float out_[]; };
-
-void main() {
-    const uint i3 = gl_WorkGroupID.z;
-    const uint i2 = gl_WorkGroupID.y;
-    const uint i1 = gl_WorkGroupID.x;
-
-    const bool is_neox = (pcs.mode & 2) != 0;
-
-    float corr_dims[2];
-    rope_yarn_corr_dims(pcs.n_dims, pcs.n_ctx_orig, pcs.freq_base, pcs.beta_fast, pcs.beta_slow, corr_dims);
-
-    const float theta_scale = pow(pcs.freq_base, -2.0/pcs.n_dims);
-
-    const int p = inB[pcs.inBOff + i2];
-
-    float theta = float(p);
-
-    if (!is_neox) {
-        for (uint i0 = 0; i0 < pcs.ne0; i0 += 2) {
-            float cos_theta, sin_theta;
-            rope_yarn(theta, pcs.freq_scale, corr_dims, i0, pcs.ext_factor, pcs.attn_factor, cos_theta, sin_theta);
-
-            theta *= theta_scale;
-
-            const uint src      = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 4) + pcs.inAOff; // Based from in
-            const uint dst_data = uint((i3*pcs.nb3  + i2*pcs.nb2  + i1*pcs.nb1  + i0*pcs.nb0)  / 4) + pcs.outOff; // Based from out_
-
-            const float x0 = inA[src];
-            const float x1 = inA[src+1];
-
-            out_[dst_data]   = x0*cos_theta - x1*sin_theta;
-            out_[dst_data+1] = x0*sin_theta + x1*cos_theta;
-        }
-    } else {
-        const float inv_ndims = -1.f/pcs.n_dims;
-        for (uint ic = 0; ic < pcs.n_dims; ic += 2) {
-            const uint cur_rot = ic;
-
-            float cos_theta, sin_theta;
-            rope_yarn(theta, pcs.freq_scale, corr_dims, cur_rot, pcs.ext_factor, pcs.attn_factor, cos_theta, sin_theta);
-
-            theta *= theta_scale;
-
-            const uint i0 = ic/2;
-
-            const uint src      = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 4) + pcs.inAOff; // Based from in
-            const uint dst_data = uint((i3*pcs.nb3  + i2*pcs.nb2  + i1*pcs.nb1  + i0*pcs.nb0)  / 4) + pcs.outOff; // Based from out_
-
-            const float x0 = inA[src];
-            const float x1 = inA[src+pcs.n_dims/2];
-
-            out_[dst_data] = x0*cos_theta - x1*sin_theta;
-            out_[dst_data+pcs.n_dims/2] = x0*sin_theta + x1*cos_theta;
-        }
-
-        for (uint ic = pcs.n_dims; ic < pcs.ne0; ic += 2) {
-            const uint i0 = ic;
-
-            const uint src = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 4) + pcs.inAOff; // Based from in
-            const uint dst_data = uint((i3*pcs.nb3  + i2*pcs.nb2  + i1*pcs.nb1  + i0*pcs.nb0) / 4) + pcs.outOff; // Based from out_
-
-            out_[dst_data + 0] = inA[src + 0];
-            out_[dst_data + 1] = inA[src + 1];
-        }
-    }
-}

ggml/src/kompute-shaders/op_scale.comp

@@ -1,19 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-layout(local_size_x = 1) in;
-
-layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
-layout(binding = 1) buffer restrict writeonly tensorOut { float out_[]; };
-
-layout(push_constant) uniform PushConstants {
-    uint inOff;
-    uint outOff;
-    float scale;
-} pcs;
-
-void main() {
-    const uint i = gl_WorkGroupID.x;
-    out_[i + pcs.outOff] = in_[i + pcs.inOff] * pcs.scale;
-}

ggml/src/kompute-shaders/op_scale_8.comp

@@ -1,23 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-layout(local_size_x = 1) in;
-
-layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
-layout(binding = 1) buffer restrict writeonly tensorOut { float out_[]; };
-
-layout(push_constant) uniform PushConstants {
-    uint inOff;
-    uint outOff;
-    float scale;
-} pcs;
-
-void main() {
-    const uint baseIndex = gl_WorkGroupID.x * 8;
-
-    for (uint x = 0; x < 8; x++) {
-        const uint i = baseIndex + x;
-        out_[i + pcs.outOff] = in_[i + pcs.inOff] * pcs.scale;
-    }
-}

ggml/src/kompute-shaders/op_silu.comp

@@ -1,22 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-layout(local_size_x = 1) in;
-
-layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
-layout(binding = 1) buffer restrict writeonly tensorOut { float out_[]; };
-layout(push_constant) uniform PushConstants {
-    uint inOff;
-    uint outOff;
-} pcs;
-
-void main() {
-    const uint baseIndex = gl_WorkGroupID.x * 4;
-
-    for (uint x = 0; x < 4; x++) {
-        const uint i = baseIndex + x;
-        const float y = in_[i + pcs.inOff];
-        out_[i + pcs.outOff] = y / (1.0 + exp(-y));
-    }
-}

ggml/src/kompute-shaders/op_softmax.comp

@@ -1,56 +0,0 @@
-// TODO: implement multi-simd softmax (llama.cpp commit e16b9fa4)
-
-#version 450
-
-#include "common.comp"
-
-layout(local_size_x_id = 0) in;
-
-layout(binding = 0) buffer restrict readonly tensorInA { float inA[]; };
-layout(binding = 1) buffer restrict readonly tensorInB { float inB[]; };
-layout(binding = 2) buffer restrict writeonly tensorOut { float out_[]; };
-
-layout(push_constant) uniform PushConstants {
-    uint inAOff;
-    uint inBOff;
-    uint outOff;
-    int ne00;
-    int ne01;
-    int ne02;
-    float scale;
-    int mask;
-} pcs;
-
-void main() {
-    if (gl_SubgroupInvocationID > 31)
-        return;
-
-    const uint i03 = gl_WorkGroupID.z;
-    const uint i02 = gl_WorkGroupID.y;
-    const uint i01 = gl_WorkGroupID.x;
-
-    const uint extra_off = i03*pcs.ne02*pcs.ne01*pcs.ne00 + i02*pcs.ne01*pcs.ne00 + i01*pcs.ne00;
-    const uint psrc0 = extra_off + pcs.inAOff; // Based from inA
-    const uint pmask = i01*pcs.ne00 + pcs.inBOff; // Based from inB
-    const uint pdst = extra_off + pcs.outOff; // Based from out_
-
-    // parallel max
-    float localMax = uintBitsToFloat(0xFF800000);
-    for (uint i00 = gl_SubgroupInvocationID.x; i00 < pcs.ne00; i00 += 32) {
-        localMax = max(localMax, inA[psrc0 + i00]*pcs.scale + (pcs.mask!=0 ? inB[pmask + i00] : 0.0f));
-    }
-    float max_ = subgroupMax(localMax);
-
-    // parallel sum
-    float localSum = 0.0f;
-    for (uint i00 = gl_SubgroupInvocationID.x; i00 < pcs.ne00; i00 += 32) {
-        const float exp_psrc0 = exp(inA[psrc0 + i00]*pcs.scale + (pcs.mask!=0 ? inB[pmask + i00] : 0.0f) - max_);
-        localSum += exp_psrc0;
-        out_[pdst + i00] = exp_psrc0;
-    }
-
-    const float sum = subgroupAdd(localSum);
-    for (uint i00 = gl_SubgroupInvocationID.x; i00 < pcs.ne00; i00 += 32) {
-        out_[pdst + i00] /= sum;
-    }
-}