@fugood/llama.node 0.3.1 → 0.3.3

package/src/llama.cpp/ggml/src/{ggml-backend.c → ggml-backend.cpp}

@@ -1,3 +1,14 @@
+// Note: porting this file to C++ is a work in progress
+
+#ifdef _WIN32
+#define WIN32_LEAN_AND_MEAN
+#ifndef NOMINMAX
+#   define NOMINMAX
+#endif
+#include <windows.h>
+#endif
+
+#include "ggml-backend.h"
 #include "ggml-backend-impl.h"
 #include "ggml-alloc.h"
 #include "ggml-impl.h"
@@ -8,9 +19,14 @@
 #include <stdio.h>
 #include <stdlib.h>
 #include <string.h>
+#include <string>
+#include <vector>
 
+#ifdef __APPLE__
+#include <sys/types.h>
+#include <sys/sysctl.h>
+#endif
 
-#define MAX(a, b) ((a) > (b) ? (a) : (b))
 
 // backend buffer type
 
@@ -18,7 +34,12 @@ const char * ggml_backend_buft_name(ggml_backend_buffer_type_t buft) {
     return buft->iface.get_name(buft);
 }
 
-GGML_CALL ggml_backend_buffer_t ggml_backend_buft_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+ggml_backend_buffer_t ggml_backend_buft_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+    if (size == 0) {
+        // return a dummy buffer for zero-sized allocations
+        return ggml_backend_buffer_init(buft, {}, NULL, 0);
+    }
+
     return buft->iface.alloc_buffer(buft, size);
 }
 
@@ -34,7 +55,7 @@ size_t ggml_backend_buft_get_max_size(ggml_backend_buffer_type_t buft) {
     return SIZE_MAX;
 }
 
-GGML_CALL size_t ggml_backend_buft_get_alloc_size(ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor) {
+size_t ggml_backend_buft_get_alloc_size(ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor) {
     // get_alloc_size is optional, defaults to ggml_nbytes
     if (buft->iface.get_alloc_size) {
         size_t size = buft->iface.get_alloc_size(buft, tensor);
@@ -51,16 +72,18 @@ bool ggml_backend_buft_is_host(ggml_backend_buffer_type_t buft) {
     return false;
 }
 
-// backend buffer
+ggml_backend_dev_t ggml_backend_buft_get_device(ggml_backend_buffer_type_t buft) {
+    return buft->device;
+}
 
-GGML_CALL ggml_backend_buffer_t ggml_backend_buffer_init(
-               ggml_backend_buffer_type_t      buft,
-        struct ggml_backend_buffer_i           iface,
-               ggml_backend_buffer_context_t   context,
-               size_t                          size) {
-    ggml_backend_buffer_t buffer = malloc(sizeof(struct ggml_backend_buffer));
+// backend buffer
 
-    (*buffer) = (struct ggml_backend_buffer) {
+ggml_backend_buffer_t ggml_backend_buffer_init(
+               ggml_backend_buffer_type_t buft,
+        struct ggml_backend_buffer_i      iface,
+               void *                     context,
+               size_t                     size) {
+    ggml_backend_buffer_t buffer = new ggml_backend_buffer {
         /* .interface = */ iface,
         /* .buft      = */ buft,
         /* .context   = */ context,
@@ -72,7 +95,7 @@ GGML_CALL ggml_backend_buffer_t ggml_backend_buffer_init(
 }
 
 const char * ggml_backend_buffer_name(ggml_backend_buffer_t buffer) {
-    return buffer->iface.get_name(buffer);
+    return ggml_backend_buft_name(ggml_backend_buffer_get_type(buffer));
 }
 
 void ggml_backend_buffer_free(ggml_backend_buffer_t buffer) {
@@ -83,7 +106,7 @@ void ggml_backend_buffer_free(ggml_backend_buffer_t buffer) {
     if (buffer->iface.free_buffer != NULL) {
         buffer->iface.free_buffer(buffer);
     }
-    free(buffer);
+    delete buffer;
 }
 
 size_t ggml_backend_buffer_get_size(ggml_backend_buffer_t buffer) {
@@ -91,6 +114,11 @@ size_t ggml_backend_buffer_get_size(ggml_backend_buffer_t buffer) {
 }
 
 void * ggml_backend_buffer_get_base(ggml_backend_buffer_t buffer) {
+    // get_base is optional if the buffer is zero-sized
+    if (buffer->size == 0) {
+        return NULL;
+    }
+
     void * base = buffer->iface.get_base(buffer);
 
     GGML_ASSERT(base != NULL && "backend buffer base cannot be NULL");
@@ -98,14 +126,23 @@ void * ggml_backend_buffer_get_base(ggml_backend_buffer_t buffer) {
     return base;
 }
 
-GGML_CALL void ggml_backend_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
+void ggml_backend_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
     // init_tensor is optional
     if (buffer->iface.init_tensor) {
         buffer->iface.init_tensor(buffer, tensor);
     }
 }
 
-size_t ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer) {
+void ggml_backend_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
+    // clear is optional if the buffer is zero-sized
+    if (buffer->size == 0) {
+        return;
+    }
+
+    buffer->iface.clear(buffer, value);
+}
+
+size_t ggml_backend_buffer_get_alignment(ggml_backend_buffer_t buffer) {
     return ggml_backend_buft_get_alignment(ggml_backend_buffer_get_type(buffer));
 }
 
@@ -117,10 +154,6 @@ size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct g
     return ggml_backend_buft_get_alloc_size(ggml_backend_buffer_get_type(buffer), tensor);
 }
 
-void ggml_backend_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
-    buffer->iface.clear(buffer, value);
-}
-
 bool ggml_backend_buffer_is_host(ggml_backend_buffer_t buffer) {
     return ggml_backend_buft_is_host(ggml_backend_buffer_get_type(buffer));
 }
@@ -181,7 +214,7 @@ void ggml_backend_free(ggml_backend_t backend) {
 }
 
 ggml_backend_buffer_type_t ggml_backend_get_default_buffer_type(ggml_backend_t backend) {
-    return backend->iface.get_default_buffer_type(backend);
+    return ggml_backend_dev_buffer_type(backend->device);
 }
 
 ggml_backend_buffer_t ggml_backend_alloc_buffer(ggml_backend_t backend, size_t size) {
@@ -218,32 +251,47 @@ void ggml_backend_tensor_get_async(ggml_backend_t backend, const struct ggml_ten
     }
 }
 
-GGML_CALL void ggml_backend_tensor_set(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+void ggml_backend_tensor_set(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
     ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
 
+    if (size == 0) {
+        return;
+    }
+
     GGML_ASSERT(buf != NULL && "tensor buffer not set");
     GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
     GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor write out of bounds");
 
-    if (!size) {
-        return;
-    }
-
     buf->iface.set_tensor(buf, tensor, data, offset, size);
 }
 
-GGML_CALL void ggml_backend_tensor_get(const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+void ggml_backend_tensor_get(const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
     ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
 
+    if (size == 0) {
+        return;
+    }
+
     GGML_ASSERT(buf != NULL && "tensor buffer not set");
     GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
     GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor read out of bounds");
 
-    if (!size) {
+    buf->iface.get_tensor(buf, tensor, data, offset, size);
+}
+
+void ggml_backend_tensor_memset(struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) {
+    ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
+
+    if (size == 0) {
         return;
     }
 
-    buf->iface.get_tensor(buf, tensor, data, offset, size);
+    GGML_ASSERT(buf != NULL && "tensor buffer not set");
+    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
+    GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor write out of bounds");
+    GGML_ASSERT(buf->iface.memset_tensor != NULL && "memset not implemented by backend buffer");
+
+    buf->iface.memset_tensor(buf, tensor, value, offset, size);
 }
 
 void ggml_backend_synchronize(ggml_backend_t backend) {
@@ -283,18 +331,19 @@ enum ggml_status ggml_backend_graph_compute_async(ggml_backend_t backend, struct
 }
 
 bool ggml_backend_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
-    return backend->iface.supports_op(backend, op);
+    return ggml_backend_dev_supports_op(backend->device, op);
 }
 
 bool ggml_backend_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft) {
-    return backend->iface.supports_buft(backend, buft);
+    return ggml_backend_dev_supports_buft(backend->device, buft);
 }
 
 bool ggml_backend_offload_op(ggml_backend_t backend, const struct ggml_tensor * op) {
-    if (backend->iface.offload_op != NULL) {
-        return backend->iface.offload_op(backend, op);
-    }
-    return false;
+    return ggml_backend_dev_offload_op(backend->device, op);
+}
+
+ggml_backend_dev_t ggml_backend_get_device(ggml_backend_t backend) {
+    return backend->device;
 }
 
 // backend copy
@@ -327,7 +376,7 @@ void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst
         ggml_backend_tensor_get(src, dst->data, 0, ggml_nbytes(src));
     } else if (!ggml_backend_buffer_copy_tensor(src, dst)) {
 #ifndef NDEBUG
-        fprintf(stderr, "%s: warning: slow copy from %s to %s\n", __func__, ggml_backend_buffer_name(src->buffer), ggml_backend_buffer_name(dst->buffer));
+        GGML_LOG_DEBUG("%s: warning: slow copy from %s to %s\n", __func__, ggml_backend_buffer_name(src->buffer), ggml_backend_buffer_name(dst->buffer));
 #endif
         size_t nbytes = ggml_nbytes(src);
         void * data = malloc(nbytes);
@@ -351,43 +400,39 @@ void ggml_backend_tensor_copy_async(ggml_backend_t backend_src, ggml_backend_t b
     }
 
     // an async copy would normally happen after all the queued operations on both backends are completed
-    // sync src, set_async dst
-    if (ggml_backend_buffer_is_host(src->buffer)) {
-        ggml_backend_synchronize(backend_src);
-        ggml_backend_tensor_set_async(backend_dst, dst, src->data, 0, ggml_nbytes(src));
-    } else {
-        ggml_backend_synchronize(backend_src);
-        ggml_backend_tensor_copy(src, dst);
-        ggml_backend_synchronize(backend_dst);
-    }
+    // to simulate the same behavior, we need to synchronize both backends first, and do a blocking copy
+    ggml_backend_synchronize(backend_src);
+    ggml_backend_synchronize(backend_dst);
+    ggml_backend_tensor_copy(src, dst);
 }
 
 // events
 
-ggml_backend_event_t ggml_backend_event_new(ggml_backend_t backend) {
-    if (backend->iface.event_new == NULL) {
+ggml_backend_event_t ggml_backend_event_new(ggml_backend_dev_t device) {
+    // null device is allowed for the transition period to the device interface
+    if (device == NULL || device->iface.event_new == NULL) {
         return NULL;
     }
-    return backend->iface.event_new(backend);
+    return device->iface.event_new(device);
 }
 
 void ggml_backend_event_free(ggml_backend_event_t event) {
     if (event == NULL) {
         return;
     }
-    event->backend->iface.event_free(event);
+    event->device->iface.event_free(event->device, event);
 }
 
-void ggml_backend_event_record(ggml_backend_event_t event) {
-    GGML_ASSERT(event->backend->iface.event_record != NULL);
+void ggml_backend_event_record(ggml_backend_event_t event, ggml_backend_t backend) {
+    GGML_ASSERT(backend->iface.event_record != NULL);
 
-    event->backend->iface.event_record(event);
+    backend->iface.event_record(backend, event);
 }
 
 void ggml_backend_event_synchronize(ggml_backend_event_t event) {
-    GGML_ASSERT(event->backend->iface.event_synchronize != NULL);
+    GGML_ASSERT(event->device->iface.event_synchronize);
 
-    event->backend->iface.event_synchronize(event);
+    event->device->iface.event_synchronize(event->device, event);
 }
 
 void ggml_backend_event_wait(ggml_backend_t backend, ggml_backend_event_t event) {
@@ -396,536 +441,88 @@ void ggml_backend_event_wait(ggml_backend_t backend, ggml_backend_event_t event)
     backend->iface.event_wait(backend, event);
 }
 
-// backend registry
-
-#define GGML_REG_MAX_BACKENDS 64
-
-struct ggml_backend_reg {
-    char name[128];
-    ggml_backend_init_fn init_fn;
-    ggml_backend_buffer_type_t default_buffer_type;
-    void * user_data;
-};
-
-static struct ggml_backend_reg ggml_backend_registry[GGML_REG_MAX_BACKENDS];
-static size_t ggml_backend_registry_count = 0;
-
-GGML_CALL static ggml_backend_t ggml_backend_reg_cpu_init(const char * params, void * user_data);
-
-GGML_CALL static void ggml_backend_registry_init(void) {
-    static bool initialized = false;
-
-    if (initialized) {
-        return;
-    }
-
-    initialized = true;
-
-    ggml_backend_register("CPU", ggml_backend_reg_cpu_init, ggml_backend_cpu_buffer_type(), NULL);
-
-    // add forward decls here to avoid including the backend headers
-#ifdef GGML_USE_CUDA
-    extern GGML_CALL void ggml_backend_cuda_reg_devices(void);
-    ggml_backend_cuda_reg_devices();
-#endif
-
-#ifdef GGML_USE_SYCL
-    extern void ggml_backend_sycl_reg_devices(void);
-    ggml_backend_sycl_reg_devices();
-#endif
-
-#ifdef GGML_USE_METAL
-    extern GGML_CALL ggml_backend_t ggml_backend_reg_metal_init(const char * params, void * user_data);
-    extern GGML_CALL ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void);
-    ggml_backend_register("Metal", ggml_backend_reg_metal_init, ggml_backend_metal_buffer_type(), NULL);
-#endif
-
-#ifdef GGML_USE_VULKAN
-    extern GGML_CALL int ggml_backend_vk_reg_devices(void);
-    ggml_backend_vk_reg_devices();
-#endif
-
-#ifdef GGML_USE_KOMPUTE
-    extern GGML_CALL void ggml_backend_kompute_reg_devices(void);
-    ggml_backend_kompute_reg_devices();
-#endif
-
-#ifdef GGML_USE_CANN
-    extern GGML_CALL int ggml_backend_cann_reg_devices(void);
-    ggml_backend_cann_reg_devices();
-#endif
-}
-
-GGML_CALL void ggml_backend_register(const char * name, ggml_backend_init_fn init_fn, ggml_backend_buffer_type_t default_buffer_type, void * user_data) {
-    GGML_ASSERT(ggml_backend_registry_count < GGML_REG_MAX_BACKENDS);
-
-    size_t id = ggml_backend_registry_count;
-
-    ggml_backend_registry[id] = (struct ggml_backend_reg) {
-        /* .name                = */ {0},
-        /* .fn                  = */ init_fn,
-        /* .default_buffer_type = */ default_buffer_type,
-        /* .user_data           = */ user_data,
-    };
-
-    snprintf(ggml_backend_registry[id].name, sizeof(ggml_backend_registry[id].name), "%s", name);
-
-#ifndef NDEBUG
-    fprintf(stderr, "%s: registered backend %s\n", __func__, name);
-#endif
-
-    ggml_backend_registry_count++;
-}
-
-size_t ggml_backend_reg_get_count(void) {
-    ggml_backend_registry_init();
+// Backend device
 
-    return ggml_backend_registry_count;
+const char * ggml_backend_dev_name(ggml_backend_dev_t device) {
+    return device->iface.get_name(device);
 }
 
-size_t ggml_backend_reg_find_by_name(const char * name) {
-    ggml_backend_registry_init();
-
-    for (size_t i = 0; i < ggml_backend_registry_count; i++) {
-        // TODO: case insensitive in a portable way
-        if (strcmp(ggml_backend_registry[i].name, name) == 0) {
-            return i;
-        }
-    }
-
-    // not found
-    return SIZE_MAX;
+const char * ggml_backend_dev_description(ggml_backend_dev_t device) {
+    return device->iface.get_description(device);
 }
 
-// init from backend:params string
-ggml_backend_t ggml_backend_reg_init_backend_from_str(const char * backend_str) {
-    ggml_backend_registry_init();
-
-    const char * params = strchr(backend_str, ':');
-    char backend_name[128];
-    if (params == NULL) {
-        snprintf(backend_name, sizeof(backend_name), "%s", backend_str);
-        params = "";
-    } else {
-        snprintf(backend_name, sizeof(backend_name), "%.*s", (int)(params - backend_str), backend_str);
-        params++;
-    }
-
-    size_t backend_i = ggml_backend_reg_find_by_name(backend_name);
-
-    if (backend_i == SIZE_MAX) {
-        fprintf(stderr, "%s: backend %s not found\n", __func__, backend_name);
-        return NULL;
-    }
-
-    return ggml_backend_reg_init_backend(backend_i, params);
+void ggml_backend_dev_memory(ggml_backend_dev_t device, size_t * free, size_t * total) {
+    device->iface.get_memory(device, free, total);
 }
 
-const char * ggml_backend_reg_get_name(size_t i) {
-    ggml_backend_registry_init();
-
-    GGML_ASSERT(i < ggml_backend_registry_count);
-    return ggml_backend_registry[i].name;
+enum ggml_backend_dev_type ggml_backend_dev_type(ggml_backend_dev_t device) {
+    return device->iface.get_type(device);
 }
 
-ggml_backend_t ggml_backend_reg_init_backend(size_t i, const char * params) {
-    ggml_backend_registry_init();
-
-    GGML_ASSERT(i < ggml_backend_registry_count);
-    return ggml_backend_registry[i].init_fn(params, ggml_backend_registry[i].user_data);
+void ggml_backend_dev_get_props(ggml_backend_dev_t device, struct ggml_backend_dev_props * props) {
+    memset(props, 0, sizeof(*props));
+    device->iface.get_props(device, props);
 }
 
-ggml_backend_buffer_type_t ggml_backend_reg_get_default_buffer_type(size_t i) {
-    ggml_backend_registry_init();
-
-    GGML_ASSERT(i < ggml_backend_registry_count);
-    return ggml_backend_registry[i].default_buffer_type;
-}
-
-ggml_backend_buffer_t ggml_backend_reg_alloc_buffer(size_t i, size_t size) {
-    ggml_backend_registry_init();
-
-    GGML_ASSERT(i < ggml_backend_registry_count);
-    return ggml_backend_buft_alloc_buffer(ggml_backend_registry[i].default_buffer_type, size);
-}
-
-// backend CPU
-
-static const size_t TENSOR_ALIGNMENT = 32; // required for mmap as gguf only guarantees 32-byte alignment
-
-GGML_CALL static const char * ggml_backend_cpu_buffer_name(ggml_backend_buffer_t buffer) {
-    return "CPU";
-
-    GGML_UNUSED(buffer);
-}
-
-GGML_CALL static void * ggml_backend_cpu_buffer_get_base(ggml_backend_buffer_t buffer) {
-    uintptr_t data = (uintptr_t)buffer->context;
-
-    // align the buffer
-    if (data % TENSOR_ALIGNMENT != 0) {
-        data = GGML_PAD(data, TENSOR_ALIGNMENT);
-    }
-
-    return (void *)data;
-}
-
-GGML_CALL static void ggml_backend_cpu_buffer_free_buffer(ggml_backend_buffer_t buffer) {
-    free(buffer->context);
-}
-
-GGML_CALL static void ggml_backend_cpu_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
-    memcpy((char *)tensor->data + offset, data, size);
-
-    GGML_UNUSED(buffer);
-}
-
-GGML_CALL static void ggml_backend_cpu_buffer_get_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
-    memcpy(data, (const char *)tensor->data + offset, size);
-
-    GGML_UNUSED(buffer);
-}
-
-GGML_CALL static bool ggml_backend_cpu_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)) {
-        memcpy(dst->data, src->data, ggml_nbytes(src));
-        return true;
-    }
-    return false;
-
-    GGML_UNUSED(buffer);
-}
-
-GGML_CALL static void ggml_backend_cpu_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
-    memset(buffer->context, value, buffer->size);
-}
-
-static struct ggml_backend_buffer_i cpu_backend_buffer_i = {
-    /* .get_name        = */ ggml_backend_cpu_buffer_name,
-    /* .free_buffer     = */ ggml_backend_cpu_buffer_free_buffer,
-    /* .get_base        = */ ggml_backend_cpu_buffer_get_base,
-    /* .init_tensor     = */ NULL, // no initialization required
-    /* .set_tensor      = */ ggml_backend_cpu_buffer_set_tensor,
-    /* .get_tensor      = */ ggml_backend_cpu_buffer_get_tensor,
-    /* .cpy_tensor      = */ ggml_backend_cpu_buffer_cpy_tensor,
-    /* .clear           = */ ggml_backend_cpu_buffer_clear,
-    /* .reset           = */ NULL,
-};
-
-// for buffers from ptr, free is not called
-static struct ggml_backend_buffer_i cpu_backend_buffer_i_from_ptr = {
-    /* .get_name        = */ ggml_backend_cpu_buffer_name,
-    /* .free_buffer     = */ NULL, // ptr is not owned by the buffer, so it does not need to be freed
-    /* .get_base        = */ ggml_backend_cpu_buffer_get_base,
-    /* .init_tensor     = */ NULL, // no initialization required
-    /* .set_tensor      = */ ggml_backend_cpu_buffer_set_tensor,
-    /* .get_tensor      = */ ggml_backend_cpu_buffer_get_tensor,
-    /* .cpy_tensor      = */ ggml_backend_cpu_buffer_cpy_tensor,
-    /* .clear           = */ ggml_backend_cpu_buffer_clear,
-    /* .reset           = */ NULL,
-};
-
-GGML_CALL static const char * ggml_backend_cpu_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
-    return "CPU";
-
-    GGML_UNUSED(buft);
-}
-
-GGML_CALL static ggml_backend_buffer_t ggml_backend_cpu_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
-    size += TENSOR_ALIGNMENT;   // malloc may return an address that is not aligned
-    void * data = malloc(size); // TODO: use GGML_ALIGNED_MALLOC (move to ggml-impl.h)
-    if (data == NULL) {
-        fprintf(stderr, "%s: failed to allocate buffer of size %zu\n", __func__, size);
-        return NULL;
-    }
-
-    return ggml_backend_buffer_init(buft, cpu_backend_buffer_i, data, size);
-}
-
-GGML_CALL static size_t ggml_backend_cpu_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
-    return TENSOR_ALIGNMENT;
-
-    GGML_UNUSED(buft);
-}
-
-GGML_CALL static bool ggml_backend_cpu_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
-    return true;
-
-    GGML_UNUSED(buft);
-}
-
-GGML_CALL ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void) {
-    static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type = {
-        /* .iface = */ {
-            /* .get_name         = */ ggml_backend_cpu_buffer_type_get_name,
-            /* .alloc_buffer     = */ ggml_backend_cpu_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;
-}
-
-#ifdef GGML_USE_CPU_HBM
-
-// buffer type HBM
-
-#include <hbwmalloc.h>
-
-GGML_CALL static const char * ggml_backend_cpu_hbm_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
-    return "CPU_HBM";
-
-    GGML_UNUSED(buft);
+ggml_backend_reg_t ggml_backend_dev_backend_reg(ggml_backend_dev_t device) {
+    return device->reg;
 }
 
-GGML_CALL static const char * ggml_backend_cpu_hbm_buffer_get_name(ggml_backend_buffer_t buf) {
-    return "CPU_HBM";
-
-    GGML_UNUSED(buf);
+ggml_backend_t ggml_backend_dev_init(ggml_backend_dev_t device, const char * params) {
+    return device->iface.init_backend(device, params);
 }
 
-GGML_CALL static void ggml_backend_cpu_hbm_buffer_free_buffer(ggml_backend_buffer_t buffer) {
-    hbw_free(buffer->context);
+ggml_backend_buffer_type_t ggml_backend_dev_buffer_type(ggml_backend_dev_t device) {
+    return device->iface.get_buffer_type(device);
 }
 
-GGML_CALL static ggml_backend_buffer_t ggml_backend_cpu_hbm_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
-    //void * ptr = hbw_malloc(size);
-    void * ptr;
-    int result = hbw_posix_memalign(&ptr, ggml_backend_cpu_buffer_type_get_alignment(buft), size);
-    if (result != 0) {
-        fprintf(stderr, "failed to allocate HBM buffer of size %zu\n", size);
+ggml_backend_buffer_type_t ggml_backend_dev_host_buffer_type(ggml_backend_dev_t device) {
+    if (device->iface.get_host_buffer_type == NULL) {
         return NULL;
     }
 
-    ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(ptr, size);
-    buffer->buft = buft;
-    buffer->iface.get_name = ggml_backend_cpu_hbm_buffer_get_name;
-    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;
+    return device->iface.get_host_buffer_type(device);
 }
-#endif
-
-struct ggml_backend_cpu_context {
-    int n_threads;
-    void * work_data;
-    size_t work_size;
-
-    ggml_abort_callback abort_callback;
-    void *              abort_callback_data;
-};
-
-GGML_CALL static const char * ggml_backend_cpu_name(ggml_backend_t backend) {
-    return "CPU";
 
-    GGML_UNUSED(backend);
+ggml_backend_buffer_t ggml_backend_dev_buffer_from_host_ptr(ggml_backend_dev_t device, void * ptr, size_t size, size_t max_tensor_size) {
+    return device->iface.buffer_from_host_ptr(device, ptr, size, max_tensor_size);
 }
 
-GGML_CALL static void ggml_backend_cpu_free(ggml_backend_t backend) {
-    struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
-    free(cpu_ctx->work_data);
-    free(cpu_ctx);
-    free(backend);
+bool ggml_backend_dev_supports_op(ggml_backend_dev_t device, const struct ggml_tensor * op) {
+    return device->iface.supports_op(device, op);
 }
 
-GGML_CALL static ggml_backend_buffer_type_t ggml_backend_cpu_get_default_buffer_type(ggml_backend_t backend) {
-    return ggml_backend_cpu_buffer_type();
-
-    GGML_UNUSED(backend);
+bool ggml_backend_dev_supports_buft(ggml_backend_dev_t device, ggml_backend_buffer_type_t buft) {
+    return device->iface.supports_buft(device, buft);
 }
 
-struct ggml_backend_plan_cpu {
-    struct ggml_cplan cplan;
-    struct ggml_cgraph cgraph;
-};
-
-GGML_CALL 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 = malloc(sizeof(struct ggml_backend_plan_cpu));
-
-    cpu_plan->cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads);
-    cpu_plan->cgraph = *cgraph; // FIXME: deep copy
-
-    if (cpu_plan->cplan.work_size > 0) {
-        cpu_plan->cplan.work_data = malloc(cpu_plan->cplan.work_size);
-        if (cpu_plan->cplan.work_data == NULL) {
-            free(cpu_plan);
-            return NULL;
-        }
+bool ggml_backend_dev_offload_op(ggml_backend_dev_t device, const struct ggml_tensor * op) {
+    if (device->iface.offload_op != NULL) {
+        return device->iface.offload_op(device, op);
     }
 
-    cpu_plan->cplan.abort_callback      = cpu_ctx->abort_callback;
-    cpu_plan->cplan.abort_callback_data = cpu_ctx->abort_callback_data;
-
-    return cpu_plan;
+    return false;
 }
 
-GGML_CALL 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;
-
-    free(cpu_plan->cplan.work_data);
-    free(cpu_plan);
+// Backend (reg)
 
-    GGML_UNUSED(backend);
+const char * ggml_backend_reg_name(ggml_backend_reg_t reg) {
+    return reg->iface.get_name(reg);
 }
 
-GGML_CALL 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);
+size_t ggml_backend_reg_dev_count(ggml_backend_reg_t reg) {
+    return reg->iface.get_device_count(reg);
 }
 
-GGML_CALL 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);
-
-    if (cpu_ctx->work_size < cplan.work_size) {
-        free(cpu_ctx->work_data);
-        cpu_ctx->work_data = malloc(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 = 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);
-}
-
-GGML_CALL static bool ggml_backend_cpu_supports_op(ggml_backend_t backend, 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 || op->src[1]->type == ggml_internal_get_type_traits(op->src[0]->type).vec_dot_type;
-        default:
-            return true;
-    }
-
-    GGML_UNUSED(backend);
-}
-
-GGML_CALL static bool ggml_backend_cpu_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft) {
-    return ggml_backend_buft_is_host(buft);
-
-    GGML_UNUSED(backend);
-}
-
-static struct ggml_backend_i cpu_backend_i = {
-    /* .get_name                = */ ggml_backend_cpu_name,
-    /* .free                    = */ ggml_backend_cpu_free,
-    /* .get_default_buffer_type = */ ggml_backend_cpu_get_default_buffer_type,
-    /* .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,
-    /* .supports_op             = */ ggml_backend_cpu_supports_op,
-    /* .supports_buft           = */ ggml_backend_cpu_supports_buft,
-    /* .offload_op              = */ NULL,
-    /* .event_new               = */ NULL,
-    /* .event_free              = */ NULL,
-    /* .event_record            = */ NULL,
-    /* .event_wait              = */ NULL,
-    /* .event_synchronize       = */ 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_dev_t ggml_backend_reg_dev_get(ggml_backend_reg_t reg, size_t index) {
+    return reg->iface.get_device(reg, index);
 }
 
-ggml_backend_t ggml_backend_cpu_init(void) {
-    struct ggml_backend_cpu_context * ctx = malloc(sizeof(struct ggml_backend_cpu_context));
-    if (ctx == NULL) {
-        return NULL;
-    }
-
-    ctx->n_threads           = GGML_DEFAULT_N_THREADS;
-    ctx->work_data           = NULL;
-    ctx->work_size           = 0;
-    ctx->abort_callback      = NULL;
-    ctx->abort_callback_data = NULL;
-
-    ggml_backend_t cpu_backend = malloc(sizeof(struct ggml_backend));
-    if (cpu_backend == NULL) {
-        free(ctx);
+void * ggml_backend_reg_get_proc_address(ggml_backend_reg_t reg, const char * name) {
+    if (!reg->iface.get_proc_address) {
         return NULL;
     }
-
-    *cpu_backend = (struct ggml_backend) {
-        /* .guid      = */ ggml_backend_cpu_guid(),
-        /* .interface = */ cpu_backend_i,
-        /* .context   = */ ctx
-    };
-    return cpu_backend;
-}
-
-GGML_CALL 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_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_CALL 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_type(), cpu_backend_buffer_i_from_ptr, ptr, size);
-}
-
-GGML_CALL static ggml_backend_t ggml_backend_reg_cpu_init(const char * params, void * user_data) {
-    return ggml_backend_cpu_init();
-
-    GGML_UNUSED(params);
-    GGML_UNUSED(user_data);
+    return reg->iface.get_proc_address(reg, name);
 }
 
 // multi-buffer buffer
@@ -935,16 +532,8 @@ struct ggml_backend_multi_buffer_context {
     size_t n_buffers;
 };
 
-typedef struct ggml_backend_multi_buffer_context * ggml_backend_multi_buffer_context_t;
-
-GGML_CALL static const char * ggml_backend_multi_buffer_get_name(ggml_backend_buffer_t buffer) {
-    ggml_backend_multi_buffer_context_t ctx = (ggml_backend_multi_buffer_context_t) buffer->context;
-
-    return ctx->buffers[0]->iface.get_name(ctx->buffers[0]);
-}
-
-GGML_CALL static void ggml_backend_multi_buffer_free_buffer(ggml_backend_buffer_t buffer) {
-    ggml_backend_multi_buffer_context_t ctx = (ggml_backend_multi_buffer_context_t) buffer->context;
+static void ggml_backend_multi_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+    ggml_backend_multi_buffer_context * ctx = (ggml_backend_multi_buffer_context *) buffer->context;
     for (size_t i = 0; i < ctx->n_buffers; i++) {
         ggml_backend_buffer_free(ctx->buffers[i]);
     }
@@ -953,31 +542,27 @@ GGML_CALL static void ggml_backend_multi_buffer_free_buffer(ggml_backend_buffer_
     free(ctx);
 }
 
-GGML_CALL static void ggml_backend_multi_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
-    ggml_backend_multi_buffer_context_t ctx = (ggml_backend_multi_buffer_context_t) buffer->context;
+static void ggml_backend_multi_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
+    ggml_backend_multi_buffer_context * ctx = (ggml_backend_multi_buffer_context *) buffer->context;
     for (size_t i = 0; i < ctx->n_buffers; i++) {
         ggml_backend_buffer_clear(ctx->buffers[i], value);
     }
 }
 
-static struct ggml_backend_buffer_i ggml_backend_multi_buffer_context_interface(void) {
-    static struct ggml_backend_buffer_i multi_backend_buffer_i = {
-        /* .get_name        = */ ggml_backend_multi_buffer_get_name,
-        /* .free_buffer     = */ ggml_backend_multi_buffer_free_buffer,
-        /* .get_base        = */ NULL,
-        /* .init_tensor     = */ NULL,
-        /* .set_tensor      = */ NULL,
-        /* .get_tensor      = */ NULL,
-        /* .cpy_tensor      = */ NULL,
-        /* .clear           = */ ggml_backend_multi_buffer_clear,
-        /* .reset           = */ NULL,
-    };
-
-    return multi_backend_buffer_i;
-}
+static const struct ggml_backend_buffer_i ggml_backend_multi_buffer_i = {
+    /* .free_buffer     = */ ggml_backend_multi_buffer_free_buffer,
+    /* .get_base        = */ NULL,
+    /* .init_tensor     = */ NULL,
+    /* .memset_tensor   = */ NULL,
+    /* .set_tensor      = */ NULL,
+    /* .get_tensor      = */ NULL,
+    /* .cpy_tensor      = */ NULL,
+    /* .clear           = */ ggml_backend_multi_buffer_clear,
+    /* .reset           = */ NULL,
+};
 
-GGML_CALL ggml_backend_buffer_t ggml_backend_multi_buffer_alloc_buffer(ggml_backend_buffer_t * buffers, size_t n_buffers) {
-    ggml_backend_multi_buffer_context_t ctx = (ggml_backend_multi_buffer_context_t) malloc(sizeof(struct ggml_backend_multi_buffer_context));
+ggml_backend_buffer_t ggml_backend_multi_buffer_alloc_buffer(ggml_backend_buffer_t * buffers, size_t n_buffers) {
+    ggml_backend_multi_buffer_context * ctx = (ggml_backend_multi_buffer_context *) malloc(sizeof(struct ggml_backend_multi_buffer_context));
     ctx->n_buffers = n_buffers;
     ctx->buffers = (ggml_backend_buffer_t *) malloc(n_buffers * sizeof(ggml_backend_buffer_t));
 
@@ -989,16 +574,16 @@ GGML_CALL ggml_backend_buffer_t ggml_backend_multi_buffer_alloc_buffer(ggml_back
         total_size += ggml_backend_buffer_get_size(buffers[i]);
     }
 
-    return ggml_backend_buffer_init(buffers[0]->buft, ggml_backend_multi_buffer_context_interface(), ctx, total_size);
+    return ggml_backend_buffer_init(buffers[0]->buft, ggml_backend_multi_buffer_i, ctx, total_size);
 }
 
-GGML_CALL bool ggml_backend_buffer_is_multi_buffer(ggml_backend_buffer_t buffer) {
-    return buffer->iface.get_name == ggml_backend_multi_buffer_get_name;
+bool ggml_backend_buffer_is_multi_buffer(ggml_backend_buffer_t buffer) {
+    return buffer->iface.free_buffer == ggml_backend_multi_buffer_free_buffer;
 }
 
-GGML_CALL void ggml_backend_multi_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage) {
+void ggml_backend_multi_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage) {
     GGML_ASSERT(ggml_backend_buffer_is_multi_buffer(buffer));
-    ggml_backend_multi_buffer_context_t ctx = (ggml_backend_multi_buffer_context_t) buffer->context;
+    ggml_backend_multi_buffer_context * ctx = (ggml_backend_multi_buffer_context *) buffer->context;
     for (size_t i = 0; i < ctx->n_buffers; i++) {
         ggml_backend_buffer_set_usage(ctx->buffers[i], usage);
     }
@@ -1023,10 +608,6 @@ static bool ggml_is_view_op(enum ggml_op op) {
 #define GGML_SCHED_MAX_BACKENDS 16
 #endif
 
-#ifndef GGML_SCHED_MAX_SPLITS
-#define GGML_SCHED_MAX_SPLITS 2048
-#endif
-
 #ifndef GGML_SCHED_MAX_SPLIT_INPUTS
 #define GGML_SCHED_MAX_SPLIT_INPUTS GGML_MAX_SRC
 #endif
@@ -1089,7 +670,7 @@ struct ggml_backend_sched {
     char * context_buffer;
     size_t context_buffer_size;
 
-    bool debug;
+    int debug;
 };
 
 #define hash_id(tensor) ggml_hash_find_or_insert(&sched->hash_set, tensor)
@@ -1108,7 +689,7 @@ static int ggml_backend_sched_backend_id(ggml_backend_sched_t sched, ggml_backen
 }
 
 static int ggml_backend_sched_backend_from_buffer(ggml_backend_sched_t sched, const struct ggml_tensor * tensor, const struct ggml_tensor * op) {
-    ggml_backend_buffer_t buffer = tensor->buffer;
+    ggml_backend_buffer_t buffer = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
     if (buffer == NULL) {
         return -1;
     }
@@ -1122,7 +703,7 @@ static int ggml_backend_sched_backend_from_buffer(ggml_backend_sched_t sched, co
     }
 
 #ifndef NDEBUG
-    fprintf(stderr, "%s: warning: no backend supports op %s with a weight with buffer type %s used in tensor %s, the weight will need to be copied\n",
+    GGML_LOG_DEBUG("%s: warning: no backend supports op %s with a weight with buffer type %s used in tensor %s, the weight will need to be copied\n",
         __func__, ggml_op_desc(tensor), ggml_backend_buffer_name(buffer), tensor->name);
 #endif
 
@@ -1130,7 +711,8 @@ static int ggml_backend_sched_backend_from_buffer(ggml_backend_sched_t sched, co
 }
 
 #if 0
-static char causes[GGML_DEFAULT_GRAPH_SIZE*16 + GGML_SCHED_MAX_SPLITS*GGML_SCHED_MAX_SPLIT_INPUTS][128]; // debug only
+#define GGML_SCHED_MAX_SPLITS_DEBUG 4096
+static char causes[GGML_DEFAULT_GRAPH_SIZE*16 + GGML_SCHED_MAX_SPLITS_DEBUG*GGML_SCHED_MAX_SPLIT_INPUTS][128]; // debug only
 #define SET_CAUSE(node, ...) sprintf(causes[hash_id(node)], __VA_ARGS__)
 #define GET_CAUSE(node) causes[hash_id(node)]
 #else
@@ -1140,8 +722,6 @@ static char causes[GGML_DEFAULT_GRAPH_SIZE*16 + GGML_SCHED_MAX_SPLITS*GGML_SCHED
 
 // returns the backend that should be used for the node based on the current locations
 static int ggml_backend_sched_backend_id_from_cur(ggml_backend_sched_t sched, struct ggml_tensor * tensor) {
-    // TODO: use supports_op to check if the backend supports the op
-
     // assign pre-allocated nodes to their backend
     int cur_backend_id = ggml_backend_sched_backend_from_buffer(sched, tensor, tensor);
     if (cur_backend_id != -1) {
@@ -1158,6 +738,11 @@ static int ggml_backend_sched_backend_id_from_cur(ggml_backend_sched_t sched, st
         }
     }
 
+    if (tensor->buffer || (tensor->view_src && tensor->view_src->buffer)) {
+        // since the tensor is pre-allocated, it cannot be moved to another backend
+        GGML_ABORT("pre-allocated tensor (%s) in a backend that cannot run the operation", tensor->name);
+    }
+
     // graph input
     if (tensor->flags & GGML_TENSOR_FLAG_INPUT) {
         cur_backend_id = sched->n_backends - 1; // last backend (assumed CPU)
@@ -1171,7 +756,9 @@ static int ggml_backend_sched_backend_id_from_cur(ggml_backend_sched_t sched, st
         if (src == NULL) {
             continue;
         }
-        if (src->buffer != NULL && src->buffer->usage == GGML_BACKEND_BUFFER_USAGE_WEIGHTS) {
+        // skip ROPE since the rope freqs tensor is too small to choose a backend based on it
+        // not an ideal solution
+        if (tensor->op != GGML_OP_ROPE && src->buffer != NULL && src->buffer->usage == GGML_BACKEND_BUFFER_USAGE_WEIGHTS) {
             int src_backend_id = ggml_backend_sched_backend_from_buffer(sched, src, tensor);
             // check if a backend with higher prio wants to offload the op
             if (src_backend_id == sched->n_backends - 1) {
@@ -1205,32 +792,34 @@ static void ggml_backend_sched_print_assignments(ggml_backend_sched_t sched, str
     for (int i = 0; i < graph->n_nodes; i++) {
         if (cur_split < sched->n_splits && i == sched->splits[cur_split].i_start) {
             ggml_backend_t split_backend = sched->backends[sched->splits[cur_split].backend_id];
-            fprintf(stderr, "\n## SPLIT #%d: %s # %d inputs: ", cur_split, ggml_backend_name(split_backend),
+            GGML_LOG_DEBUG("\n## SPLIT #%d: %s # %d inputs: ", cur_split, ggml_backend_name(split_backend),
                 sched->splits[cur_split].n_inputs);
             for (int j = 0; j < sched->splits[cur_split].n_inputs; j++) {
-                fprintf(stderr, "[%s (%5.5s)] ", sched->splits[cur_split].inputs[j]->name,
+                GGML_LOG_DEBUG("[%s (%5.5s)] ", sched->splits[cur_split].inputs[j]->name,
                     fmt_size(ggml_nbytes(sched->splits[cur_split].inputs[j])));
             }
-            fprintf(stderr, "\n");
+            GGML_LOG_DEBUG("\n");
             cur_split++;
         }
         struct ggml_tensor * node = graph->nodes[i];
         if (ggml_is_view_op(node->op)) {
             continue;
         }
-        ggml_backend_t tensor_backend = ggml_backend_sched_get_tensor_backend(sched, node);
-        fprintf(stderr, "node #%3d (%10.10s): %20.20s (%5.5s) [%5.5s %8.8s]:", i, ggml_op_name(node->op), node->name,
-            fmt_size(ggml_nbytes(node)), tensor_backend ? ggml_backend_name(tensor_backend) : "NULL", GET_CAUSE(node));
-        for (int j = 0; j < GGML_MAX_SRC; j++) {
-            struct ggml_tensor * src = node->src[j];
-            if (src == NULL) {
-                continue;
+        if (sched->debug > 1) {
+            ggml_backend_t tensor_backend = ggml_backend_sched_get_tensor_backend(sched, node);
+            GGML_LOG_DEBUG("node #%3d (%10.10s): %20.20s (%5.5s) [%5.5s %8.8s]:", i, ggml_op_name(node->op), node->name,
+                fmt_size(ggml_nbytes(node)), tensor_backend ? ggml_backend_name(tensor_backend) : "NULL", GET_CAUSE(node));
+            for (int j = 0; j < GGML_MAX_SRC; j++) {
+                struct ggml_tensor * src = node->src[j];
+                if (src == NULL) {
+                    continue;
+                }
+                ggml_backend_t src_backend = ggml_backend_sched_get_tensor_backend(sched, src);
+                GGML_LOG_DEBUG(" %20.20s (%5.5s) [%5.5s %8.8s]", src->name,
+                    fmt_size(ggml_nbytes(src)), src_backend ? ggml_backend_name(src_backend) : "NULL", GET_CAUSE(src));
             }
-            ggml_backend_t src_backend = ggml_backend_sched_get_tensor_backend(sched, src);
-            fprintf(stderr, " %20.20s (%5.5s) [%5.5s %8.8s]", src->name,
-                fmt_size(ggml_nbytes(src)), src_backend ? ggml_backend_name(src_backend) : "NULL", GET_CAUSE(src));
+            GGML_LOG_DEBUG("\n");
         }
-        fprintf(stderr, "\n");
     }
 }
 
@@ -1295,6 +884,9 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
     for (int i = 0; i < graph->n_nodes; i++) {
         struct ggml_tensor * node = graph->nodes[i];
         int * node_backend_id = &tensor_backend_id(node);
+        if (ggml_is_view_op(node->op)) {
+            continue;
+        }
         // do not overwrite user assignments
         if (*node_backend_id == -1) {
             *node_backend_id = ggml_backend_sched_backend_id_from_cur(sched, node);
@@ -1522,11 +1114,11 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
                     if (src == NULL) {
                         continue;
                     }
-                    // check if a weight is on a different backend
+                    // check if a weight is on a different and incompatible backend
                     // by starting a new split, the memory of the previously offloaded weights can be reused
                     if (src->buffer != NULL && src->buffer->usage == GGML_BACKEND_BUFFER_USAGE_WEIGHTS) {
                         int src_backend_id = tensor_backend_id(src);
-                        if (src_backend_id != cur_backend_id) {
+                        if (src_backend_id != cur_backend_id && !ggml_backend_sched_buffer_supported(sched, src, cur_backend_id)) {
                             need_new_split = true;
                             break;
                         }
@@ -1538,7 +1130,6 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
                         int src_backend_id = sched->hv_tensor_backend_ids[id];
                         bool supported = ggml_backend_sched_buffer_supported(sched, src, cur_backend_id);
                         if (src_backend_id != cur_backend_id && tensor_id_copy(id, cur_backend_id, 0) == NULL && !supported) {
-                            //printf("starting new split because of too many inputs: node %s, input %s\n", node->name, src->name);
                             need_new_split = true;
                             break;
                         }
@@ -1551,10 +1142,10 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
                 i_split++;
                 if (i_split >= sched->splits_capacity) {
                     sched->splits_capacity *= 2;
-                    sched->splits = realloc(sched->splits, sched->splits_capacity * sizeof(struct ggml_backend_sched_split));
+                    sched->splits = (ggml_backend_sched_split *)
+                        realloc(sched->splits, sched->splits_capacity * sizeof(struct ggml_backend_sched_split));
                     GGML_ASSERT(sched->splits != NULL);
                 }
-                GGML_ASSERT(i_split < GGML_SCHED_MAX_SPLITS);
                 split = &sched->splits[i_split];
                 split->backend_id = node_backend_id;
                 split->i_start = i;
@@ -1638,11 +1229,11 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
         sched->prev_leaf_backend_ids = tmp;
     }
 
-    int graph_size = graph->n_nodes + sched->n_splits*GGML_SCHED_MAX_SPLIT_INPUTS*2;
+    int graph_size = std::max(graph->n_nodes, graph->n_leafs) + sched->n_splits*GGML_SCHED_MAX_SPLIT_INPUTS*2*sched->n_copies;
     if (sched->graph.size < graph_size) {
         sched->graph.size = graph_size;
-        sched->graph.nodes = realloc(sched->graph.nodes, graph_size * sizeof(struct ggml_tensor *));
-        sched->graph.leafs = realloc(sched->graph.leafs, graph_size * sizeof(struct ggml_tensor *));
+        sched->graph.nodes = (ggml_tensor **) realloc(sched->graph.nodes, graph_size * sizeof(struct ggml_tensor *));
+        sched->graph.leafs = (ggml_tensor **) realloc(sched->graph.leafs, graph_size * sizeof(struct ggml_tensor *));
         GGML_ASSERT(sched->graph.nodes != NULL);
         GGML_ASSERT(sched->graph.leafs != NULL);
     }
@@ -1690,6 +1281,7 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
             for (int c = 0; c < sched->n_copies; c++) {
                 struct ggml_tensor * input_cpy = tensor_id_copy(id, backend_id, c);
                 sched->leaf_backend_ids[graph_copy->n_leafs] = backend_id;
+                assert(graph_copy->size > graph_copy->n_leafs);
                 graph_copy->leafs[graph_copy->n_leafs++] = input_cpy;
             }
         }
@@ -1703,6 +1295,7 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
                 for (int c = 0; c < sched->n_copies; c++) {
                     struct ggml_tensor * input_cpy = tensor_id_copy(id, backend_id, c);
                     sched->leaf_backend_ids[graph_copy->n_leafs] = backend_id;
+                    assert(graph_copy->size > graph_copy->n_leafs);
                     graph_copy->leafs[graph_copy->n_leafs++] = input_cpy;
                 }
             }
@@ -1713,6 +1306,7 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
     for (int i = 0; i < graph->n_leafs; i++) {
         struct ggml_tensor * leaf = graph->leafs[i];
         sched->leaf_backend_ids[graph_copy->n_leafs] = tensor_backend_id(leaf);
+        assert(graph_copy->size > graph_copy->n_leafs);
         graph_copy->leafs[graph_copy->n_leafs++] = leaf;
     }
 }
@@ -1741,11 +1335,11 @@ static bool ggml_backend_sched_alloc_splits(ggml_backend_sched_t sched) {
         // the re-allocation may cause the split inputs to be moved to a different address
         ggml_backend_sched_synchronize(sched);
 #ifndef NDEBUG
-        fprintf(stderr, "%s: failed to allocate graph, reserving (backend_ids_changed = %d)\n", __func__, backend_ids_changed);
+        GGML_LOG_DEBUG("%s: failed to allocate graph, reserving (backend_ids_changed = %d)\n", __func__, backend_ids_changed);
 #endif
         ggml_gallocr_reserve_n(sched->galloc, &sched->graph, sched->node_backend_ids, sched->leaf_backend_ids);
         if (!ggml_gallocr_alloc_graph(sched->galloc, &sched->graph)) {
-            fprintf(stderr, "%s: failed to allocate graph\n", __func__);
+            GGML_LOG_ERROR("%s: failed to allocate graph\n", __func__);
             return false;
         }
     }
@@ -1782,7 +1376,17 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
                 } else {
                     ggml_backend_synchronize(split_backend);
                 }
-                ggml_backend_tensor_copy_async(input_backend, split_backend, input, input_cpy);
+                // try async copy, but if not possible, we can still use a sync copy without synchronizing the dst backend, since we handle the synchronization here with multiple copies and events
+                // TODO: add public function to facilitate this, since applications do not have direct access to the backend interface
+                if (!split_backend->iface.cpy_tensor_async || !split_backend->iface.cpy_tensor_async(input_backend, split_backend, input, input_cpy)) {
+                    ggml_backend_synchronize(input_backend);
+                    if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
+                        ggml_backend_event_synchronize(sched->events[split_backend_id][sched->cur_copy]);
+                    } else {
+                        ggml_backend_synchronize(split_backend);
+                    }
+                    ggml_backend_tensor_copy(input, input_cpy);
+                }
             }
         }
 
@@ -1828,7 +1432,7 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
         // record the event of this copy
         if (split->n_inputs > 0) {
             if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
-                ggml_backend_event_record(sched->events[split_backend_id][sched->cur_copy]);
+                ggml_backend_event_record(sched->events[split_backend_id][sched->cur_copy], split_backend);
             }
         }
     }
@@ -1846,40 +1450,43 @@ 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 = calloc(1, sizeof(struct ggml_backend_sched));
+    struct ggml_backend_sched * sched = (ggml_backend_sched *) calloc(1, sizeof(struct ggml_backend_sched));
 
-    sched->debug = getenv("GGML_SCHED_DEBUG") != NULL;
+    const char * GGML_SCHED_DEBUG = getenv("GGML_SCHED_DEBUG");
+    sched->debug = GGML_SCHED_DEBUG ? atoi(GGML_SCHED_DEBUG) : 0;
     sched->n_backends = n_backends;
     sched->n_copies = parallel ? GGML_SCHED_MAX_COPIES : 1;
 
     // initialize hash table
     // FIXME: needs to be size*2 to account for leafs (do it in graph_split instead)
     sched->hash_set    = ggml_hash_set_new(graph_size);
-    sched->hv_tensor_backend_ids = malloc(sched->hash_set.size * sizeof(sched->hv_tensor_backend_ids[0]));
-    sched->hv_tensor_copies      = malloc(sched->hash_set.size * sched->n_backends * sched->n_copies * sizeof(struct ggml_tensor *));
+    sched->hv_tensor_backend_ids = (int *) malloc(sched->hash_set.size * sizeof(sched->hv_tensor_backend_ids[0]));
+    sched->hv_tensor_copies      = (ggml_tensor **) malloc(sched->hash_set.size * sched->n_backends * sched->n_copies * sizeof(struct ggml_tensor *));
 
-    const size_t nodes_size = graph_size + GGML_SCHED_MAX_SPLITS*GGML_SCHED_MAX_SPLIT_INPUTS*2;
-    sched->node_backend_ids = calloc(nodes_size, sizeof(sched->node_backend_ids[0]));
-    sched->leaf_backend_ids = calloc(nodes_size, sizeof(sched->leaf_backend_ids[0]));
-    sched->prev_node_backend_ids = calloc(nodes_size, sizeof(sched->prev_node_backend_ids[0]));
-    sched->prev_leaf_backend_ids = calloc(nodes_size, sizeof(sched->prev_leaf_backend_ids[0]));
+    const size_t ggml_sched_max_splits = graph_size; // at most there is one split for each node in the graph
+    const size_t nodes_size = graph_size + ggml_sched_max_splits*GGML_SCHED_MAX_SPLIT_INPUTS*2;
+    sched->node_backend_ids = (int *) calloc(nodes_size, sizeof(sched->node_backend_ids[0]));
+    sched->leaf_backend_ids = (int *) calloc(nodes_size, sizeof(sched->leaf_backend_ids[0]));
+    sched->prev_node_backend_ids = (int *) calloc(nodes_size, sizeof(sched->prev_node_backend_ids[0]));
+    sched->prev_leaf_backend_ids = (int *) calloc(nodes_size, sizeof(sched->prev_leaf_backend_ids[0]));
 
-    sched->context_buffer_size = GGML_SCHED_MAX_SPLITS*GGML_SCHED_MAX_SPLIT_INPUTS*2*sizeof(struct ggml_tensor) + ggml_graph_overhead_custom(graph_size, false);
-    sched->context_buffer = malloc(sched->context_buffer_size);
+    sched->context_buffer_size = ggml_sched_max_splits*GGML_SCHED_MAX_SPLIT_INPUTS*2*sizeof(struct ggml_tensor) + ggml_graph_overhead_custom(graph_size, false);
+    sched->context_buffer = (char *) malloc(sched->context_buffer_size);
 
     const int initial_splits_capacity = 16;
-    sched->splits = calloc(initial_splits_capacity, sizeof(sched->splits[0]));
+    sched->splits = (ggml_backend_sched_split *) calloc(initial_splits_capacity, sizeof(sched->splits[0]));
     sched->splits_capacity = initial_splits_capacity;
 
     for (int b = 0; b < n_backends; b++) {
         sched->backends[b] = backends[b];
         sched->bufts[b] = bufts ? bufts[b] : ggml_backend_get_default_buffer_type(backends[b]);
         GGML_ASSERT(ggml_backend_supports_buft(backends[b], sched->bufts[b]));
+
         if (sched->n_copies > 1) {
             for (int c = 0; c < sched->n_copies; c++) {
-                sched->events[b][c] = ggml_backend_event_new(backends[b]);
+                sched->events[b][c] = ggml_backend_event_new(backends[b]->device);
             }
         }
     }
@@ -1932,12 +1539,13 @@ bool ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph *
 
     ggml_backend_sched_split_graph(sched, measure_graph);
 
+    ggml_backend_sched_synchronize(sched);
+
     if (!ggml_gallocr_reserve_n(sched->galloc, &sched->graph, sched->node_backend_ids, sched->leaf_backend_ids)) {
         return false;
     }
 
     ggml_backend_sched_reset(sched);
-    ggml_backend_sched_synchronize(sched);
 
     return true;
 }
@@ -2115,8 +1723,8 @@ static void graph_copy_init_tensor(struct ggml_hash_set * hash_set, struct ggml_
 
 struct ggml_backend_graph_copy ggml_backend_graph_copy(ggml_backend_t backend, struct ggml_cgraph * graph) {
     struct ggml_hash_set hash_set = ggml_hash_set_new(graph->visited_hash_set.size);
-    struct ggml_tensor ** node_copies = calloc(hash_set.size, sizeof(node_copies[0])); // NOLINT
-    bool * node_init = calloc(hash_set.size, sizeof(node_init[0]));
+    struct ggml_tensor ** node_copies = (ggml_tensor **) calloc(hash_set.size, sizeof(node_copies[0])); // NOLINT
+    bool * node_init = (bool *) calloc(hash_set.size, sizeof(node_init[0]));
 
     struct ggml_init_params params = {
         /* .mem_size   = */ ggml_tensor_overhead()*hash_set.size + ggml_graph_overhead_custom(graph->size, false),
@@ -2128,13 +1736,13 @@ struct ggml_backend_graph_copy ggml_backend_graph_copy(ggml_backend_t backend, s
     struct ggml_context * ctx_unallocated = ggml_init(params);
 
     if (ctx_allocated == NULL || ctx_unallocated == NULL) {
-        fprintf(stderr, "failed to allocate context for graph copy\n");
+        GGML_LOG_ERROR("%s: failed to allocate context for graph copy\n", __func__);
         ggml_hash_set_free(&hash_set);
         free(node_copies);
         free(node_init);
         ggml_free(ctx_allocated);
         ggml_free(ctx_unallocated);
-        return (struct ggml_backend_graph_copy) {
+        return {
             /* .buffer           = */ NULL,
             /* .ctx_allocated    = */ NULL,
             /* .ctx_unallocated  = */ NULL,
@@ -2151,13 +1759,13 @@ struct ggml_backend_graph_copy ggml_backend_graph_copy(ggml_backend_t backend, s
     // allocate nodes
     ggml_backend_buffer_t buffer = ggml_backend_alloc_ctx_tensors(ctx_allocated, backend);
     if (buffer == NULL) {
-        fprintf(stderr, "failed to allocate buffer for graph copy\n");
+        GGML_LOG_ERROR("%s: failed to allocate buffer for graph copy\n", __func__);
         ggml_hash_set_free(&hash_set);
         free(node_copies);
         free(node_init);
         ggml_free(ctx_allocated);
         ggml_free(ctx_unallocated);
-        return (struct ggml_backend_graph_copy) {
+        return {
             /* .buffer           = */ NULL,
             /* .ctx_allocated    = */ NULL,
             /* .ctx_unallocated  = */ NULL,
@@ -2186,7 +1794,7 @@ struct ggml_backend_graph_copy ggml_backend_graph_copy(ggml_backend_t backend, s
     free(node_copies);
     free(node_init);
 
-    return (struct ggml_backend_graph_copy) {
+    return {
         /* .buffer           = */ buffer,
         /* .ctx_allocated    = */ ctx_allocated,
         /* .ctx_unallocated  = */ ctx_unallocated,
@@ -2238,3 +1846,154 @@ bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t
 
     return true;
 }
+
+// CPU backend - buffer
+
+static void * ggml_backend_cpu_buffer_get_base(ggml_backend_buffer_t buffer) {
+    uintptr_t data = (uintptr_t)buffer->context;
+
+    // align the buffer
+    if (data % TENSOR_ALIGNMENT != 0) {
+        data = GGML_PAD(data, TENSOR_ALIGNMENT);
+    }
+
+    return (void *)data;
+}
+
+static void ggml_backend_cpu_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+    ggml_aligned_free(buffer->context, buffer->size);
+}
+
+static void ggml_backend_cpu_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_cpu_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+    memcpy((char *)tensor->data + offset, data, size);
+
+    GGML_UNUSED(buffer);
+}
+
+static void ggml_backend_cpu_buffer_get_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+    memcpy(data, (const char *)tensor->data + offset, size);
+
+    GGML_UNUSED(buffer);
+}
+
+static bool ggml_backend_cpu_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)) {
+        memcpy(dst->data, src->data, ggml_nbytes(src));
+        return true;
+    }
+    return false;
+
+    GGML_UNUSED(buffer);
+}
+
+static void ggml_backend_cpu_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
+    memset(buffer->context, value, buffer->size);
+}
+
+static const struct ggml_backend_buffer_i ggml_backend_cpu_buffer_i = {
+    /* .free_buffer     = */ ggml_backend_cpu_buffer_free_buffer,
+    /* .get_base        = */ ggml_backend_cpu_buffer_get_base,
+    /* .init_tensor     = */ NULL, // no initialization required
+    /* .memset_tensor   = */ ggml_backend_cpu_buffer_memset_tensor,
+    /* .set_tensor      = */ ggml_backend_cpu_buffer_set_tensor,
+    /* .get_tensor      = */ ggml_backend_cpu_buffer_get_tensor,
+    /* .cpy_tensor      = */ ggml_backend_cpu_buffer_cpy_tensor,
+    /* .clear           = */ ggml_backend_cpu_buffer_clear,
+    /* .reset           = */ NULL,
+};
+
+static const struct ggml_backend_buffer_i ggml_backend_cpu_buffer_from_ptr_i = {
+    /* .free_buffer     = */ NULL, // ptr is not owned by the buffer, so it does not need to be freed
+    /* .get_base        = */ ggml_backend_cpu_buffer_get_base,
+    /* .init_tensor     = */ NULL, // no initialization required
+    /* .memset_tensor   = */ ggml_backend_cpu_buffer_memset_tensor,
+    /* .set_tensor      = */ ggml_backend_cpu_buffer_set_tensor,
+    /* .get_tensor      = */ ggml_backend_cpu_buffer_get_tensor,
+    /* .cpy_tensor      = */ ggml_backend_cpu_buffer_cpy_tensor,
+    /* .clear           = */ ggml_backend_cpu_buffer_clear,
+    /* .reset           = */ NULL,
+};
+
+// 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";
+
+    GGML_UNUSED(buft);
+}
+
+static ggml_backend_buffer_t ggml_backend_cpu_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+    void * data = ggml_aligned_malloc(size);
+
+    if (data == NULL) {
+        GGML_LOG_ERROR("%s: failed to allocate buffer of size %zu\n", __func__, size);
+        return NULL;
+    }
+
+    return ggml_backend_buffer_init(buft, ggml_backend_cpu_buffer_i, data, size);
+}
+
+static size_t ggml_backend_cpu_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
+    return TENSOR_ALIGNMENT;
+
+    GGML_UNUSED(buft);
+}
+
+static bool ggml_backend_cpu_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
+    return true;
+
+    GGML_UNUSED(buft);
+}
+
+ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void) {
+    static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type = {
+        /* .iface   = */ {
+            /* .get_name         = */ ggml_backend_cpu_buffer_type_get_name,
+            /* .alloc_buffer     = */ ggml_backend_cpu_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,
+        },
+        /* .device  = */ NULL, // FIXME ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
+        /* .context = */ NULL,
+    };
+
+    return &ggml_backend_cpu_buffer_type;
+}
+
+static const char * ggml_backend_cpu_buffer_from_ptr_type_get_name(ggml_backend_buffer_type_t buft) {
+    return "CPU_Mapped";
+
+    GGML_UNUSED(buft);
+}
+
+static ggml_backend_buffer_type_t ggml_backend_cpu_buffer_from_ptr_type(void) {
+    static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type = {
+        /* .iface   = */ {
+            /* .get_name         = */ ggml_backend_cpu_buffer_from_ptr_type_get_name,
+            /* .alloc_buffer     = */ ggml_backend_cpu_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,
+        },
+        /* .device  = */ NULL, // FIXME ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
+        /* .context = */ NULL,
+    };
+
+    return &ggml_backend_cpu_buffer_type;
+}
+
+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);
+}