llama : initial ggml-backend integration (#4520)

* llama : initial ggml-backend integration

* add ggml-metal

* cuda backend can be used though ggml-backend with LLAMA_GGML_BACKEND_CUDA_TEST
access all tensor data with ggml_backend_tensor_get/set

* add ggml_backend_buffer_clear
zero-init KV cache buffer

* add ggml_backend_buffer_is_hos, used to avoid copies if possible when accesing tensor data

* disable gpu backends with ngl 0

* more accurate mlock

* unmap offloaded part of the model

* use posix_fadvise64(.., POSIX_FADV_SEQUENTIAL) to improve performance with mmap

* update quantize and lora

* update session copy/set to use ggml-backend

ggml-ci

* use posix_fadvise instead of posix_fadvise64

* ggml_backend_alloc_ctx_tensors_from_buft : remove old print

* llama_mmap::align_offset : use pointers instead of references for out parameters

* restore progress_callback behavior

* move final progress_callback call to load_all_data

* cuda : fix fprintf format string (minor)

* do not offload scales

* llama_mmap : avoid unmapping the same fragments again in the destructor

* remove unnecessary unmap

* metal : add default log function that prints to stderr, cleanup code

ggml-ci

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

Makefile

@@ -65,7 +65,7 @@ test: $(TEST_TARGETS)
 			./$$test_target; \
 		fi; \
 		if [ $$? -ne 0 ]; then \
-			printf 'Test $$test_target FAILED!\n\n' $$test_target; \
+			printf 'Test %s FAILED!\n\n' $$test_target; \
 			failures=$$(( failures + 1 )); \
 		else \
 			printf 'Test %s passed.\n\n' $$test_target; \

ggml-alloc.c

@@ -449,11 +449,10 @@ static void init_view(ggml_gallocr_t galloc, struct ggml_tensor * view, bool upd
     if (update_backend) {
         view->backend = view->view_src->backend;
     }
-    view->buffer  = view->view_src->buffer;
+    // views are initialized in the alloc buffer rather than the view_src buffer
+    view->buffer  = alloc->buffer;
     view->data    = (char *)view->view_src->data + view->view_offs;
 
-    // FIXME: the view should be initialized by the owning buffer, but currently this breaks the CUDA backend
-    // due to the ggml_tensor_extra_gpu ring buffer overwriting the KV cache extras
     assert(ggml_tallocr_is_measure(alloc) || !view->buffer || view->buffer->buft == alloc->buffer->buft);
 
     if (!alloc->measure) {
@@ -736,6 +735,10 @@ void ggml_allocr_set_parse_seq(ggml_allocr_t alloc, const int * list, int n) {
 }
 
 void ggml_allocr_free(ggml_allocr_t alloc) {
+    if (alloc == NULL) {
+        return;
+    }
+
     ggml_gallocr_free(alloc->galloc);
     ggml_tallocr_free(alloc->talloc);
     free(alloc);
@@ -775,7 +778,7 @@ ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_conte
     }
 
     if (nbytes == 0) {
-        fprintf(stderr, "%s: no tensors to allocate\n", __func__);
+        // all the tensors in the context are already allocated
         return NULL;
     }
 
@@ -789,6 +792,11 @@ ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_conte
             } else {
                 ggml_backend_view_init(buffer, t);
             }
+        } else {
+            if (t->view_src != NULL) {
+                // view of a pre-allocated tensor
+                ggml_backend_view_init(buffer, t);
+            }
         }
     }
 

ggml-backend-impl.h

@@ -20,6 +20,9 @@ extern "C" {
         size_t                (*get_alignment)   (ggml_backend_buffer_type_t buft); // tensor alignment
         size_t                (*get_alloc_size)  (ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor); // data size needed to allocate the tensor, including padding
         bool                  (*supports_backend)(ggml_backend_buffer_type_t buft, ggml_backend_t backend); // check if the buffer type is usable by the backend
+        // check if tensor data is in host memory
+        // should be equivalent to supports_backend(buft, ggml_backend_cpu_init())
+        bool                  (*is_host)         (ggml_backend_buffer_type_t buft);
     };
 
     struct ggml_backend_buffer_type {
@@ -40,6 +43,7 @@ extern "C" {
         // (optional) copy tensor between different buffer-type, allow for single-copy tranfers
         void   (*cpy_tensor_from)(ggml_backend_buffer_t buffer, struct ggml_tensor * src, struct ggml_tensor * dst);
         void   (*cpy_tensor_to)  (ggml_backend_buffer_t buffer, struct ggml_tensor * src, struct ggml_tensor * dst);
+        void   (*clear)          (ggml_backend_buffer_t buffer, uint8_t value);
     };
 
     struct ggml_backend_buffer {

ggml-backend.c

@@ -35,6 +35,13 @@ bool ggml_backend_buft_supports_backend(ggml_backend_buffer_type_t buft, ggml_ba
     return buft->iface.supports_backend(buft, backend);
 }
 
+bool ggml_backend_buft_is_host(ggml_backend_buffer_type_t buft) {
+    if (buft->iface.is_host) {
+        return buft->iface.is_host(buft);
+    }
+    return false;
+}
+
 // backend buffer
 
 ggml_backend_buffer_t ggml_backend_buffer_init(
@@ -94,6 +101,14 @@ 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_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_type(buffer));
+}
+
 ggml_backend_buffer_type_t ggml_backend_buffer_type(ggml_backend_buffer_t buffer) {
     return buffer->buft;
 }
@@ -378,7 +393,6 @@ static void * ggml_backend_cpu_buffer_get_base(ggml_backend_buffer_t buffer) {
 
 static void ggml_backend_cpu_buffer_free_buffer(ggml_backend_buffer_t buffer) {
     free(buffer->context);
-    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) {
@@ -411,6 +425,10 @@ static void ggml_backend_cpu_buffer_cpy_tensor_to(ggml_backend_buffer_t buffer,
     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 struct ggml_backend_buffer_i cpu_backend_buffer_i = {
     /* .free_buffer     = */ ggml_backend_cpu_buffer_free_buffer,
     /* .get_base        = */ ggml_backend_cpu_buffer_get_base,
@@ -419,6 +437,7 @@ static struct ggml_backend_buffer_i cpu_backend_buffer_i = {
     /* .get_tensor      = */ ggml_backend_cpu_buffer_get_tensor,
     /* .cpy_tensor_from = */ ggml_backend_cpu_buffer_cpy_tensor_from,
     /* .cpy_tensor_to   = */ ggml_backend_cpu_buffer_cpy_tensor_to,
+    /* .clear           = */ ggml_backend_cpu_buffer_clear,
 };
 
 // for buffers from ptr, free is not called
@@ -430,6 +449,7 @@ static struct ggml_backend_buffer_i cpu_backend_buffer_i_from_ptr = {
     /* .get_tensor      = */ ggml_backend_cpu_buffer_get_tensor,
     /* .cpy_tensor_from = */ ggml_backend_cpu_buffer_cpy_tensor_from,
     /* .cpy_tensor_to   = */ ggml_backend_cpu_buffer_cpy_tensor_to,
+    /* .clear           = */ ggml_backend_cpu_buffer_clear,
 };
 
 static const size_t TENSOR_ALIGNMENT = 64; // should be enough for AVX 512
@@ -455,20 +475,70 @@ static bool ggml_backend_cpu_buffer_type_supports_backend(ggml_backend_buffer_ty
     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_buffer_type_cpu = {
+    static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type = {
         /* .iface = */ {
             /* .alloc_buffer     = */ ggml_backend_cpu_buffer_type_alloc_buffer,
             /* .get_alignment    = */ ggml_backend_cpu_buffer_type_get_alignment,
             /* .get_alloc_size   = */ NULL, // defaults to ggml_nbytes
             /* .supports_backend = */ ggml_backend_cpu_buffer_type_supports_backend,
+            /* .is_host          = */ ggml_backend_cpu_buffer_type_is_host,
         },
         /* .context = */ NULL,
     };
 
-    return &ggml_backend_buffer_type_cpu;
+    return &ggml_backend_cpu_buffer_type;
 }
 
+#ifdef GGML_USE_CPU_HBM
+
+// buffer type HBM
+
+#include <hbwmalloc.h>
+
+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 = 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);
+        return NULL;
+    }
+
+    // FIXME: this is a hack to avoid having to implement a new buffer type
+    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() {
+    static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type_hbm = {
+        /* .iface    = */ {
+            /* .alloc_buffer     = */ ggml_backend_cpu_hbm_buffer_type_alloc_buffer,
+            /* .get_alignment    = */ ggml_backend_cpu_buffer_type_get_alignment,
+            /* .get_alloc_size   = */ NULL, // defaults to ggml_nbytes
+            /* .supports_backend = */ ggml_backend_cpu_buffer_type_supports_backend,
+            /* .is_host          = */ ggml_backend_cpu_buffer_type_is_host,
+        },
+        /* .context  = */ NULL,
+    };
+
+    return &ggml_backend_cpu_buffer_type_hbm;
+}
+#endif
+
 struct ggml_backend_cpu_context {
     int n_threads;
     void * work_data;
@@ -505,7 +575,7 @@ static ggml_backend_graph_plan_t ggml_backend_cpu_graph_plan_create(ggml_backend
     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;
+    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);
@@ -1180,7 +1250,7 @@ void ggml_backend_sched_set_node_backend(ggml_backend_sched_t sched, struct ggml
 // utils
 void ggml_backend_view_init(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
     GGML_ASSERT(tensor->buffer == NULL);
-    GGML_ASSERT(tensor->data == NULL);
+    //GGML_ASSERT(tensor->data == NULL); // views of pre-allocted tensors may have the data set, but still need to be initialized
     GGML_ASSERT(tensor->view_src != NULL);
     GGML_ASSERT(tensor->view_src->buffer != NULL);
     GGML_ASSERT(tensor->view_src->data != NULL);

ggml-backend.h

@@ -21,6 +21,7 @@ extern "C" {
     GGML_API size_t ggml_backend_buft_get_alignment (ggml_backend_buffer_type_t buft);
     GGML_API size_t ggml_backend_buft_get_alloc_size(ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor);
     GGML_API bool ggml_backend_buft_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend);
+    GGML_API bool ggml_backend_buft_is_host         (ggml_backend_buffer_type_t buft);
 
     // buffer
     GGML_API void   ggml_backend_buffer_free          (ggml_backend_buffer_t buffer);
@@ -29,6 +30,8 @@ extern "C" {
     GGML_API void   ggml_backend_buffer_init_tensor   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
     GGML_API size_t ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer);
     GGML_API size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+    GGML_API void   ggml_backend_buffer_clear         (ggml_backend_buffer_t buffer, uint8_t value);
+    GGML_API bool   ggml_backend_buffer_is_host       (ggml_backend_buffer_t buffer);
     GGML_API ggml_backend_buffer_type_t ggml_backend_buffer_type(ggml_backend_buffer_t buffer);
 
     //
@@ -76,6 +79,10 @@ extern "C" {
 
     GGML_API ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void);
 
+#ifdef GGML_USE_CPU_HBM
+    GGML_API ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void);
+#endif
+
     //
     // Backend registry
     //

ggml-cuda.cu

@@ -9226,11 +9226,10 @@ void ggml_cuda_assign_scratch_offset(struct ggml_tensor * tensor, size_t offset)
 
     ggml_tensor_extra_gpu * extra = ggml_cuda_alloc_temp_tensor_extra();
 
-    const bool inplace = (tensor->src[0] != nullptr && tensor->src[0]->data == tensor->data) ||
+    const bool inplace = tensor->view_src != nullptr;
-        tensor->op == GGML_OP_VIEW;
 
-    if (inplace && (tensor->src[0]->backend == GGML_BACKEND_GPU || tensor->src[0]->backend == GGML_BACKEND_GPU_SPLIT)) {
+    if (inplace && (tensor->view_src->backend == GGML_BACKEND_GPU || tensor->view_src->backend == GGML_BACKEND_GPU_SPLIT)) {
-        ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu * ) tensor->src[0]->extra;
+        ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu * ) tensor->view_src->extra;
         char * src0_ddc = (char *) src0_extra->data_device[g_main_device];
         size_t view_offset = 0;
         if (tensor->op == GGML_OP_VIEW) {
@@ -9317,7 +9316,7 @@ bool ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_
     if (tensor->op == GGML_OP_MUL_MAT) {
         if (tensor->src[0]->ne[3] != tensor->src[1]->ne[3]) {
 #ifndef NDEBUG
-            fprintf(stderr, "%s: cannot compute %s: src0->ne[3] = " PRId64 ", src1->ne[3] = " PRId64 " - fallback to CPU\n", __func__, tensor->name, tensor->src[0]->ne[3], tensor->src[1]->ne[3]);
+            fprintf(stderr, "%s: cannot compute %s: src0->ne[3] = %" PRId64 ", src1->ne[3] = %" PRId64 " - fallback to CPU\n", __func__, tensor->name, tensor->src[0]->ne[3], tensor->src[1]->ne[3]);
 #endif
             return false;
         }
@@ -9523,7 +9522,7 @@ static void ggml_backend_cuda_buffer_init_tensor(ggml_backend_buffer_t buffer, g
     ggml_backend_buffer_context_cuda * ctx = (ggml_backend_buffer_context_cuda *)buffer->context;
 
     if (tensor->view_src != NULL && tensor->view_offs == 0) {
-        assert(tensor->view_src->buffer->buft == buffer->buft); // TODO
+        assert(tensor->view_src->buffer->buft == buffer->buft);
         tensor->backend = tensor->view_src->backend;
         tensor->extra = tensor->view_src->extra;
         return;
@@ -9554,23 +9553,34 @@ static void ggml_backend_cuda_buffer_init_tensor(ggml_backend_buffer_t buffer, g
 }
 
 static void ggml_backend_cuda_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
-    GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor write out of bounds");
-    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
     GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
 
-    CUDA_CHECK(cudaMemcpy((char *)tensor->data + offset, data, size, cudaMemcpyHostToDevice));
+    ggml_backend_buffer_context_cuda * ctx = (ggml_backend_buffer_context_cuda *)buffer->context;
 
-    UNUSED(buffer);
+    ggml_cuda_set_device(ctx->device);
+    CUDA_CHECK(cudaDeviceSynchronize());
+
+    CUDA_CHECK(cudaMemcpy((char *)tensor->data + offset, data, size, cudaMemcpyHostToDevice));
 }
 
 static void ggml_backend_cuda_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
-    GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor read out of bounds");
-    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
     GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
 
-    CUDA_CHECK(cudaMemcpy(data, (const char *)tensor->data + offset, size, cudaMemcpyDeviceToHost));
+    ggml_backend_buffer_context_cuda * ctx = (ggml_backend_buffer_context_cuda *)buffer->context;
 
-    UNUSED(buffer);
+    ggml_cuda_set_device(ctx->device);
+    CUDA_CHECK(cudaDeviceSynchronize());
+
+    CUDA_CHECK(cudaMemcpy(data, (const char *)tensor->data + offset, size, cudaMemcpyDeviceToHost));
+}
+
+static void ggml_backend_cuda_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
+    ggml_backend_buffer_context_cuda * ctx = (ggml_backend_buffer_context_cuda *)buffer->context;
+
+    ggml_cuda_set_device(ctx->device);
+    CUDA_CHECK(cudaDeviceSynchronize());
+
+    CUDA_CHECK(cudaMemset(ctx->dev_ptr, value, buffer->size));
 }
 
 static struct ggml_backend_buffer_i cuda_backend_buffer_interface = {
@@ -9581,6 +9591,7 @@ static struct ggml_backend_buffer_i cuda_backend_buffer_interface = {
     /* .get_tensor      = */ ggml_backend_cuda_buffer_get_tensor,
     /* .cpy_tensor_from = */ NULL,
     /* .cpy_tensor_to   = */ NULL,
+    /* .clear           = */ ggml_backend_cuda_buffer_clear,
 };
 
 // cuda buffer type
@@ -9632,35 +9643,36 @@ static bool ggml_backend_cuda_buffer_type_supports_backend(ggml_backend_buffer_t
     UNUSED(buft);
 }
 
-static ggml_backend_buffer_type_i cuda_backend_buffer_type_interface = {
+static ggml_backend_buffer_type_i ggml_backend_cuda_buffer_type_interface = {
     /* .alloc_buffer     = */ ggml_backend_cuda_buffer_type_alloc_buffer,
     /* .get_alignment    = */ ggml_backend_cuda_buffer_type_get_alignment,
     /* .get_alloc_size   = */ ggml_backend_cuda_buffer_type_get_alloc_size,
     /* .supports_backend = */ ggml_backend_cuda_buffer_type_supports_backend,
+    /* .is_host          = */ nullptr,
 };
 
 ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device) {
-    static struct ggml_backend_buffer_type ggml_backend_buffer_type_cuda[GGML_CUDA_MAX_DEVICES];
+    static struct ggml_backend_buffer_type ggml_backend_cuda_buffer_types[GGML_CUDA_MAX_DEVICES];
-    static bool ggml_backend_buffer_type_cuda_initialized = false;
+
-    if (!ggml_backend_buffer_type_cuda_initialized) {
+    static bool ggml_backend_cuda_buffer_type_initialized = false;
+
+    if (!ggml_backend_cuda_buffer_type_initialized) {
         for (int i = 0; i < GGML_CUDA_MAX_DEVICES; i++) {
-            ggml_backend_buffer_type_cuda[i] = {
+            ggml_backend_cuda_buffer_types[i] = {
-                /* .iface    = */ cuda_backend_buffer_type_interface,
+                /* .iface    = */ ggml_backend_cuda_buffer_type_interface,
                 /* .context  = */ (ggml_backend_buffer_type_context_t) (intptr_t) i,
             };
         }
-        ggml_backend_buffer_type_cuda_initialized = true;
+        ggml_backend_cuda_buffer_type_initialized = true;
     }
 
-    return &ggml_backend_buffer_type_cuda[device];
+    return &ggml_backend_cuda_buffer_types[device];
 }
 
 // host buffer type
 
 static void ggml_backend_cuda_host_buffer_free_buffer(ggml_backend_buffer_t buffer) {
-    ggml_backend_buffer_context_cuda * ctx = (ggml_backend_buffer_context_cuda *)buffer->context;
+    CUDA_CHECK(cudaFreeHost(buffer->context));
-    CUDA_CHECK(cudaFreeHost(ctx->dev_ptr));
-    delete ctx;
 }
 
 static ggml_backend_buffer_t ggml_backend_cuda_host_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
@@ -9673,24 +9685,21 @@ static ggml_backend_buffer_t ggml_backend_cuda_host_buffer_type_alloc_buffer(ggm
     buffer->iface.free_buffer = ggml_backend_cuda_host_buffer_free_buffer;
 
     return buffer;
-
-    UNUSED(buft);
 }
 
-struct ggml_backend_buffer_type_i cuda_backend_host_buffer_type_interface = {
+ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type() {
+    static struct ggml_backend_buffer_type ggml_backend_cuda_buffer_type_host = {
+        /* .iface    = */ {
             /* .alloc_buffer     = */ ggml_backend_cuda_host_buffer_type_alloc_buffer,
             /* .get_alignment    = */ ggml_backend_cpu_buffer_type()->iface.get_alignment,
             /* .get_alloc_size   = */ ggml_backend_cpu_buffer_type()->iface.get_alloc_size,
             /* .supports_backend = */ ggml_backend_cpu_buffer_type()->iface.supports_backend,
-};
+            /* .is_host          = */ ggml_backend_cpu_buffer_type()->iface.is_host,
-
+        },
-ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type() {
-    static struct ggml_backend_buffer_type ggml_backend_buffer_type_cuda_host = {
-        /* .iface    = */ cuda_backend_host_buffer_type_interface,
         /* .context  = */ nullptr,
     };
 
-    return &ggml_backend_buffer_type_cuda_host;
+    return &ggml_backend_cuda_buffer_type_host;
 }
 
 // backend
@@ -9722,8 +9731,6 @@ static void ggml_backend_cuda_set_tensor_async(ggml_backend_t backend, ggml_tens
     ggml_backend_context_cuda * cuda_ctx = (ggml_backend_context_cuda *)backend->context;
 
     GGML_ASSERT(tensor->buffer->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device) && "unsupported buffer type");
-    GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor write out of bounds");
-    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
     GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
 
     CUDA_CHECK(cudaMemcpyAsync((char *)tensor->data + offset, data, size, cudaMemcpyHostToDevice, g_cudaStreams[cuda_ctx->device][0]));
@@ -9733,8 +9740,6 @@ static void ggml_backend_cuda_get_tensor_async(ggml_backend_t backend, const ggm
     ggml_backend_context_cuda * cuda_ctx = (ggml_backend_context_cuda *)backend->context;
 
     GGML_ASSERT(tensor->buffer->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device) && "unsupported buffer type");
-    GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor read out of bounds");
-    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
     GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
 
     CUDA_CHECK(cudaMemcpyAsync(data, (const char *)tensor->data + offset, size, cudaMemcpyDeviceToHost, g_cudaStreams[cuda_ctx->device][0]));

ggml-metal.h

@@ -98,7 +98,10 @@ GGML_API ggml_backend_t ggml_backend_metal_init(void);
 
 GGML_API bool ggml_backend_is_metal(ggml_backend_t backend);
 
+GGML_API ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size);
+
 GGML_API void ggml_backend_metal_set_n_cb(ggml_backend_t backend, int n_cb);
+
 GGML_API ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void);
 
 // helper to check if the device supports a specific family

ggml-metal.m

@@ -180,7 +180,15 @@ struct ggml_metal_context {
 @implementation GGMLMetalClass
 @end
 
-ggml_log_callback ggml_metal_log_callback = NULL;
+
+static void ggml_metal_default_log_callback(enum ggml_log_level level, const char * msg, void * user_data) {
+    fprintf(stderr, "%s", msg);
+
+    UNUSED(level);
+    UNUSED(user_data);
+}
+
+ggml_log_callback ggml_metal_log_callback = ggml_metal_default_log_callback;
 void * ggml_metal_log_user_data = NULL;
 
 void ggml_metal_log_set_callback(ggml_log_callback log_callback, void * user_data) {
@@ -607,12 +615,24 @@ int * ggml_metal_get_concur_list(struct ggml_metal_context * ctx) {
 }
 
 // temporarily defined here for compatibility between ggml-backend and the old API
-struct ggml_backend_metal_buffer_context {
+
+struct ggml_backend_metal_buffer {
     void   * data;
+    size_t   size;
 
     id<MTLBuffer> metal;
 };
 
+struct ggml_backend_metal_buffer_context {
+    void * all_data;
+    size_t all_size;
+    bool owned;
+
+    // multiple buffers are used only to avoid the maximum buffer size limitation when using mmap
+    int n_buffers;
+    struct ggml_backend_metal_buffer buffers[GGML_METAL_MAX_BUFFERS];
+};
+
 // finds the Metal buffer that contains the tensor data on the GPU device
 // the assumption is that there is 1-to-1 mapping between the host and device memory buffers, so we can find the
 // Metal buffer based on the host memory pointer
@@ -622,17 +642,29 @@ static id<MTLBuffer> ggml_metal_get_buffer(struct ggml_metal_context * ctx, stru
 
     const int64_t tsize = ggml_nbytes(t);
 
+    ggml_backend_buffer_t buffer = t->view_src ? t->view_src->buffer : t->buffer;
+
     // compatibility with ggml-backend
-    if (t->buffer && t->buffer->buft == ggml_backend_metal_buffer_type()) {
+    if (buffer && buffer->buft == ggml_backend_metal_buffer_type()) {
-        struct ggml_backend_metal_buffer_context * buf_ctx = (struct ggml_backend_metal_buffer_context *) t->buffer->context;
+        struct ggml_backend_metal_buffer_context * buf_ctx = (struct ggml_backend_metal_buffer_context *) buffer->context;
 
-        const int64_t ioffs = (int64_t) t->data - (int64_t) buf_ctx->data;
+        // find the view that contains the tensor fully
-
+        for (int i = 0; i < buf_ctx->n_buffers; ++i) {
-        GGML_ASSERT(ioffs >= 0 && ioffs + tsize <= (int64_t) t->buffer->size);
+            const int64_t ioffs = (int64_t) t->data - (int64_t) buf_ctx->buffers[i].data;
 
+            //GGML_METAL_LOG_INFO("ioffs = %10ld, tsize = %10ld, sum = %10ld, buf_ctx->buffers[%d].size = %10ld\n", ioffs, tsize, ioffs + tsize, i, buf_ctx->buffers[i].size);
+            if (ioffs >= 0 && ioffs + tsize <= (int64_t) buf_ctx->buffers[i].size) {
                 *offs = (size_t) ioffs;
 
-        return buf_ctx->metal;
+                //GGML_METAL_LOG_INFO("%s: tensor '%16s', offs = %8ld\n", __func__, t->name, *offs);
+
+                return buf_ctx->buffers[i].metal;
+            }
+        }
+
+        GGML_METAL_LOG_ERROR("%s: error: tensor '%s' buffer is nil\n", __func__, t->name);
+
+        return nil;
     }
 
     // find the view that contains the tensor fully
@@ -2361,6 +2393,7 @@ void ggml_metal_graph_compute(
 
 // backend interface
 
+// default buffer
 static id<MTLDevice> g_backend_device = nil;
 static int g_backend_device_ref_count = 0;
 
@@ -2388,34 +2421,31 @@ static void ggml_backend_metal_free_device(void) {
 static void * ggml_backend_metal_buffer_get_base(ggml_backend_buffer_t buffer) {
     struct ggml_backend_metal_buffer_context * ctx = (struct ggml_backend_metal_buffer_context *)buffer->context;
 
-    return ctx->data;
+    return ctx->all_data;
 }
 
 static void ggml_backend_metal_buffer_free_buffer(ggml_backend_buffer_t buffer) {
     struct ggml_backend_metal_buffer_context * ctx = (struct ggml_backend_metal_buffer_context *)buffer->context;
 
-    [ctx->metal release];
+    for (int i = 0; i < ctx->n_buffers; i++) {
+        [ctx->buffers[i].metal release];
+    }
     ggml_backend_metal_free_device();
 
-    free(ctx->data);
+    if (ctx->owned) {
-    free(ctx);
+        free(ctx->all_data);
+    }
 
-    UNUSED(buffer);
+    free(ctx);
 }
 
 static void ggml_backend_metal_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
-    GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor write out of bounds");
-    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
-
     memcpy((char *)tensor->data + offset, data, size);
 
     UNUSED(buffer);
 }
 
 static void ggml_backend_metal_buffer_get_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
-    GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor read out of bounds");
-    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
-
     memcpy(data, (const char *)tensor->data + offset, size);
 
     UNUSED(buffer);
@@ -2433,7 +2463,13 @@ static void ggml_backend_metal_buffer_cpy_tensor_to(ggml_backend_buffer_t buffer
     UNUSED(buffer);
 }
 
-static struct ggml_backend_buffer_i metal_backend_buffer_i = {
+static void ggml_backend_metal_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
+    struct ggml_backend_metal_buffer_context * ctx = (struct ggml_backend_metal_buffer_context *)buffer->context;
+
+    memset(ctx->all_data, value, ctx->all_size);
+}
+
+static struct ggml_backend_buffer_i ggml_backend_metal_buffer_i = {
     /* .free_buffer     = */ ggml_backend_metal_buffer_free_buffer,
     /* .get_base        = */ ggml_backend_metal_buffer_get_base,
     /* .init_tensor     = */ NULL,
@@ -2441,8 +2477,11 @@ static struct ggml_backend_buffer_i metal_backend_buffer_i = {
     /* .get_tensor      = */ ggml_backend_metal_buffer_get_tensor,
     /* .cpy_tensor_from = */ ggml_backend_metal_buffer_cpy_tensor_from,
     /* .cpy_tensor_to   = */ ggml_backend_metal_buffer_cpy_tensor_to,
+    /* .clear           = */ ggml_backend_metal_buffer_clear,
 };
 
+// default buffer type
+
 static ggml_backend_buffer_t ggml_backend_metal_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
     struct ggml_backend_metal_buffer_context * ctx = malloc(sizeof(struct ggml_backend_metal_buffer_context));
 
@@ -2453,13 +2492,46 @@ static ggml_backend_buffer_t ggml_backend_metal_buffer_type_alloc_buffer(ggml_ba
         size_aligned += (size_page - (size_aligned % size_page));
     }
 
-    ctx->data  = ggml_metal_host_malloc(size);
+    id<MTLDevice> device = ggml_backend_metal_get_device();
-    ctx->metal = [ggml_backend_metal_get_device() newBufferWithBytesNoCopy:ctx->data
+
+    ctx->all_data = ggml_metal_host_malloc(size_aligned);
+    ctx->all_size = size_aligned;
+    ctx->owned = true;
+    ctx->n_buffers = 1;
+
+    ctx->buffers[0].data = ctx->all_data;
+    ctx->buffers[0].size = size;
+    ctx->buffers[0].metal = [device newBufferWithBytesNoCopy:ctx->all_data
                     length:size_aligned
                     options:MTLResourceStorageModeShared
                     deallocator:nil];
 
-    return ggml_backend_buffer_init(buft, metal_backend_buffer_i, ctx, size);
+    if (ctx->buffers[0].metal == nil) {
+        GGML_METAL_LOG_ERROR("%s: error: failed to allocate buffer, size = %8.2f MiB\n", __func__, size_aligned / 1024.0 / 1024.0);
+        free(ctx);
+        ggml_backend_metal_free_device();
+        return NULL;
+    }
+
+    GGML_METAL_LOG_INFO("%s: allocated buffer, size = %8.2f MiB", __func__, size_aligned / 1024.0 / 1024.0);
+
+
+#if TARGET_OS_OSX
+    GGML_METAL_LOG_INFO(", (%8.2f / %8.2f)",
+            device.currentAllocatedSize / 1024.0 / 1024.0,
+            device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);
+
+    if (device.currentAllocatedSize > device.recommendedMaxWorkingSetSize) {
+        GGML_METAL_LOG_WARN("%s: warning: current allocated size is greater than the recommended max working set size\n", __func__);
+    } else {
+        GGML_METAL_LOG_INFO("\n");
+    }
+#else
+    GGML_METAL_LOG_INFO(", (%8.2f)\n", device.currentAllocatedSize / 1024.0 / 1024.0);
+#endif
+
+
+    return ggml_backend_buffer_init(buft, ggml_backend_metal_buffer_i, ctx, size);
 }
 
 static size_t ggml_backend_metal_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
@@ -2470,7 +2542,13 @@ static size_t ggml_backend_metal_buffer_type_get_alignment(ggml_backend_buffer_t
 static bool ggml_backend_metal_buffer_type_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend) {
     return ggml_backend_is_metal(backend) || ggml_backend_is_cpu(backend);
 
-    GGML_UNUSED(buft);
+    UNUSED(buft);
+}
+
+static bool ggml_backend_metal_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
+    return true;
+
+    UNUSED(buft);
 }
 
 ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void) {
@@ -2480,6 +2558,7 @@ ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void) {
             /* .get_alignment    = */ ggml_backend_metal_buffer_type_get_alignment,
             /* .get_alloc_size   = */ NULL, // defaults to ggml_nbytes
             /* .supports_backend = */ ggml_backend_metal_buffer_type_supports_backend,
+            /* .is_host          = */ ggml_backend_metal_buffer_type_is_host,
         },
         /* .context = */ NULL,
     };
@@ -2487,6 +2566,87 @@ ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void) {
     return &ggml_backend_buffer_type_metal;
 }
 
+// buffer from ptr
+
+ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size) {
+    struct ggml_backend_metal_buffer_context * ctx = malloc(sizeof(struct ggml_backend_metal_buffer_context));
+
+    ctx->all_data = data;
+    ctx->all_size = size;
+    ctx->owned = false;
+    ctx->n_buffers = 0;
+
+    const size_t size_page = sysconf(_SC_PAGESIZE);
+    size_t size_aligned = size;
+    if ((size_aligned % size_page) != 0) {
+        size_aligned += (size_page - (size_aligned % size_page));
+    }
+
+    id<MTLDevice> device = ggml_backend_metal_get_device();
+
+    // the buffer fits into the max buffer size allowed by the device
+    if (size_aligned <= device.maxBufferLength) {
+        ctx->buffers[ctx->n_buffers].data = data;
+        ctx->buffers[ctx->n_buffers].size = size;
+
+        ctx->buffers[ctx->n_buffers].metal = [device newBufferWithBytesNoCopy:data length:size_aligned options:MTLResourceStorageModeShared deallocator:nil];
+
+        if (ctx->buffers[ctx->n_buffers].metal == nil) {
+            GGML_METAL_LOG_ERROR("%s: error: failed to allocate buffer, size = %8.2f MiB\n", __func__, size_aligned / 1024.0 / 1024.0);
+            return false;
+        }
+
+        GGML_METAL_LOG_INFO("%s: allocated buffer, size = %8.2f MiB", __func__, size_aligned / 1024.0 / 1024.0);
+
+        ++ctx->n_buffers;
+    } else {
+        // this overlap between the views will guarantee that the tensor with the maximum size will fully fit into
+        // one of the views
+        const size_t size_ovlp = ((max_size + size_page - 1) / size_page + 1) * size_page; // round-up 2 pages just in case
+        const size_t size_step = device.maxBufferLength - size_ovlp;
+        const size_t size_view = device.maxBufferLength;
+
+        for (size_t i = 0; i < size; i += size_step) {
+            const size_t size_step_aligned = (i + size_view <= size) ? size_view : (size_aligned - i);
+
+            ctx->buffers[ctx->n_buffers].data = (void *) ((uint8_t *) data + i);
+            ctx->buffers[ctx->n_buffers].size = size_step_aligned;
+
+            ctx->buffers[ctx->n_buffers].metal = [device newBufferWithBytesNoCopy:(void *) ((uint8_t *) data + i) length:size_step_aligned options:MTLResourceStorageModeShared deallocator:nil];
+
+            if (ctx->buffers[ctx->n_buffers].metal == nil) {
+                GGML_METAL_LOG_ERROR("%s: error: failed to allocate buffer, size = %8.2f MiB\n", __func__, size_step_aligned / 1024.0 / 1024.0);
+                return false;
+            }
+
+            GGML_METAL_LOG_INFO("%s: allocated buffer, size = %8.2f MiB, offs = %12ld", __func__, size_step_aligned / 1024.0 / 1024.0, i);
+            if (i + size_step < size) {
+                GGML_METAL_LOG_INFO("\n");
+            }
+
+            ++ctx->n_buffers;
+        }
+    }
+
+#if TARGET_OS_OSX
+    GGML_METAL_LOG_INFO(", (%8.2f / %8.2f)",
+            device.currentAllocatedSize / 1024.0 / 1024.0,
+            device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);
+
+    if (device.currentAllocatedSize > device.recommendedMaxWorkingSetSize) {
+        GGML_METAL_LOG_WARN("%s: warning: current allocated size is greater than the recommended max working set size\n", __func__);
+    } else {
+        GGML_METAL_LOG_INFO("\n");
+    }
+#else
+    GGML_METAL_LOG_INFO(", (%8.2f)\n", device.currentAllocatedSize / 1024.0 / 1024.0);
+#endif
+
+    return ggml_backend_buffer_init(ggml_backend_metal_buffer_type(), ggml_backend_metal_buffer_i, ctx, size);
+}
+
+// backend
+
 static const char * ggml_backend_metal_name(ggml_backend_t backend) {
     return "Metal";
 
@@ -2499,10 +2659,6 @@ static void ggml_backend_metal_free(ggml_backend_t backend) {
     free(backend);
 }
 
-static void ggml_backend_metal_synchronize(ggml_backend_t backend) {
-    UNUSED(backend);
-}
-
 static ggml_backend_buffer_type_t ggml_backend_metal_get_default_buffer_type(ggml_backend_t backend) {
     return ggml_backend_metal_buffer_type();
 
@@ -2529,25 +2685,15 @@ static struct ggml_backend_i metal_backend_i = {
     /* .get_tensor_async        = */ NULL,
     /* .cpy_tensor_from_async   = */ NULL,
     /* .cpy_tensor_to_async     = */ NULL,
-    /* .synchronize             = */ ggml_backend_metal_synchronize,
+    /* .synchronize             = */ NULL,
-    /* .graph_plan_create       = */ NULL, // the metal implementation does not require creating graph plans atm
+    /* .graph_plan_create       = */ NULL,
     /* .graph_plan_free         = */ NULL,
     /* .graph_plan_compute      = */ NULL,
     /* .graph_compute           = */ ggml_backend_metal_graph_compute,
     /* .supports_op             = */ ggml_backend_metal_supports_op,
 };
 
-// TODO: make a common log callback for all backends in ggml-backend
-static void ggml_backend_log_callback(enum ggml_log_level level, const char * msg, void * user_data) {
-    fprintf(stderr, "%s", msg);
-
-    UNUSED(level);
-    UNUSED(user_data);
-}
-
 ggml_backend_t ggml_backend_metal_init(void) {
-    ggml_metal_log_set_callback(ggml_backend_log_callback, NULL);
-
     struct ggml_metal_context * ctx = ggml_metal_init(GGML_DEFAULT_N_THREADS);
 
     if (ctx == NULL) {

ggml.c

@@ -2383,20 +2383,8 @@ size_t ggml_get_mem_size(const struct ggml_context * ctx) {
 size_t ggml_get_max_tensor_size(const struct ggml_context * ctx) {
     size_t max_size = 0;
 
-    struct ggml_object * obj = ctx->objects_begin;
+    for (struct ggml_tensor * tensor = ggml_get_first_tensor(ctx); tensor != NULL; tensor = ggml_get_next_tensor(ctx, tensor)) {
-
+        max_size = MAX(max_size, ggml_nbytes(tensor));
-    while (obj != NULL) {
-        if (obj->type == GGML_OBJECT_TENSOR) {
-            struct ggml_tensor * tensor = (struct ggml_tensor *) ((char *) ctx->mem_buffer + obj->offs);
-
-            const size_t size = ggml_nbytes(tensor);
-
-            if (max_size < size) {
-                max_size = size;
-            }
-        }
-
-        obj = obj->next;
     }
 
     return max_size;
@@ -3093,7 +3081,7 @@ struct ggml_tensor * ggml_view_tensor(
     return result;
 }
 
-struct ggml_tensor * ggml_get_first_tensor(struct ggml_context * ctx) {
+struct ggml_tensor * ggml_get_first_tensor(const struct ggml_context * ctx) {
     struct ggml_object * obj = ctx->objects_begin;
 
     char * const mem_buffer = ctx->mem_buffer;
@@ -3109,7 +3097,7 @@ struct ggml_tensor * ggml_get_first_tensor(struct ggml_context * ctx) {
     return NULL;
 }
 
-struct ggml_tensor * ggml_get_next_tensor(struct ggml_context * ctx, struct ggml_tensor * tensor) {
+struct ggml_tensor * ggml_get_next_tensor(const struct ggml_context * ctx, struct ggml_tensor * tensor) {
     struct ggml_object * obj = (struct ggml_object *) ((char *)tensor - GGML_OBJECT_SIZE);
     obj = obj->next;
 
@@ -19213,6 +19201,10 @@ char * gguf_get_tensor_name(const struct gguf_context * ctx, int i) {
     return ctx->infos[i].name.data;
 }
 
+enum ggml_type gguf_get_tensor_type(const struct gguf_context * ctx, int i) {
+    return ctx->infos[i].type;
+}
+
 // returns the index
 static int gguf_get_or_add_key(struct gguf_context * ctx, const char * key) {
     const int idx = gguf_find_key(ctx, key);

ggml.h

@@ -735,8 +735,8 @@ extern "C" {
     GGML_API struct ggml_tensor * ggml_view_tensor(struct ggml_context * ctx, struct ggml_tensor * src);
 
     // Context tensor enumeration and lookup
-    GGML_API struct ggml_tensor * ggml_get_first_tensor(struct ggml_context * ctx);
+    GGML_API struct ggml_tensor * ggml_get_first_tensor(const struct ggml_context * ctx);
-    GGML_API struct ggml_tensor * ggml_get_next_tensor (struct ggml_context * ctx, struct ggml_tensor * tensor);
+    GGML_API struct ggml_tensor * ggml_get_next_tensor (const struct ggml_context * ctx, struct ggml_tensor * tensor);
     GGML_API struct ggml_tensor * ggml_get_tensor(struct ggml_context * ctx, const char * name);
 
     GGML_API struct ggml_tensor * ggml_set_zero(struct ggml_tensor * tensor);
@@ -2139,6 +2139,7 @@ extern "C" {
     GGML_API int            gguf_find_tensor      (const struct gguf_context * ctx, const char * name);
     GGML_API size_t         gguf_get_tensor_offset(const struct gguf_context * ctx, int i);
     GGML_API char *         gguf_get_tensor_name  (const struct gguf_context * ctx, int i);
+    GGML_API enum ggml_type gguf_get_tensor_type  (const struct gguf_context * ctx, int i);
 
     // overrides existing values or adds a new one
     GGML_API void gguf_set_val_u8  (struct gguf_context * ctx, const char * key, uint8_t  val);