sync : ggml (backend v2, k-quants, CUDA opts, Metal opts, etc.) (#1422)

* sync : ggml (backend v2, k-quants, CUDA opts, Metal opts, etc.)

* metal : allow env metal variable to override resource path (#1415)

* Allow env variable to override resource path

* Update ggml-metal.m

---------

Co-authored-by: Georgi Gerganov <[email protected]>

* sync : restore common / main from `master`

* sync : restore whisper from `master`

* talk-llama : update to latest llama.cpp

* ruby : fix build

* ggml : fix 32-bit ARM build

* ggml : fix MIN / MAX macro collisions + update ios bindings

* ggml : fix ifdefs and MIN / MAX again

* exampels : fix Obj-C and Swift examples

* ggml : fix 32-bit ARM compatibility

* ggml : one more attempt to fix 32-bit ARM compat

* whisper : fix support for larger graphs

---------

Co-authored-by: Chris Raethke <[email protected]>

CMakeLists.txt

@@ -464,6 +464,10 @@ add_library(${TARGET}
     ggml.c
     ggml-alloc.h
     ggml-alloc.c
+    ggml-backend.h
+    ggml-backend.c
+    ggml-quants.h
+    ggml-quants.c
     ${GGML_SOURCES_METAL}
     ${GGML_SOURCES_CUDA}
     ${GGML_SOURCES_OPENCL}

Makefile

@@ -301,7 +301,13 @@ ggml.o: ggml.c ggml.h ggml-cuda.h
 ggml-alloc.o: ggml-alloc.c ggml.h ggml-alloc.h
 	$(CC)  $(CFLAGS)   -c $< -o $@
 
-WHISPER_OBJ += ggml-alloc.o
+ggml-backend.o: ggml-backend.c ggml.h ggml-backend.h
+	$(CC)  $(CFLAGS)   -c $< -o $@
+
+ggml-quants.o: ggml-quants.c ggml.h ggml-quants.h
+	$(CC)  $(CFLAGS)   -c $< -o $@
+
+WHISPER_OBJ += ggml-alloc.o ggml-backend.o ggml-quants.o
 
 whisper.o: whisper.cpp whisper.h ggml.h ggml-cuda.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@

bindings/ios

@@ -1 +1 @@
-Subproject commit 22a9eef021afc67f2154bc9811ed620b26299d1b
+Subproject commit 44b39fd4ec616a9ce66635e36045372d03dd45e0

bindings/ruby/ext/extconf.rb

@@ -3,8 +3,14 @@ system("cp #{File.join(File.dirname(__FILE__),'..','..','..','whisper.cpp')} .")
 system("cp #{File.join(File.dirname(__FILE__),'..','..','..','whisper.h')} .")
 system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml.h')} .")
 system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml.c')} .")
+system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-impl.h')} .")
 system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-alloc.h')} .")
 system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-alloc.c')} .")
+system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-backend-impl.h')} .")
+system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-backend.h')} .")
+system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-backend.c')} .")
+system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-quants.h')} .")
+system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-quants.c')} .")
 system("cp #{File.join(File.dirname(__FILE__),'..','..','..','examples','dr_wav.h')} .")
 
 

bindings/ruby/ext/ggml-backend-impl.h

@@ -0,0 +1,87 @@
+#pragma once
+
+// ggml-backend internal header
+
+#include "ggml-backend.h"
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+    //
+    // Backend buffer
+    //
+
+    typedef void * ggml_backend_buffer_context_t;
+
+    struct ggml_backend_buffer_i {
+        void   (*free_buffer)   (ggml_backend_buffer_t buffer);
+        void * (*get_base)      (ggml_backend_buffer_t buffer); // get base pointer
+        size_t (*get_alloc_size)(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); // pre-allocation callback
+        void   (*init_tensor)   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); // post-allocation callback
+        void   (*free_tensor)   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); // pre-free callback
+    };
+
+    struct ggml_backend_buffer {
+        struct ggml_backend_buffer_i iface;
+
+        ggml_backend_t                backend;
+        ggml_backend_buffer_context_t context;
+
+        size_t size;
+    };
+
+    GGML_API ggml_backend_buffer_t ggml_backend_buffer_init(
+            struct ggml_backend                  * backend,
+            struct ggml_backend_buffer_i           iface,
+                   ggml_backend_buffer_context_t   context,
+                   size_t                          size);
+
+    //
+    // Backend
+    //
+
+    typedef void * ggml_backend_context_t;
+
+    struct ggml_backend_i {
+        const char * (*get_name)(ggml_backend_t backend);
+
+        void (*free)(ggml_backend_t backend);
+
+        // buffer allocation
+        ggml_backend_buffer_t (*alloc_buffer)(ggml_backend_t backend, size_t size);
+
+        // get buffer alignment
+        size_t (*get_alignment)(ggml_backend_t backend);
+
+        // tensor data access
+        // these functions can be asynchronous, helper functions are provided for synchronous access that automatically call synchronize
+        void (*set_tensor_async)(ggml_backend_t backend,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
+        void (*get_tensor_async)(ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
+        void (*synchronize)     (ggml_backend_t backend);
+
+        // (optional) copy tensor between different backends, allow for single-copy tranfers
+        void (*cpy_tensor_from)(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst);
+        void (*cpy_tensor_to)  (ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst);
+
+        // compute graph with a plan
+        ggml_backend_graph_plan_t (*graph_plan_create) (ggml_backend_t backend, struct ggml_cgraph * cgraph);
+        void                      (*graph_plan_free)   (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
+        void                      (*graph_plan_compute)(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
+
+        // compute graph without a plan
+        void (*graph_compute)(ggml_backend_t backend, struct ggml_cgraph * cgraph);
+
+        // check if the backend supports an operation
+        bool (*supports_op)(ggml_backend_t backend, const struct ggml_tensor * op);
+    };
+
+    struct ggml_backend {
+        struct ggml_backend_i iface;
+
+        ggml_backend_context_t context;
+    };
+
+#ifdef  __cplusplus
+}
+#endif

bindings/ruby/ext/ggml-backend.c

@@ -0,0 +1,950 @@
+#include "ggml-backend-impl.h"
+#include "ggml-alloc.h"
+#include "ggml-impl.h"
+
+#include <assert.h>
+#include <limits.h>
+#include <stdarg.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#define UNUSED GGML_UNUSED
+
+#define MAX(a, b) ((a) > (b) ? (a) : (b))
+
+// backend buffer
+
+ggml_backend_buffer_t ggml_backend_buffer_init(
+        struct ggml_backend                  * backend,
+        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));
+
+    GGML_ASSERT(iface.get_base != NULL);
+
+    (*buffer) = (struct ggml_backend_buffer) {
+        /* .interface = */ iface,
+        /* .backend   = */ backend,
+        /* .context   = */ context,
+        /* .size      = */ size,
+    };
+
+    return buffer;
+}
+
+void ggml_backend_buffer_free(ggml_backend_buffer_t buffer) {
+    if (buffer == NULL) {
+        return;
+    }
+
+    if (buffer->iface.free_buffer != NULL) {
+        buffer->iface.free_buffer(buffer);
+    }
+    free(buffer);
+}
+
+size_t ggml_backend_buffer_get_alignment(ggml_backend_buffer_t buffer) {
+    return ggml_backend_get_alignment(buffer->backend);
+}
+
+size_t ggml_backend_buffer_get_size(ggml_backend_buffer_t buffer) {
+    return buffer->size;
+}
+
+void * ggml_backend_buffer_get_base(ggml_backend_buffer_t buffer) {
+    void * base = buffer->iface.get_base(buffer);
+
+    GGML_ASSERT(base != NULL && "backend buffer base cannot be NULL");
+
+    return base;
+}
+
+size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
+    // get_alloc_size is optional, defaults to ggml_nbytes
+    if (buffer->iface.get_alloc_size) {
+        return buffer->iface.get_alloc_size(buffer, tensor);
+    }
+    return ggml_nbytes(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);
+    }
+}
+
+void ggml_backend_buffer_free_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
+    // free_tensor is optional
+    if (buffer->iface.free_tensor) {
+        buffer->iface.free_tensor(buffer, tensor);
+    }
+}
+
+// backend
+
+ggml_backend_t ggml_get_backend(const struct ggml_tensor * tensor) {
+    return tensor->buffer ? tensor->buffer->backend : NULL;
+}
+
+const char * ggml_backend_name(ggml_backend_t backend) {
+    if (backend == NULL) {
+        return "NULL";
+    }
+    return backend->iface.get_name(backend);
+}
+
+void ggml_backend_free(ggml_backend_t backend) {
+    if (backend == NULL) {
+        return;
+    }
+
+    backend->iface.free(backend);
+}
+
+ggml_backend_buffer_t ggml_backend_alloc_buffer(ggml_backend_t backend, size_t size) {
+    return backend->iface.alloc_buffer(backend, size);
+}
+
+size_t ggml_backend_get_alignment(ggml_backend_t backend) {
+    return backend->iface.get_alignment(backend);
+}
+
+void ggml_backend_tensor_set_async(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+    ggml_get_backend(tensor)->iface.set_tensor_async(ggml_get_backend(tensor), tensor, data, offset, size);
+}
+
+void ggml_backend_tensor_get_async(const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+    ggml_get_backend(tensor)->iface.get_tensor_async(ggml_get_backend(tensor), tensor, data, offset, size);
+}
+
+void ggml_backend_tensor_set(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+    ggml_backend_t backend = ggml_get_backend(tensor);
+
+    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
+    GGML_ASSERT(backend != NULL && "tensor backend not set");
+
+    backend->iface.set_tensor_async(backend, tensor, data, offset, size);
+    backend->iface.synchronize(backend);
+}
+
+void ggml_backend_tensor_get(const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+    ggml_backend_t backend = ggml_get_backend(tensor);
+
+    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
+    GGML_ASSERT(backend != NULL && "tensor backend not set");
+
+    backend->iface.get_tensor_async(backend, tensor, data, offset, size);
+    backend->iface.synchronize(backend);
+}
+
+void ggml_backend_synchronize(ggml_backend_t backend) {
+    backend->iface.synchronize(backend);
+}
+
+ggml_backend_graph_plan_t ggml_backend_graph_plan_create(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+    return backend->iface.graph_plan_create(backend, cgraph);
+}
+
+void ggml_backend_graph_plan_free(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
+    backend->iface.graph_plan_free(backend, plan);
+}
+
+void ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
+    backend->iface.graph_plan_compute(backend, plan);
+}
+
+void ggml_backend_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+    backend->iface.graph_compute(backend, cgraph);
+}
+
+bool ggml_backend_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
+    return backend->iface.supports_op(backend, op);
+}
+
+// backend copy
+
+static bool ggml_are_same_layout(const struct ggml_tensor * a, const struct ggml_tensor * b) {
+    if (a->type != b->type) {
+        return false;
+    }
+    for (int i = 0; i < GGML_MAX_DIMS; i++) {
+        if (a->ne[i] != b->ne[i]) {
+            return false;
+        }
+        if (a->nb[i] != b->nb[i]) {
+            return false;
+        }
+    }
+    return true;
+}
+
+void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst) {
+    //printf("src: %s ne: [%d %d %d %d] nb: [%d %d %d %d]\n", src->name, (int)src->ne[0], (int)src->ne[1], (int)src->ne[2], (int)src->ne[3], (int)src->nb[0], (int)src->nb[1], (int)src->nb[2], (int)src->nb[3]);
+    //printf("dst: %s ne: [%d %d %d %d] nb: [%d %d %d %d]\n", dst->name, (int)dst->ne[0], (int)dst->ne[1], (int)dst->ne[2], (int)dst->ne[3], (int)dst->nb[0], (int)dst->nb[1], (int)dst->nb[2], (int)dst->nb[3]);
+    GGML_ASSERT(ggml_are_same_layout(src, dst) && "cannot copy tensors with different layouts");
+
+    // fprintf(stderr, "cpy tensor %s from %s to %s (%lu bytes)\n", src->name, ggml_backend_name(src->backend), ggml_backend_name(dst->backend), ggml_nbytes(src));
+
+    if (src == dst) {
+        return;
+    }
+
+    // TODO: allow backends to support copy to/from same backend
+
+    if (ggml_get_backend(dst)->iface.cpy_tensor_from != NULL) {
+        ggml_get_backend(dst)->iface.cpy_tensor_from(ggml_get_backend(dst)->context, src, dst);
+    } else if (ggml_get_backend(src)->iface.cpy_tensor_to != NULL) {
+        ggml_get_backend(src)->iface.cpy_tensor_to(ggml_get_backend(src)->context, src, dst);
+    } else {
+        // shouldn't be hit when copying from/to CPU
+        #ifndef NDEBUG
+        fprintf(stderr, "ggml_backend_tensor_copy: neither cpy_tensor_from nor cpy_tensor_to are implemented for backends %s and %s, falling back to get/set\n", ggml_backend_name(src->buffer->backend), ggml_backend_name(dst->buffer->backend));
+        #endif
+        size_t nbytes = ggml_nbytes(src);
+        void * data = malloc(nbytes);
+        ggml_backend_tensor_get(src, data, 0, nbytes);
+        ggml_backend_tensor_set(dst, data, 0, nbytes);
+        free(data);
+    }
+}
+
+// backend CPU
+
+struct ggml_backend_cpu_context {
+    int n_threads;
+    void * work_data;
+    size_t work_size;
+};
+
+static const char * ggml_backend_cpu_name(ggml_backend_t backend) {
+    return "CPU";
+
+    UNUSED(backend);
+}
+
+static void ggml_backend_cpu_free(ggml_backend_t backend) {
+    struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
+    free(cpu_ctx->work_data);
+    free(cpu_ctx);
+    free(backend);
+}
+
+static void * ggml_backend_cpu_buffer_get_base(ggml_backend_buffer_t buffer) {
+    return (void *)buffer->context;
+}
+
+static void ggml_backend_cpu_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+    free(buffer->context);
+    UNUSED(buffer);
+}
+
+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,
+    /* .get_alloc_size = */ NULL, // defaults to ggml_nbytes
+    /* .init_tensor    = */ NULL, // no initialization required
+    /* .free_tensor    = */ NULL, // no cleanup required
+};
+
+// for buffers from ptr, free is not called
+static struct ggml_backend_buffer_i cpu_backend_buffer_i_from_ptr = {
+    /* .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,
+    /* .get_alloc_size = */ NULL, // defaults to ggml_nbytes
+    /* .init_tensor    = */ NULL,
+    /* .free_tensor    = */ NULL,
+};
+
+static const size_t TENSOR_ALIGNMENT = 64; // should be enough for AVX 512
+
+static ggml_backend_buffer_t ggml_backend_cpu_alloc_buffer(ggml_backend_t backend, size_t size) {
+    size += TENSOR_ALIGNMENT;   // malloc may return an address that is not aligned
+    void * data = malloc(size); // TODO: maybe use GGML_ALIGNED_MALLOC?
+
+    GGML_ASSERT(data != NULL && "failed to allocate buffer");
+
+    return ggml_backend_buffer_init(backend, cpu_backend_buffer_i, data, size);
+}
+
+static size_t ggml_backend_cpu_get_alignment(ggml_backend_t backend) {
+    return TENSOR_ALIGNMENT;
+    UNUSED(backend);
+}
+
+static void ggml_backend_cpu_set_tensor_async(ggml_backend_t backend, 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(backend);
+}
+
+static void ggml_backend_cpu_get_tensor_async(ggml_backend_t backend, 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(backend);
+}
+
+static void ggml_backend_cpu_synchronize(ggml_backend_t backend) {
+    UNUSED(backend);
+}
+
+static void ggml_backend_cpu_cpy_tensor_from(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst) {
+    ggml_backend_tensor_get(src, dst->data, 0, ggml_nbytes(src));
+
+    UNUSED(backend);
+}
+
+static void ggml_backend_cpu_cpy_tensor_to(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst) {
+    ggml_backend_tensor_set(dst, src->data, 0, ggml_nbytes(src));
+
+    UNUSED(backend);
+}
+
+struct ggml_backend_plan_cpu {
+    struct ggml_cplan cplan;
+    struct ggml_cgraph cgraph;
+};
+
+static ggml_backend_graph_plan_t ggml_backend_cpu_graph_plan_create(ggml_backend_t backend, 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;
+
+    if (cpu_plan->cplan.work_size > 0) {
+        cpu_plan->cplan.work_data = malloc(cpu_plan->cplan.work_size);
+    }
+
+    return cpu_plan;
+}
+
+static void ggml_backend_cpu_graph_plan_free(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
+    struct ggml_backend_plan_cpu * cpu_plan = (struct ggml_backend_plan_cpu *)plan;
+
+    free(cpu_plan->cplan.work_data);
+    free(cpu_plan);
+
+    UNUSED(backend);
+}
+
+static void 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;
+
+    ggml_graph_compute(&cpu_plan->cgraph, &cpu_plan->cplan);
+
+    UNUSED(backend);
+}
+
+static void 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) {
+        // TODO: may be faster to free and use malloc to avoid the copy
+        cpu_ctx->work_data = realloc(cpu_ctx->work_data, cplan.work_size);
+        cpu_ctx->work_size = cplan.work_size;
+    }
+
+    cplan.work_data = cpu_ctx->work_data;
+
+    ggml_graph_compute(cgraph, &cplan);
+}
+
+static bool ggml_backend_cpu_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
+    return true;
+    UNUSED(backend);
+    UNUSED(op);
+}
+
+static struct ggml_backend_i cpu_backend_i = {
+    /* .get_name            = */ ggml_backend_cpu_name,
+    /* .free                = */ ggml_backend_cpu_free,
+    /* .alloc_buffer        = */ ggml_backend_cpu_alloc_buffer,
+    /* .get_alignment       = */ ggml_backend_cpu_get_alignment,
+    /* .set_tensor_async    = */ ggml_backend_cpu_set_tensor_async,
+    /* .get_tensor_async    = */ ggml_backend_cpu_get_tensor_async,
+    /* .synchronize         = */ ggml_backend_cpu_synchronize,
+    /* .cpy_tensor_from     = */ ggml_backend_cpu_cpy_tensor_from,
+    /* .cpy_tensor_to       = */ ggml_backend_cpu_cpy_tensor_to,
+    /* .graph_plan_create   = */ ggml_backend_cpu_graph_plan_create,
+    /* .graph_plan_free     = */ ggml_backend_cpu_graph_plan_free,
+    /* .graph_plan_compute  = */ ggml_backend_cpu_graph_plan_compute,
+    /* .graph_compute       = */ ggml_backend_cpu_graph_compute,
+    /* .supports_op         = */ ggml_backend_cpu_supports_op,
+};
+
+ggml_backend_t ggml_backend_cpu_init(void) {
+    struct ggml_backend_cpu_context * ctx = malloc(sizeof(struct ggml_backend_cpu_context));
+
+    ctx->n_threads = GGML_DEFAULT_N_THREADS;
+    ctx->work_data = NULL;
+    ctx->work_size = 0;
+
+    ggml_backend_t cpu_backend = malloc(sizeof(struct ggml_backend));
+
+    *cpu_backend = (struct ggml_backend) {
+        /* .interface = */ cpu_backend_i,
+        /* .context   = */ ctx
+    };
+    return cpu_backend;
+}
+
+bool ggml_backend_is_cpu(ggml_backend_t backend) {
+    return backend->iface.get_name == ggml_backend_cpu_name;
+}
+
+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;
+}
+
+ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(ggml_backend_t backend_cpu, void * ptr, size_t size) {
+    return ggml_backend_buffer_init(backend_cpu, cpu_backend_buffer_i_from_ptr, ptr, size);
+}
+
+// scheduler
+
+#define GGML_MAX_BACKENDS 4
+#define GGML_MAX_SPLITS 256
+#define GGML_MAX_SPLIT_INPUTS 16
+
+struct ggml_backend_sched_split {
+    ggml_tallocr_t tallocr;
+    int i_start;
+    int i_end;
+    struct ggml_tensor * inputs[GGML_MAX_SPLIT_INPUTS];
+    int n_inputs;
+    struct ggml_cgraph * graph;
+};
+
+struct ggml_backend_sched {
+    int n_backends;
+    ggml_backend_t backends[GGML_MAX_BACKENDS];
+    ggml_tallocr_t  tallocs[GGML_MAX_BACKENDS];
+
+    ggml_gallocr_t galloc;
+
+    struct ggml_hash_set    hash_set;
+    ggml_tallocr_t *        node_talloc;                     // [hash_set.size]
+    struct ggml_tensor * (* node_copies)[GGML_MAX_BACKENDS]; // [hash_set.size][GGML_MAX_BACKENDS]
+
+    struct ggml_cgraph * graph;
+    struct ggml_backend_sched_split splits[GGML_MAX_SPLITS];
+    int n_splits;
+
+    struct ggml_context * ctx;
+
+    // align context_buffer to GGML_MEM_ALIGN
+    #ifdef _MSC_VER
+    __declspec(align(GGML_MEM_ALIGN))
+    #else
+    __attribute__((aligned(GGML_MEM_ALIGN)))
+    #endif
+    char context_buffer[GGML_MAX_SPLITS*GGML_MAX_SPLIT_INPUTS*sizeof(struct ggml_tensor) + GGML_MAX_SPLITS*sizeof(struct ggml_cgraph)];
+};
+
+#define hash_id(node) ggml_hash_find_or_insert(sched->hash_set, node)
+#define node_allocr(node) sched->node_talloc[hash_id(node)]
+
+static bool ggml_is_view_op(enum ggml_op op) {
+    return op == GGML_OP_VIEW || op == GGML_OP_RESHAPE || op == GGML_OP_PERMUTE || op == GGML_OP_TRANSPOSE;
+}
+
+// returns the priority of the backend, lower is better
+static int sched_backend_prio(ggml_backend_sched_t sched, ggml_backend_t backend) {
+    for (int i = 0; i < sched->n_backends; i++) {
+        if (sched->backends[i] == backend) {
+            return i;
+        }
+    }
+    return INT_MAX;
+}
+
+static int sched_allocr_prio(ggml_backend_sched_t sched, ggml_tallocr_t allocr) {
+    for (int i = 0; i < sched->n_backends; i++) {
+        if (sched->tallocs[i] == allocr) {
+            return i;
+        }
+    }
+    return INT_MAX;
+}
+
+// returns the backend that should be used for the node based on the current locations
+char causes[GGML_DEFAULT_GRAPH_SIZE*4 + GGML_MAX_SPLITS*GGML_MAX_SPLIT_INPUTS][128]; // debug, remove
+static ggml_backend_t sched_backend_from_cur(ggml_backend_sched_t sched, struct ggml_tensor * node) {
+    // if the dst tensor is already allocated in a buffer, we must assume that it is critical to keep it there
+    // ie. kv cache updates
+    // note that this doesn't allow fallback to CPU. need to add output tensors to the splits to copy the data back to the original backend.
+    // dst
+    ggml_backend_t cur_backend = ggml_get_backend(node);
+    if (cur_backend != NULL) {
+        sprintf(causes[hash_id(node)], "1.dst");
+        return cur_backend;
+    }
+
+    // view_src
+    if (node->view_src != NULL && ggml_get_backend(node->view_src) != NULL) {
+        sprintf(causes[hash_id(node)], "1.vsrc");
+        return ggml_get_backend(node->view_src);
+    }
+
+    // src
+    int cur_prio = INT_MAX;
+    size_t cur_size = 0;
+
+    for (int i = 0; i < GGML_MAX_SRC; i++) {
+        const struct ggml_tensor * src = node->src[i];
+        if (src == NULL) {
+            break;
+        }
+        ggml_backend_t src_backend = ggml_get_backend(src);
+        if (src_backend != NULL) {
+            int src_prio = sched_backend_prio(sched, src_backend);
+            size_t src_size = ggml_nbytes(src);
+            if (src_prio < cur_prio && src_size >= cur_size) {
+                cur_prio = src_prio;
+                cur_size = src_size;
+                cur_backend = src_backend;
+                sprintf(causes[hash_id(node)], "1.src%d", i);
+            }
+        }
+    }
+    return cur_backend;
+}
+
+static char * fmt_size(size_t size) {
+    static char buffer[128];
+    if (size >= 1024*1024) {
+        sprintf(buffer, "%zuM", size/1024/1024);
+    } else {
+        sprintf(buffer, "%zuK", size/1024);
+    }
+    return buffer;
+}
+
+static void sched_print_assignments(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
+    int cur_split = 0;
+    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 = ggml_tallocr_get_buffer(sched->splits[cur_split].tallocr)->backend;
+            fprintf(stderr, "\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, fmt_size(ggml_nbytes(sched->splits[cur_split].inputs[j])));
+            }
+            fprintf(stderr, "\n");
+            cur_split++;
+        }
+        struct ggml_tensor * node = graph->nodes[i];
+        if (ggml_is_view_op(node->op)) {
+            continue;
+        }
+        ggml_tallocr_t node_allocr = node_allocr(node);
+        ggml_backend_t node_backend = node_allocr ? ggml_tallocr_get_buffer(node_allocr)->backend : NULL;
+        fprintf(stderr, "node #%3d (%10.10s): %20.20s (%4.4s) [%4.4s %8.8s]:", i, ggml_op_name(node->op), node->name, fmt_size(ggml_nbytes(node)), node_allocr ? ggml_backend_name(node_backend) : "NULL", causes[hash_id(node)]);
+        for (int j = 0; j < GGML_MAX_SRC; j++) {
+            struct ggml_tensor * src = node->src[j];
+            if (src == NULL) {
+                break;
+            }
+            ggml_tallocr_t src_allocr = node_allocr(src);
+            ggml_backend_t src_backend = src_allocr ? ggml_tallocr_get_buffer(src_allocr)->backend : NULL;
+            fprintf(stderr, " %20.20s (%4.4s) [%4.4s %8.8s]", src->name, fmt_size(ggml_nbytes(src)), src_backend ? ggml_backend_name(src_backend) : "NULL", causes[hash_id(src)]);
+        }
+        fprintf(stderr, "\n");
+    }
+}
+
+// creates a copy of the tensor with the same memory layout
+static struct ggml_tensor * ggml_dup_tensor_layout(struct ggml_context * ctx, const struct ggml_tensor * tensor) {
+    struct ggml_tensor * dup = ggml_dup_tensor(ctx, tensor);
+    for (int i = 0; i < GGML_MAX_DIMS; i++) {
+        dup->nb[i] = tensor->nb[i];
+    }
+    return dup;
+}
+
+// assigns backends to ops and splits the graph into subgraphs that can be computed on the same backend
+// TODO: merge passes
+static void sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
+    // reset state
+    size_t hash_size = sched->hash_set.size;
+    memset(sched->hash_set.keys, 0, sizeof(sched->hash_set.keys[0]) * hash_size);
+    memset(sched->node_talloc,   0, sizeof(sched->node_talloc[0])   * hash_size);
+    memset(sched->node_copies,   0, sizeof(sched->node_copies[0])   * hash_size);
+    sched->n_splits = 0;
+
+    struct ggml_init_params params = {
+        /*.mem_size =   */ sizeof(sched->context_buffer),
+        /*.mem_buffer = */ sched->context_buffer,
+        /*.no_alloc =   */ true
+    };
+
+    if (sched->ctx != NULL) {
+        ggml_free(sched->ctx);
+    }
+
+    sched->ctx = ggml_init(params);
+
+    // pass 1: assign backends to ops with allocated inputs
+    for (int i = 0; i < graph->n_leafs; i++) {
+        struct ggml_tensor * leaf = graph->leafs[i];
+        if (node_allocr(leaf) != NULL) {
+            // do not overwrite user assignments
+            continue;
+        }
+        ggml_backend_t leaf_backend = ggml_get_backend(leaf);
+        if (leaf_backend == NULL && leaf->view_src != NULL) {
+            leaf_backend = ggml_get_backend(leaf->view_src);
+        }
+        if (leaf_backend != NULL) {
+            node_allocr(leaf) = ggml_backend_sched_get_tallocr(sched, leaf_backend);
+        }
+    }
+
+    for (int i = 0; i < graph->n_nodes; i++) {
+        struct ggml_tensor * node = graph->nodes[i];
+        if (node_allocr(node) != NULL) {
+            // do not overwrite user assignments
+            continue;
+        }
+        ggml_backend_t node_backend = sched_backend_from_cur(sched, node);
+        if (node_backend != NULL) {
+            node_allocr(node) = ggml_backend_sched_get_tallocr(sched, node_backend);
+        }
+    }
+    //printf("PASS 1 ASSIGNMENTS\n"); sched_print_assignments(sched, graph);
+
+    // pass 2: assign backends to ops from current assignments
+    // TODO:
+    //  - reuse sched_backend_from_cur
+    for (int i = 0; i < graph->n_nodes; i++) {
+        struct ggml_tensor * node = graph->nodes[i];
+        ggml_tallocr_t node_allocr = node_allocr(node);
+        if (node_allocr == NULL) {
+            int    cur_prio = INT_MAX;
+            size_t cur_size = 0;
+            for (int j = 0; j < GGML_MAX_SRC; j++) {
+                struct ggml_tensor * src = node->src[j];
+                if (src == NULL) {
+                    break;
+                }
+                ggml_tallocr_t src_allocr = node_allocr(src);
+                if (src_allocr != NULL) {
+                    int    src_prio = sched_allocr_prio(sched, src_allocr);
+                    size_t src_size = ggml_nbytes(src);
+                    if (src_prio < cur_prio && src_size >= cur_size) {
+                        cur_prio = src_prio;
+                        cur_size = src_size;
+                        node_allocr = src_allocr;
+                        sprintf(causes[hash_id(node)], "2.src%d", j);
+                    }
+                }
+            }
+            if (node_allocr != NULL) {
+                node_allocr(node) = node_allocr;
+            }
+        }
+    }
+    //printf("PASS 2 ASSIGNMENTS\n"); sched_print_assignments(sched, graph);
+
+    // pass 3: assign backends to remaining src from dst (should only be leafs)
+    for (int i = 0; i < graph->n_nodes; i++) {
+        struct ggml_tensor * node = graph->nodes[i];
+        ggml_tallocr_t node_allocr = node_allocr(node);
+        for (int j = 0; j < GGML_MAX_SRC; j++) {
+            struct ggml_tensor * src = node->src[j];
+            if (src == NULL) {
+                break;
+            }
+            ggml_tallocr_t src_allocr = node_allocr(src);
+            if (src_allocr == NULL) {
+                node_allocr(src) = node_allocr;
+            }
+        }
+    }
+    //printf("PASS 3 ASSIGNMENTS\n"); sched_print_assignments(sched, graph);
+
+    // pass 4: split graph, find tensors that need to be copied
+    // TODO:
+    //  - when switching from a less preferred backend to a more preferred backend, check if it is possible to move the switch to an earlier point for the same cost
+    // find first backend
+    int cur_split = 0;
+    for (int i = 0; i < graph->n_nodes; i++) {
+        struct ggml_tensor * node = graph->nodes[i];
+        if (node->view_src == NULL) {
+            sched->splits[0].tallocr = node_allocr(node);
+            break;
+        }
+    }
+    sched->splits[0].i_start = 0;
+    sched->splits[0].n_inputs = 0;
+    memset(sched->splits[0].inputs, 0, sizeof(sched->splits[0].inputs)); //HACK
+    ggml_tallocr_t cur_allocr = sched->splits[0].tallocr;
+    size_t cur_backend_id = sched_allocr_prio(sched, cur_allocr);
+    for (int i = 0; i < graph->n_nodes; i++) {
+        struct ggml_tensor * node = graph->nodes[i];
+
+        if (ggml_is_view_op(node->op)) {
+            continue;
+        }
+
+        ggml_tallocr_t node_allocr = node_allocr(node);
+
+        if (node_allocr != cur_allocr) {
+            sched->splits[cur_split].i_end = i;
+            cur_split++;
+            GGML_ASSERT(cur_split < GGML_MAX_SPLITS);
+            sched->splits[cur_split].tallocr = node_allocr;
+            sched->splits[cur_split].i_start = i;
+            sched->splits[cur_split].n_inputs = 0;
+            memset(sched->splits[cur_split].inputs, 0, sizeof(sched->splits[cur_split].inputs)); //HACK
+            cur_allocr = node_allocr;
+            cur_backend_id = sched_allocr_prio(sched, cur_allocr);
+        }
+
+        // find inputs that are not on the same backend
+        for (int j = 0; j < GGML_MAX_SRC; j++) {
+            struct ggml_tensor * src = node->src[j];
+            if (src == NULL) {
+                break;
+            }
+            ggml_tallocr_t src_allocr = node_allocr(src);
+            if (src_allocr != node_allocr) {
+                int n_inputs = sched->splits[cur_split].n_inputs++;
+                GGML_ASSERT(n_inputs < GGML_MAX_SPLIT_INPUTS);
+                sched->splits[cur_split].inputs[n_inputs] = (struct ggml_tensor *)src;
+
+                // create copies
+                size_t id = hash_id(src);
+                if (sched->node_copies[id][cur_backend_id] == NULL) {
+                    struct ggml_tensor * tensor_copy = ggml_dup_tensor_layout(sched->ctx, src);
+                    sched->node_copies[id][cur_backend_id] = tensor_copy;
+                    node_allocr(tensor_copy) = cur_allocr;
+                    ggml_backend_t backend = ggml_tallocr_get_buffer(cur_allocr)->backend;
+                    ggml_format_name(tensor_copy, "%s#%s", ggml_backend_name(backend), src->name);
+                }
+                node->src[j] = sched->node_copies[id][cur_backend_id];
+            }
+        }
+    }
+    sched->splits[cur_split].i_end = graph->n_nodes;
+    sched->n_splits = cur_split + 1;
+
+    //fprintf(stderr, "PASS 4 ASSIGNMENTS\n"); sched_print_assignments(sched, graph); fflush(stdout);
+
+#if 1
+    // sanity check: all sources should have the same backend as the node
+    for (int i = 0; i < graph->n_nodes; i++) {
+        struct ggml_tensor * node = graph->nodes[i];
+        ggml_tallocr_t node_allocr = node_allocr(node);
+        if (node_allocr == NULL) {
+            fprintf(stderr, "!!!!!!! %s has no backend\n", node->name);
+        }
+        for (int j = 0; j < GGML_MAX_SRC; j++) {
+            struct ggml_tensor * src = node->src[j];
+            if (src == NULL) {
+                break;
+            }
+            ggml_tallocr_t src_allocr = node_allocr(src);
+            if (src_allocr != node_allocr /* && src_backend != NULL */) { // ignore nulls for now
+                fprintf(stderr, "!!!! %s has backend %s, src %d (%s) has backend %s\n",
+                    node->name, node_allocr ? ggml_backend_name(ggml_tallocr_get_buffer(node_allocr)->backend) : "NULL",
+                    j, src->name, src_allocr ? ggml_backend_name(ggml_tallocr_get_buffer(src_allocr)->backend) : "NULL");
+            }
+        }
+    }
+#endif
+
+    // create copies of the graph for each split
+    // FIXME: avoid this copy, pass split inputs to ggml_gallocr_alloc_graph_n in some other way
+    struct ggml_cgraph * graph_copy = ggml_new_graph_custom(sched->ctx, graph->n_nodes + sched->n_splits*GGML_MAX_SPLIT_INPUTS, false);
+    for (int i = 0; i < sched->n_splits; i++) {
+        struct ggml_backend_sched_split * split = &sched->splits[i];
+        split->graph = ggml_graph_view(sched->ctx, graph, split->i_start, split->i_end);
+
+        // add inputs to the graph copy so that they are allocated by ggml-alloc at the start of the split
+        for (int j = 0; j < split->n_inputs; j++) {
+            struct ggml_tensor * input = split->inputs[j];
+            struct ggml_tensor * input_cpy = sched->node_copies[hash_id(input)][sched_allocr_prio(sched, split->tallocr)];
+            input_cpy->src[0] = input;
+            graph_copy->nodes[graph_copy->n_nodes++] = input_cpy;
+        }
+
+        for (int j = split->i_start; j < split->i_end; j++) {
+            graph_copy->nodes[graph_copy->n_nodes++] = graph->nodes[j];
+        }
+    }
+    sched->graph = graph_copy;
+}
+
+static void sched_alloc_splits(ggml_backend_sched_t sched) {
+    ggml_gallocr_alloc_graph_n(
+        sched->galloc,
+        sched->graph,
+        sched->hash_set,
+        sched->node_talloc);
+}
+
+static void sched_compute_splits(ggml_backend_sched_t sched) {
+    uint64_t copy_us[GGML_MAX_BACKENDS] = {0};
+    uint64_t compute_us[GGML_MAX_BACKENDS] = {0};
+
+    struct ggml_backend_sched_split * splits = sched->splits;
+
+    for (int i = 0; i < sched->n_splits; i++) {
+        struct ggml_backend_sched_split * split = &splits[i];
+        ggml_backend_t split_backend = ggml_tallocr_get_buffer(split->tallocr)->backend;
+        int split_backend_id = sched_backend_prio(sched, split_backend);
+
+        // copy the input tensors to the split backend
+        uint64_t copy_start_us = ggml_time_us();
+        for (int j = 0; j < split->n_inputs; j++) {
+            struct ggml_tensor * input_cpy = sched->node_copies[hash_id(split->inputs[j])][sched_backend_prio(sched, split_backend)];
+            if (split->inputs[j]->buffer == NULL) {
+                if (split->inputs[j]->view_src == NULL) {
+                    fprintf(stderr, "input %s has no buffer and no view_src\n", split->inputs[j]->name);
+                    exit(1);
+                }
+                struct ggml_tensor * view = split->inputs[j];
+                view->backend = view->view_src->backend;
+                view->buffer  = view->view_src->buffer;
+                view->data    = (char *)view->view_src->data + view->view_offs;
+                ggml_backend_buffer_init_tensor(ggml_backend_sched_get_buffer(sched, view->buffer->backend), view);
+            }
+            if (input_cpy->buffer == NULL) {
+                fprintf(stderr, "input_cpy %s has no buffer\n", input_cpy->name);
+                exit(1);
+            }
+            GGML_ASSERT(split->inputs[j]->buffer->backend != input_cpy->buffer->backend);
+            GGML_ASSERT(input_cpy->buffer->backend == split_backend);
+            ggml_backend_tensor_copy(split->inputs[j], input_cpy);
+        }
+        // ggml_backend_synchronize(split_backend);
+        int64_t copy_end_us = ggml_time_us();
+        copy_us[split_backend_id] += copy_end_us - copy_start_us;
+
+#if 0
+        char split_filename[GGML_MAX_NAME];
+        snprintf(split_filename, GGML_MAX_NAME, "split_%i_%s.dot", i, ggml_backend_name(split_backend));
+        ggml_graph_dump_dot(split->graph, NULL, split_filename);
+#endif
+
+        uint64_t compute_start_us = ggml_time_us();
+        ggml_backend_graph_compute(split_backend, split->graph);
+        // ggml_backend_synchronize(split_backend);
+        uint64_t compute_end_us = ggml_time_us();
+        compute_us[split_backend_id] += compute_end_us - compute_start_us;
+    }
+
+#if 0
+    // per-backend timings
+    fprintf(stderr, "sched_compute_splits times (%d splits):\n", sched->n_splits);
+    for (int i = 0; i < sched->n_backends; i++) {
+        if (copy_us[i] > 0 || compute_us[i] > 0) {
+            fprintf(stderr, "\t%5.5s: %lu us copy, %lu us compute\n", ggml_backend_name(sched->backends[i]), copy_us[i], compute_us[i]);
+        }
+    }
+#endif
+}
+
+static void sched_reset(ggml_backend_sched_t sched) {
+    for (int i = 0; i < sched->n_backends; i++) {
+        ggml_tallocr_reset(sched->tallocs[i]);
+    }
+}
+
+ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, int n_backends) {
+    GGML_ASSERT(n_backends <= GGML_MAX_BACKENDS);
+
+    struct ggml_backend_sched * sched = malloc(sizeof(struct ggml_backend_sched));
+    memset(sched, 0, sizeof(struct ggml_backend_sched));
+
+    fprintf(stderr, "ggml_backend_sched size: %lu KB\n", sizeof(struct ggml_backend_sched)/1024);
+
+    sched->n_backends = n_backends;
+    for (int i = 0; i < n_backends; i++) {
+        sched->backends[i] = backends[i];
+    }
+
+    sched->galloc = ggml_gallocr_new();
+
+    // init measure allocs for each backend
+    for (int i = 0; i < n_backends; i++) {
+        sched->tallocs[i] = ggml_tallocr_new_measure_from_backend(backends[i]);
+    }
+
+    return sched;
+}
+
+void ggml_backend_sched_free(ggml_backend_sched_t sched) {
+    if (sched == NULL) {
+        return;
+    }
+    for (int i = 0; i < sched->n_backends; i++) {
+        ggml_tallocr_free(sched->tallocs[i]);
+    }
+    ggml_gallocr_free(sched->galloc);
+    free(sched->hash_set.keys);
+    free(sched->node_talloc);
+    free(sched->node_copies);
+    free(sched);
+}
+
+void ggml_backend_sched_init_measure(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph) {
+    // initialize hash tables
+    size_t hash_size = measure_graph->visited_hash_table.size + GGML_MAX_SPLITS*GGML_MAX_SPLIT_INPUTS;
+    sched->hash_set.size = hash_size;
+    sched->hash_set.keys = malloc(sizeof(sched->hash_set.keys[0]) * hash_size);
+    sched->node_talloc   = malloc(sizeof(sched->node_talloc[0])   * hash_size);
+    sched->node_copies   = malloc(sizeof(sched->node_copies[0])   * hash_size);
+
+    sched_split_graph(sched, measure_graph);
+    sched_alloc_splits(sched);
+
+    // allocate buffers and reset allocators
+    for (int i = 0; i < sched->n_backends; i++) {
+        size_t size = ggml_tallocr_max_size(sched->tallocs[i]);
+        ggml_tallocr_free(sched->tallocs[i]);
+        sched->tallocs[i] = ggml_tallocr_new_from_backend(sched->backends[i], size);
+    }
+
+    sched_reset(sched);
+}
+
+void ggml_backend_sched_graph_compute(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
+    GGML_ASSERT(sched->hash_set.size >= graph->visited_hash_table.size + GGML_MAX_SPLITS*GGML_MAX_SPLIT_INPUTS);
+
+    sched_split_graph(sched, graph);
+    sched_alloc_splits(sched);
+    sched_compute_splits(sched);
+    sched_reset(sched);
+}
+
+ggml_tallocr_t ggml_backend_sched_get_tallocr(ggml_backend_sched_t sched, ggml_backend_t backend) {
+    int backend_index = sched_backend_prio(sched, backend);
+    return sched->tallocs[backend_index];
+}
+
+ggml_backend_buffer_t ggml_backend_sched_get_buffer(ggml_backend_sched_t sched, ggml_backend_t backend) {
+    int backend_index = sched_backend_prio(sched, backend);
+    return ggml_tallocr_get_buffer(sched->tallocs[backend_index]);
+}
+
+void ggml_backend_sched_set_node_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend) {
+    int backend_index = sched_backend_prio(sched, backend);
+    GGML_ASSERT(backend_index >= 0 && backend_index < sched->n_backends);
+    node_allocr(node) = sched->tallocs[backend_index];
+}

bindings/ruby/ext/ggml-backend.h

@@ -0,0 +1,136 @@
+#pragma once
+
+#include "ggml.h"
+#include "ggml-alloc.h"
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+    //
+    // Backend buffer
+    //
+
+    struct ggml_backend_buffer;
+    typedef struct ggml_backend_buffer * ggml_backend_buffer_t;
+
+    // backend buffer functions
+    GGML_API void   ggml_backend_buffer_free          (ggml_backend_buffer_t buffer);
+    GGML_API size_t ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer);
+    GGML_API void * ggml_backend_buffer_get_base      (ggml_backend_buffer_t buffer);
+    GGML_API size_t ggml_backend_buffer_get_size      (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_init_tensor   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+    GGML_API void   ggml_backend_buffer_free_tensor   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+
+    //
+    // Backend
+    //
+
+    struct ggml_backend;
+    typedef struct ggml_backend * ggml_backend_t;
+    typedef void * ggml_backend_graph_plan_t;
+
+    GGML_API ggml_backend_t ggml_get_backend(const struct ggml_tensor * tensor);
+
+    GGML_API const char * ggml_backend_name(ggml_backend_t backend);
+    GGML_API void         ggml_backend_free(ggml_backend_t backend);
+
+    GGML_API ggml_backend_buffer_t ggml_backend_alloc_buffer(ggml_backend_t backend, size_t size);
+
+    GGML_API size_t ggml_backend_get_alignment(ggml_backend_t backend);
+
+    GGML_API void ggml_backend_tensor_set_async(      struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
+    GGML_API void ggml_backend_tensor_get_async(const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
+
+    GGML_API void ggml_backend_tensor_set(      struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
+    GGML_API void ggml_backend_tensor_get(const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
+
+    GGML_API void ggml_backend_synchronize(ggml_backend_t backend);
+
+    GGML_API ggml_backend_graph_plan_t ggml_backend_graph_plan_create (ggml_backend_t backend, struct ggml_cgraph * cgraph);
+
+    GGML_API void ggml_backend_graph_plan_free   (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
+    GGML_API void ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
+    GGML_API void ggml_backend_graph_compute     (ggml_backend_t backend, struct ggml_cgraph * cgraph);
+    GGML_API bool ggml_backend_supports_op       (ggml_backend_t backend, const struct ggml_tensor * op);
+
+    // tensor copy between different backends
+    GGML_API void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst);
+
+    //
+    // CPU backend
+    //
+
+    GGML_API ggml_backend_t ggml_backend_cpu_init(void);
+
+    GGML_API bool ggml_backend_is_cpu(ggml_backend_t backend);
+    GGML_API void ggml_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads);
+
+    // Create a backend buffer from an existing pointer
+    GGML_API ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(ggml_backend_t backend_cpu, void * ptr, size_t size);
+
+
+    //
+    // Backend scheduler
+    //
+
+    // The backend scheduler allows for multiple backends to be used together
+    // Handles compute buffer allocation, assignment of tensors to backends, and copying of tensors between backends
+    // The backends are selected based on:
+    // - the backend that supports the operation
+    // - the location of the pre-allocated tensors (e.g. the weights)
+    /*
+      Example usage:
+
+        sched = ggml_backend_sched_new({backend_gpu, backend_gpu2, backend_cpu}, num_backends);
+        // sched is initialized with measure allocators and cannot be used until allocated with a measure graph
+
+        // initialize buffers from a measure graph
+        measure_graph = build_graph(sched); // use the allocr to allocate inputs as needed
+
+        // in build_graph:
+        build_graph(...) {
+            // allocating tensors in a specific backend (optional, recommended: pre-allocate inputs in a different buffer)
+            alloc_cpu = ggml_backend_sched_get_allocr(sched, backend_cpu);
+            ggml_allocr_alloc(alloc_cpu, tensor);
+
+            // manually assigning nodes to a backend (optional, shouldn't be needed in most cases)
+            struct ggml_tensor * node = ggml_mul_mat(ctx, ...);
+            ggml_backend_sched_set_node_backend(sched, node, backend_gpu);
+        }
+
+        // allocate backend buffers from measure graph
+        ggml_backend_sched_init_measure(sched, measure_graph);
+
+        // the scheduler is now ready to compute graphs
+
+        // compute
+        graph = build_graph(sched);
+        ggml_backend_sched_graph_compute(sched, graph);
+    */
+
+    struct ggml_backend_sched;
+    typedef struct ggml_backend_sched * ggml_backend_sched_t;
+
+    // Initialize a backend scheduler
+    GGML_API ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, int n_backends);
+
+    GGML_API void ggml_backend_sched_free(ggml_backend_sched_t sched);
+
+    // Initialize backend buffers from a measure graph
+    GGML_API void ggml_backend_sched_init_measure(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph);
+
+    GGML_API ggml_tallocr_t        ggml_backend_sched_get_tallocr(ggml_backend_sched_t sched, ggml_backend_t backend);
+    GGML_API ggml_backend_buffer_t ggml_backend_sched_get_buffer (ggml_backend_sched_t sched, ggml_backend_t backend);
+
+    GGML_API void ggml_backend_sched_set_node_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend);
+
+    // Allocate a graph on the backend scheduler
+    GGML_API void ggml_backend_sched_graph_compute(
+            ggml_backend_sched_t sched,
+            struct ggml_cgraph * graph);
+
+#ifdef  __cplusplus
+}
+#endif

bindings/ruby/ext/ggml-impl.h

@@ -0,0 +1,249 @@
+#pragma once
+
+#include "ggml.h"
+
+// GGML internal header
+
+#include <assert.h>
+#include <stddef.h>
+#include <stdbool.h>
+#include <string.h> // memcpy
+#include <math.h>   // fabsf
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// static_assert should be a #define, but if it's not,
+// fall back to the _Static_assert C11 keyword.
+// if C99 - static_assert is noop
+// ref: https://stackoverflow.com/a/53923785/4039976
+#ifndef static_assert
+#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201100L)
+#define static_assert(cond, msg) _Static_assert(cond, msg)
+#else
+#define static_assert(cond, msg) struct global_scope_noop_trick
+#endif
+#endif
+
+// __FMA__ and __F16C__ are not defined in MSVC, however they are implied with AVX2/AVX512
+#if defined(_MSC_VER) && (defined(__AVX2__) || defined(__AVX512F__))
+#ifndef __FMA__
+#define __FMA__
+#endif
+#ifndef __F16C__
+#define __F16C__
+#endif
+#ifndef __SSE3__
+#define __SSE3__
+#endif
+#endif
+
+#undef MIN
+#undef MAX
+
+#define MIN(a, b) ((a) < (b) ? (a) : (b))
+#define MAX(a, b) ((a) > (b) ? (a) : (b))
+
+// 16-bit float
+// on Arm, we use __fp16
+// on x86, we use uint16_t
+#if defined(__ARM_NEON) && !defined(_MSC_VER)
+
+// if YCM cannot find <arm_neon.h>, make a symbolic link to it, for example:
+//
+//   $ ln -sfn /Library/Developer/CommandLineTools/usr/lib/clang/13.1.6/include/arm_neon.h ./src/
+//
+#include <arm_neon.h>
+
+#define GGML_COMPUTE_FP16_TO_FP32(x) ((float) (x))
+#define GGML_COMPUTE_FP32_TO_FP16(x) (x)
+
+#define GGML_FP16_TO_FP32(x) ((float) (x))
+#define GGML_FP32_TO_FP16(x) (x)
+
+#else
+
+#ifdef __wasm_simd128__
+#include <wasm_simd128.h>
+#else
+#ifdef __POWER9_VECTOR__
+#include <altivec.h>
+#undef bool
+#define bool _Bool
+#else
+#if defined(_MSC_VER) || defined(__MINGW32__)
+#include <intrin.h>
+#else
+#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__) || defined(__SSE3__)
+#if !defined(__riscv)
+#include <immintrin.h>
+#endif
+#endif
+#endif
+#endif
+#endif
+
+#ifdef __riscv_v_intrinsic
+#include <riscv_vector.h>
+#endif
+
+#ifdef __F16C__
+
+#ifdef _MSC_VER
+#define GGML_COMPUTE_FP16_TO_FP32(x) _mm_cvtss_f32(_mm_cvtph_ps(_mm_cvtsi32_si128(x)))
+#define GGML_COMPUTE_FP32_TO_FP16(x) _mm_extract_epi16(_mm_cvtps_ph(_mm_set_ss(x), 0), 0)
+#else
+#define GGML_COMPUTE_FP16_TO_FP32(x) _cvtsh_ss(x)
+#define GGML_COMPUTE_FP32_TO_FP16(x) _cvtss_sh(x, 0)
+#endif
+
+#elif defined(__POWER9_VECTOR__)
+
+#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
+#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
+/* the inline asm below is about 12% faster than the lookup method */
+#define GGML_FP16_TO_FP32(x) GGML_COMPUTE_FP16_TO_FP32(x)
+#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
+
+static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
+    register float f;
+    register double d;
+    __asm__(
+        "mtfprd %0,%2\n"
+        "xscvhpdp %0,%0\n"
+        "frsp %1,%0\n" :
+        /* temp */ "=d"(d),
+        /* out */  "=f"(f):
+        /* in */   "r"(h));
+    return f;
+}
+
+static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
+    register double d;
+    register ggml_fp16_t r;
+    __asm__( /* xscvdphp can work on double or single precision */
+        "xscvdphp %0,%2\n"
+        "mffprd %1,%0\n" :
+        /* temp */ "=d"(d),
+        /* out */  "=r"(r):
+        /* in */   "f"(f));
+    return r;
+}
+
+#else
+
+// FP16 <-> FP32
+// ref: https://github.com/Maratyszcza/FP16
+
+static inline float fp32_from_bits(uint32_t w) {
+    union {
+        uint32_t as_bits;
+        float as_value;
+    } fp32;
+    fp32.as_bits = w;
+    return fp32.as_value;
+}
+
+static inline uint32_t fp32_to_bits(float f) {
+    union {
+        float as_value;
+        uint32_t as_bits;
+    } fp32;
+    fp32.as_value = f;
+    return fp32.as_bits;
+}
+
+static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
+    const uint32_t w = (uint32_t) h << 16;
+    const uint32_t sign = w & UINT32_C(0x80000000);
+    const uint32_t two_w = w + w;
+
+    const uint32_t exp_offset = UINT32_C(0xE0) << 23;
+#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)
+    const float exp_scale = 0x1.0p-112f;
+#else
+    const float exp_scale = fp32_from_bits(UINT32_C(0x7800000));
+#endif
+    const float normalized_value = fp32_from_bits((two_w >> 4) + exp_offset) * exp_scale;
+
+    const uint32_t magic_mask = UINT32_C(126) << 23;
+    const float magic_bias = 0.5f;
+    const float denormalized_value = fp32_from_bits((two_w >> 17) | magic_mask) - magic_bias;
+
+    const uint32_t denormalized_cutoff = UINT32_C(1) << 27;
+    const uint32_t result = sign |
+        (two_w < denormalized_cutoff ? fp32_to_bits(denormalized_value) : fp32_to_bits(normalized_value));
+    return fp32_from_bits(result);
+}
+
+static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
+#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)
+    const float scale_to_inf = 0x1.0p+112f;
+    const float scale_to_zero = 0x1.0p-110f;
+#else
+    const float scale_to_inf = fp32_from_bits(UINT32_C(0x77800000));
+    const float scale_to_zero = fp32_from_bits(UINT32_C(0x08800000));
+#endif
+    float base = (fabsf(f) * scale_to_inf) * scale_to_zero;
+
+    const uint32_t w = fp32_to_bits(f);
+    const uint32_t shl1_w = w + w;
+    const uint32_t sign = w & UINT32_C(0x80000000);
+    uint32_t bias = shl1_w & UINT32_C(0xFF000000);
+    if (bias < UINT32_C(0x71000000)) {
+        bias = UINT32_C(0x71000000);
+    }
+
+    base = fp32_from_bits((bias >> 1) + UINT32_C(0x07800000)) + base;
+    const uint32_t bits = fp32_to_bits(base);
+    const uint32_t exp_bits = (bits >> 13) & UINT32_C(0x00007C00);
+    const uint32_t mantissa_bits = bits & UINT32_C(0x00000FFF);
+    const uint32_t nonsign = exp_bits + mantissa_bits;
+    return (sign >> 16) | (shl1_w > UINT32_C(0xFF000000) ? UINT16_C(0x7E00) : nonsign);
+}
+
+#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
+#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
+
+#endif // __F16C__
+
+#endif // __ARM_NEON
+
+// precomputed f32 table for f16 (256 KB)
+// defined in ggml.c, initialized in ggml_init()
+extern float ggml_table_f32_f16[1 << 16];
+
+// On ARM NEON, it's quicker to directly convert x -> x instead of calling into ggml_lookup_fp16_to_fp32,
+// so we define GGML_FP16_TO_FP32 and GGML_FP32_TO_FP16 elsewhere for NEON.
+// This is also true for POWER9.
+#if !defined(GGML_FP16_TO_FP32) || !defined(GGML_FP32_TO_FP16)
+
+inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
+    uint16_t s;
+    memcpy(&s, &f, sizeof(uint16_t));
+    return ggml_table_f32_f16[s];
+}
+
+#define GGML_FP16_TO_FP32(x) ggml_lookup_fp16_to_fp32(x)
+#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
+
+#endif
+
+#define GGML_HASHTABLE_FULL ((size_t)-1)
+#define GGML_HASHTABLE_ALREADY_EXISTS ((size_t)-2)
+
+bool   ggml_hash_contains      (const struct ggml_hash_set hash_set, struct ggml_tensor * key);
+
+// returns GGML_HASHTABLE_FULL if table is full, otherwise the current index of the key or where it should be inserted
+size_t ggml_hash_find          (const struct ggml_hash_set hash_set, struct ggml_tensor * key);
+
+// returns GGML_HAHSHTABLE_ALREADY_EXISTS if key already exists, index otherwise, asserts if table is full
+size_t ggml_hash_insert        (      struct ggml_hash_set hash_set, struct ggml_tensor * key);
+
+// return index, asserts if table is full
+size_t ggml_hash_find_or_insert(      struct ggml_hash_set hash_set, struct ggml_tensor * key);
+
+#ifdef __cplusplus
+}
+#endif

bindings/ruby/ext/ggml-quants.h

@@ -0,0 +1,224 @@
+#pragma once
+
+#include "ggml-impl.h"
+
+// GGML internal header
+
+#include <stdint.h>
+#include <stddef.h>
+
+#define QK4_0 32
+typedef struct {
+    ggml_fp16_t d;          // delta
+    uint8_t qs[QK4_0 / 2];  // nibbles / quants
+} block_q4_0;
+static_assert(sizeof(block_q4_0) == sizeof(ggml_fp16_t) + QK4_0 / 2, "wrong q4_0 block size/padding");
+
+#define QK4_1 32
+typedef struct {
+    ggml_fp16_t d;          // delta
+    ggml_fp16_t m;          // min
+    uint8_t qs[QK4_1 / 2];  // nibbles / quants
+} block_q4_1;
+static_assert(sizeof(block_q4_1) == 2 * sizeof(ggml_fp16_t) + QK4_1 / 2, "wrong q4_1 block size/padding");
+
+#define QK5_0 32
+typedef struct {
+    ggml_fp16_t d;         // delta
+    uint8_t qh[4];         // 5-th bit of quants
+    uint8_t qs[QK5_0 / 2]; // nibbles / quants
+} block_q5_0;
+static_assert(sizeof(block_q5_0) == sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_0 / 2, "wrong q5_0 block size/padding");
+
+#define QK5_1 32
+typedef struct {
+    ggml_fp16_t d;         // delta
+    ggml_fp16_t m;         // min
+    uint8_t qh[4];         // 5-th bit of quants
+    uint8_t qs[QK5_1 / 2]; // nibbles / quants
+} block_q5_1;
+static_assert(sizeof(block_q5_1) == 2 * sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_1 / 2, "wrong q5_1 block size/padding");
+
+#define QK8_0 32
+typedef struct {
+    ggml_fp16_t d;         // delta
+    int8_t  qs[QK8_0];     // quants
+} block_q8_0;
+static_assert(sizeof(block_q8_0) == sizeof(ggml_fp16_t) + QK8_0, "wrong q8_0 block size/padding");
+
+#define QK8_1 32
+typedef struct {
+    float d;               // delta
+    float s;               // d * sum(qs[i])
+    int8_t  qs[QK8_1];     // quants
+} block_q8_1;
+static_assert(sizeof(block_q8_1) == 2*sizeof(float) + QK8_1, "wrong q8_1 block size/padding");
+
+//
+// Super-block quantization structures
+//
+
+// Super-block size
+#ifdef GGML_QKK_64
+#define QK_K 64
+#define K_SCALE_SIZE 4
+#else
+#define QK_K 256
+#define K_SCALE_SIZE 12
+#endif
+
+// 2-bit quantization
+// weight is represented as x = a * q + b
+// 16 blocks of 16 elements each
+// Effectively 2.5625 bits per weight
+typedef struct {
+    uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
+    uint8_t qs[QK_K/4];      // quants
+    ggml_fp16_t d;           // super-block scale for quantized scales
+    ggml_fp16_t dmin;        // super-block scale for quantized mins
+} block_q2_K;
+static_assert(sizeof(block_q2_K) == 2*sizeof(ggml_fp16_t) + QK_K/16 + QK_K/4, "wrong q2_K block size/padding");
+
+// 3-bit quantization
+// weight is represented as x = a * q
+// 16 blocks of 16 elements each
+// Effectively 3.4375 bits per weight
+#ifdef GGML_QKK_64
+typedef struct {
+    uint8_t hmask[QK_K/8];     // quants - high bit
+    uint8_t qs[QK_K/4];        // quants - low 2 bits
+    uint8_t scales[2];
+    ggml_fp16_t d;             // super-block scale
+} block_q3_K;
+static_assert(sizeof(block_q3_K) == sizeof(ggml_fp16_t) + QK_K / 4 + QK_K / 8 + 2, "wrong q3_K block size/padding");
+#else
+typedef struct {
+    uint8_t hmask[QK_K/8];     // quants - high bit
+    uint8_t qs[QK_K/4];        // quants - low 2 bits
+    uint8_t scales[12];        // scales, quantized with 6 bits
+    ggml_fp16_t d;             // super-block scale
+} block_q3_K;
+static_assert(sizeof(block_q3_K) == sizeof(ggml_fp16_t) + QK_K / 4 + QK_K / 8 + 12, "wrong q3_K block size/padding");
+#endif
+
+// 4-bit quantization
+// 8 blocks of 32 elements each
+// weight is represented as x = a * q + b
+// Effectively 4.5 bits per weight
+#ifdef GGML_QKK_64
+typedef struct {
+    ggml_fp16_t d[2];          // super-block scales/mins
+    uint8_t scales[2];         // 4-bit block scales/mins
+    uint8_t qs[QK_K/2];        // 4--bit quants
+} block_q4_K;
+static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_fp16_t) + QK_K/2 + 2, "wrong q4_K block size/padding");
+#else
+typedef struct {
+    ggml_fp16_t d;             // super-block scale for quantized scales
+    ggml_fp16_t dmin;          // super-block scale for quantized mins
+    uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
+    uint8_t qs[QK_K/2];        // 4--bit quants
+} block_q4_K;
+static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_fp16_t) + K_SCALE_SIZE + QK_K/2, "wrong q4_K block size/padding");
+#endif
+
+// 5-bit quantization
+// 8 blocks of 32 elements each
+// weight is represented as x = a * q + b
+// Effectively 5.5 bits per weight
+#ifdef GGML_QKK_64
+typedef struct {
+    ggml_fp16_t d;               // super-block scale
+    int8_t  scales[QK_K/16];     // 8-bit block scales
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+} block_q5_K;
+static_assert(sizeof(block_q5_K) == sizeof(ggml_fp16_t) + QK_K/2 + QK_K/8 + QK_K/16, "wrong q5_K block size/padding");
+#else
+typedef struct {
+    ggml_fp16_t d;               // super-block scale for quantized scales
+    ggml_fp16_t dmin;            // super-block scale for quantized mins
+    uint8_t scales[K_SCALE_SIZE];   // scales and mins, quantized with 6 bits
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+} block_q5_K;
+static_assert(sizeof(block_q5_K) == 2*sizeof(ggml_fp16_t) + K_SCALE_SIZE + QK_K/2 + QK_K/8, "wrong q5_K block size/padding");
+#endif
+
+// 6-bit quantization
+// weight is represented as x = a * q
+// 16 blocks of 16 elements each
+// Effectively 6.5625 bits per weight
+typedef struct {
+    uint8_t ql[QK_K/2];      // quants, lower 4 bits
+    uint8_t qh[QK_K/4];      // quants, upper 2 bits
+    int8_t  scales[QK_K/16]; // scales, quantized with 8 bits
+    ggml_fp16_t d;           // super-block scale
+} block_q6_K;
+static_assert(sizeof(block_q6_K) == sizeof(ggml_fp16_t) + QK_K / 16 + 3*QK_K/4, "wrong q6_K block size/padding");
+
+// This is only used for intermediate quantization and dot products
+typedef struct {
+    float   d;              // delta
+    int8_t  qs[QK_K];       // quants
+    int16_t bsums[QK_K/16]; // sum of quants in groups of 16
+} block_q8_K;
+static_assert(sizeof(block_q8_K) == sizeof(float) + QK_K + QK_K/16*sizeof(int16_t), "wrong q8_K block size/padding");
+
+
+// Quantization
+void quantize_row_q4_0_reference(const float * restrict x, block_q4_0 * restrict y, int k);
+void quantize_row_q4_1_reference(const float * restrict x, block_q4_1 * restrict y, int k);
+void quantize_row_q5_0_reference(const float * restrict x, block_q5_0 * restrict y, int k);
+void quantize_row_q5_1_reference(const float * restrict x, block_q5_1 * restrict y, int k);
+void quantize_row_q8_0_reference(const float * restrict x, block_q8_0 * restrict y, int k);
+void quantize_row_q8_1_reference(const float * restrict x, block_q8_1 * restrict y, int k);
+
+void quantize_row_q2_K_reference(const float * restrict x, block_q2_K * restrict y, int k);
+void quantize_row_q3_K_reference(const float * restrict x, block_q3_K * restrict y, int k);
+void quantize_row_q4_K_reference(const float * restrict x, block_q4_K * restrict y, int k);
+void quantize_row_q5_K_reference(const float * restrict x, block_q5_K * restrict y, int k);
+void quantize_row_q6_K_reference(const float * restrict x, block_q6_K * restrict y, int k);
+void quantize_row_q8_K_reference(const float * restrict x, block_q8_K * restrict y, int k);
+
+void quantize_row_q4_0(const float * restrict x, void * restrict y, int k);
+void quantize_row_q4_1(const float * restrict x, void * restrict y, int k);
+void quantize_row_q5_0(const float * restrict x, void * restrict y, int k);
+void quantize_row_q5_1(const float * restrict x, void * restrict y, int k);
+void quantize_row_q8_0(const float * restrict x, void * restrict y, int k);
+void quantize_row_q8_1(const float * restrict x, void * restrict y, int k);
+
+void quantize_row_q2_K(const float * restrict x, void * restrict y, int k);
+void quantize_row_q3_K(const float * restrict x, void * restrict y, int k);
+void quantize_row_q4_K(const float * restrict x, void * restrict y, int k);
+void quantize_row_q5_K(const float * restrict x, void * restrict y, int k);
+void quantize_row_q6_K(const float * restrict x, void * restrict y, int k);
+void quantize_row_q8_K(const float * restrict x, void * restrict y, int k);
+
+// Dequantization
+void dequantize_row_q4_0(const block_q4_0 * restrict x, float * restrict y, int k);
+void dequantize_row_q4_1(const block_q4_1 * restrict x, float * restrict y, int k);
+void dequantize_row_q5_0(const block_q5_0 * restrict x, float * restrict y, int k);
+void dequantize_row_q5_1(const block_q5_1 * restrict x, float * restrict y, int k);
+void dequantize_row_q8_0(const block_q8_0 * restrict x, float * restrict y, int k);
+//void dequantize_row_q8_1(const block_q8_1 * restrict x, float * restrict y, int k);
+
+void dequantize_row_q2_K(const block_q2_K * restrict x, float * restrict y, int k);
+void dequantize_row_q3_K(const block_q3_K * restrict x, float * restrict y, int k);
+void dequantize_row_q4_K(const block_q4_K * restrict x, float * restrict y, int k);
+void dequantize_row_q5_K(const block_q5_K * restrict x, float * restrict y, int k);
+void dequantize_row_q6_K(const block_q6_K * restrict x, float * restrict y, int k);
+void dequantize_row_q8_K(const block_q8_K * restrict x, float * restrict y, int k);
+
+// Dot product
+void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q5_0_q8_0(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+
+void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);

examples/common.cpp

@@ -38,12 +38,12 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
             params.seed = std::stoi(get_next_arg(i, argc, argv, arg, params));
         } else if (arg == "-t" || arg == "--threads") {
             params.n_threads = std::stoi(get_next_arg(i, argc, argv, arg, params));
-        } else if (arg == "-ngl" || arg == "--gpu-layers" || arg == "--n-gpu-layers") {
-            params.n_gpu_layers = std::stoi(get_next_arg(i, argc, argv, arg, params));
         } else if (arg == "-p" || arg == "--prompt") {
             params.prompt = get_next_arg(i, argc, argv, arg, params);
         } else if (arg == "-n" || arg == "--n_predict") {
             params.n_predict = std::stoi(get_next_arg(i, argc, argv, arg, params));
+        } else if (arg == "-np" || arg == "--n_parallel") {
+            params.n_parallel = std::stoi(get_next_arg(i, argc, argv, arg, params));
         } else if (arg == "--top_k") {
             params.top_k = std::stoi(get_next_arg(i, argc, argv, arg, params));
         } else if (arg == "--top_p") {
@@ -56,6 +56,12 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
             params.repeat_penalty = std::stof(get_next_arg(i, argc, argv, arg, params));
         } else if (arg == "-b" || arg == "--batch_size") {
             params.n_batch= std::stoi(get_next_arg(i, argc, argv, arg, params));
+        } else if (arg == "-c" || arg == "--context") {
+            params.n_ctx= std::stoi(get_next_arg(i, argc, argv, arg, params));
+        } else if (arg == "-ngl" || arg == "--gpu-layers" || arg == "--n-gpu-layers") {
+            params.n_gpu_layers = std::stoi(get_next_arg(i, argc, argv, arg, params));
+        } else if (arg == "--ignore-eos") {
+            params.ignore_eos = true;
         } else if (arg == "-m" || arg == "--model") {
             params.model = get_next_arg(i, argc, argv, arg, params);
         } else if (arg == "-i" || arg == "--interactive") {
@@ -97,7 +103,6 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     fprintf(stderr, "  -h, --help            show this help message and exit\n");
     fprintf(stderr, "  -s SEED, --seed SEED  RNG seed (default: -1)\n");
     fprintf(stderr, "  -t N, --threads N     number of threads to use during computation (default: %d)\n", params.n_threads);
-    fprintf(stderr, "  -ngl N, --gpu-layers N  number of layers to offload to GPU on supported models (default: %d)\n", params.n_gpu_layers);
     fprintf(stderr, "  -p PROMPT, --prompt PROMPT\n");
     fprintf(stderr, "                        prompt to start generation with (default: random)\n");
     fprintf(stderr, "  -f FNAME, --file FNAME\n");
@@ -111,6 +116,9 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     fprintf(stderr, "  --repeat-last-n N     last n tokens to consider for penalize (default: %d, 0 = disabled)\n", params.repeat_last_n);
     fprintf(stderr, "  --repeat-penalty N    penalize repeat sequence of tokens (default: %.2f, 1.0 = disabled)\n", (double)params.repeat_penalty);
     fprintf(stderr, "  -b N, --batch_size N  batch size for prompt processing (default: %d)\n", params.n_batch);
+    fprintf(stderr, "  -c N, --context N     context / KV cache size (default: %d)\n", params.n_ctx);
+    fprintf(stderr, "  --ignore-eos          ignore EOS token during generation\n");
+    fprintf(stderr, "  -ngl N, --gpu-layers N  number of layers to offload to GPU on supported models (default: %d)\n", params.n_gpu_layers);
     fprintf(stderr, "  -m FNAME, --model FNAME\n");
     fprintf(stderr, "                        model path (default: %s)\n", params.model.c_str());
     fprintf(stderr, "\n");

examples/common.h

@@ -17,10 +17,15 @@
 //
 
 struct gpt_params {
-    int32_t seed      = -1;  // RNG seed
-    int32_t n_threads = std::min(4, (int32_t) std::thread::hardware_concurrency());
-    int32_t n_predict = 200; // new tokens to predict
-    int32_t n_batch   = 8;   // batch size for prompt processing
+    int32_t seed         = -1;   // RNG seed
+    int32_t n_threads    = std::min(4, (int32_t) std::thread::hardware_concurrency());
+    int32_t n_predict    = 200;  // new tokens to predict
+    int32_t n_parallel   = 1;    // number of parallel streams
+    int32_t n_batch      = 8;    // batch size for prompt processing
+    int32_t n_ctx        = 2048; // context size (this is the KV cache max size)
+    int32_t n_gpu_layers = 0;    // number of layers to offlload to the GPU
+
+    bool ignore_eos = false; // ignore EOS token when generating text
 
     // sampling parameters
     int32_t top_k          = 40;
@@ -35,8 +40,6 @@ struct gpt_params {
 
     bool    interactive      = false;
     int32_t interactive_port = -1;
-
-    int32_t n_gpu_layers     = 0;
 };
 
 bool gpt_params_parse(int argc, char ** argv, gpt_params & params);

examples/talk-llama/CMakeLists.txt

@@ -7,7 +7,14 @@ if (WHISPER_SDL2)
 
     # TODO: this is temporary
     #       need to export ggml symbols for MSVC, but too lazy ..
-    add_executable(${TARGET} talk-llama.cpp llama.cpp ../common.cpp ../common-sdl.cpp ../../ggml.c ../../ggml-alloc.c ../../whisper.cpp)
+    add_executable(${TARGET}
+        talk-llama.cpp
+        llama.cpp
+        ../common.cpp
+        ../common-sdl.cpp
+        ../../ggml.c
+        ../../ggml-alloc.c
+        ../../whisper.cpp)
 
     target_include_directories(${TARGET} PRIVATE ${SDL2_INCLUDE_DIRS} ../../)
     target_link_libraries(${TARGET} PRIVATE ${SDL2_LIBRARIES} ${CMAKE_THREAD_LIBS_INIT})

examples/talk-llama/llama.h

@@ -37,10 +37,12 @@
 
 #define LLAMA_DEFAULT_SEED 0xFFFFFFFF
 
+#define LLAMA_MAX_RNG_STATE (64*1024)
+
 #define LLAMA_FILE_MAGIC_GGSN 0x6767736eu // 'ggsn'
 
 #define LLAMA_SESSION_MAGIC   LLAMA_FILE_MAGIC_GGSN
-#define LLAMA_SESSION_VERSION 1
+#define LLAMA_SESSION_VERSION 2
 
 #if defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST) || defined(GGML_USE_METAL)
 // Defined when llama.cpp is compiled with support for offloading model layers to GPU.
@@ -60,13 +62,9 @@ extern "C" {
     struct llama_model;
     struct llama_context;
 
-    typedef int llama_token;
-
-    enum llama_log_level {
-        LLAMA_LOG_LEVEL_ERROR = 2,
-        LLAMA_LOG_LEVEL_WARN  = 3,
-        LLAMA_LOG_LEVEL_INFO  = 4
-    };
+    typedef int32_t llama_pos;
+    typedef int32_t llama_token;
+    typedef int32_t llama_seq_id;
 
     enum llama_vocab_type {
         LLAMA_VOCAB_TYPE_SPM = 0, // SentencePiece
@@ -86,28 +84,36 @@ extern "C" {
     // model file types
     enum llama_ftype {
         LLAMA_FTYPE_ALL_F32              = 0,
-        LLAMA_FTYPE_MOSTLY_F16           = 1, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q4_0          = 2, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q4_1          = 3, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16 = 4, // tok_embeddings.weight and output.weight are F16
-        // LLAMA_FTYPE_MOSTLY_Q4_2       = 5, // support has been removed
-        // LLAMA_FTYPE_MOSTLY_Q4_3       = 6, // support has been removed
-        LLAMA_FTYPE_MOSTLY_Q8_0          = 7, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q5_0          = 8, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q5_1          = 9, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q2_K          = 10,// except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q3_K_S        = 11,// except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q3_K_M        = 12,// except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q3_K_L        = 13,// except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q4_K_S        = 14,// except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q4_K_M        = 15,// except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q5_K_S        = 16,// except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q5_K_M        = 17,// except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q6_K          = 18,// except 1d tensors
+        LLAMA_FTYPE_MOSTLY_F16           = 1,  // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q4_0          = 2,  // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q4_1          = 3,  // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16 = 4,  // tok_embeddings.weight and output.weight are F16
+        // LLAMA_FTYPE_MOSTLY_Q4_2       = 5,  // support has been removed
+        // LLAMA_FTYPE_MOSTLY_Q4_3       = 6,  // support has been removed
+        LLAMA_FTYPE_MOSTLY_Q8_0          = 7,  // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q5_0          = 8,  // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q5_1          = 9,  // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q2_K          = 10, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q3_K_S        = 11, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q3_K_M        = 12, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q3_K_L        = 13, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q4_K_S        = 14, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q4_K_M        = 15, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q5_K_S        = 16, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q5_K_M        = 17, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q6_K          = 18, // except 1d tensors
 
         LLAMA_FTYPE_GUESSED = 1024, // not specified in the model file
     };
 
+    enum llama_rope_scaling_type {
+        LLAMA_ROPE_SCALING_UNSPECIFIED = -1,
+        LLAMA_ROPE_SCALING_NONE        = 0,
+        LLAMA_ROPE_SCALING_LINEAR      = 1,
+        LLAMA_ROPE_SCALING_YARN        = 2,
+        LLAMA_ROPE_SCALING_MAX_VALUE   = LLAMA_ROPE_SCALING_YARN,
+    };
+
     typedef struct llama_token_data {
         llama_token id; // token id
         float logit;    // log-odds of the token
@@ -122,41 +128,75 @@ extern "C" {
 
     typedef void (*llama_progress_callback)(float progress, void *ctx);
 
-    struct llama_context_params {
-        uint32_t seed;         // RNG seed, -1 for random
-        int32_t  n_ctx;        // text context
-        int32_t  n_batch;      // prompt processing batch size
-        int32_t  n_gpu_layers; // number of layers to store in VRAM
-        int32_t  main_gpu;     // the GPU that is used for scratch and small tensors
-
+    // Input data for llama_decode
+    // A llama_batch object can contain input about one or many sequences
+    // The provided arrays (i.e. token, embd, pos, etc.) must have size of n_tokens
+    //
+    // - token  : the token ids of the input (used when embd is NULL)
+    // - embd   : token embeddings (i.e. float vector of size n_embd) (used when token is NULL)
+    // - pos    : the positions of the respective token in the sequence
+    // - seq_id : the sequence to which the respective token belongs
+    // - logits : if zero, the logits for the respective token will not be output
+    //
+    typedef struct llama_batch {
+        int32_t n_tokens;
+
+        llama_token  *  token;
+        float        *  embd;
+        llama_pos    *  pos;
+        int32_t      *  n_seq_id;
+        llama_seq_id ** seq_id;
+        int8_t       *  logits;
+
+        // NOTE: helpers for smooth API transition - can be deprecated in the future
+        //       for future-proof code, use the above fields instead and ignore everything below
+        //
+        // pos[i] = all_pos_0 + i*all_pos_1
+        //
+        llama_pos    all_pos_0;  // used if pos == NULL
+        llama_pos    all_pos_1;  // used if pos == NULL
+        llama_seq_id all_seq_id; // used if seq_id == NULL
+    } llama_batch;
+
+    struct llama_model_params {
+        int32_t n_gpu_layers; // number of layers to store in VRAM
+        int32_t main_gpu;     // the GPU that is used for scratch and small tensors
         const float * tensor_split; // how to split layers across multiple GPUs (size: LLAMA_MAX_DEVICES)
 
-        // ref: https://github.com/ggerganov/llama.cpp/pull/2054
-        float    rope_freq_base;  // RoPE base frequency
-        float    rope_freq_scale; // RoPE frequency scaling factor
-
         // called with a progress value between 0 and 1, pass NULL to disable
         llama_progress_callback progress_callback;
         // context pointer passed to the progress callback
         void * progress_callback_user_data;
 
         // Keep the booleans together to avoid misalignment during copy-by-value.
-        bool low_vram;   // if true, reduce VRAM usage at the cost of performance
-        bool mul_mat_q;  // if true, use experimental mul_mat_q kernels
-        bool f16_kv;     // use fp16 for KV cache
-        bool logits_all; // the llama_eval() call computes all logits, not just the last one
         bool vocab_only; // only load the vocabulary, no weights
         bool use_mmap;   // use mmap if possible
         bool use_mlock;  // force system to keep model in RAM
-        bool embedding;  // embedding mode only
     };
 
-    // Signature for logging events
-    // Note that text includes the new line character at the end for most events.
-    // If your logging mechanism cannot handle that, check if the last character is '\n' and strip it
-    // if it exists.
-    // It might not exist for progress report where '.' is output repeatedly.
-    typedef void (*llama_log_callback)(enum llama_log_level level, const char * text, void * user_data);
+    struct llama_context_params {
+        uint32_t seed;              // RNG seed, -1 for random
+        uint32_t n_ctx;             // text context, 0 = from model
+        uint32_t n_batch;           // prompt processing maximum batch size
+        uint32_t n_threads;         // number of threads to use for generation
+        uint32_t n_threads_batch;   // number of threads to use for batch processing
+        int8_t   rope_scaling_type; // RoPE scaling type, from `enum llama_rope_scaling_type`
+
+        // ref: https://github.com/ggerganov/llama.cpp/pull/2054
+        float    rope_freq_base;   // RoPE base frequency, 0 = from model
+        float    rope_freq_scale;  // RoPE frequency scaling factor, 0 = from model
+        float    yarn_ext_factor;  // YaRN extrapolation mix factor, NaN = from model
+        float    yarn_attn_factor; // YaRN magnitude scaling factor
+        float    yarn_beta_fast;   // YaRN low correction dim
+        float    yarn_beta_slow;   // YaRN high correction dim
+        uint32_t yarn_orig_ctx;    // YaRN original context size
+
+        // Keep the booleans together to avoid misalignment during copy-by-value.
+        bool mul_mat_q;  // if true, use experimental mul_mat_q kernels (DEPRECATED - always true)
+        bool f16_kv;     // use fp16 for KV cache, fp32 otherwise
+        bool logits_all; // the llama_eval() call computes all logits, not just the last one
+        bool embedding;  // embedding mode only
+    };
 
     // model quantization parameters
     typedef struct llama_model_quantize_params {
@@ -165,6 +205,7 @@ extern "C" {
         bool allow_requantize;       // allow quantizing non-f32/f16 tensors
         bool quantize_output_tensor; // quantize output.weight
         bool only_copy;              // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored
+        bool pure;                   // disable k-quant mixtures and quantize all tensors to the same type
     } llama_model_quantize_params;
 
     // grammar types
@@ -215,6 +256,8 @@ extern "C" {
         int32_t n_eval;
     };
 
+    // Helpers for getting default parameters
+    LLAMA_API struct llama_model_params llama_model_default_params(void);
     LLAMA_API struct llama_context_params llama_context_default_params(void);
     LLAMA_API struct llama_model_quantize_params llama_model_quantize_default_params(void);
 
@@ -228,7 +271,7 @@ extern "C" {
 
     LLAMA_API struct llama_model * llama_load_model_from_file(
                              const char * path_model,
-            struct llama_context_params   params);
+            struct llama_model_params     params);
 
     LLAMA_API void llama_free_model(struct llama_model * model);
 
@@ -245,25 +288,31 @@ extern "C" {
     LLAMA_API bool llama_mmap_supported (void);
     LLAMA_API bool llama_mlock_supported(void);
 
-    LLAMA_API int llama_n_vocab    (const struct llama_context * ctx);
+    LLAMA_API const struct llama_model * llama_get_model(const struct llama_context * ctx);
+
     LLAMA_API int llama_n_ctx      (const struct llama_context * ctx);
-    LLAMA_API int llama_n_ctx_train(const struct llama_context * ctx);
-    LLAMA_API int llama_n_embd     (const struct llama_context * ctx);
 
-    LLAMA_API enum llama_vocab_type llama_vocab_type(const struct llama_context * ctx);
+    LLAMA_API enum llama_vocab_type llama_vocab_type(const struct llama_model * model);
+
+    LLAMA_API int llama_n_vocab    (const struct llama_model * model);
+    LLAMA_API int llama_n_ctx_train(const struct llama_model * model);
+    LLAMA_API int llama_n_embd     (const struct llama_model * model);
 
-    LLAMA_API int llama_model_n_vocab    (const struct llama_model * model);
-    LLAMA_API int llama_model_n_ctx      (const struct llama_model * model);
-    LLAMA_API int llama_model_n_ctx_train(const struct llama_model * model);
-    LLAMA_API int llama_model_n_embd     (const struct llama_model * model);
+    // Get the model's RoPE frequency scaling factor
+    LLAMA_API float llama_rope_freq_scale_train(const struct llama_model * model);
 
     // Get a string describing the model type
     LLAMA_API int llama_model_desc(const struct llama_model * model, char * buf, size_t buf_size);
+
     // Returns the total size of all the tensors in the model in bytes
     LLAMA_API uint64_t llama_model_size(const struct llama_model * model);
+
     // Returns the total number of parameters in the model
     LLAMA_API uint64_t llama_model_n_params(const struct llama_model * model);
 
+    // Get a llama model tensor
+    LLAMA_API struct ggml_tensor * llama_get_model_tensor(struct llama_model * model, const char * name);
+
     // Returns 0 on success
     LLAMA_API int llama_model_quantize(
             const char * fname_inp,
@@ -279,21 +328,70 @@ extern "C" {
     LLAMA_API DEPRECATED(int llama_apply_lora_from_file(
             struct llama_context * ctx,
                       const char * path_lora,
+                           float   scale,
                       const char * path_base_model,
                              int   n_threads),
-            "please use llama_model_apply_lora_from_file instead");
+            "use llama_model_apply_lora_from_file instead");
 
     LLAMA_API int llama_model_apply_lora_from_file(
             const struct llama_model * model,
-                          const char * path_lora,
-                          const char * path_base_model,
-                                 int   n_threads);
+                      const char * path_lora,
+                           float   scale,
+                      const char * path_base_model,
+                             int   n_threads);
+
+    //
+    // KV cache
+    //
 
     // Returns the number of tokens in the KV cache
-    LLAMA_API int llama_get_kv_cache_token_count(const struct llama_context * ctx);
+    LLAMA_API DEPRECATED(int llama_get_kv_cache_token_count(const struct llama_context * ctx),
+            "avoid using this, it will be removed in the future, instead - count the tokens in user code");
+
+    // Clear the KV cache
+    LLAMA_API void llama_kv_cache_clear(
+            struct llama_context * ctx);
+
+    // Removes all tokens that belong to the specified sequence and have positions in [p0, p1)
+    // seq_id < 0 : match any sequence
+    // p0 < 0     : [0,  p1]
+    // p1 < 0     : [p0, inf)
+    LLAMA_API void llama_kv_cache_seq_rm(
+            struct llama_context * ctx,
+                    llama_seq_id   seq_id,
+                       llama_pos   p0,
+                       llama_pos   p1);
+
+    // Copy all tokens that belong to the specified sequence to another sequence
+    // Note that this does not allocate extra KV cache memory - it simply assigns the tokens to the new sequence
+    // p0 < 0 : [0,  p1]
+    // p1 < 0 : [p0, inf)
+    LLAMA_API void llama_kv_cache_seq_cp(
+            struct llama_context * ctx,
+                    llama_seq_id   seq_id_src,
+                    llama_seq_id   seq_id_dst,
+                       llama_pos   p0,
+                       llama_pos   p1);
 
-    // Sets the current rng seed.
-    LLAMA_API void llama_set_rng_seed(struct llama_context * ctx, uint32_t seed);
+    // Removes all tokens that do not belong to the specified sequence
+    LLAMA_API void llama_kv_cache_seq_keep(
+            struct llama_context * ctx,
+                    llama_seq_id   seq_id);
+
+    // Adds relative position "delta" to all tokens that belong to the specified sequence and have positions in [p0, p1)
+    // If the KV cache is RoPEd, the KV data is updated accordingly
+    // p0 < 0 : [0,  p1]
+    // p1 < 0 : [p0, inf)
+    LLAMA_API void llama_kv_cache_seq_shift(
+            struct llama_context * ctx,
+                    llama_seq_id   seq_id,
+                       llama_pos   p0,
+                       llama_pos   p1,
+                       llama_pos   delta);
+
+    //
+    // State / sessions
+    //
 
     // Returns the maximum size in bytes of the state (rng, logits, embedding
     // and kv_cache) - will often be smaller after compacting tokens
@@ -302,48 +400,104 @@ extern "C" {
     // Copies the state to the specified destination address.
     // Destination needs to have allocated enough memory.
     // Returns the number of bytes copied
-    LLAMA_API size_t llama_copy_state_data(struct llama_context * ctx, uint8_t * dst);
+    LLAMA_API size_t llama_copy_state_data(
+            struct llama_context * ctx,
+                         uint8_t * dst);
 
     // Set the state reading from the specified address
     // Returns the number of bytes read
-    LLAMA_API size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src);
+    LLAMA_API size_t llama_set_state_data(
+            struct llama_context * ctx,
+                         uint8_t * src);
 
     // Save/load session file
-    LLAMA_API bool llama_load_session_file(struct llama_context * ctx, const char * path_session, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out);
-    LLAMA_API bool llama_save_session_file(struct llama_context * ctx, const char * path_session, const llama_token * tokens, size_t n_token_count);
+    LLAMA_API bool llama_load_session_file(
+            struct llama_context * ctx,
+                      const char * path_session,
+                     llama_token * tokens_out,
+                          size_t   n_token_capacity,
+                          size_t * n_token_count_out);
 
-    // Run the llama inference to obtain the logits and probabilities for the next token.
+    LLAMA_API bool llama_save_session_file(
+            struct llama_context * ctx,
+                      const char * path_session,
+               const llama_token * tokens,
+                          size_t   n_token_count);
+
+    //
+    // Decoding
+    //
+
+    // Run the llama inference to obtain the logits and probabilities for the next token(s).
     // tokens + n_tokens is the provided batch of new tokens to process
     // n_past is the number of tokens to use from previous eval calls
     // Returns 0 on success
-    LLAMA_API int llama_eval(
+    // DEPRECATED: use llama_decode() instead
+    LLAMA_API DEPRECATED(int llama_eval(
             struct llama_context * ctx,
-               const llama_token * tokens,
-                             int   n_tokens,
-                             int   n_past,
-                             int   n_threads);
+                     llama_token * tokens,
+                         int32_t   n_tokens,
+                             int   n_past),
+            "use llama_decode() instead");
 
     // Same as llama_eval, but use float matrix input directly.
-    LLAMA_API int llama_eval_embd(
+    // DEPRECATED: use llama_decode() instead
+    LLAMA_API DEPRECATED(int llama_eval_embd(
             struct llama_context * ctx,
-                     const float * embd,
-                             int   n_tokens,
-                             int   n_past,
-                             int   n_threads);
+                           float * embd,
+                         int32_t   n_tokens,
+                             int   n_past),
+            "use llama_decode() instead");
+
+    // Return batch for single sequence of tokens starting at pos_0
+    //
+    // NOTE: this is a helper function to facilitate transition to the new batch API - avoid using it
+    //
+    LLAMA_API struct llama_batch llama_batch_get_one(
+                  llama_token * tokens,
+                      int32_t   n_tokens,
+                    llama_pos   pos_0,
+                 llama_seq_id   seq_id);
+
+    // Allocates a batch of tokens on the heap that can hold a maximum of n_tokens
+    // Each token can be assigned up to n_seq_max sequence ids
+    // The batch has to be freed with llama_batch_free()
+    // If embd != 0, llama_batch.embd will be allocated with size of n_tokens * embd * sizeof(float)
+    // Otherwise, llama_batch.token will be allocated to store n_tokens llama_token
+    // The rest of the llama_batch members are allocated with size n_tokens
+    // All members are left uninitialized
+    LLAMA_API struct llama_batch llama_batch_init(
+            int32_t n_tokens,
+            int32_t embd,
+            int32_t n_seq_max);
+
+    // Frees a batch of tokens allocated with llama_batch_init()
+    LLAMA_API void llama_batch_free(struct llama_batch batch);
+
+    // Positive return values does not mean a fatal error, but rather a warning.
+    //   0 - success
+    //   1 - could not find a KV slot for the batch (try reducing the size of the batch or increase the context)
+    // < 0 - error
+    LLAMA_API int llama_decode(
+            struct llama_context * ctx,
+              struct llama_batch   batch);
 
-    // Export a static computation graph for context of 511 and batch size of 1
-    // NOTE: since this functionality is mostly for debugging and demonstration purposes, we hardcode these
-    //       parameters here to keep things simple
-    // IMPORTANT: do not use for anything else other than debugging and testing!
-    LLAMA_API int llama_eval_export(struct llama_context * ctx, const char * fname);
+    // Set the number of threads used for decoding
+    // n_threads is the number of threads used for generation (single token)
+    // n_threads_batch is the number of threads used for prompt and batch processing (multiple tokens)
+    LLAMA_API void llama_set_n_threads(struct llama_context * ctx, uint32_t n_threads, uint32_t n_threads_batch);
 
     // Token logits obtained from the last call to llama_eval()
     // The logits for the last token are stored in the last row
-    // Can be mutated in order to change the probabilities of the next token
-    // Rows: n_tokens
+    // Logits for which llama_batch.logits[i] == 0 are undefined
+    // Rows: n_tokens provided with llama_batch
     // Cols: n_vocab
     LLAMA_API float * llama_get_logits(struct llama_context * ctx);
 
+    // Logits for the ith token. Equivalent to:
+    // llama_get_logits(ctx) + i*n_vocab
+    LLAMA_API float * llama_get_logits_ith(struct llama_context * ctx, int32_t i);
+
     // Get the embeddings for the input
     // shape: [n_embd] (1-dimensional)
     LLAMA_API float * llama_get_embeddings(struct llama_context * ctx);
@@ -352,50 +506,47 @@ extern "C" {
     // Vocab
     //
 
-    LLAMA_API const char * llama_token_get_text(const struct llama_context * ctx, llama_token token);
+    LLAMA_API const char * llama_token_get_text(const struct llama_model * model, llama_token token);
 
-    LLAMA_API float llama_token_get_score(const struct llama_context * ctx, llama_token token);
+    LLAMA_API float llama_token_get_score(const struct llama_model * model, llama_token token);
 
-    LLAMA_API enum llama_token_type llama_token_get_type(const struct llama_context * ctx, llama_token token);
+    LLAMA_API enum llama_token_type llama_token_get_type(const struct llama_model * model, llama_token token);
 
     // Special tokens
-    LLAMA_API llama_token llama_token_bos(const struct llama_context * ctx);  // beginning-of-sentence
-    LLAMA_API llama_token llama_token_eos(const struct llama_context * ctx);  // end-of-sentence
-    LLAMA_API llama_token llama_token_nl (const struct llama_context * ctx);  // next-line
+    LLAMA_API llama_token llama_token_bos(const struct llama_model * model); // beginning-of-sentence
+    LLAMA_API llama_token llama_token_eos(const struct llama_model * model); // end-of-sentence
+    LLAMA_API llama_token llama_token_nl (const struct llama_model * model); // next-line
+
+    // codellama infill tokens
+    LLAMA_API llama_token llama_token_prefix(const struct llama_model * model); // Beginning of infill prefix
+    LLAMA_API llama_token llama_token_middle(const struct llama_model * model); // Beginning of infill middle
+    LLAMA_API llama_token llama_token_suffix(const struct llama_model * model); // Beginning of infill suffix
+    LLAMA_API llama_token llama_token_eot   (const struct llama_model * model); // End of infill middle
 
     //
     // Tokenization
     //
 
-    // Convert the provided text into tokens.
-    // The tokens pointer must be large enough to hold the resulting tokens.
-    // Returns the number of tokens on success, no more than n_max_tokens
-    // Returns a negative number on failure - the number of tokens that would have been returned
+    /// @details Convert the provided text into tokens.
+    /// @param tokens The tokens pointer must be large enough to hold the resulting tokens.
+    /// @return Returns the number of tokens on success, no more than n_max_tokens
+    /// @return Returns a negative number on failure - the number of tokens that would have been returned
+    /// @param special Allow tokenizing special and/or control tokens which otherwise are not exposed and treated as plaintext.
+    ///                Does not insert a leading space.
     LLAMA_API int llama_tokenize(
-            struct llama_context * ctx,
-                      const char * text,
-                     llama_token * tokens,
-                             int   n_max_tokens,
-                            bool   add_bos);
-
-    LLAMA_API int llama_tokenize_with_model(
         const struct llama_model * model,
                       const char * text,
+                             int   text_len,
                      llama_token * tokens,
                              int   n_max_tokens,
-                            bool   add_bos);
+                            bool   add_bos,
+                            bool   special);
 
     // Token Id -> Piece.
     // Uses the vocabulary in the provided context.
     // Does not write null terminator to the buffer.
     // User code is responsible to remove the leading whitespace of the first non-BOS token when decoding multiple tokens.
     LLAMA_API int llama_token_to_piece(
-            const struct llama_context * ctx,
-                           llama_token   token,
-                                  char * buf,
-                                  int    length);
-
-    LLAMA_API int llama_token_to_piece_with_model(
               const struct llama_model * model,
                            llama_token   token,
                                   char * buf,
@@ -418,11 +569,19 @@ extern "C" {
     // Sampling functions
     //
 
-    /// @details Repetition penalty described in CTRL academic paper https://arxiv.org/abs/1909.05858, with negative logit fix.
-    LLAMA_API void llama_sample_repetition_penalty(struct llama_context * ctx, llama_token_data_array * candidates, const llama_token * last_tokens, size_t last_tokens_size, float penalty);
+    // Sets the current rng seed.
+    LLAMA_API void llama_set_rng_seed(struct llama_context * ctx, uint32_t seed);
 
+    /// @details Repetition penalty described in CTRL academic paper https://arxiv.org/abs/1909.05858, with negative logit fix.
     /// @details Frequency and presence penalties described in OpenAI API https://platform.openai.com/docs/api-reference/parameter-details.
-    LLAMA_API void llama_sample_frequency_and_presence_penalties(struct llama_context * ctx, llama_token_data_array * candidates, const llama_token * last_tokens, size_t last_tokens_size, float alpha_frequency, float alpha_presence);
+    LLAMA_API void llama_sample_repetition_penalties(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates,
+               const llama_token * last_tokens,
+                          size_t   penalty_last_n,
+                           float   penalty_repeat,
+                           float   penalty_freq,
+                           float   penalty_present);
 
     /// @details Apply classifier-free guidance to the logits as described in academic paper "Stay on topic with Classifier-Free Guidance" https://arxiv.org/abs/2306.17806
     /// @param candidates A vector of `llama_token_data` containing the candidate tokens, the logits must be directly extracted from the original generation context without being sorted.
@@ -435,23 +594,61 @@ extern "C" {
                              float   scale);
 
     /// @details Sorts candidate tokens by their logits in descending order and calculate probabilities based on logits.
-    LLAMA_API void llama_sample_softmax(struct llama_context * ctx, llama_token_data_array * candidates);
+    LLAMA_API void llama_sample_softmax(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates);
 
     /// @details Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
-    LLAMA_API void llama_sample_top_k(struct llama_context * ctx, llama_token_data_array * candidates, int k, size_t min_keep);
+    LLAMA_API void llama_sample_top_k(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates,
+                             int   k,
+                          size_t   min_keep);
 
     /// @details Nucleus sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
-    LLAMA_API void llama_sample_top_p(struct llama_context * ctx, llama_token_data_array * candidates, float p, size_t min_keep);
+    LLAMA_API void llama_sample_top_p(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates,
+                           float   p,
+                          size_t   min_keep);
+
+    /// @details Minimum P sampling as described in https://github.com/ggerganov/llama.cpp/pull/3841
+    LLAMA_API void llama_sample_min_p(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates,
+                           float   p,
+                          size_t   min_keep);
 
     /// @details Tail Free Sampling described in https://www.trentonbricken.com/Tail-Free-Sampling/.
-    LLAMA_API void llama_sample_tail_free(struct llama_context * ctx, llama_token_data_array * candidates, float z, size_t min_keep);
+    LLAMA_API void llama_sample_tail_free(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates,
+                           float   z,
+                          size_t   min_keep);
 
     /// @details Locally Typical Sampling implementation described in the paper https://arxiv.org/abs/2202.00666.
-    LLAMA_API void llama_sample_typical(struct llama_context * ctx, llama_token_data_array * candidates, float p, size_t min_keep);
-    LLAMA_API void llama_sample_temperature(struct llama_context * ctx, llama_token_data_array * candidates, float temp);
+    LLAMA_API void llama_sample_typical(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates,
+                           float   p,
+                          size_t   min_keep);
+
+    LLAMA_API void llama_sample_temp(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates,
+                           float   temp);
+
+    LLAMA_API DEPRECATED(void llama_sample_temperature(
+                struct llama_context * ctx,
+              llama_token_data_array * candidates,
+                               float   temp),
+            "use llama_sample_temp instead");
 
     /// @details Apply constraints from grammar
-    LLAMA_API void llama_sample_grammar(struct llama_context * ctx, llama_token_data_array * candidates, const struct llama_grammar * grammar);
+    LLAMA_API void llama_sample_grammar(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates,
+      const struct llama_grammar * grammar);
 
     /// @details Mirostat 1.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words.
     /// @param candidates A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text.
@@ -459,23 +656,42 @@ extern "C" {
     /// @param eta The learning rate used to update `mu` based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause `mu` to be updated more quickly, while a smaller learning rate will result in slower updates.
     /// @param m The number of tokens considered in the estimation of `s_hat`. This is an arbitrary value that is used to calculate `s_hat`, which in turn helps to calculate the value of `k`. In the paper, they use `m = 100`, but you can experiment with different values to see how it affects the performance of the algorithm.
     /// @param mu Maximum cross-entropy. This value is initialized to be twice the target cross-entropy (`2 * tau`) and is updated in the algorithm based on the error between the target and observed surprisal.
-    LLAMA_API llama_token llama_sample_token_mirostat(struct llama_context * ctx, llama_token_data_array * candidates, float tau, float eta, int m, float * mu);
+    LLAMA_API llama_token llama_sample_token_mirostat(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates,
+                           float   tau,
+                           float   eta,
+                             int   m,
+                           float * mu);
 
     /// @details Mirostat 2.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words.
     /// @param candidates A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text.
     /// @param tau  The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text.
     /// @param eta The learning rate used to update `mu` based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause `mu` to be updated more quickly, while a smaller learning rate will result in slower updates.
     /// @param mu Maximum cross-entropy. This value is initialized to be twice the target cross-entropy (`2 * tau`) and is updated in the algorithm based on the error between the target and observed surprisal.
-    LLAMA_API llama_token llama_sample_token_mirostat_v2(struct llama_context * ctx, llama_token_data_array * candidates, float tau, float eta, float * mu);
+    LLAMA_API llama_token llama_sample_token_mirostat_v2(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates,
+                           float   tau,
+                           float   eta,
+                           float * mu);
 
     /// @details Selects the token with the highest probability.
-    LLAMA_API llama_token llama_sample_token_greedy(struct llama_context * ctx, llama_token_data_array * candidates);
+    ///          Does not compute the token probabilities. Use llama_sample_softmax() instead.
+    LLAMA_API llama_token llama_sample_token_greedy(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates);
 
     /// @details Randomly selects a token from the candidates based on their probabilities.
-    LLAMA_API llama_token llama_sample_token(struct llama_context * ctx, llama_token_data_array * candidates);
+    LLAMA_API llama_token llama_sample_token(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates);
 
     /// @details Accepts the sampled token into the grammar
-    LLAMA_API void llama_grammar_accept_token(struct llama_context * ctx, struct llama_grammar * grammar, llama_token token);
+    LLAMA_API void llama_grammar_accept_token(
+            struct llama_context * ctx,
+            struct llama_grammar * grammar,
+                     llama_token   token);
 
     //
     // Beam search
@@ -483,9 +699,10 @@ extern "C" {
 
     struct llama_beam_view {
         const llama_token * tokens;
+
         size_t n_tokens;
-        float p;   // Cumulative beam probability (renormalized relative to all beams)
-        bool eob;  // Callback should set this to true when a beam is at end-of-beam.
+        float  p;        // Cumulative beam probability (renormalized relative to all beams)
+        bool   eob;      // Callback should set this to true when a beam is at end-of-beam.
     };
 
     // Passed to beam_search_callback function.
@@ -494,9 +711,10 @@ extern "C" {
     // These pointers are valid only during the synchronous callback, so should not be saved.
     struct llama_beams_state {
         struct llama_beam_view * beam_views;
+
         size_t n_beams;               // Number of elements in beam_views[].
         size_t common_prefix_length;  // Current max length of prefix tokens shared by all beams.
-        bool last_call;               // True iff this is the last callback invocation.
+        bool   last_call;             // True iff this is the last callback invocation.
     };
 
     // Type of pointer to the beam_search_callback function.
@@ -511,11 +729,17 @@ extern "C" {
     /// @param n_beams Number of beams to use.
     /// @param n_past Number of tokens already evaluated.
     /// @param n_predict Maximum number of tokens to predict. EOS may occur earlier.
-    /// @param n_threads Number of threads as passed to llama_eval().
-    LLAMA_API void llama_beam_search(struct llama_context * ctx, llama_beam_search_callback_fn_t callback, void * callback_data, size_t n_beams, int n_past, int n_predict, int n_threads);
+    LLAMA_API void llama_beam_search(
+                   struct llama_context * ctx,
+        llama_beam_search_callback_fn_t   callback,
+                                   void * callback_data,
+                                 size_t   n_beams,
+                                    int   n_past,
+                                    int   n_predict);
 
     // Performance information
     LLAMA_API struct llama_timings llama_get_timings(struct llama_context * ctx);
+
     LLAMA_API void llama_print_timings(struct llama_context * ctx);
     LLAMA_API void llama_reset_timings(struct llama_context * ctx);
 
@@ -524,7 +748,7 @@ extern "C" {
 
     // Set callback for all future logging events.
     // If this is not called, or NULL is supplied, everything is output on stderr.
-    LLAMA_API void llama_log_set(llama_log_callback log_callback, void * user_data);
+    LLAMA_API void llama_log_set(ggml_log_callback log_callback, void * user_data);
 
     LLAMA_API void llama_dump_timing_info_yaml(FILE * stream, const struct llama_context * ctx);
 
@@ -540,7 +764,9 @@ extern "C" {
 
 struct ggml_tensor;
 
-const std::vector<std::pair<std::string, struct ggml_tensor *>>& llama_internal_get_tensor_map(struct llama_context * ctx);
+const std::vector<std::pair<std::string, struct ggml_tensor *>> & llama_internal_get_tensor_map(
+    struct llama_context * ctx
+);
 
 #endif // LLAMA_API_INTERNAL
 

examples/talk-llama/talk-llama.cpp

@@ -16,21 +16,28 @@
 #include <regex>
 
 std::vector<llama_token> llama_tokenize(struct llama_context * ctx, const std::string & text, bool add_bos) {
-    // initialize to prompt numer of chars, since n_tokens <= n_prompt_chars
-    std::vector<llama_token> res(text.size() + (int)add_bos);
-    int n = llama_tokenize(ctx, text.c_str(), res.data(), res.size(), add_bos);
-    assert(n >= 0);
-    res.resize(n);
+    auto * model = llama_get_model(ctx);
 
-    return res;
+    // upper limit for the number of tokens
+    int n_tokens = text.length() + add_bos;
+    std::vector<llama_token> result(n_tokens);
+    n_tokens = llama_tokenize(model, text.data(), text.length(), result.data(), result.size(), add_bos, false);
+    if (n_tokens < 0) {
+        result.resize(-n_tokens);
+        int check = llama_tokenize(model, text.data(), text.length(), result.data(), result.size(), add_bos, false);
+        GGML_ASSERT(check == -n_tokens);
+    } else {
+        result.resize(n_tokens);
+    }
+    return result;
 }
 
 std::string llama_token_to_piece(const struct llama_context * ctx, llama_token token) {
     std::vector<char> result(8, 0);
-    const int n_tokens = llama_token_to_piece(ctx, token, result.data(), result.size());
+    const int n_tokens = llama_token_to_piece(llama_get_model(ctx), token, result.data(), result.size());
     if (n_tokens < 0) {
         result.resize(-n_tokens);
-        int check = llama_token_to_piece(ctx, token, result.data(), result.size());
+        int check = llama_token_to_piece(llama_get_model(ctx), token, result.data(), result.size());
         GGML_ASSERT(check == -n_tokens);
     } else {
         result.resize(n_tokens);
@@ -251,16 +258,19 @@ int main(int argc, char ** argv) {
 
     llama_backend_init(true);
 
-    auto lparams = llama_context_default_params();
+    auto lmparams = llama_model_default_params();
 
-    // tune these to your liking
-    lparams.n_ctx      = 2048;
-    lparams.seed       = 1;
-    lparams.f16_kv     = true;
+    struct llama_model * model_llama = llama_load_model_from_file(params.model_llama.c_str(), lmparams);
 
-    struct llama_model * model_llama = llama_load_model_from_file(params.model_llama.c_str(), lparams);
+    llama_context_params lcparams = llama_context_default_params();
+
+    // tune these to your liking
+    lcparams.n_ctx      = 2048;
+    lcparams.seed       = 1;
+    lcparams.f16_kv     = true;
+    lcparams.n_threads  = params.n_threads;
 
-    struct llama_context * ctx_llama = llama_new_context_with_model(model_llama, lparams);
+    struct llama_context * ctx_llama = llama_new_context_with_model(model_llama, lcparams);
 
     // print some info about the processing
     {
@@ -356,7 +366,7 @@ int main(int argc, char ** argv) {
         if (fp != NULL) {
             std::fclose(fp);
 
-            session_tokens.resize(lparams.n_ctx);
+            session_tokens.resize(llama_n_ctx(ctx_llama));
             size_t n_token_count_out = 0;
             if (!llama_load_session_file(ctx_llama, path_session.c_str(), session_tokens.data(), session_tokens.capacity(), &n_token_count_out)) {
                 fprintf(stderr, "%s: error: failed to load session file '%s'\n", __func__, path_session.c_str());
@@ -378,7 +388,7 @@ int main(int argc, char ** argv) {
     printf("\n");
     printf("%s : initializing - please wait ...\n", __func__);
 
-    if (llama_eval(ctx_llama, embd_inp.data(), embd_inp.size(), 0, params.n_threads)) {
+    if (llama_eval(ctx_llama, embd_inp.data(), embd_inp.size(), 0)) {
         fprintf(stderr, "%s : failed to eval\n", __func__);
         return 1;
     }
@@ -561,7 +571,7 @@ int main(int argc, char ** argv) {
                             n_session_consumed = session_tokens.size();
                         }
 
-                        if (llama_eval(ctx_llama, embd.data(), embd.size(), n_past, params.n_threads)) {
+                        if (llama_eval(ctx_llama, embd.data(), embd.size(), n_past)) {
                             fprintf(stderr, "%s : failed to eval\n", __func__);
                             return 1;
                         }
@@ -593,9 +603,9 @@ int main(int argc, char ** argv) {
 
                         {
                             auto logits = llama_get_logits(ctx_llama);
-                            auto n_vocab = llama_n_vocab(ctx_llama);
+                            auto n_vocab = llama_n_vocab(model_llama);
 
-                            logits[llama_token_eos(ctx_llama)] = 0;
+                            logits[llama_token_eos(model_llama)] = 0;
 
                             std::vector<llama_token_data> candidates;
                             candidates.reserve(n_vocab);
@@ -606,13 +616,13 @@ int main(int argc, char ** argv) {
                             llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
 
                             // apply repeat penalty
-                            const float nl_logit = logits[llama_token_nl(ctx_llama)];
+                            const float nl_logit = logits[llama_token_nl(model_llama)];
 
-                            llama_sample_repetition_penalty(ctx_llama, &candidates_p,
+                            llama_sample_repetition_penalties(ctx_llama, &candidates_p,
                                     embd_inp.data() + std::max(0, n_past - repeat_last_n),
-                                    repeat_last_n, repeat_penalty);
+                                    repeat_last_n, repeat_penalty, 0.0, 0.0f);
 
-                            logits[llama_token_nl(ctx_llama)] = nl_logit;
+                            logits[llama_token_nl(model_llama)] = nl_logit;
 
                             if (temp <= 0) {
                                 // Greedy sampling
@@ -621,12 +631,12 @@ int main(int argc, char ** argv) {
                                 // Temperature sampling
                                 llama_sample_top_k(ctx_llama, &candidates_p, top_k, 1);
                                 llama_sample_top_p(ctx_llama, &candidates_p, top_p, 1);
-                                llama_sample_temperature(ctx_llama, &candidates_p, temp);
+                                llama_sample_temp (ctx_llama, &candidates_p, temp);
                                 id = llama_sample_token(ctx_llama, &candidates_p);
                             }
                         }
 
-                        if (id != llama_token_eos(ctx_llama)) {
+                        if (id != llama_token_eos(model_llama)) {
                             // add it to the context
                             embd.push_back(id);
 

examples/talk-llama/unicode.h

@@ -0,0 +1,462 @@
+#pragma once
+
+#include <cassert>
+#include <stdexcept>
+#include <vector>
+#include <unordered_map>
+
+static const std::vector<std::pair<uint32_t, uint32_t>> digit_ranges = {
+{0x30, 0x39}, {0xB2, 0xB3}, {0xB9, 0xB9}, {0x660, 0x669}, {0x6F0, 0x6F9}, {0x7C0, 0x7C9}, {0x966, 0x96F}, {0x9E6, 0x9EF}, {0xA66, 0xA6F}, {0xAE6, 0xAEF}, {0xB66, 0xB6F}, {0xBE6, 0xBEF}, {0xC66, 0xC6F},
+{0xCE6, 0xCEF}, {0xD66, 0xD6F}, {0xDE6, 0xDEF}, {0xE50, 0xE59}, {0xED0, 0xED9}, {0xF20, 0xF29}, {0x1040, 0x1049}, {0x1090, 0x1099}, {0x1369, 0x1371}, {0x17E0, 0x17E9}, {0x1810, 0x1819}, {0x1946, 0x194F},
+{0x19D0, 0x19DA}, {0x1A80, 0x1A89}, {0x1A90, 0x1A99}, {0x1B50, 0x1B59}, {0x1BB0, 0x1BB9}, {0x1C40, 0x1C49}, {0x1C50, 0x1C59}, {0x2070, 0x2070}, {0x2074, 0x2079}, {0x2080, 0x2089}, {0x2460, 0x2468},
+{0x2474, 0x247C}, {0x2488, 0x2490}, {0x24EA, 0x24EA}, {0x24F5, 0x24FD}, {0x24FF, 0x24FF}, {0x2776, 0x277E}, {0x2780, 0x2788}, {0x278A, 0x2792}, {0xA620, 0xA629}, {0xA8D0, 0xA8D9}, {0xA900, 0xA909},
+{0xA9D0, 0xA9D9}, {0xA9F0, 0xA9F9}, {0xAA50, 0xAA59}, {0xABF0, 0xABF9}, {0xFF10, 0xFF19}, {0x104A0, 0x104A9}, {0x10A40, 0x10A43}, {0x10D30, 0x10D39}, {0x10E60, 0x10E68}, {0x11052, 0x1105A},
+{0x11066, 0x1106F}, {0x110F0, 0x110F9}, {0x11136, 0x1113F}, {0x111D0, 0x111D9}, {0x112F0, 0x112F9}, {0x11450, 0x11459}, {0x114D0, 0x114D9}, {0x11650, 0x11659}, {0x116C0, 0x116C9}, {0x11730, 0x11739},
+{0x118E0, 0x118E9}, {0x11950, 0x11959}, {0x11C50, 0x11C59}, {0x11D50, 0x11D59}, {0x11DA0, 0x11DA9}, {0x16A60, 0x16A69}, {0x16B50, 0x16B59}, {0x1D7CE, 0x1D7FF}, {0x1E140, 0x1E149}, {0x1E2F0, 0x1E2F9},
+{0x1E950, 0x1E959}, {0x1F100, 0x1F10A}, {0x1FBF0, 0x1FBF9},
+};
+
+static const std::vector<std::pair<uint32_t, uint32_t>> letter_ranges = {
+{0x41, 0x5A}, {0x61, 0x7A}, {0xAA, 0xAA}, {0xB5, 0xB5}, {0xBA, 0xBA}, {0xC0, 0xD6}, {0xD8, 0xF6}, {0xF8, 0x2C1}, {0x2C6, 0x2D1}, {0x2E0, 0x2E4}, {0x2EC, 0x2EC}, {0x2EE, 0x2EE}, {0x370, 0x374},
+{0x376, 0x377}, {0x37A, 0x37D}, {0x37F, 0x37F}, {0x386, 0x386}, {0x388, 0x38A}, {0x38C, 0x38C}, {0x38E, 0x3A1}, {0x3A3, 0x3F5}, {0x3F7, 0x481}, {0x48A, 0x52F}, {0x531, 0x556}, {0x559, 0x559},
+{0x560, 0x588}, {0x5D0, 0x5EA}, {0x5EF, 0x5F2}, {0x620, 0x64A}, {0x66E, 0x66F}, {0x671, 0x6D3}, {0x6D5, 0x6D5}, {0x6E5, 0x6E6}, {0x6EE, 0x6EF}, {0x6FA, 0x6FC}, {0x6FF, 0x6FF}, {0x710, 0x710},
+{0x712, 0x72F}, {0x74D, 0x7A5}, {0x7B1, 0x7B1}, {0x7CA, 0x7EA}, {0x7F4, 0x7F5}, {0x7FA, 0x7FA}, {0x800, 0x815}, {0x81A, 0x81A}, {0x824, 0x824}, {0x828, 0x828}, {0x840, 0x858}, {0x860, 0x86A},
+{0x8A0, 0x8B4}, {0x8B6, 0x8C7}, {0x904, 0x939}, {0x93D, 0x93D}, {0x950, 0x950}, {0x958, 0x961}, {0x971, 0x980}, {0x985, 0x98C}, {0x98F, 0x990}, {0x993, 0x9A8}, {0x9AA, 0x9B0}, {0x9B2, 0x9B2},
+{0x9B6, 0x9B9}, {0x9BD, 0x9BD}, {0x9CE, 0x9CE}, {0x9DC, 0x9DD}, {0x9DF, 0x9E1}, {0x9F0, 0x9F1}, {0x9FC, 0x9FC}, {0xA05, 0xA0A}, {0xA0F, 0xA10}, {0xA13, 0xA28}, {0xA2A, 0xA30}, {0xA32, 0xA33},
+{0xA35, 0xA36}, {0xA38, 0xA39}, {0xA59, 0xA5C}, {0xA5E, 0xA5E}, {0xA72, 0xA74}, {0xA85, 0xA8D}, {0xA8F, 0xA91}, {0xA93, 0xAA8}, {0xAAA, 0xAB0}, {0xAB2, 0xAB3}, {0xAB5, 0xAB9}, {0xABD, 0xABD},
+{0xAD0, 0xAD0}, {0xAE0, 0xAE1}, {0xAF9, 0xAF9}, {0xB05, 0xB0C}, {0xB0F, 0xB10}, {0xB13, 0xB28}, {0xB2A, 0xB30}, {0xB32, 0xB33}, {0xB35, 0xB39}, {0xB3D, 0xB3D}, {0xB5C, 0xB5D}, {0xB5F, 0xB61},
+{0xB71, 0xB71}, {0xB83, 0xB83}, {0xB85, 0xB8A}, {0xB8E, 0xB90}, {0xB92, 0xB95}, {0xB99, 0xB9A}, {0xB9C, 0xB9C}, {0xB9E, 0xB9F}, {0xBA3, 0xBA4}, {0xBA8, 0xBAA}, {0xBAE, 0xBB9}, {0xBD0, 0xBD0},
+{0xC05, 0xC0C}, {0xC0E, 0xC10}, {0xC12, 0xC28}, {0xC2A, 0xC39}, {0xC3D, 0xC3D}, {0xC58, 0xC5A}, {0xC60, 0xC61}, {0xC80, 0xC80}, {0xC85, 0xC8C}, {0xC8E, 0xC90}, {0xC92, 0xCA8}, {0xCAA, 0xCB3},
+{0xCB5, 0xCB9}, {0xCBD, 0xCBD}, {0xCDE, 0xCDE}, {0xCE0, 0xCE1}, {0xCF1, 0xCF2}, {0xD04, 0xD0C}, {0xD0E, 0xD10}, {0xD12, 0xD3A}, {0xD3D, 0xD3D}, {0xD4E, 0xD4E}, {0xD54, 0xD56}, {0xD5F, 0xD61},
+{0xD7A, 0xD7F}, {0xD85, 0xD96}, {0xD9A, 0xDB1}, {0xDB3, 0xDBB}, {0xDBD, 0xDBD}, {0xDC0, 0xDC6}, {0xE01, 0xE30}, {0xE32, 0xE33}, {0xE40, 0xE46}, {0xE81, 0xE82}, {0xE84, 0xE84}, {0xE86, 0xE8A},
+{0xE8C, 0xEA3}, {0xEA5, 0xEA5}, {0xEA7, 0xEB0}, {0xEB2, 0xEB3}, {0xEBD, 0xEBD}, {0xEC0, 0xEC4}, {0xEC6, 0xEC6}, {0xEDC, 0xEDF}, {0xF00, 0xF00}, {0xF40, 0xF47}, {0xF49, 0xF6C}, {0xF88, 0xF8C},
+{0x1000, 0x102A}, {0x103F, 0x103F}, {0x1050, 0x1055}, {0x105A, 0x105D}, {0x1061, 0x1061}, {0x1065, 0x1066}, {0x106E, 0x1070}, {0x1075, 0x1081}, {0x108E, 0x108E}, {0x10A0, 0x10C5}, {0x10C7, 0x10C7},
+{0x10CD, 0x10CD}, {0x10D0, 0x10FA}, {0x10FC, 0x1248}, {0x124A, 0x124D}, {0x1250, 0x1256}, {0x1258, 0x1258}, {0x125A, 0x125D}, {0x1260, 0x1288}, {0x128A, 0x128D}, {0x1290, 0x12B0}, {0x12B2, 0x12B5},
+{0x12B8, 0x12BE}, {0x12C0, 0x12C0}, {0x12C2, 0x12C5}, {0x12C8, 0x12D6}, {0x12D8, 0x1310}, {0x1312, 0x1315}, {0x1318, 0x135A}, {0x1380, 0x138F}, {0x13A0, 0x13F5}, {0x13F8, 0x13FD}, {0x1401, 0x166C},
+{0x166F, 0x167F}, {0x1681, 0x169A}, {0x16A0, 0x16EA}, {0x16F1, 0x16F8}, {0x1700, 0x170C}, {0x170E, 0x1711}, {0x1720, 0x1731}, {0x1740, 0x1751}, {0x1760, 0x176C}, {0x176E, 0x1770}, {0x1780, 0x17B3},
+{0x17D7, 0x17D7}, {0x17DC, 0x17DC}, {0x1820, 0x1878}, {0x1880, 0x1884}, {0x1887, 0x18A8}, {0x18AA, 0x18AA}, {0x18B0, 0x18F5}, {0x1900, 0x191E}, {0x1950, 0x196D}, {0x1970, 0x1974}, {0x1980, 0x19AB},
+{0x19B0, 0x19C9}, {0x1A00, 0x1A16}, {0x1A20, 0x1A54}, {0x1AA7, 0x1AA7}, {0x1B05, 0x1B33}, {0x1B45, 0x1B4B}, {0x1B83, 0x1BA0}, {0x1BAE, 0x1BAF}, {0x1BBA, 0x1BE5}, {0x1C00, 0x1C23}, {0x1C4D, 0x1C4F},
+{0x1C5A, 0x1C7D}, {0x1C80, 0x1C88}, {0x1C90, 0x1CBA}, {0x1CBD, 0x1CBF}, {0x1CE9, 0x1CEC}, {0x1CEE, 0x1CF3}, {0x1CF5, 0x1CF6}, {0x1CFA, 0x1CFA}, {0x1D00, 0x1DBF}, {0x1E00, 0x1F15}, {0x1F18, 0x1F1D},
+{0x1F20, 0x1F45}, {0x1F48, 0x1F4D}, {0x1F50, 0x1F57}, {0x1F59, 0x1F59}, {0x1F5B, 0x1F5B}, {0x1F5D, 0x1F5D}, {0x1F5F, 0x1F7D}, {0x1F80, 0x1FB4}, {0x1FB6, 0x1FBC}, {0x1FBE, 0x1FBE}, {0x1FC2, 0x1FC4},
+{0x1FC6, 0x1FCC}, {0x1FD0, 0x1FD3}, {0x1FD6, 0x1FDB}, {0x1FE0, 0x1FEC}, {0x1FF2, 0x1FF4}, {0x1FF6, 0x1FFC}, {0x2071, 0x2071}, {0x207F, 0x207F}, {0x2090, 0x209C}, {0x2102, 0x2102}, {0x2107, 0x2107},
+{0x210A, 0x2113}, {0x2115, 0x2115}, {0x2119, 0x211D}, {0x2124, 0x2124}, {0x2126, 0x2126}, {0x2128, 0x2128}, {0x212A, 0x212D}, {0x212F, 0x2139}, {0x213C, 0x213F}, {0x2145, 0x2149}, {0x214E, 0x214E},
+{0x2183, 0x2184}, {0x2C00, 0x2C2E}, {0x2C30, 0x2C5E}, {0x2C60, 0x2CE4}, {0x2CEB, 0x2CEE}, {0x2CF2, 0x2CF3}, {0x2D00, 0x2D25}, {0x2D27, 0x2D27}, {0x2D2D, 0x2D2D}, {0x2D30, 0x2D67}, {0x2D6F, 0x2D6F},
+{0x2D80, 0x2D96}, {0x2DA0, 0x2DA6}, {0x2DA8, 0x2DAE}, {0x2DB0, 0x2DB6}, {0x2DB8, 0x2DBE}, {0x2DC0, 0x2DC6}, {0x2DC8, 0x2DCE}, {0x2DD0, 0x2DD6}, {0x2DD8, 0x2DDE}, {0x2E2F, 0x2E2F}, {0x3005, 0x3006},
+{0x3031, 0x3035}, {0x303B, 0x303C}, {0x3041, 0x3096}, {0x309D, 0x309F}, {0x30A1, 0x30FA}, {0x30FC, 0x30FF}, {0x3105, 0x312F}, {0x3131, 0x318E}, {0x31A0, 0x31BF}, {0x31F0, 0x31FF}, {0x3400, 0x4DBF},
+{0x4E00, 0x9FFC}, {0xA000, 0xA48C}, {0xA4D0, 0xA4FD}, {0xA500, 0xA60C}, {0xA610, 0xA61F}, {0xA62A, 0xA62B}, {0xA640, 0xA66E}, {0xA67F, 0xA69D}, {0xA6A0, 0xA6E5}, {0xA717, 0xA71F}, {0xA722, 0xA788},
+{0xA78B, 0xA7BF}, {0xA7C2, 0xA7CA}, {0xA7F5, 0xA801}, {0xA803, 0xA805}, {0xA807, 0xA80A}, {0xA80C, 0xA822}, {0xA840, 0xA873}, {0xA882, 0xA8B3}, {0xA8F2, 0xA8F7}, {0xA8FB, 0xA8FB}, {0xA8FD, 0xA8FE},
+{0xA90A, 0xA925}, {0xA930, 0xA946}, {0xA960, 0xA97C}, {0xA984, 0xA9B2}, {0xA9CF, 0xA9CF}, {0xA9E0, 0xA9E4}, {0xA9E6, 0xA9EF}, {0xA9FA, 0xA9FE}, {0xAA00, 0xAA28}, {0xAA40, 0xAA42}, {0xAA44, 0xAA4B},
+{0xAA60, 0xAA76}, {0xAA7A, 0xAA7A}, {0xAA7E, 0xAAAF}, {0xAAB1, 0xAAB1}, {0xAAB5, 0xAAB6}, {0xAAB9, 0xAABD}, {0xAAC0, 0xAAC0}, {0xAAC2, 0xAAC2}, {0xAADB, 0xAADD}, {0xAAE0, 0xAAEA}, {0xAAF2, 0xAAF4},
+{0xAB01, 0xAB06}, {0xAB09, 0xAB0E}, {0xAB11, 0xAB16}, {0xAB20, 0xAB26}, {0xAB28, 0xAB2E}, {0xAB30, 0xAB5A}, {0xAB5C, 0xAB69}, {0xAB70, 0xABE2}, {0xAC00, 0xD7A3}, {0xD7B0, 0xD7C6}, {0xD7CB, 0xD7FB},
+{0xF900, 0xFA6D}, {0xFA70, 0xFAD9}, {0xFB00, 0xFB06}, {0xFB13, 0xFB17}, {0xFB1D, 0xFB1D}, {0xFB1F, 0xFB28}, {0xFB2A, 0xFB36}, {0xFB38, 0xFB3C}, {0xFB3E, 0xFB3E}, {0xFB40, 0xFB41}, {0xFB43, 0xFB44},
+{0xFB46, 0xFBB1}, {0xFBD3, 0xFD3D}, {0xFD50, 0xFD8F}, {0xFD92, 0xFDC7}, {0xFDF0, 0xFDFB}, {0xFE70, 0xFE74}, {0xFE76, 0xFEFC}, {0xFF21, 0xFF3A}, {0xFF41, 0xFF5A}, {0xFF66, 0xFFBE}, {0xFFC2, 0xFFC7},
+{0xFFCA, 0xFFCF}, {0xFFD2, 0xFFD7}, {0xFFDA, 0xFFDC}, {0x10000, 0x1000B}, {0x1000D, 0x10026}, {0x10028, 0x1003A}, {0x1003C, 0x1003D}, {0x1003F, 0x1004D}, {0x10050, 0x1005D}, {0x10080, 0x100FA},
+{0x10280, 0x1029C}, {0x102A0, 0x102D0}, {0x10300, 0x1031F}, {0x1032D, 0x10340}, {0x10342, 0x10349}, {0x10350, 0x10375}, {0x10380, 0x1039D}, {0x103A0, 0x103C3}, {0x103C8, 0x103CF}, {0x10400, 0x1049D},
+{0x104B0, 0x104D3}, {0x104D8, 0x104FB}, {0x10500, 0x10527}, {0x10530, 0x10563}, {0x10600, 0x10736}, {0x10740, 0x10755}, {0x10760, 0x10767}, {0x10800, 0x10805}, {0x10808, 0x10808}, {0x1080A, 0x10835},
+{0x10837, 0x10838}, {0x1083C, 0x1083C}, {0x1083F, 0x10855}, {0x10860, 0x10876}, {0x10880, 0x1089E}, {0x108E0, 0x108F2}, {0x108F4, 0x108F5}, {0x10900, 0x10915}, {0x10920, 0x10939}, {0x10980, 0x109B7},
+{0x109BE, 0x109BF}, {0x10A00, 0x10A00}, {0x10A10, 0x10A13}, {0x10A15, 0x10A17}, {0x10A19, 0x10A35}, {0x10A60, 0x10A7C}, {0x10A80, 0x10A9C}, {0x10AC0, 0x10AC7}, {0x10AC9, 0x10AE4}, {0x10B00, 0x10B35},
+{0x10B40, 0x10B55}, {0x10B60, 0x10B72}, {0x10B80, 0x10B91}, {0x10C00, 0x10C48}, {0x10C80, 0x10CB2}, {0x10CC0, 0x10CF2}, {0x10D00, 0x10D23}, {0x10E80, 0x10EA9}, {0x10EB0, 0x10EB1}, {0x10F00, 0x10F1C},
+{0x10F27, 0x10F27}, {0x10F30, 0x10F45}, {0x10FB0, 0x10FC4}, {0x10FE0, 0x10FF6}, {0x11003, 0x11037}, {0x11083, 0x110AF}, {0x110D0, 0x110E8}, {0x11103, 0x11126}, {0x11144, 0x11144}, {0x11147, 0x11147},
+{0x11150, 0x11172}, {0x11176, 0x11176}, {0x11183, 0x111B2}, {0x111C1, 0x111C4}, {0x111DA, 0x111DA}, {0x111DC, 0x111DC}, {0x11200, 0x11211}, {0x11213, 0x1122B}, {0x11280, 0x11286}, {0x11288, 0x11288},
+{0x1128A, 0x1128D}, {0x1128F, 0x1129D}, {0x1129F, 0x112A8}, {0x112B0, 0x112DE}, {0x11305, 0x1130C}, {0x1130F, 0x11310}, {0x11313, 0x11328}, {0x1132A, 0x11330}, {0x11332, 0x11333}, {0x11335, 0x11339},
+{0x1133D, 0x1133D}, {0x11350, 0x11350}, {0x1135D, 0x11361}, {0x11400, 0x11434}, {0x11447, 0x1144A}, {0x1145F, 0x11461}, {0x11480, 0x114AF}, {0x114C4, 0x114C5}, {0x114C7, 0x114C7}, {0x11580, 0x115AE},
+{0x115D8, 0x115DB}, {0x11600, 0x1162F}, {0x11644, 0x11644}, {0x11680, 0x116AA}, {0x116B8, 0x116B8}, {0x11700, 0x1171A}, {0x11800, 0x1182B}, {0x118A0, 0x118DF}, {0x118FF, 0x11906}, {0x11909, 0x11909},
+{0x1190C, 0x11913}, {0x11915, 0x11916}, {0x11918, 0x1192F}, {0x1193F, 0x1193F}, {0x11941, 0x11941}, {0x119A0, 0x119A7}, {0x119AA, 0x119D0}, {0x119E1, 0x119E1}, {0x119E3, 0x119E3}, {0x11A00, 0x11A00},
+{0x11A0B, 0x11A32}, {0x11A3A, 0x11A3A}, {0x11A50, 0x11A50}, {0x11A5C, 0x11A89}, {0x11A9D, 0x11A9D}, {0x11AC0, 0x11AF8}, {0x11C00, 0x11C08}, {0x11C0A, 0x11C2E}, {0x11C40, 0x11C40}, {0x11C72, 0x11C8F},
+{0x11D00, 0x11D06}, {0x11D08, 0x11D09}, {0x11D0B, 0x11D30}, {0x11D46, 0x11D46}, {0x11D60, 0x11D65}, {0x11D67, 0x11D68}, {0x11D6A, 0x11D89}, {0x11D98, 0x11D98}, {0x11EE0, 0x11EF2}, {0x11FB0, 0x11FB0},
+{0x12000, 0x12399}, {0x12480, 0x12543}, {0x13000, 0x1342E}, {0x14400, 0x14646}, {0x16800, 0x16A38}, {0x16A40, 0x16A5E}, {0x16AD0, 0x16AED}, {0x16B00, 0x16B2F}, {0x16B40, 0x16B43}, {0x16B63, 0x16B77},
+{0x16B7D, 0x16B8F}, {0x16E40, 0x16E7F}, {0x16F00, 0x16F4A}, {0x16F50, 0x16F50}, {0x16F93, 0x16F9F}, {0x16FE0, 0x16FE1}, {0x16FE3, 0x16FE3}, {0x17000, 0x187F7}, {0x18800, 0x18CD5}, {0x18D00, 0x18D08},
+{0x1B000, 0x1B11E}, {0x1B150, 0x1B152}, {0x1B164, 0x1B167}, {0x1B170, 0x1B2FB}, {0x1BC00, 0x1BC6A}, {0x1BC70, 0x1BC7C}, {0x1BC80, 0x1BC88}, {0x1BC90, 0x1BC99}, {0x1D400, 0x1D454}, {0x1D456, 0x1D49C},
+{0x1D49E, 0x1D49F}, {0x1D4A2, 0x1D4A2}, {0x1D4A5, 0x1D4A6}, {0x1D4A9, 0x1D4AC}, {0x1D4AE, 0x1D4B9}, {0x1D4BB, 0x1D4BB}, {0x1D4BD, 0x1D4C3}, {0x1D4C5, 0x1D505}, {0x1D507, 0x1D50A}, {0x1D50D, 0x1D514},
+{0x1D516, 0x1D51C}, {0x1D51E, 0x1D539}, {0x1D53B, 0x1D53E}, {0x1D540, 0x1D544}, {0x1D546, 0x1D546}, {0x1D54A, 0x1D550}, {0x1D552, 0x1D6A5}, {0x1D6A8, 0x1D6C0}, {0x1D6C2, 0x1D6DA}, {0x1D6DC, 0x1D6FA},
+{0x1D6FC, 0x1D714}, {0x1D716, 0x1D734}, {0x1D736, 0x1D74E}, {0x1D750, 0x1D76E}, {0x1D770, 0x1D788}, {0x1D78A, 0x1D7A8}, {0x1D7AA, 0x1D7C2}, {0x1D7C4, 0x1D7CB}, {0x1E100, 0x1E12C}, {0x1E137, 0x1E13D},
+{0x1E14E, 0x1E14E}, {0x1E2C0, 0x1E2EB}, {0x1E800, 0x1E8C4}, {0x1E900, 0x1E943}, {0x1E94B, 0x1E94B}, {0x1EE00, 0x1EE03}, {0x1EE05, 0x1EE1F}, {0x1EE21, 0x1EE22}, {0x1EE24, 0x1EE24}, {0x1EE27, 0x1EE27},
+{0x1EE29, 0x1EE32}, {0x1EE34, 0x1EE37}, {0x1EE39, 0x1EE39}, {0x1EE3B, 0x1EE3B}, {0x1EE42, 0x1EE42}, {0x1EE47, 0x1EE47}, {0x1EE49, 0x1EE49}, {0x1EE4B, 0x1EE4B}, {0x1EE4D, 0x1EE4F}, {0x1EE51, 0x1EE52},
+{0x1EE54, 0x1EE54}, {0x1EE57, 0x1EE57}, {0x1EE59, 0x1EE59}, {0x1EE5B, 0x1EE5B}, {0x1EE5D, 0x1EE5D}, {0x1EE5F, 0x1EE5F}, {0x1EE61, 0x1EE62}, {0x1EE64, 0x1EE64}, {0x1EE67, 0x1EE6A}, {0x1EE6C, 0x1EE72},
+{0x1EE74, 0x1EE77}, {0x1EE79, 0x1EE7C}, {0x1EE7E, 0x1EE7E}, {0x1EE80, 0x1EE89}, {0x1EE8B, 0x1EE9B}, {0x1EEA1, 0x1EEA3}, {0x1EEA5, 0x1EEA9}, {0x1EEAB, 0x1EEBB}, {0x20000, 0x2A6DD}, {0x2A700, 0x2B734},
+{0x2B740, 0x2B81D}, {0x2B820, 0x2CEA1}, {0x2CEB0, 0x2EBE0}, {0x2F800, 0x2FA1D}, {0x30000, 0x3134A},
+};
+
+static const std::vector<std::pair<uint32_t, uint32_t>> whitespace_ranges = {
+{0x9, 0xD}, {0x1C, 0x20}, {0x85, 0x85}, {0xA0, 0xA0}, {0x1680, 0x1680}, {0x2000, 0x200A}, {0x2028, 0x2029}, {0x202F, 0x202F}, {0x205F, 0x205F}, {0x3000, 0x3000},
+};
+
+static const std::vector<std::pair<uint32_t, uint32_t>> accent_mark_ranges = {
+{0x300, 0x36F}, {0x483, 0x489}, {0x591, 0x5BD}, {0x5BF, 0x5BF}, {0x5C1, 0x5C2}, {0x5C4, 0x5C5}, {0x5C7, 0x5C7}, {0x610, 0x61A}, {0x64B, 0x65F}, {0x670, 0x670}, {0x6D6, 0x6DC}, {0x6DF, 0x6E4},
+{0x6E7, 0x6E8}, {0x6EA, 0x6ED}, {0x711, 0x711}, {0x730, 0x74A}, {0x7A6, 0x7B0}, {0x7EB, 0x7F3}, {0x7FD, 0x7FD}, {0x816, 0x819}, {0x81B, 0x823}, {0x825, 0x827}, {0x829, 0x82D}, {0x859, 0x85B},
+{0x8D3, 0x8E1}, {0x8E3, 0x903}, {0x93A, 0x93C}, {0x93E, 0x94F}, {0x951, 0x957}, {0x962, 0x963}, {0x981, 0x983}, {0x9BC, 0x9BC}, {0x9BE, 0x9C4}, {0x9C7, 0x9C8}, {0x9CB, 0x9CD}, {0x9D7, 0x9D7},
+{0x9E2, 0x9E3}, {0x9FE, 0x9FE}, {0xA01, 0xA03}, {0xA3C, 0xA3C}, {0xA3E, 0xA42}, {0xA47, 0xA48}, {0xA4B, 0xA4D}, {0xA51, 0xA51}, {0xA70, 0xA71}, {0xA75, 0xA75}, {0xA81, 0xA83}, {0xABC, 0xABC},
+{0xABE, 0xAC5}, {0xAC7, 0xAC9}, {0xACB, 0xACD}, {0xAE2, 0xAE3}, {0xAFA, 0xAFF}, {0xB01, 0xB03}, {0xB3C, 0xB3C}, {0xB3E, 0xB44}, {0xB47, 0xB48}, {0xB4B, 0xB4D}, {0xB55, 0xB57}, {0xB62, 0xB63},
+{0xB82, 0xB82}, {0xBBE, 0xBC2}, {0xBC6, 0xBC8}, {0xBCA, 0xBCD}, {0xBD7, 0xBD7}, {0xC00, 0xC04}, {0xC3E, 0xC44}, {0xC46, 0xC48}, {0xC4A, 0xC4D}, {0xC55, 0xC56}, {0xC62, 0xC63}, {0xC81, 0xC83},
+{0xCBC, 0xCBC}, {0xCBE, 0xCC4}, {0xCC6, 0xCC8}, {0xCCA, 0xCCD}, {0xCD5, 0xCD6}, {0xCE2, 0xCE3}, {0xD00, 0xD03}, {0xD3B, 0xD3C}, {0xD3E, 0xD44}, {0xD46, 0xD48}, {0xD4A, 0xD4D}, {0xD57, 0xD57},
+{0xD62, 0xD63}, {0xD81, 0xD83}, {0xDCA, 0xDCA}, {0xDCF, 0xDD4}, {0xDD6, 0xDD6}, {0xDD8, 0xDDF}, {0xDF2, 0xDF3}, {0xE31, 0xE31}, {0xE34, 0xE3A}, {0xE47, 0xE4E}, {0xEB1, 0xEB1}, {0xEB4, 0xEBC},
+{0xEC8, 0xECD}, {0xF18, 0xF19}, {0xF35, 0xF35}, {0xF37, 0xF37}, {0xF39, 0xF39}, {0xF3E, 0xF3F}, {0xF71, 0xF84}, {0xF86, 0xF87}, {0xF8D, 0xF97}, {0xF99, 0xFBC}, {0xFC6, 0xFC6}, {0x102B, 0x103E},
+{0x1056, 0x1059}, {0x105E, 0x1060}, {0x1062, 0x1064}, {0x1067, 0x106D}, {0x1071, 0x1074}, {0x1082, 0x108D}, {0x108F, 0x108F}, {0x109A, 0x109D}, {0x135D, 0x135F}, {0x1712, 0x1714}, {0x1732, 0x1734},
+{0x1752, 0x1753}, {0x1772, 0x1773}, {0x17B4, 0x17D3}, {0x17DD, 0x17DD}, {0x180B, 0x180D}, {0x1885, 0x1886}, {0x18A9, 0x18A9}, {0x1920, 0x192B}, {0x1930, 0x193B}, {0x1A17, 0x1A1B}, {0x1A55, 0x1A5E},
+{0x1A60, 0x1A7C}, {0x1A7F, 0x1A7F}, {0x1AB0, 0x1AC0}, {0x1B00, 0x1B04}, {0x1B34, 0x1B44}, {0x1B6B, 0x1B73}, {0x1B80, 0x1B82}, {0x1BA1, 0x1BAD}, {0x1BE6, 0x1BF3}, {0x1C24, 0x1C37}, {0x1CD0, 0x1CD2},
+{0x1CD4, 0x1CE8}, {0x1CED, 0x1CED}, {0x1CF4, 0x1CF4}, {0x1CF7, 0x1CF9}, {0x1DC0, 0x1DF9}, {0x1DFB, 0x1DFF}, {0x20D0, 0x20F0}, {0x2CEF, 0x2CF1}, {0x2D7F, 0x2D7F}, {0x2DE0, 0x2DFF}, {0x302A, 0x302F},
+{0x3099, 0x309A}, {0xA66F, 0xA672}, {0xA674, 0xA67D}, {0xA69E, 0xA69F}, {0xA6F0, 0xA6F1}, {0xA802, 0xA802}, {0xA806, 0xA806}, {0xA80B, 0xA80B}, {0xA823, 0xA827}, {0xA82C, 0xA82C}, {0xA880, 0xA881},
+{0xA8B4, 0xA8C5}, {0xA8E0, 0xA8F1}, {0xA8FF, 0xA8FF}, {0xA926, 0xA92D}, {0xA947, 0xA953}, {0xA980, 0xA983}, {0xA9B3, 0xA9C0}, {0xA9E5, 0xA9E5}, {0xAA29, 0xAA36}, {0xAA43, 0xAA43}, {0xAA4C, 0xAA4D},
+{0xAA7B, 0xAA7D}, {0xAAB0, 0xAAB0}, {0xAAB2, 0xAAB4}, {0xAAB7, 0xAAB8}, {0xAABE, 0xAABF}, {0xAAC1, 0xAAC1}, {0xAAEB, 0xAAEF}, {0xAAF5, 0xAAF6}, {0xABE3, 0xABEA}, {0xABEC, 0xABED}, {0xFB1E, 0xFB1E},
+{0xFE00, 0xFE0F}, {0xFE20, 0xFE2F}, {0x101FD, 0x101FD}, {0x102E0, 0x102E0}, {0x10376, 0x1037A}, {0x10A01, 0x10A03}, {0x10A05, 0x10A06}, {0x10A0C, 0x10A0F}, {0x10A38, 0x10A3A}, {0x10A3F, 0x10A3F},
+{0x10AE5, 0x10AE6}, {0x10D24, 0x10D27}, {0x10EAB, 0x10EAC}, {0x10F46, 0x10F50}, {0x11000, 0x11002}, {0x11038, 0x11046}, {0x1107F, 0x11082}, {0x110B0, 0x110BA}, {0x11100, 0x11102}, {0x11127, 0x11134},
+{0x11145, 0x11146}, {0x11173, 0x11173}, {0x11180, 0x11182}, {0x111B3, 0x111C0}, {0x111C9, 0x111CC}, {0x111CE, 0x111CF}, {0x1122C, 0x11237}, {0x1123E, 0x1123E}, {0x112DF, 0x112EA}, {0x11300, 0x11303},
+{0x1133B, 0x1133C}, {0x1133E, 0x11344}, {0x11347, 0x11348}, {0x1134B, 0x1134D}, {0x11357, 0x11357}, {0x11362, 0x11363}, {0x11366, 0x1136C}, {0x11370, 0x11374}, {0x11435, 0x11446}, {0x1145E, 0x1145E},
+{0x114B0, 0x114C3}, {0x115AF, 0x115B5}, {0x115B8, 0x115C0}, {0x115DC, 0x115DD}, {0x11630, 0x11640}, {0x116AB, 0x116B7}, {0x1171D, 0x1172B}, {0x1182C, 0x1183A}, {0x11930, 0x11935}, {0x11937, 0x11938},
+{0x1193B, 0x1193E}, {0x11940, 0x11940}, {0x11942, 0x11943}, {0x119D1, 0x119D7}, {0x119DA, 0x119E0}, {0x119E4, 0x119E4}, {0x11A01, 0x11A0A}, {0x11A33, 0x11A39}, {0x11A3B, 0x11A3E}, {0x11A47, 0x11A47},
+{0x11A51, 0x11A5B}, {0x11A8A, 0x11A99}, {0x11C2F, 0x11C36}, {0x11C38, 0x11C3F}, {0x11C92, 0x11CA7}, {0x11CA9, 0x11CB6}, {0x11D31, 0x11D36}, {0x11D3A, 0x11D3A}, {0x11D3C, 0x11D3D}, {0x11D3F, 0x11D45},
+{0x11D47, 0x11D47}, {0x11D8A, 0x11D8E}, {0x11D90, 0x11D91}, {0x11D93, 0x11D97}, {0x11EF3, 0x11EF6}, {0x16AF0, 0x16AF4}, {0x16B30, 0x16B36}, {0x16F4F, 0x16F4F}, {0x16F51, 0x16F87}, {0x16F8F, 0x16F92},
+{0x16FE4, 0x16FE4}, {0x16FF0, 0x16FF1}, {0x1BC9D, 0x1BC9E}, {0x1D165, 0x1D169}, {0x1D16D, 0x1D172}, {0x1D17B, 0x1D182}, {0x1D185, 0x1D18B}, {0x1D1AA, 0x1D1AD}, {0x1D242, 0x1D244}, {0x1DA00, 0x1DA36},
+{0x1DA3B, 0x1DA6C}, {0x1DA75, 0x1DA75}, {0x1DA84, 0x1DA84}, {0x1DA9B, 0x1DA9F}, {0x1DAA1, 0x1DAAF}, {0x1E000, 0x1E006}, {0x1E008, 0x1E018}, {0x1E01B, 0x1E021}, {0x1E023, 0x1E024}, {0x1E026, 0x1E02A},
+{0x1E130, 0x1E136}, {0x1E2EC, 0x1E2EF}, {0x1E8D0, 0x1E8D6}, {0x1E944, 0x1E94A}, {0xE0100, 0xE01EF},
+};
+
+static const std::vector<std::pair<uint32_t, uint32_t>> punctuation_ranges = {
+{0x21, 0x23}, {0x25, 0x2A}, {0x2C, 0x2F}, {0x3A, 0x3B}, {0x3F, 0x40}, {0x5B, 0x5D}, {0x5F, 0x5F}, {0x7B, 0x7B}, {0x7D, 0x7D}, {0xA1, 0xA1}, {0xA7, 0xA7}, {0xAB, 0xAB}, {0xB6, 0xB7}, {0xBB, 0xBB},
+{0xBF, 0xBF}, {0x37E, 0x37E}, {0x387, 0x387}, {0x55A, 0x55F}, {0x589, 0x58A}, {0x5BE, 0x5BE}, {0x5C0, 0x5C0}, {0x5C3, 0x5C3}, {0x5C6, 0x5C6}, {0x5F3, 0x5F4}, {0x609, 0x60A}, {0x60C, 0x60D},
+{0x61B, 0x61B}, {0x61E, 0x61F}, {0x66A, 0x66D}, {0x6D4, 0x6D4}, {0x700, 0x70D}, {0x7F7, 0x7F9}, {0x830, 0x83E}, {0x85E, 0x85E}, {0x964, 0x965}, {0x970, 0x970}, {0x9FD, 0x9FD}, {0xA76, 0xA76},
+{0xAF0, 0xAF0}, {0xC77, 0xC77}, {0xC84, 0xC84}, {0xDF4, 0xDF4}, {0xE4F, 0xE4F}, {0xE5A, 0xE5B}, {0xF04, 0xF12}, {0xF14, 0xF14}, {0xF3A, 0xF3D}, {0xF85, 0xF85}, {0xFD0, 0xFD4}, {0xFD9, 0xFDA},
+{0x104A, 0x104F}, {0x10FB, 0x10FB}, {0x1360, 0x1368}, {0x1400, 0x1400}, {0x166E, 0x166E}, {0x169B, 0x169C}, {0x16EB, 0x16ED}, {0x1735, 0x1736}, {0x17D4, 0x17D6}, {0x17D8, 0x17DA}, {0x1800, 0x180A},
+{0x1944, 0x1945}, {0x1A1E, 0x1A1F}, {0x1AA0, 0x1AA6}, {0x1AA8, 0x1AAD}, {0x1B5A, 0x1B60}, {0x1BFC, 0x1BFF}, {0x1C3B, 0x1C3F}, {0x1C7E, 0x1C7F}, {0x1CC0, 0x1CC7}, {0x1CD3, 0x1CD3}, {0x2010, 0x2027},
+{0x2030, 0x2043}, {0x2045, 0x2051}, {0x2053, 0x205E}, {0x207D, 0x207E}, {0x208D, 0x208E}, {0x2308, 0x230B}, {0x2329, 0x232A}, {0x2768, 0x2775}, {0x27C5, 0x27C6}, {0x27E6, 0x27EF}, {0x2983, 0x2998},
+{0x29D8, 0x29DB}, {0x29FC, 0x29FD}, {0x2CF9, 0x2CFC}, {0x2CFE, 0x2CFF}, {0x2D70, 0x2D70}, {0x2E00, 0x2E2E}, {0x2E30, 0x2E4F}, {0x2E52, 0x2E52}, {0x3001, 0x3003}, {0x3008, 0x3011}, {0x3014, 0x301F},
+{0x3030, 0x3030}, {0x303D, 0x303D}, {0x30A0, 0x30A0}, {0x30FB, 0x30FB}, {0xA4FE, 0xA4FF}, {0xA60D, 0xA60F}, {0xA673, 0xA673}, {0xA67E, 0xA67E}, {0xA6F2, 0xA6F7}, {0xA874, 0xA877}, {0xA8CE, 0xA8CF},
+{0xA8F8, 0xA8FA}, {0xA8FC, 0xA8FC}, {0xA92E, 0xA92F}, {0xA95F, 0xA95F}, {0xA9C1, 0xA9CD}, {0xA9DE, 0xA9DF}, {0xAA5C, 0xAA5F}, {0xAADE, 0xAADF}, {0xAAF0, 0xAAF1}, {0xABEB, 0xABEB}, {0xFD3E, 0xFD3F},
+{0xFE10, 0xFE19}, {0xFE30, 0xFE52}, {0xFE54, 0xFE61}, {0xFE63, 0xFE63}, {0xFE68, 0xFE68}, {0xFE6A, 0xFE6B}, {0xFF01, 0xFF03}, {0xFF05, 0xFF0A}, {0xFF0C, 0xFF0F}, {0xFF1A, 0xFF1B}, {0xFF1F, 0xFF20},
+{0xFF3B, 0xFF3D}, {0xFF3F, 0xFF3F}, {0xFF5B, 0xFF5B}, {0xFF5D, 0xFF5D}, {0xFF5F, 0xFF65}, {0x10100, 0x10102}, {0x1039F, 0x1039F}, {0x103D0, 0x103D0}, {0x1056F, 0x1056F}, {0x10857, 0x10857},
+{0x1091F, 0x1091F}, {0x1093F, 0x1093F}, {0x10A50, 0x10A58}, {0x10A7F, 0x10A7F}, {0x10AF0, 0x10AF6}, {0x10B39, 0x10B3F}, {0x10B99, 0x10B9C}, {0x10EAD, 0x10EAD}, {0x10F55, 0x10F59}, {0x11047, 0x1104D},
+{0x110BB, 0x110BC}, {0x110BE, 0x110C1}, {0x11140, 0x11143}, {0x11174, 0x11175}, {0x111C5, 0x111C8}, {0x111CD, 0x111CD}, {0x111DB, 0x111DB}, {0x111DD, 0x111DF}, {0x11238, 0x1123D}, {0x112A9, 0x112A9},
+{0x1144B, 0x1144F}, {0x1145A, 0x1145B}, {0x1145D, 0x1145D}, {0x114C6, 0x114C6}, {0x115C1, 0x115D7}, {0x11641, 0x11643}, {0x11660, 0x1166C}, {0x1173C, 0x1173E}, {0x1183B, 0x1183B}, {0x11944, 0x11946},
+{0x119E2, 0x119E2}, {0x11A3F, 0x11A46}, {0x11A9A, 0x11A9C}, {0x11A9E, 0x11AA2}, {0x11C41, 0x11C45}, {0x11C70, 0x11C71}, {0x11EF7, 0x11EF8}, {0x11FFF, 0x11FFF}, {0x12470, 0x12474}, {0x16A6E, 0x16A6F},
+{0x16AF5, 0x16AF5}, {0x16B37, 0x16B3B}, {0x16B44, 0x16B44}, {0x16E97, 0x16E9A}, {0x16FE2, 0x16FE2}, {0x1BC9F, 0x1BC9F}, {0x1DA87, 0x1DA8B}, {0x1E95E, 0x1E95F},
+};
+
+static const std::vector<std::pair<uint32_t, uint32_t>> symbol_ranges = {
+{0x24, 0x24}, {0x2B, 0x2B}, {0x3C, 0x3E}, {0x5E, 0x5E}, {0x60, 0x60}, {0x7C, 0x7C}, {0x7E, 0x7E}, {0xA2, 0xA6}, {0xA8, 0xA9}, {0xAC, 0xAC}, {0xAE, 0xB1}, {0xB4, 0xB4}, {0xB8, 0xB8}, {0xD7, 0xD7},
+{0xF7, 0xF7}, {0x2C2, 0x2C5}, {0x2D2, 0x2DF}, {0x2E5, 0x2EB}, {0x2ED, 0x2ED}, {0x2EF, 0x2FF}, {0x375, 0x375}, {0x384, 0x385}, {0x3F6, 0x3F6}, {0x482, 0x482}, {0x58D, 0x58F}, {0x606, 0x608},
+{0x60B, 0x60B}, {0x60E, 0x60F}, {0x6DE, 0x6DE}, {0x6E9, 0x6E9}, {0x6FD, 0x6FE}, {0x7F6, 0x7F6}, {0x7FE, 0x7FF}, {0x9F2, 0x9F3}, {0x9FA, 0x9FB}, {0xAF1, 0xAF1}, {0xB70, 0xB70}, {0xBF3, 0xBFA},
+{0xC7F, 0xC7F}, {0xD4F, 0xD4F}, {0xD79, 0xD79}, {0xE3F, 0xE3F}, {0xF01, 0xF03}, {0xF13, 0xF13}, {0xF15, 0xF17}, {0xF1A, 0xF1F}, {0xF34, 0xF34}, {0xF36, 0xF36}, {0xF38, 0xF38}, {0xFBE, 0xFC5},
+{0xFC7, 0xFCC}, {0xFCE, 0xFCF}, {0xFD5, 0xFD8}, {0x109E, 0x109F}, {0x1390, 0x1399}, {0x166D, 0x166D}, {0x17DB, 0x17DB}, {0x1940, 0x1940}, {0x19DE, 0x19FF}, {0x1B61, 0x1B6A}, {0x1B74, 0x1B7C},
+{0x1FBD, 0x1FBD}, {0x1FBF, 0x1FC1}, {0x1FCD, 0x1FCF}, {0x1FDD, 0x1FDF}, {0x1FED, 0x1FEF}, {0x1FFD, 0x1FFE}, {0x2044, 0x2044}, {0x2052, 0x2052}, {0x207A, 0x207C}, {0x208A, 0x208C}, {0x20A0, 0x20BF},
+{0x2100, 0x2101}, {0x2103, 0x2106}, {0x2108, 0x2109}, {0x2114, 0x2114}, {0x2116, 0x2118}, {0x211E, 0x2123}, {0x2125, 0x2125}, {0x2127, 0x2127}, {0x2129, 0x2129}, {0x212E, 0x212E}, {0x213A, 0x213B},
+{0x2140, 0x2144}, {0x214A, 0x214D}, {0x214F, 0x214F}, {0x218A, 0x218B}, {0x2190, 0x2307}, {0x230C, 0x2328}, {0x232B, 0x2426}, {0x2440, 0x244A}, {0x249C, 0x24E9}, {0x2500, 0x2767}, {0x2794, 0x27C4},
+{0x27C7, 0x27E5}, {0x27F0, 0x2982}, {0x2999, 0x29D7}, {0x29DC, 0x29FB}, {0x29FE, 0x2B73}, {0x2B76, 0x2B95}, {0x2B97, 0x2BFF}, {0x2CE5, 0x2CEA}, {0x2E50, 0x2E51}, {0x2E80, 0x2E99}, {0x2E9B, 0x2EF3},
+{0x2F00, 0x2FD5}, {0x2FF0, 0x2FFB}, {0x3004, 0x3004}, {0x3012, 0x3013}, {0x3020, 0x3020}, {0x3036, 0x3037}, {0x303E, 0x303F}, {0x309B, 0x309C}, {0x3190, 0x3191}, {0x3196, 0x319F}, {0x31C0, 0x31E3},
+{0x3200, 0x321E}, {0x322A, 0x3247}, {0x3250, 0x3250}, {0x3260, 0x327F}, {0x328A, 0x32B0}, {0x32C0, 0x33FF}, {0x4DC0, 0x4DFF}, {0xA490, 0xA4C6}, {0xA700, 0xA716}, {0xA720, 0xA721}, {0xA789, 0xA78A},
+{0xA828, 0xA82B}, {0xA836, 0xA839}, {0xAA77, 0xAA79}, {0xAB5B, 0xAB5B}, {0xAB6A, 0xAB6B}, {0xFB29, 0xFB29}, {0xFBB2, 0xFBC1}, {0xFDFC, 0xFDFD}, {0xFE62, 0xFE62}, {0xFE64, 0xFE66}, {0xFE69, 0xFE69},
+{0xFF04, 0xFF04}, {0xFF0B, 0xFF0B}, {0xFF1C, 0xFF1E}, {0xFF3E, 0xFF3E}, {0xFF40, 0xFF40}, {0xFF5C, 0xFF5C}, {0xFF5E, 0xFF5E}, {0xFFE0, 0xFFE6}, {0xFFE8, 0xFFEE}, {0xFFFC, 0xFFFD}, {0x10137, 0x1013F},
+{0x10179, 0x10189}, {0x1018C, 0x1018E}, {0x10190, 0x1019C}, {0x101A0, 0x101A0}, {0x101D0, 0x101FC}, {0x10877, 0x10878}, {0x10AC8, 0x10AC8}, {0x1173F, 0x1173F}, {0x11FD5, 0x11FF1}, {0x16B3C, 0x16B3F},
+{0x16B45, 0x16B45}, {0x1BC9C, 0x1BC9C}, {0x1D000, 0x1D0F5}, {0x1D100, 0x1D126}, {0x1D129, 0x1D164}, {0x1D16A, 0x1D16C}, {0x1D183, 0x1D184}, {0x1D18C, 0x1D1A9}, {0x1D1AE, 0x1D1E8}, {0x1D200, 0x1D241},
+{0x1D245, 0x1D245}, {0x1D300, 0x1D356}, {0x1D6C1, 0x1D6C1}, {0x1D6DB, 0x1D6DB}, {0x1D6FB, 0x1D6FB}, {0x1D715, 0x1D715}, {0x1D735, 0x1D735}, {0x1D74F, 0x1D74F}, {0x1D76F, 0x1D76F}, {0x1D789, 0x1D789},
+{0x1D7A9, 0x1D7A9}, {0x1D7C3, 0x1D7C3}, {0x1D800, 0x1D9FF}, {0x1DA37, 0x1DA3A}, {0x1DA6D, 0x1DA74}, {0x1DA76, 0x1DA83}, {0x1DA85, 0x1DA86}, {0x1E14F, 0x1E14F}, {0x1E2FF, 0x1E2FF}, {0x1ECAC, 0x1ECAC},
+{0x1ECB0, 0x1ECB0}, {0x1ED2E, 0x1ED2E}, {0x1EEF0, 0x1EEF1}, {0x1F000, 0x1F02B}, {0x1F030, 0x1F093}, {0x1F0A0, 0x1F0AE}, {0x1F0B1, 0x1F0BF}, {0x1F0C1, 0x1F0CF}, {0x1F0D1, 0x1F0F5}, {0x1F10D, 0x1F1AD},
+{0x1F1E6, 0x1F202}, {0x1F210, 0x1F23B}, {0x1F240, 0x1F248}, {0x1F250, 0x1F251}, {0x1F260, 0x1F265}, {0x1F300, 0x1F6D7}, {0x1F6E0, 0x1F6EC}, {0x1F6F0, 0x1F6FC}, {0x1F700, 0x1F773}, {0x1F780, 0x1F7D8},
+{0x1F7E0, 0x1F7EB}, {0x1F800, 0x1F80B}, {0x1F810, 0x1F847}, {0x1F850, 0x1F859}, {0x1F860, 0x1F887}, {0x1F890, 0x1F8AD}, {0x1F8B0, 0x1F8B1}, {0x1F900, 0x1F978}, {0x1F97A, 0x1F9CB}, {0x1F9CD, 0x1FA53},
+{0x1FA60, 0x1FA6D}, {0x1FA70, 0x1FA74}, {0x1FA78, 0x1FA7A}, {0x1FA80, 0x1FA86}, {0x1FA90, 0x1FAA8}, {0x1FAB0, 0x1FAB6}, {0x1FAC0, 0x1FAC2}, {0x1FAD0, 0x1FAD6}, {0x1FB00, 0x1FB92}, {0x1FB94, 0x1FBCA},
+};
+
+static const std::vector<std::pair<uint32_t, uint32_t>> control_ranges = {
+{0x0, 0x8}, {0xE, 0x1B}, {0x7F, 0x84}, {0x86, 0x9F}, {0xAD, 0xAD}, {0x378, 0x379}, {0x380, 0x383}, {0x38B, 0x38B}, {0x38D, 0x38D}, {0x3A2, 0x3A2}, {0x530, 0x530}, {0x557, 0x558}, {0x58B, 0x58C},
+{0x590, 0x590}, {0x5C8, 0x5CF}, {0x5EB, 0x5EE}, {0x5F5, 0x605}, {0x61C, 0x61D}, {0x6DD, 0x6DD}, {0x70E, 0x70F}, {0x74B, 0x74C}, {0x7B2, 0x7BF}, {0x7FB, 0x7FC}, {0x82E, 0x82F}, {0x83F, 0x83F},
+{0x85C, 0x85D}, {0x85F, 0x85F}, {0x86B, 0x89F}, {0x8B5, 0x8B5}, {0x8C8, 0x8D2}, {0x8E2, 0x8E2}, {0x984, 0x984}, {0x98D, 0x98E}, {0x991, 0x992}, {0x9A9, 0x9A9}, {0x9B1, 0x9B1}, {0x9B3, 0x9B5},
+{0x9BA, 0x9BB}, {0x9C5, 0x9C6}, {0x9C9, 0x9CA}, {0x9CF, 0x9D6}, {0x9D8, 0x9DB}, {0x9DE, 0x9DE}, {0x9E4, 0x9E5}, {0x9FF, 0xA00}, {0xA04, 0xA04}, {0xA0B, 0xA0E}, {0xA11, 0xA12}, {0xA29, 0xA29},
+{0xA31, 0xA31}, {0xA34, 0xA34}, {0xA37, 0xA37}, {0xA3A, 0xA3B}, {0xA3D, 0xA3D}, {0xA43, 0xA46}, {0xA49, 0xA4A}, {0xA4E, 0xA50}, {0xA52, 0xA58}, {0xA5D, 0xA5D}, {0xA5F, 0xA65}, {0xA77, 0xA80},
+{0xA84, 0xA84}, {0xA8E, 0xA8E}, {0xA92, 0xA92}, {0xAA9, 0xAA9}, {0xAB1, 0xAB1}, {0xAB4, 0xAB4}, {0xABA, 0xABB}, {0xAC6, 0xAC6}, {0xACA, 0xACA}, {0xACE, 0xACF}, {0xAD1, 0xADF}, {0xAE4, 0xAE5},
+{0xAF2, 0xAF8}, {0xB00, 0xB00}, {0xB04, 0xB04}, {0xB0D, 0xB0E}, {0xB11, 0xB12}, {0xB29, 0xB29}, {0xB31, 0xB31}, {0xB34, 0xB34}, {0xB3A, 0xB3B}, {0xB45, 0xB46}, {0xB49, 0xB4A}, {0xB4E, 0xB54},
+{0xB58, 0xB5B}, {0xB5E, 0xB5E}, {0xB64, 0xB65}, {0xB78, 0xB81}, {0xB84, 0xB84}, {0xB8B, 0xB8D}, {0xB91, 0xB91}, {0xB96, 0xB98}, {0xB9B, 0xB9B}, {0xB9D, 0xB9D}, {0xBA0, 0xBA2}, {0xBA5, 0xBA7},
+{0xBAB, 0xBAD}, {0xBBA, 0xBBD}, {0xBC3, 0xBC5}, {0xBC9, 0xBC9}, {0xBCE, 0xBCF}, {0xBD1, 0xBD6}, {0xBD8, 0xBE5}, {0xBFB, 0xBFF}, {0xC0D, 0xC0D}, {0xC11, 0xC11}, {0xC29, 0xC29}, {0xC3A, 0xC3C},
+{0xC45, 0xC45}, {0xC49, 0xC49}, {0xC4E, 0xC54}, {0xC57, 0xC57}, {0xC5B, 0xC5F}, {0xC64, 0xC65}, {0xC70, 0xC76}, {0xC8D, 0xC8D}, {0xC91, 0xC91}, {0xCA9, 0xCA9}, {0xCB4, 0xCB4}, {0xCBA, 0xCBB},
+{0xCC5, 0xCC5}, {0xCC9, 0xCC9}, {0xCCE, 0xCD4}, {0xCD7, 0xCDD}, {0xCDF, 0xCDF}, {0xCE4, 0xCE5}, {0xCF0, 0xCF0}, {0xCF3, 0xCFF}, {0xD0D, 0xD0D}, {0xD11, 0xD11}, {0xD45, 0xD45}, {0xD49, 0xD49},
+{0xD50, 0xD53}, {0xD64, 0xD65}, {0xD80, 0xD80}, {0xD84, 0xD84}, {0xD97, 0xD99}, {0xDB2, 0xDB2}, {0xDBC, 0xDBC}, {0xDBE, 0xDBF}, {0xDC7, 0xDC9}, {0xDCB, 0xDCE}, {0xDD5, 0xDD5}, {0xDD7, 0xDD7},
+{0xDE0, 0xDE5}, {0xDF0, 0xDF1}, {0xDF5, 0xE00}, {0xE3B, 0xE3E}, {0xE5C, 0xE80}, {0xE83, 0xE83}, {0xE85, 0xE85}, {0xE8B, 0xE8B}, {0xEA4, 0xEA4}, {0xEA6, 0xEA6}, {0xEBE, 0xEBF}, {0xEC5, 0xEC5},
+{0xEC7, 0xEC7}, {0xECE, 0xECF}, {0xEDA, 0xEDB}, {0xEE0, 0xEFF}, {0xF48, 0xF48}, {0xF6D, 0xF70}, {0xF98, 0xF98}, {0xFBD, 0xFBD}, {0xFCD, 0xFCD}, {0xFDB, 0xFFF}, {0x10C6, 0x10C6}, {0x10C8, 0x10CC},
+{0x10CE, 0x10CF}, {0x1249, 0x1249}, {0x124E, 0x124F}, {0x1257, 0x1257}, {0x1259, 0x1259}, {0x125E, 0x125F}, {0x1289, 0x1289}, {0x128E, 0x128F}, {0x12B1, 0x12B1}, {0x12B6, 0x12B7}, {0x12BF, 0x12BF},
+{0x12C1, 0x12C1}, {0x12C6, 0x12C7}, {0x12D7, 0x12D7}, {0x1311, 0x1311}, {0x1316, 0x1317}, {0x135B, 0x135C}, {0x137D, 0x137F}, {0x139A, 0x139F}, {0x13F6, 0x13F7}, {0x13FE, 0x13FF}, {0x169D, 0x169F},
+{0x16F9, 0x16FF}, {0x170D, 0x170D}, {0x1715, 0x171F}, {0x1737, 0x173F}, {0x1754, 0x175F}, {0x176D, 0x176D}, {0x1771, 0x1771}, {0x1774, 0x177F}, {0x17DE, 0x17DF}, {0x17EA, 0x17EF}, {0x17FA, 0x17FF},
+{0x180E, 0x180F}, {0x181A, 0x181F}, {0x1879, 0x187F}, {0x18AB, 0x18AF}, {0x18F6, 0x18FF}, {0x191F, 0x191F}, {0x192C, 0x192F}, {0x193C, 0x193F}, {0x1941, 0x1943}, {0x196E, 0x196F}, {0x1975, 0x197F},
+{0x19AC, 0x19AF}, {0x19CA, 0x19CF}, {0x19DB, 0x19DD}, {0x1A1C, 0x1A1D}, {0x1A5F, 0x1A5F}, {0x1A7D, 0x1A7E}, {0x1A8A, 0x1A8F}, {0x1A9A, 0x1A9F}, {0x1AAE, 0x1AAF}, {0x1AC1, 0x1AFF}, {0x1B4C, 0x1B4F},
+{0x1B7D, 0x1B7F}, {0x1BF4, 0x1BFB}, {0x1C38, 0x1C3A}, {0x1C4A, 0x1C4C}, {0x1C89, 0x1C8F}, {0x1CBB, 0x1CBC}, {0x1CC8, 0x1CCF}, {0x1CFB, 0x1CFF}, {0x1DFA, 0x1DFA}, {0x1F16, 0x1F17}, {0x1F1E, 0x1F1F},
+{0x1F46, 0x1F47}, {0x1F4E, 0x1F4F}, {0x1F58, 0x1F58}, {0x1F5A, 0x1F5A}, {0x1F5C, 0x1F5C}, {0x1F5E, 0x1F5E}, {0x1F7E, 0x1F7F}, {0x1FB5, 0x1FB5}, {0x1FC5, 0x1FC5}, {0x1FD4, 0x1FD5}, {0x1FDC, 0x1FDC},
+{0x1FF0, 0x1FF1}, {0x1FF5, 0x1FF5}, {0x1FFF, 0x1FFF}, {0x200B, 0x200F}, {0x202A, 0x202E}, {0x2060, 0x206F}, {0x2072, 0x2073}, {0x208F, 0x208F}, {0x209D, 0x209F}, {0x20C0, 0x20CF}, {0x20F1, 0x20FF},
+{0x218C, 0x218F}, {0x2427, 0x243F}, {0x244B, 0x245F}, {0x2B74, 0x2B75}, {0x2B96, 0x2B96}, {0x2C2F, 0x2C2F}, {0x2C5F, 0x2C5F}, {0x2CF4, 0x2CF8}, {0x2D26, 0x2D26}, {0x2D28, 0x2D2C}, {0x2D2E, 0x2D2F},
+{0x2D68, 0x2D6E}, {0x2D71, 0x2D7E}, {0x2D97, 0x2D9F}, {0x2DA7, 0x2DA7}, {0x2DAF, 0x2DAF}, {0x2DB7, 0x2DB7}, {0x2DBF, 0x2DBF}, {0x2DC7, 0x2DC7}, {0x2DCF, 0x2DCF}, {0x2DD7, 0x2DD7}, {0x2DDF, 0x2DDF},
+{0x2E53, 0x2E7F}, {0x2E9A, 0x2E9A}, {0x2EF4, 0x2EFF}, {0x2FD6, 0x2FEF}, {0x2FFC, 0x2FFF}, {0x3040, 0x3040}, {0x3097, 0x3098}, {0x3100, 0x3104}, {0x3130, 0x3130}, {0x318F, 0x318F}, {0x31E4, 0x31EF},
+{0x321F, 0x321F}, {0x9FFD, 0x9FFF}, {0xA48D, 0xA48F}, {0xA4C7, 0xA4CF}, {0xA62C, 0xA63F}, {0xA6F8, 0xA6FF}, {0xA7C0, 0xA7C1}, {0xA7CB, 0xA7F4}, {0xA82D, 0xA82F}, {0xA83A, 0xA83F}, {0xA878, 0xA87F},
+{0xA8C6, 0xA8CD}, {0xA8DA, 0xA8DF}, {0xA954, 0xA95E}, {0xA97D, 0xA97F}, {0xA9CE, 0xA9CE}, {0xA9DA, 0xA9DD}, {0xA9FF, 0xA9FF}, {0xAA37, 0xAA3F}, {0xAA4E, 0xAA4F}, {0xAA5A, 0xAA5B}, {0xAAC3, 0xAADA},
+{0xAAF7, 0xAB00}, {0xAB07, 0xAB08}, {0xAB0F, 0xAB10}, {0xAB17, 0xAB1F}, {0xAB27, 0xAB27}, {0xAB2F, 0xAB2F}, {0xAB6C, 0xAB6F}, {0xABEE, 0xABEF}, {0xABFA, 0xABFF}, {0xD7A4, 0xD7AF}, {0xD7C7, 0xD7CA},
+{0xD7FC, 0xF8FF}, {0xFA6E, 0xFA6F}, {0xFADA, 0xFAFF}, {0xFB07, 0xFB12}, {0xFB18, 0xFB1C}, {0xFB37, 0xFB37}, {0xFB3D, 0xFB3D}, {0xFB3F, 0xFB3F}, {0xFB42, 0xFB42}, {0xFB45, 0xFB45}, {0xFBC2, 0xFBD2},
+{0xFD40, 0xFD4F}, {0xFD90, 0xFD91}, {0xFDC8, 0xFDEF}, {0xFDFE, 0xFDFF}, {0xFE1A, 0xFE1F}, {0xFE53, 0xFE53}, {0xFE67, 0xFE67}, {0xFE6C, 0xFE6F}, {0xFE75, 0xFE75}, {0xFEFD, 0xFF00}, {0xFFBF, 0xFFC1},
+{0xFFC8, 0xFFC9}, {0xFFD0, 0xFFD1}, {0xFFD8, 0xFFD9}, {0xFFDD, 0xFFDF}, {0xFFE7, 0xFFE7}, {0xFFEF, 0xFFFB}, {0xFFFE, 0xFFFF}, {0x1000C, 0x1000C}, {0x10027, 0x10027}, {0x1003B, 0x1003B},
+{0x1003E, 0x1003E}, {0x1004E, 0x1004F}, {0x1005E, 0x1007F}, {0x100FB, 0x100FF}, {0x10103, 0x10106}, {0x10134, 0x10136}, {0x1018F, 0x1018F}, {0x1019D, 0x1019F}, {0x101A1, 0x101CF}, {0x101FE, 0x1027F},
+{0x1029D, 0x1029F}, {0x102D1, 0x102DF}, {0x102FC, 0x102FF}, {0x10324, 0x1032C}, {0x1034B, 0x1034F}, {0x1037B, 0x1037F}, {0x1039E, 0x1039E}, {0x103C4, 0x103C7}, {0x103D6, 0x103FF}, {0x1049E, 0x1049F},
+{0x104AA, 0x104AF}, {0x104D4, 0x104D7}, {0x104FC, 0x104FF}, {0x10528, 0x1052F}, {0x10564, 0x1056E}, {0x10570, 0x105FF}, {0x10737, 0x1073F}, {0x10756, 0x1075F}, {0x10768, 0x107FF}, {0x10806, 0x10807},
+{0x10809, 0x10809}, {0x10836, 0x10836}, {0x10839, 0x1083B}, {0x1083D, 0x1083E}, {0x10856, 0x10856}, {0x1089F, 0x108A6}, {0x108B0, 0x108DF}, {0x108F3, 0x108F3}, {0x108F6, 0x108FA}, {0x1091C, 0x1091E},
+{0x1093A, 0x1093E}, {0x10940, 0x1097F}, {0x109B8, 0x109BB}, {0x109D0, 0x109D1}, {0x10A04, 0x10A04}, {0x10A07, 0x10A0B}, {0x10A14, 0x10A14}, {0x10A18, 0x10A18}, {0x10A36, 0x10A37}, {0x10A3B, 0x10A3E},
+{0x10A49, 0x10A4F}, {0x10A59, 0x10A5F}, {0x10AA0, 0x10ABF}, {0x10AE7, 0x10AEA}, {0x10AF7, 0x10AFF}, {0x10B36, 0x10B38}, {0x10B56, 0x10B57}, {0x10B73, 0x10B77}, {0x10B92, 0x10B98}, {0x10B9D, 0x10BA8},
+{0x10BB0, 0x10BFF}, {0x10C49, 0x10C7F}, {0x10CB3, 0x10CBF}, {0x10CF3, 0x10CF9}, {0x10D28, 0x10D2F}, {0x10D3A, 0x10E5F}, {0x10E7F, 0x10E7F}, {0x10EAA, 0x10EAA}, {0x10EAE, 0x10EAF}, {0x10EB2, 0x10EFF},
+{0x10F28, 0x10F2F}, {0x10F5A, 0x10FAF}, {0x10FCC, 0x10FDF}, {0x10FF7, 0x10FFF}, {0x1104E, 0x11051}, {0x11070, 0x1107E}, {0x110BD, 0x110BD}, {0x110C2, 0x110CF}, {0x110E9, 0x110EF}, {0x110FA, 0x110FF},
+{0x11135, 0x11135}, {0x11148, 0x1114F}, {0x11177, 0x1117F}, {0x111E0, 0x111E0}, {0x111F5, 0x111FF}, {0x11212, 0x11212}, {0x1123F, 0x1127F}, {0x11287, 0x11287}, {0x11289, 0x11289}, {0x1128E, 0x1128E},
+{0x1129E, 0x1129E}, {0x112AA, 0x112AF}, {0x112EB, 0x112EF}, {0x112FA, 0x112FF}, {0x11304, 0x11304}, {0x1130D, 0x1130E}, {0x11311, 0x11312}, {0x11329, 0x11329}, {0x11331, 0x11331}, {0x11334, 0x11334},
+{0x1133A, 0x1133A}, {0x11345, 0x11346}, {0x11349, 0x1134A}, {0x1134E, 0x1134F}, {0x11351, 0x11356}, {0x11358, 0x1135C}, {0x11364, 0x11365}, {0x1136D, 0x1136F}, {0x11375, 0x113FF}, {0x1145C, 0x1145C},
+{0x11462, 0x1147F}, {0x114C8, 0x114CF}, {0x114DA, 0x1157F}, {0x115B6, 0x115B7}, {0x115DE, 0x115FF}, {0x11645, 0x1164F}, {0x1165A, 0x1165F}, {0x1166D, 0x1167F}, {0x116B9, 0x116BF}, {0x116CA, 0x116FF},
+{0x1171B, 0x1171C}, {0x1172C, 0x1172F}, {0x11740, 0x117FF}, {0x1183C, 0x1189F}, {0x118F3, 0x118FE}, {0x11907, 0x11908}, {0x1190A, 0x1190B}, {0x11914, 0x11914}, {0x11917, 0x11917}, {0x11936, 0x11936},
+{0x11939, 0x1193A}, {0x11947, 0x1194F}, {0x1195A, 0x1199F}, {0x119A8, 0x119A9}, {0x119D8, 0x119D9}, {0x119E5, 0x119FF}, {0x11A48, 0x11A4F}, {0x11AA3, 0x11ABF}, {0x11AF9, 0x11BFF}, {0x11C09, 0x11C09},
+{0x11C37, 0x11C37}, {0x11C46, 0x11C4F}, {0x11C6D, 0x11C6F}, {0x11C90, 0x11C91}, {0x11CA8, 0x11CA8}, {0x11CB7, 0x11CFF}, {0x11D07, 0x11D07}, {0x11D0A, 0x11D0A}, {0x11D37, 0x11D39}, {0x11D3B, 0x11D3B},
+{0x11D3E, 0x11D3E}, {0x11D48, 0x11D4F}, {0x11D5A, 0x11D5F}, {0x11D66, 0x11D66}, {0x11D69, 0x11D69}, {0x11D8F, 0x11D8F}, {0x11D92, 0x11D92}, {0x11D99, 0x11D9F}, {0x11DAA, 0x11EDF}, {0x11EF9, 0x11FAF},
+{0x11FB1, 0x11FBF}, {0x11FF2, 0x11FFE}, {0x1239A, 0x123FF}, {0x1246F, 0x1246F}, {0x12475, 0x1247F}, {0x12544, 0x12FFF}, {0x1342F, 0x143FF}, {0x14647, 0x167FF}, {0x16A39, 0x16A3F}, {0x16A5F, 0x16A5F},
+{0x16A6A, 0x16A6D}, {0x16A70, 0x16ACF}, {0x16AEE, 0x16AEF}, {0x16AF6, 0x16AFF}, {0x16B46, 0x16B4F}, {0x16B5A, 0x16B5A}, {0x16B62, 0x16B62}, {0x16B78, 0x16B7C}, {0x16B90, 0x16E3F}, {0x16E9B, 0x16EFF},
+{0x16F4B, 0x16F4E}, {0x16F88, 0x16F8E}, {0x16FA0, 0x16FDF}, {0x16FE5, 0x16FEF}, {0x16FF2, 0x16FFF}, {0x187F8, 0x187FF}, {0x18CD6, 0x18CFF}, {0x18D09, 0x1AFFF}, {0x1B11F, 0x1B14F}, {0x1B153, 0x1B163},
+{0x1B168, 0x1B16F}, {0x1B2FC, 0x1BBFF}, {0x1BC6B, 0x1BC6F}, {0x1BC7D, 0x1BC7F}, {0x1BC89, 0x1BC8F}, {0x1BC9A, 0x1BC9B}, {0x1BCA0, 0x1CFFF}, {0x1D0F6, 0x1D0FF}, {0x1D127, 0x1D128}, {0x1D173, 0x1D17A},
+{0x1D1E9, 0x1D1FF}, {0x1D246, 0x1D2DF}, {0x1D2F4, 0x1D2FF}, {0x1D357, 0x1D35F}, {0x1D379, 0x1D3FF}, {0x1D455, 0x1D455}, {0x1D49D, 0x1D49D}, {0x1D4A0, 0x1D4A1}, {0x1D4A3, 0x1D4A4}, {0x1D4A7, 0x1D4A8},
+{0x1D4AD, 0x1D4AD}, {0x1D4BA, 0x1D4BA}, {0x1D4BC, 0x1D4BC}, {0x1D4C4, 0x1D4C4}, {0x1D506, 0x1D506}, {0x1D50B, 0x1D50C}, {0x1D515, 0x1D515}, {0x1D51D, 0x1D51D}, {0x1D53A, 0x1D53A}, {0x1D53F, 0x1D53F},
+{0x1D545, 0x1D545}, {0x1D547, 0x1D549}, {0x1D551, 0x1D551}, {0x1D6A6, 0x1D6A7}, {0x1D7CC, 0x1D7CD}, {0x1DA8C, 0x1DA9A}, {0x1DAA0, 0x1DAA0}, {0x1DAB0, 0x1DFFF}, {0x1E007, 0x1E007}, {0x1E019, 0x1E01A},
+{0x1E022, 0x1E022}, {0x1E025, 0x1E025}, {0x1E02B, 0x1E0FF}, {0x1E12D, 0x1E12F}, {0x1E13E, 0x1E13F}, {0x1E14A, 0x1E14D}, {0x1E150, 0x1E2BF}, {0x1E2FA, 0x1E2FE}, {0x1E300, 0x1E7FF}, {0x1E8C5, 0x1E8C6},
+{0x1E8D7, 0x1E8FF}, {0x1E94C, 0x1E94F}, {0x1E95A, 0x1E95D}, {0x1E960, 0x1EC70}, {0x1ECB5, 0x1ED00}, {0x1ED3E, 0x1EDFF}, {0x1EE04, 0x1EE04}, {0x1EE20, 0x1EE20}, {0x1EE23, 0x1EE23}, {0x1EE25, 0x1EE26},
+{0x1EE28, 0x1EE28}, {0x1EE33, 0x1EE33}, {0x1EE38, 0x1EE38}, {0x1EE3A, 0x1EE3A}, {0x1EE3C, 0x1EE41}, {0x1EE43, 0x1EE46}, {0x1EE48, 0x1EE48}, {0x1EE4A, 0x1EE4A}, {0x1EE4C, 0x1EE4C}, {0x1EE50, 0x1EE50},
+{0x1EE53, 0x1EE53}, {0x1EE55, 0x1EE56}, {0x1EE58, 0x1EE58}, {0x1EE5A, 0x1EE5A}, {0x1EE5C, 0x1EE5C}, {0x1EE5E, 0x1EE5E}, {0x1EE60, 0x1EE60}, {0x1EE63, 0x1EE63}, {0x1EE65, 0x1EE66}, {0x1EE6B, 0x1EE6B},
+{0x1EE73, 0x1EE73}, {0x1EE78, 0x1EE78}, {0x1EE7D, 0x1EE7D}, {0x1EE7F, 0x1EE7F}, {0x1EE8A, 0x1EE8A}, {0x1EE9C, 0x1EEA0}, {0x1EEA4, 0x1EEA4}, {0x1EEAA, 0x1EEAA}, {0x1EEBC, 0x1EEEF}, {0x1EEF2, 0x1EFFF},
+{0x1F02C, 0x1F02F}, {0x1F094, 0x1F09F}, {0x1F0AF, 0x1F0B0}, {0x1F0C0, 0x1F0C0}, {0x1F0D0, 0x1F0D0}, {0x1F0F6, 0x1F0FF}, {0x1F1AE, 0x1F1E5}, {0x1F203, 0x1F20F}, {0x1F23C, 0x1F23F}, {0x1F249, 0x1F24F},
+{0x1F252, 0x1F25F}, {0x1F266, 0x1F2FF}, {0x1F6D8, 0x1F6DF}, {0x1F6ED, 0x1F6EF}, {0x1F6FD, 0x1F6FF}, {0x1F774, 0x1F77F}, {0x1F7D9, 0x1F7DF}, {0x1F7EC, 0x1F7FF}, {0x1F80C, 0x1F80F}, {0x1F848, 0x1F84F},
+{0x1F85A, 0x1F85F}, {0x1F888, 0x1F88F}, {0x1F8AE, 0x1F8AF}, {0x1F8B2, 0x1F8FF}, {0x1F979, 0x1F979}, {0x1F9CC, 0x1F9CC}, {0x1FA54, 0x1FA5F}, {0x1FA6E, 0x1FA6F}, {0x1FA75, 0x1FA77}, {0x1FA7B, 0x1FA7F},
+{0x1FA87, 0x1FA8F}, {0x1FAA9, 0x1FAAF}, {0x1FAB7, 0x1FABF}, {0x1FAC3, 0x1FACF}, {0x1FAD7, 0x1FAFF}, {0x1FB93, 0x1FB93}, {0x1FBCB, 0x1FBEF}, {0x1FBFA, 0x1FFFF}, {0x2A6DE, 0x2A6FF}, {0x2B735, 0x2B73F},
+{0x2B81E, 0x2B81F}, {0x2CEA2, 0x2CEAF}, {0x2EBE1, 0x2F7FF}, {0x2FA1E, 0x2FFFF}, {0x3134B, 0xE00FF}, {0xE01F0, 0x10FFFF},
+};
+
+static std::string codepoint_to_utf8(uint32_t cp) {
+    std::string result;
+    if (/* 0x00 <= cp && */ cp <= 0x7f) {
+        result.push_back(cp);
+    }
+    else if (0x80 <= cp && cp <= 0x7ff) {
+        result.push_back(0xc0 | ((cp >> 6) & 0x1f));
+        result.push_back(0x80 | (cp & 0x3f));
+    }
+    else if (0x800 <= cp && cp <= 0xffff) {
+        result.push_back(0xe0 | ((cp >> 12) & 0x0f));
+        result.push_back(0x80 | ((cp >> 6) & 0x3f));
+        result.push_back(0x80 | (cp & 0x3f));
+    }
+    else if (0x10000 <= cp && cp <= 0x10ffff) {
+        result.push_back(0xf0 | ((cp >> 18) & 0x07));
+        result.push_back(0x80 | ((cp >> 12) & 0x3f));
+        result.push_back(0x80 | ((cp >> 6) & 0x3f));
+        result.push_back(0x80 | (cp & 0x3f));
+    }
+    else {
+        throw std::invalid_argument("invalid codepoint");
+    }
+    return result;
+}
+
+static std::string codepoints_to_utf8(const std::vector<uint32_t> & cps) {
+    std::string result;
+    for (size_t i = 0; i < cps.size(); ++i) {
+        result.append(codepoint_to_utf8(cps[i]));
+    }
+    return result;
+}
+
+static uint32_t codepoint_from_utf8(const std::string & utf8, size_t & offset) {
+    assert(offset < utf8.size());
+    if (!(utf8[offset + 0] & 0x80)) {
+        auto result = utf8[offset + 0];
+        offset += 1;
+        return result;
+    }
+    else if (!(utf8[offset + 0] & 0x40)) {
+        throw std::invalid_argument("invalid character");
+    }
+    else if (!(utf8[offset + 0] & 0x20)) {
+        if (offset + 1 >= utf8.size() || ! ((utf8[offset + 1] & 0xc0) == 0x80))
+            throw std::invalid_argument("invalid character");
+        auto result = ((utf8[offset + 0] & 0x1f) << 6) | (utf8[offset + 1] & 0x3f);
+        offset += 2;
+        return result;
+    }
+    else if (!(utf8[offset + 0] & 0x10)) {
+        if (offset + 2 >= utf8.size() || ! ((utf8[offset + 1] & 0xc0) == 0x80) || ! ((utf8[offset + 2] & 0xc0) == 0x80))
+            throw std::invalid_argument("invalid character");
+        auto result = ((utf8[offset + 0] & 0x0f) << 12) | ((utf8[offset + 1] & 0x3f) << 6) | (utf8[offset + 2] & 0x3f);
+        offset += 3;
+        return result;
+    }
+    else if (!(utf8[offset + 0] & 0x08)) {
+        if (offset + 3 >= utf8.size() || ! ((utf8[offset + 1] & 0xc0) == 0x80) || ! ((utf8[offset + 2] & 0xc0) == 0x80) || !((utf8[offset + 3] & 0xc0) == 0x80))
+            throw std::invalid_argument("invalid character");
+        auto result = ((utf8[offset + 0] & 0x07) << 18) | ((utf8[offset + 1] & 0x3f) << 12) | ((utf8[offset + 2] & 0x3f) << 6) | (utf8[offset + 3] & 0x3f);
+        offset += 4;
+        return result;
+    }
+    throw std::invalid_argument("invalid string");
+}
+
+static std::vector<uint32_t> codepoints_from_utf8(const std::string & utf8) {
+    std::vector<uint32_t> result;
+    size_t offset = 0;
+    while (offset < utf8.size()) {
+        result.push_back(codepoint_from_utf8(utf8, offset));
+    }
+    return result;
+}
+
+static std::vector<uint16_t> codepoint_to_utf16(uint32_t cp) {
+    std::vector<uint16_t> result;
+    if (/* 0x0000 <= cp && */ cp <= 0xffff) {
+        result.emplace_back(cp);
+    }
+    else if (0x10000 <= cp && cp <= 0x10ffff) {
+        result.emplace_back(0xd800 | ((cp - 0x10000) >> 10));
+        result.emplace_back(0xdc00 | ((cp - 0x10000) & 0x03ff));
+    }
+    else {
+        throw std::invalid_argument("invalid codepoint");
+    }
+    return result;
+}
+
+static std::vector<uint16_t> codepoints_to_utf16(const std::vector<uint32_t> & cps) {
+    std::vector<uint16_t> result;
+    for (size_t i = 0; i < cps.size(); ++i) {
+        auto temp = codepoint_to_utf16(cps[i]);
+        result.insert(result.end(), temp.begin(), temp.end());
+    }
+    return result;
+}
+
+static uint32_t codepoint_from_utf16(const std::vector<uint16_t> & utf16, size_t & offset) {
+    assert(offset < utf16.size());
+    if (((utf16[0] >> 10) << 10) != 0xd800) {
+        auto result = utf16[offset + 0];
+        offset += 1;
+        return result;
+    }
+    else {
+        if (offset + 1 >= utf16.size() || !((utf16[1] & 0xdc00) == 0xdc00))
+            throw std::invalid_argument("invalid character");
+        auto result = 0x10000 + (((utf16[0] & 0x03ff) << 10) | (utf16[1] & 0x03ff));
+        offset += 2;
+        return result;
+    }
+    throw std::invalid_argument("invalid string");
+}
+
+static std::vector<uint32_t> codepoints_from_utf16(const std::vector<uint16_t> & utf16) {
+    std::vector<uint32_t> result;
+    size_t offset = 0;
+    while (offset < utf16.size())
+        result.push_back(codepoint_from_utf16(utf16, offset));
+    return result;
+}
+
+#define CODEPOINT_TYPE_UNIDENTIFIED 0
+#define CODEPOINT_TYPE_DIGIT 1
+#define CODEPOINT_TYPE_LETTER 2
+#define CODEPOINT_TYPE_WHITESPACE 3
+#define CODEPOINT_TYPE_ACCENT_MARK 4
+#define CODEPOINT_TYPE_PUNCTUATION 5
+#define CODEPOINT_TYPE_SYMBOL 6
+#define CODEPOINT_TYPE_CONTROL 7
+
+static std::unordered_map<uint32_t, int> codepoint_type_map() {
+    std::unordered_map<uint32_t, int> codepoint_types;
+    for (auto p : digit_ranges) {
+        for(auto i = p.first; i <= p.second; ++ i)
+            codepoint_types[i] = CODEPOINT_TYPE_DIGIT;
+    }
+    for(auto p : letter_ranges) {
+        for(auto i = p.first; i <= p.second; ++ i)
+            codepoint_types[i] = CODEPOINT_TYPE_LETTER;
+    }
+    for(auto p : whitespace_ranges) {
+        for(auto i = p.first; i <= p.second; ++ i)
+            codepoint_types[i] = CODEPOINT_TYPE_WHITESPACE;
+    }
+    for(auto p : accent_mark_ranges) {
+        for(auto i = p.first; i <= p.second; ++ i)
+            codepoint_types[i] = CODEPOINT_TYPE_ACCENT_MARK;
+    }
+    for(auto p : punctuation_ranges) {
+        for(auto i = p.first; i <= p.second; ++ i)
+            codepoint_types[i] = CODEPOINT_TYPE_PUNCTUATION;
+    }
+    for (auto p : symbol_ranges) {
+        for (auto i = p.first; i <= p.second; ++i)
+            codepoint_types[i] = CODEPOINT_TYPE_SYMBOL;
+    }
+    for(auto p : control_ranges) {
+        for(auto i = p.first; i <= p.second; ++ i)
+            codepoint_types[i] = CODEPOINT_TYPE_CONTROL;
+    }
+    return codepoint_types;
+}
+
+static int codepoint_type(uint32_t cp) {
+    static std::unordered_map<uint32_t, int> codepoint_types = codepoint_type_map();
+    return codepoint_types[cp];
+}
+
+static int codepoint_type(const std::string & utf8) {
+    if (utf8.length() == 0)
+        return CODEPOINT_TYPE_UNIDENTIFIED;
+    size_t offset = 0;
+    return codepoint_type(codepoint_from_utf8(utf8, offset));
+}
+
+static std::unordered_map<uint8_t, std::string> bytes_to_unicode_map_bpe() {
+    std::unordered_map<uint8_t, std::string> map;
+    for (int ch = u'!'; ch <= u'~'; ++ch) {
+        assert(0 <= ch && ch < 256);
+        map[ch] = codepoint_to_utf8(ch);
+    }
+    for (int ch = u'¡'; ch <= u'¬'; ++ch) {
+        assert(0 <= ch && ch < 256);
+        map[ch] = codepoint_to_utf8(ch);
+    }
+    for (int ch = u'®'; ch <= u'ÿ'; ++ch) {
+        assert(0 <= ch && ch < 256);
+        map[ch] = codepoint_to_utf8(ch);
+    }
+    auto n = 0;
+    for (int ch = 0; ch < 256; ++ch) {
+        if (map.find(ch) == map.end()) {
+            map[ch] = codepoint_to_utf8(256 + n);
+            ++n;
+        }
+    }
+    return map;
+}
+
+static std::string bytes_to_unicode_bpe(uint8_t byte) {
+    static std::unordered_map<uint8_t, std::string> map = bytes_to_unicode_map_bpe();
+    return map.at(byte);
+}
+
+static std::unordered_map<std::string, uint8_t> unicode_to_bytes_map_bpe() {
+    std::unordered_map<std::string, uint8_t> map;
+    for (int ch = u'!'; ch <= u'~'; ++ch) {
+        assert(0 <= ch && ch < 256);
+        map[codepoint_to_utf8(ch)] = ch;
+    }
+    for (int ch = u'¡'; ch <= u'¬'; ++ch) {
+        assert(0 <= ch && ch < 256);
+        map[codepoint_to_utf8(ch)] = ch;
+    }
+    for (int ch = u'®'; ch <= u'ÿ'; ++ch) {
+        assert(0 <= ch && ch < 256);
+        map[codepoint_to_utf8(ch)] = ch;
+    }
+    auto n = 0;
+    for (int ch = 0; ch < 256; ++ch) {
+        if (map.find(codepoint_to_utf8(ch)) == map.end()) {
+            map[codepoint_to_utf8(256 + n)] = ch;
+            ++n;
+        }
+    }
+    return map;
+}
+
+static uint8_t unicode_to_bytes_bpe(const std::string & utf8) {
+    static std::unordered_map<std::string, uint8_t> map = unicode_to_bytes_map_bpe();
+    return map.at(utf8);
+}
+

examples/whisper.android/app/src/main/jni/whisper/CMakeLists.txt

@@ -9,6 +9,8 @@ set(
         SOURCE_FILES
         ${WHISPER_LIB_DIR}/ggml.c
         ${WHISPER_LIB_DIR}/ggml-alloc.c
+        ${WHISPER_LIB_DIR}/ggml-backend.c
+        ${WHISPER_LIB_DIR}/ggml-quants.c
         ${WHISPER_LIB_DIR}/whisper.cpp
         ${CMAKE_SOURCE_DIR}/jni.c
 )

examples/whisper.objc/whisper.objc.xcodeproj/project.pbxproj

@@ -20,6 +20,8 @@
 		18627C9429052C4900BD2A04 /* whisper.cpp in Sources */ = {isa = PBXBuildFile; fileRef = 18627C9329052C4900BD2A04 /* whisper.cpp */; settings = {COMPILER_FLAGS = "-DWHISPER_USE_COREML"; }; };
 		18627C9629052C5800BD2A04 /* ggml.c in Sources */ = {isa = PBXBuildFile; fileRef = 18627C9529052C5800BD2A04 /* ggml.c */; settings = {COMPILER_FLAGS = "-DGGML_USE_ACCELERATE"; }; };
 		18627C9B29052CFF00BD2A04 /* ggml-base.en.bin in Resources */ = {isa = PBXBuildFile; fileRef = 18627C9A29052CFF00BD2A04 /* ggml-base.en.bin */; };
+		18ABE15A2AF556340044A204 /* ggml-backend.c in Sources */ = {isa = PBXBuildFile; fileRef = 18ABE1572AF556340044A204 /* ggml-backend.c */; };
+		18ABE15B2AF556340044A204 /* ggml-quants.c in Sources */ = {isa = PBXBuildFile; fileRef = 18ABE1592AF556340044A204 /* ggml-quants.c */; };
 		7FE3424B2A0C3FA20015A058 /* whisper-encoder-impl.m in Sources */ = {isa = PBXBuildFile; fileRef = 7FE342452A0C3FA20015A058 /* whisper-encoder-impl.m */; };
 		7FE3424C2A0C3FA20015A058 /* whisper-encoder.mm in Sources */ = {isa = PBXBuildFile; fileRef = 7FE342472A0C3FA20015A058 /* whisper-encoder.mm */; };
 		7FE3424D2A0C3FA20015A058 /* whisper-decoder-impl.m in Sources */ = {isa = PBXBuildFile; fileRef = 7FE3424A2A0C3FA20015A058 /* whisper-decoder-impl.m */; };
@@ -61,6 +63,12 @@
 		18627C9529052C5800BD2A04 /* ggml.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; name = ggml.c; path = ../../../ggml.c; sourceTree = "<group>"; };
 		18627C9729052C6600BD2A04 /* ggml.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = ggml.h; path = ../../../ggml.h; sourceTree = "<group>"; };
 		18627C9A29052CFF00BD2A04 /* ggml-base.en.bin */ = {isa = PBXFileReference; lastKnownFileType = archive.macbinary; name = "ggml-base.en.bin"; path = "../../../models/ggml-base.en.bin"; sourceTree = "<group>"; };
+		18ABE1542AF556340044A204 /* ggml-quants.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = "ggml-quants.h"; path = "../../../ggml-quants.h"; sourceTree = "<group>"; };
+		18ABE1552AF556340044A204 /* ggml-backend.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = "ggml-backend.h"; path = "../../../ggml-backend.h"; sourceTree = "<group>"; };
+		18ABE1562AF556340044A204 /* ggml-backend-impl.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = "ggml-backend-impl.h"; path = "../../../ggml-backend-impl.h"; sourceTree = "<group>"; };
+		18ABE1572AF556340044A204 /* ggml-backend.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; name = "ggml-backend.c"; path = "../../../ggml-backend.c"; sourceTree = "<group>"; };
+		18ABE1582AF556340044A204 /* ggml-impl.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = "ggml-impl.h"; path = "../../../ggml-impl.h"; sourceTree = "<group>"; };
+		18ABE1592AF556340044A204 /* ggml-quants.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; name = "ggml-quants.c"; path = "../../../ggml-quants.c"; sourceTree = "<group>"; };
 		7FE342452A0C3FA20015A058 /* whisper-encoder-impl.m */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.objc; path = "whisper-encoder-impl.m"; sourceTree = "<group>"; };
 		7FE342462A0C3FA20015A058 /* whisper-encoder.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; path = "whisper-encoder.h"; sourceTree = "<group>"; };
 		7FE342472A0C3FA20015A058 /* whisper-encoder.mm */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.cpp.objcpp; path = "whisper-encoder.mm"; sourceTree = "<group>"; };
@@ -100,6 +108,12 @@
 		18627C7829052BDF00BD2A04 /* whisper.objc */ = {
 			isa = PBXGroup;
 			children = (
+				18ABE1562AF556340044A204 /* ggml-backend-impl.h */,
+				18ABE1572AF556340044A204 /* ggml-backend.c */,
+				18ABE1552AF556340044A204 /* ggml-backend.h */,
+				18ABE1582AF556340044A204 /* ggml-impl.h */,
+				18ABE1592AF556340044A204 /* ggml-quants.c */,
+				18ABE1542AF556340044A204 /* ggml-quants.h */,
 				1844471D2AB2195F007D6BFE /* ggml-metal.metal */,
 				1844471B2AB21655007D6BFE /* ggml-metal.m */,
 				184447182AB211A2007D6BFE /* ggml-alloc.c */,
@@ -214,12 +228,14 @@
 			buildActionMask = 2147483647;
 			files = (
 				18627C8129052BDF00BD2A04 /* ViewController.m in Sources */,
+				18ABE15B2AF556340044A204 /* ggml-quants.c in Sources */,
 				7FE3424C2A0C3FA20015A058 /* whisper-encoder.mm in Sources */,
 				18627C9429052C4900BD2A04 /* whisper.cpp in Sources */,
 				18627C9629052C5800BD2A04 /* ggml.c in Sources */,
 				18627C7B29052BDF00BD2A04 /* AppDelegate.m in Sources */,
 				7FE3424D2A0C3FA20015A058 /* whisper-decoder-impl.m in Sources */,
 				1844471A2AB211A2007D6BFE /* ggml-alloc.c in Sources */,
+				18ABE15A2AF556340044A204 /* ggml-backend.c in Sources */,
 				18627C8C29052BE000BD2A04 /* main.m in Sources */,
 				18627C7E29052BDF00BD2A04 /* SceneDelegate.m in Sources */,
 				1844471C2AB21655007D6BFE /* ggml-metal.m in Sources */,

examples/whisper.swiftui/whisper.swiftui.xcodeproj/project.pbxproj

@@ -20,6 +20,8 @@
 		0AAC5DCC29539EB1003032C3 /* ggml.c in Sources */ = {isa = PBXBuildFile; fileRef = 0AAC5DC929539EB0003032C3 /* ggml.c */; settings = {COMPILER_FLAGS = "-DGGML_USE_ACCELERATE -Wno-shorten-64-to-32"; }; };
 		0AAC5DCE2953A05C003032C3 /* WhisperState.swift in Sources */ = {isa = PBXBuildFile; fileRef = 0AAC5DCD2953A05C003032C3 /* WhisperState.swift */; };
 		0AAC5DD12953A394003032C3 /* LibWhisper.swift in Sources */ = {isa = PBXBuildFile; fileRef = 0AAC5DD02953A394003032C3 /* LibWhisper.swift */; };
+		18ABE1522AF555FA0044A204 /* ggml-backend.c in Sources */ = {isa = PBXBuildFile; fileRef = 18ABE14C2AF555FA0044A204 /* ggml-backend.c */; };
+		18ABE1532AF555FA0044A204 /* ggml-quants.c in Sources */ = {isa = PBXBuildFile; fileRef = 18ABE1512AF555FA0044A204 /* ggml-quants.c */; };
 		18AED4812AB21F2B009D854F /* ggml-alloc.c in Sources */ = {isa = PBXBuildFile; fileRef = 18AED47F2AB21F2B009D854F /* ggml-alloc.c */; };
 /* End PBXBuildFile section */
 
@@ -42,6 +44,12 @@
 		0AAC5DCA29539EB0003032C3 /* ggml.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; path = ggml.h; sourceTree = "<group>"; };
 		0AAC5DCD2953A05C003032C3 /* WhisperState.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = WhisperState.swift; sourceTree = "<group>"; };
 		0AAC5DD02953A394003032C3 /* LibWhisper.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = LibWhisper.swift; sourceTree = "<group>"; };
+		18ABE14C2AF555FA0044A204 /* ggml-backend.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; path = "ggml-backend.c"; sourceTree = "<group>"; };
+		18ABE14D2AF555FA0044A204 /* ggml-backend.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; path = "ggml-backend.h"; sourceTree = "<group>"; };
+		18ABE14E2AF555FA0044A204 /* ggml-backend-impl.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; path = "ggml-backend-impl.h"; sourceTree = "<group>"; };
+		18ABE14F2AF555FA0044A204 /* ggml-quants.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; path = "ggml-quants.h"; sourceTree = "<group>"; };
+		18ABE1502AF555FA0044A204 /* ggml-impl.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; path = "ggml-impl.h"; sourceTree = "<group>"; };
+		18ABE1512AF555FA0044A204 /* ggml-quants.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; path = "ggml-quants.c"; sourceTree = "<group>"; };
 		18AED47F2AB21F2B009D854F /* ggml-alloc.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; path = "ggml-alloc.c"; sourceTree = "<group>"; };
 		18AED4802AB21F2B009D854F /* ggml-alloc.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; path = "ggml-alloc.h"; sourceTree = "<group>"; };
 /* End PBXFileReference section */
@@ -127,6 +135,12 @@
 		0AAC5DC529539E89003032C3 /* whisper.cpp */ = {
 			isa = PBXGroup;
 			children = (
+				18ABE14E2AF555FA0044A204 /* ggml-backend-impl.h */,
+				18ABE14C2AF555FA0044A204 /* ggml-backend.c */,
+				18ABE14D2AF555FA0044A204 /* ggml-backend.h */,
+				18ABE1502AF555FA0044A204 /* ggml-impl.h */,
+				18ABE1512AF555FA0044A204 /* ggml-quants.c */,
+				18ABE14F2AF555FA0044A204 /* ggml-quants.h */,
 				18AED47F2AB21F2B009D854F /* ggml-alloc.c */,
 				18AED4802AB21F2B009D854F /* ggml-alloc.h */,
 				0AAC5DC929539EB0003032C3 /* ggml.c */,
@@ -242,12 +256,14 @@
 				0AAC5D9D29539CCF003032C3 /* ContentView.swift in Sources */,
 				0AAC5D9B29539CCF003032C3 /* WhisperCppDemoApp.swift in Sources */,
 				0AAC5DCC29539EB1003032C3 /* ggml.c in Sources */,
+				18ABE1532AF555FA0044A204 /* ggml-quants.c in Sources */,
 				0AAC5DCE2953A05C003032C3 /* WhisperState.swift in Sources */,
 				0AAC5DD12953A394003032C3 /* LibWhisper.swift in Sources */,
 				0AA7514C2953B569001EE061 /* RiffWaveUtils.swift in Sources */,
 				0AAC5DCB29539EB1003032C3 /* whisper.cpp in Sources */,
 				0AA7514E2953D958001EE061 /* Recorder.swift in Sources */,
 				18AED4812AB21F2B009D854F /* ggml-alloc.c in Sources */,
+				18ABE1522AF555FA0044A204 /* ggml-backend.c in Sources */,
 			);
 			runOnlyForDeploymentPostprocessing = 0;
 		};

extra/sync-ggml.sh

@@ -1,20 +1,30 @@
 #!/bin/bash
 
-cp -rpv ../ggml/src/ggml.c                ./ggml.c
-cp -rpv ../ggml/src/ggml-alloc.c          ./ggml-alloc.c
-cp -rpv ../ggml/src/ggml-cuda.h           ./ggml-cuda.h
-cp -rpv ../ggml/src/ggml-cuda.cu          ./ggml-cuda.cu
-cp -rpv ../ggml/src/ggml-opencl.h         ./ggml-opencl.h
-cp -rpv ../ggml/src/ggml-opencl.cpp       ./ggml-opencl.cpp
-cp -rpv ../ggml/src/ggml-metal.h          ./ggml-metal.h
-cp -rpv ../ggml/src/ggml-metal.m          ./ggml-metal.m
-cp -rpv ../ggml/src/ggml-metal.metal      ./ggml-metal.metal
-cp -rpv ../ggml/include/ggml/ggml.h       ./ggml.h
-cp -rpv ../ggml/include/ggml/ggml-alloc.h ./ggml-alloc.h
-cp -rpv ../ggml/examples/common.h         ./examples/common.h
-cp -rpv ../ggml/examples/common.cpp       ./examples/common.cpp
-cp -rpv ../ggml/examples/common-ggml.h    ./examples/common-ggml.h
-cp -rpv ../ggml/examples/common-ggml.cpp  ./examples/common-ggml.cpp
+cp -rpv ../ggml/src/ggml.c              ./ggml.c
+cp -rpv ../ggml/src/ggml-alloc.c        ./ggml-alloc.c
+cp -rpv ../ggml/src/ggml-backend-impl.h ./ggml-backend-impl.h
+cp -rpv ../ggml/src/ggml-backend.c      ./ggml-backend.c
+cp -rpv ../ggml/src/ggml-cuda.cu        ./ggml-cuda.cu
+cp -rpv ../ggml/src/ggml-cuda.h         ./ggml-cuda.h
+cp -rpv ../ggml/src/ggml-impl.h         ./ggml-impl.h
+cp -rpv ../ggml/src/ggml-metal.h        ./ggml-metal.h
+cp -rpv ../ggml/src/ggml-metal.m        ./ggml-metal.m
+cp -rpv ../ggml/src/ggml-metal.metal    ./ggml-metal.metal
+#cp -rpv ../ggml/src/ggml-mpi.h          ./ggml-mpi.h
+#cp -rpv ../ggml/src/ggml-mpi.c          ./ggml-mpi.c
+cp -rpv ../ggml/src/ggml-opencl.cpp     ./ggml-opencl.cpp
+cp -rpv ../ggml/src/ggml-opencl.h       ./ggml-opencl.h
+cp -rpv ../ggml/src/ggml-quants.c       ./ggml-quants.c
+cp -rpv ../ggml/src/ggml-quants.h       ./ggml-quants.h
+
+cp -rpv ../ggml/include/ggml/ggml.h         ./ggml.h
+cp -rpv ../ggml/include/ggml/ggml-alloc.h   ./ggml-alloc.h
+cp -rpv ../ggml/include/ggml/ggml-backend.h ./ggml-backend.h
+
+cp -rpv ../ggml/examples/common.h        ./examples/common.h
+cp -rpv ../ggml/examples/common.cpp      ./examples/common.cpp
+cp -rpv ../ggml/examples/common-ggml.h   ./examples/common-ggml.h
+cp -rpv ../ggml/examples/common-ggml.cpp ./examples/common-ggml.cpp
 
 cp -rpv ../ggml/examples/whisper/whisper.h    ./whisper.h
 cp -rpv ../ggml/examples/whisper/whisper.cpp  ./whisper.cpp

ggml-alloc.c

@@ -1,69 +1,21 @@
 #include "ggml-alloc.h"
+#include "ggml-backend-impl.h"
 #include "ggml.h"
+#include "ggml-impl.h"
 #include <assert.h>
+#include <limits.h>
 #include <stdarg.h>
 #include <stdio.h>
 #include <stdlib.h>
 #include <string.h>
 
-#ifdef __has_include
-    #if __has_include(<unistd.h>)
-        #include <unistd.h>
-        #if defined(_POSIX_MAPPED_FILES)
-            #include <sys/types.h>
-            #include <sys/mman.h>
-        #endif
-    #endif
-#endif
-
-#if defined(_WIN32)
-    #define WIN32_LEAN_AND_MEAN
-    #ifndef NOMINMAX
-        #define NOMINMAX
-    #endif
-    #include <windows.h>
-    #include <memoryapi.h>
-#endif
-
-
-#define UNUSED(x) (void)(x)
 #define MAX(a, b) ((a) > (b) ? (a) : (b))
-#define GGML_MAX_CONCUR (2*GGML_MAX_NODES)
+#define MAX_FREE_BLOCKS 256
 
 //#define GGML_ALLOCATOR_DEBUG
 
-//#define AT_PRINTF printf
-#define AT_PRINTF(...) ((void)0)
-
-struct hash_node {
-    struct ggml_tensor * t;
-    int n_children;
-    int n_views;
-};
-
-static size_t hash(void * p) {
-    return (size_t)p % GGML_GRAPH_HASHTABLE_SIZE;
-}
-
-static struct hash_node * hash_get(struct hash_node hash_table[], struct ggml_tensor * t) {
-    size_t h = hash(t);
-
-    // linear probing
-    size_t i = h;
-    while (hash_table[i].t != NULL) {
-        if (hash_table[i].t == t) {
-            return &hash_table[i];
-        }
-        i = (i + 1) % GGML_GRAPH_HASHTABLE_SIZE;
-        if (i == h) {
-            // hash table is full
-            GGML_ASSERT(false);
-        }
-    }
-
-    hash_table[i].t = t;
-    return &hash_table[i];
-}
+//#define AT_PRINTF(...) fprintf(stderr, __VA_ARGS__)
+#define AT_PRINTF(...)
 
 // TODO: GGML_PAD ?
 static size_t aligned_offset(const void * buffer, size_t offset, size_t alignment) {
@@ -77,19 +29,18 @@ struct free_block {
     size_t size;
 };
 
-#define MAX_FREE_BLOCKS 128
-
-struct ggml_allocr {
-    void * data;
-    size_t size;
+struct ggml_tallocr {
+    struct ggml_backend_buffer * buffer;
+    bool buffer_owned;
+    void * base;
     size_t alignment;
+
     int n_free_blocks;
     struct free_block free_blocks[MAX_FREE_BLOCKS];
-    struct hash_node hash_table[GGML_GRAPH_HASHTABLE_SIZE];
+
     size_t max_size;
+
     bool measure;
-    int parse_seq[GGML_MAX_CONCUR];
-    int parse_seq_len;
 
 #ifdef GGML_ALLOCATOR_DEBUG
     struct ggml_tensor * allocated_tensors[1024];
@@ -97,7 +48,7 @@ struct ggml_allocr {
 };
 
 #ifdef GGML_ALLOCATOR_DEBUG
-static void add_allocated_tensor(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
+static void add_allocated_tensor(ggml_tallocr_t alloc, struct ggml_tensor * tensor) {
     for (int i = 0; i < 1024; i++) {
         if (alloc->allocated_tensors[i] == NULL) {
             alloc->allocated_tensors[i] = tensor;
@@ -106,7 +57,7 @@ static void add_allocated_tensor(struct ggml_allocr * alloc, struct ggml_tensor
     }
     GGML_ASSERT(!"out of allocated_tensors");
 }
-static void remove_allocated_tensor(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
+static void remove_allocated_tensor(ggml_tallocr_t alloc, struct ggml_tensor * tensor) {
     for (int i = 0; i < 1024; i++) {
         if (alloc->allocated_tensors[i] == tensor ||
             (alloc->allocated_tensors[i] != NULL && alloc->allocated_tensors[i]->data == tensor->data)) {
@@ -119,28 +70,20 @@ static void remove_allocated_tensor(struct ggml_allocr * alloc, struct ggml_tens
 }
 #endif
 
-static size_t ggml_allocr_get_alloc_size(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
-    return ggml_nbytes(tensor);
-
-    UNUSED(alloc);
-}
-
 // check if a tensor is allocated by this buffer
-static bool ggml_allocr_is_own(struct ggml_allocr * alloc, const struct ggml_tensor * tensor) {
-    void * ptr = tensor->data;
-    return ptr >= alloc->data && (char *)ptr < (char *)alloc->data + alloc->max_size;
+static bool ggml_tallocr_is_own(ggml_tallocr_t alloc, const struct ggml_tensor * tensor) {
+    return tensor->buffer == alloc->buffer;
 }
 
 static bool ggml_is_view(struct ggml_tensor * t) {
     return t->view_src != NULL;
 }
 
-void ggml_allocr_alloc(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
-#ifdef GGML_ALLOCATOR_DEBUG
+void ggml_tallocr_alloc(ggml_tallocr_t alloc, struct ggml_tensor * tensor) {
     GGML_ASSERT(!ggml_is_view(tensor)); // views generally get data pointer from one of their sources
     GGML_ASSERT(tensor->data == NULL); // avoid allocating tensor which already has memory allocated
-#endif
-    size_t size = ggml_allocr_get_alloc_size(alloc, tensor);
+
+    size_t size = ggml_backend_buffer_get_alloc_size(alloc->buffer, tensor);
     size = aligned_offset(NULL, size, alloc->alignment);
 
     AT_PRINTF("%s: allocating %s (%zu bytes) - ", __func__, tensor->name, size);
@@ -187,6 +130,10 @@ void ggml_allocr_alloc(struct ggml_allocr * alloc, struct ggml_tensor * tensor)
     }
 
     tensor->data = addr;
+    tensor->buffer = alloc->buffer;
+    if (!alloc->measure) {
+        ggml_backend_buffer_init_tensor(alloc->buffer, tensor);
+    }
 
 #ifdef GGML_ALLOCATOR_DEBUG
     add_allocated_tensor(alloc, tensor);
@@ -202,23 +149,28 @@ void ggml_allocr_alloc(struct ggml_allocr * alloc, struct ggml_tensor * tensor)
     }
 #endif
 
-    alloc->max_size = MAX(alloc->max_size, (char*)addr - (char*)alloc->data + size);
+    alloc->max_size = MAX(alloc->max_size, (char*)addr - (char*)alloc->base + size);
 }
 
 // this is a very naive implementation, but for our case the number of free blocks should be very small
-static void ggml_allocr_free_tensor(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
-    void * ptr = tensor->data;
-
-    if (ggml_allocr_is_own(alloc, tensor) == false) {
+static void ggml_tallocr_free_tensor(ggml_tallocr_t alloc, struct ggml_tensor * tensor) {
+    if (ggml_tallocr_is_own(alloc, tensor) == false) {
         // the tensor was not allocated in this buffer
         // this can happen because the graph allocator will try to free weights and other tensors from different buffers
         // the easiest way to deal with this is just to ignore it
+        // AT_PRINTF("ignoring %s (their buffer: %p, our buffer: %p)\n", tensor->name, (void *)tensor->buffer, (void *)alloc->buffer);
         return;
     }
 
-    size_t size = ggml_allocr_get_alloc_size(alloc, tensor);
+    void * ptr = tensor->data;
+
+    size_t size = ggml_backend_buffer_get_alloc_size(alloc->buffer, tensor);
     size = aligned_offset(NULL, size, alloc->alignment);
-    AT_PRINTF("%s: freeing %s (%zu bytes) - n_free_blocks = %d\n", __func__, tensor->name, size, alloc->n_free_blocks);
+    AT_PRINTF("%s: freeing %s at %p (%zu bytes) - n_free_blocks = %d\n", __func__, tensor->name, ptr, size, alloc->n_free_blocks);
+
+    if (!alloc->measure) {
+        ggml_backend_buffer_free_tensor(alloc->buffer, tensor);
+    }
 
 #ifdef GGML_ALLOCATOR_DEBUG
     remove_allocated_tensor(alloc, tensor);
@@ -272,136 +224,180 @@ static void ggml_allocr_free_tensor(struct ggml_allocr * alloc, struct ggml_tens
     alloc->n_free_blocks++;
 }
 
-void ggml_allocr_set_parse_seq(struct ggml_allocr * alloc, const int * list, int n) {
-    for (int i = 0; i < n; i++) {
-        alloc->parse_seq[i] = list[i];
+void ggml_tallocr_reset(ggml_tallocr_t alloc) {
+    alloc->n_free_blocks = 1;
+    size_t align_offset = aligned_offset(alloc->base, 0, alloc->alignment);
+    alloc->free_blocks[0].addr = (char *)alloc->base + align_offset;
+
+    if (alloc->measure) {
+        alloc->free_blocks[0].size = SIZE_MAX/2; // restrict maximum size of a measure allocator to half size_t max to avoid overflows
+    } else {
+        alloc->free_blocks[0].size = ggml_backend_buffer_get_size(alloc->buffer) - align_offset;
     }
-    alloc->parse_seq_len = n;
 }
 
-void ggml_allocr_reset(struct ggml_allocr * alloc) {
-    alloc->n_free_blocks = 1;
-    size_t align_offset = aligned_offset(alloc->data, 0, alloc->alignment);
-    alloc->free_blocks[0].addr = (char *)alloc->data + align_offset;
-    alloc->free_blocks[0].size = alloc->size - align_offset;
-}
+ggml_tallocr_t ggml_tallocr_new(void * data, size_t size, size_t alignment) {
+    struct ggml_backend_buffer * buffer = ggml_backend_cpu_buffer_from_ptr(NULL, data, size);
 
-struct ggml_allocr * ggml_allocr_new(void * data, size_t size, size_t alignment) {
-    struct ggml_allocr * alloc = (struct ggml_allocr *)malloc(sizeof(struct ggml_allocr) /* + n_free_blocks * sizeof(struct free_block) */);
+    ggml_tallocr_t alloc = (ggml_tallocr_t)malloc(sizeof(struct ggml_tallocr));
 
-    *alloc = (struct ggml_allocr){
-        /*.data          = */ data,
-        /*.size          = */ size,
+    *alloc = (struct ggml_tallocr) {
+        /*.buffer        = */ buffer,
+        /*.buffer_owned  = */ true,
+        /*.base          = */ ggml_backend_buffer_get_base(buffer),
         /*.alignment     = */ alignment,
         /*.n_free_blocks = */ 0,
         /*.free_blocks   = */ {{0}},
-        /*.hash_table    = */ {{0}},
         /*.max_size      = */ 0,
         /*.measure       = */ false,
-        /*.parse_seq     = */ {0},
-        /*.parse_seq_len = */ 0,
 #ifdef GGML_ALLOCATOR_DEBUG
         /*.allocated_tensors = */ {0},
 #endif
     };
 
-    ggml_allocr_reset(alloc);
+    ggml_tallocr_reset(alloc);
 
     return alloc;
 }
 
-// OS specific functions to allocate and free uncommitted virtual memory
-static void * alloc_vmem(size_t size) {
-#if defined(_WIN32)
-    return VirtualAlloc(NULL, size, MEM_RESERVE, PAGE_NOACCESS);
-#elif defined(_POSIX_MAPPED_FILES)
-    void * ptr = mmap(NULL, size, PROT_NONE, MAP_PRIVATE | MAP_ANON, -1, 0);
-    if (ptr == MAP_FAILED) {
-        return NULL;
-    }
-    return ptr;
-#else
-    // use a fixed address for other platforms
-    uintptr_t base_addr = (uintptr_t)-size - 0x100;
-    return (void *)base_addr;
-#endif
-}
+ggml_tallocr_t ggml_tallocr_new_measure(size_t alignment) {
+    ggml_tallocr_t alloc = ggml_tallocr_new((void *)0x1000, SIZE_MAX/2, alignment);
+    alloc->measure = true;
 
-static void free_vmem(void * base_addr, size_t size) {
-#if defined(_WIN32)
-    VirtualFree(base_addr, 0, MEM_RELEASE);
-    UNUSED(size);
-#elif defined(_POSIX_MAPPED_FILES)
-    munmap(base_addr, size);
-#else
-    // nothing to do
-    UNUSED(base_addr);
-    UNUSED(size);
-#endif
+    return alloc;
 }
 
-// allocate uncommitted virtual memory to measure the size of the graph
-static void alloc_measure_vmem(void ** base_addr, size_t * size) {
-    // 128GB for 64-bit, 1GB for 32-bit
-    *size = sizeof(void *) == 4 ? 1ULL<<30 : 1ULL<<37;
-    do {
-        *base_addr = alloc_vmem(*size);
-        if (*base_addr != NULL) {
-            AT_PRINTF("allocated %.2f GB of virtual memory for measure buffer at %p\n", *size / 1024.0 / 1024.0 / 1024.0, *base_addr);
-            return;
-        }
-        // try again with half the size
-        *size /= 2;
-    } while (*size > 0);
+ggml_tallocr_t ggml_tallocr_new_measure_from_backend(struct ggml_backend * backend) {
+    // create a backend buffer to get the correct tensor allocation sizes
+    ggml_backend_buffer_t buffer = ggml_backend_alloc_buffer(backend, 1);
 
-    GGML_ASSERT(!"failed to allocate virtual memory for measure buffer");
+    // TODO: move alloc initialization to a common ggml_tallocr_new_impl function
+    ggml_tallocr_t alloc = ggml_tallocr_new_from_buffer(buffer);
+    alloc->buffer_owned = true;
+    alloc->measure = true;
+    ggml_tallocr_reset(alloc);
+    return alloc;
 }
 
-static void free_measure_vmem(void * base_addr, size_t size) {
-    free_vmem(base_addr, size);
+ggml_tallocr_t ggml_tallocr_new_from_backend(struct ggml_backend * backend, size_t size) {
+    ggml_backend_buffer_t buffer = ggml_backend_alloc_buffer(backend, size);
+    ggml_tallocr_t alloc = ggml_tallocr_new_from_buffer(buffer);
+    alloc->buffer_owned = true;
+    return alloc;
 }
 
-struct ggml_allocr * ggml_allocr_new_measure(size_t alignment) {
-    struct ggml_allocr * alloc = (struct ggml_allocr *)malloc(sizeof(struct ggml_allocr) /* + n_free_blocks * sizeof(struct free_block) */);
+ggml_tallocr_t ggml_tallocr_new_from_buffer(struct ggml_backend_buffer * buffer) {
+    ggml_tallocr_t alloc = (ggml_tallocr_t)malloc(sizeof(struct ggml_tallocr));
 
-    void * base_addr;
-    size_t size;
-
-    alloc_measure_vmem(&base_addr, &size);
-
-    *alloc = (struct ggml_allocr){
-        /*.data          = */ base_addr,
-        /*.size          = */ size,
-        /*.alignment     = */ alignment,
+    *alloc = (struct ggml_tallocr) {
+        /*.buffer        = */ buffer,
+        /*.buffer_owned  = */ false,
+        /*.base          = */ ggml_backend_buffer_get_base(buffer),
+        /*.alignment     = */ ggml_backend_buffer_get_alignment(buffer),
         /*.n_free_blocks = */ 0,
         /*.free_blocks   = */ {{0}},
-        /*.hash_table    = */ {{0}},
         /*.max_size      = */ 0,
-        /*.measure       = */ true,
-        /*.parse_seq     = */ {0},
-        /*.parse_seq_len = */ 0,
+        /*.measure       = */ false,
 #ifdef GGML_ALLOCATOR_DEBUG
         /*.allocated_tensors = */ {0},
 #endif
     };
 
-    ggml_allocr_reset(alloc);
+    ggml_tallocr_reset(alloc);
 
     return alloc;
 }
 
-void ggml_allocr_free(struct ggml_allocr * alloc) {
-    if (alloc->measure) {
-        free_measure_vmem(alloc->data, alloc->size);
+struct ggml_backend_buffer * ggml_tallocr_get_buffer(ggml_tallocr_t alloc) {
+    return alloc->buffer;
+}
+
+void ggml_tallocr_free(ggml_tallocr_t alloc) {
+    if (alloc == NULL) {
+        return;
+    }
+
+    if (alloc->buffer_owned) {
+        ggml_backend_buffer_free(alloc->buffer);
     }
     free(alloc);
 }
 
-bool ggml_allocr_is_measure(struct ggml_allocr * alloc) {
+bool ggml_tallocr_is_measure(ggml_tallocr_t alloc) {
     return alloc->measure;
 }
 
-//////////// compute graph allocator
+size_t ggml_tallocr_max_size(ggml_tallocr_t alloc) {
+    return alloc->max_size;
+}
+
+// graph allocator
+
+struct hash_node {
+    int n_children;
+    int n_views;
+};
+
+struct ggml_gallocr {
+    ggml_tallocr_t talloc;
+    struct ggml_hash_set hash_set;
+    struct hash_node * hash_values;
+    size_t hash_values_size;
+    ggml_tallocr_t * hash_allocs;
+    int * parse_seq;
+    int parse_seq_len;
+};
+
+ggml_gallocr_t ggml_gallocr_new(void) {
+    ggml_gallocr_t galloc = (ggml_gallocr_t)malloc(sizeof(struct ggml_gallocr));
+
+    *galloc = (struct ggml_gallocr) {
+        /*.talloc           = */ NULL,
+        /*.hash_set         = */ {0},
+        /*.hash_values      = */ NULL,
+        /*.hash_values_size = */ 0,
+        /*.hash_allocs      = */ NULL,
+        /*.parse_seq        = */ NULL,
+        /*.parse_seq_len    = */ 0,
+    };
+
+    return galloc;
+}
+
+void ggml_gallocr_free(ggml_gallocr_t galloc) {
+    if (galloc == NULL) {
+        return;
+    }
+
+    if (galloc->hash_set.keys != NULL) {
+        free(galloc->hash_set.keys);
+    }
+    if (galloc->hash_values != NULL) {
+        free(galloc->hash_values);
+    }
+    if (galloc->hash_allocs != NULL) {
+        free(galloc->hash_allocs);
+    }
+    if (galloc->parse_seq != NULL) {
+        free(galloc->parse_seq);
+    }
+    free(galloc);
+}
+
+void ggml_gallocr_set_parse_seq(ggml_gallocr_t galloc, const int * list, int n) {
+    free(galloc->parse_seq);
+    galloc->parse_seq = malloc(sizeof(int) * n);
+
+    for (int i = 0; i < n; i++) {
+        galloc->parse_seq[i] = list[i];
+    }
+    galloc->parse_seq_len = n;
+}
+
+static struct hash_node * hash_get(ggml_gallocr_t galloc, struct ggml_tensor * t) {
+    size_t i = ggml_hash_find_or_insert(galloc->hash_set, t);
+    return &galloc->hash_values[i];
+}
 
 static bool ggml_are_same_layout(const struct ggml_tensor * a, const struct ggml_tensor * b) {
     if (a->type != b->type) {
@@ -435,7 +431,6 @@ static bool ggml_op_can_inplace(enum ggml_op op) {
         case GGML_OP_ROPE:
         case GGML_OP_RMS_NORM:
         case GGML_OP_SOFT_MAX:
-        case GGML_OP_CONT:
             return true;
 
         default:
@@ -443,12 +438,38 @@ static bool ggml_op_can_inplace(enum ggml_op op) {
     }
 }
 
-static void allocate_node(struct ggml_allocr * alloc, struct ggml_tensor * node) {
-    struct hash_node * ht = alloc->hash_table;
+static ggml_tallocr_t node_tallocr(ggml_gallocr_t galloc, struct ggml_tensor * node) {
+    if (galloc->talloc != NULL) {
+        return galloc->talloc;
+    }
+
+    return galloc->hash_allocs[ggml_hash_find_or_insert(galloc->hash_set, node)];
+}
+
+static void init_view(ggml_gallocr_t galloc, struct ggml_tensor * view) {
+    ggml_tallocr_t alloc = node_tallocr(galloc, view);
+
+    //printf("init_view: %s from src %s\n", view->name, view->view_src->name);
+    GGML_ASSERT(view->view_src != NULL && view->view_src->data != NULL);
+    view->backend = view->view_src->backend;
+    view->buffer  = view->view_src->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->backend == alloc->buffer->backend);
+
+    if (!alloc->measure) {
+        ggml_backend_buffer_init_tensor(alloc->buffer, view);
+    }
+}
+
+static void allocate_node(ggml_gallocr_t galloc, struct ggml_tensor * node) {
+    ggml_tallocr_t alloc = node_tallocr(galloc, node);
+
     if (node->data == NULL) {
         if (ggml_is_view(node)) {
-            assert(node->view_src->data != NULL);
-            node->data = (char *)node->view_src->data + node->view_offs;
+            init_view(galloc, node);
         } else {
             // see if we can reuse a parent's buffer (inplace)
             if (ggml_op_can_inplace(node->op)) {
@@ -459,16 +480,16 @@ static void allocate_node(struct ggml_allocr * alloc, struct ggml_tensor * node)
                     }
 
                     // if the node's data is external, then we cannot re-use it
-                    if (ggml_allocr_is_own(alloc, parent) == false) {
+                    if (ggml_tallocr_is_own(alloc, parent) == false) {
                         AT_PRINTF("not reusing parent %s for %s as %p is external\n", parent->name, node->name, parent->data);
                         continue;
                     }
 
-                    struct hash_node * p_hn = hash_get(ht, parent);
+                    struct hash_node * p_hn = hash_get(galloc, parent);
                     if (parent->data != NULL && p_hn->n_children == 1 && p_hn->n_views == 0 && ggml_are_same_layout(node, parent)) {
                         if (ggml_is_view(parent)) {
                             struct ggml_tensor * view_src = parent->view_src;
-                            struct hash_node * view_src_hn = hash_get(ht, view_src);
+                            struct hash_node * view_src_hn = hash_get(galloc, view_src);
                             if (view_src_hn->n_views == 1 && view_src_hn->n_children == 0 && view_src->data == parent->data) {
                                 // TODO: the offset of the view parent must be kept to ensure that the op doesn't overwrite
                                 // the parent's data that it will need later (same layout requirement). the problem is that then
@@ -476,158 +497,270 @@ static void allocate_node(struct ggml_allocr * alloc, struct ggml_tensor * node)
                                 // adding a view_src pointer to the tensor would solve this and simplify the code dealing with views
                                 // for now, we only reuse the parent's data if the offset is zero (view_src->data == parent->data)
                                 AT_PRINTF("reusing view parent %s (%s) for %s\n", parent->name, view_src->name, node->name);
-                                node->data = parent->data;
+                                node->view_src = view_src;
+                                view_src_hn->n_views += 1;
+                                init_view(galloc, node);
                                 return;
                             }
                         }
                         else {
                             AT_PRINTF("reusing parent %s for %s\n", parent->name, node->name);
-                            node->data = parent->data;
+                            node->view_src = parent;
+                            p_hn->n_views += 1;
+                            init_view(galloc, node);
                             return;
                         }
                     }
                 }
             }
-            ggml_allocr_alloc(alloc, node);
+            ggml_tallocr_alloc(alloc, node);
         }
     }
 }
 
-static size_t ggml_allocr_alloc_graph_tensors_n(
-    struct ggml_allocr * alloc,
-    struct ggml_cgraph ** graphs, int n_graphs,
-    struct ggml_tensor *** inputs, struct ggml_tensor *** outputs) {
+static void free_node(ggml_gallocr_t galloc, struct ggml_tensor * node) {
+    ggml_tallocr_t alloc = node_tallocr(galloc, node);
 
-    // reset hash table
-    struct hash_node * ht = alloc->hash_table;
-    memset(ht, 0, sizeof(struct hash_node) * GGML_GRAPH_HASHTABLE_SIZE);
+    ggml_tallocr_free_tensor(alloc, node);
+}
+
+static void ggml_tallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgraph * gf) {
+    const int * parse_seq     = galloc->parse_seq;
+    int         parse_seq_len = galloc->parse_seq_len;
 
     // count number of children and views
-    for (int g = 0; g < n_graphs; g++) {
-        struct ggml_cgraph * gf = graphs[g];
-        for (int i = 0; i < gf->n_nodes; i++) {
-            struct ggml_tensor * node = gf->nodes[i];
+    for (int i = 0; i < gf->n_nodes; i++) {
+        struct ggml_tensor * node = gf->nodes[i];
+
+        if (ggml_is_view(node)) {
+            struct ggml_tensor * view_src = node->view_src;
+            hash_get(galloc, view_src)->n_views += 1;
+            if (node->buffer == NULL && node->data != NULL) {
+                // view of a pre-allocated tensor, didn't call init_view() yet
+                init_view(galloc, node);
+            }
+        }
 
-            if (ggml_is_view(node)) {
-                struct ggml_tensor * view_src = node->view_src;
-                hash_get(ht, view_src)->n_views += 1;
+        for (int j = 0; j < GGML_MAX_SRC; j++) {
+            struct ggml_tensor * parent = node->src[j];
+            if (parent == NULL) {
+                break;
             }
+            hash_get(galloc, parent)->n_children += 1;
+            if (ggml_is_view(parent) && parent->buffer == NULL && parent->data != NULL) {
+                init_view(galloc, parent);
+            }
+        }
+   }
 
+    // allocate tensors
+    // if we have parse_seq then we allocate nodes following the list, and we only free nodes at barriers
+    int last_barrier_pos = 0;
+    int n_nodes = parse_seq_len ? parse_seq_len : gf->n_nodes;
+
+    for (int ind = 0; ind < n_nodes; ind++) {
+        // allocate a node if there is no parse_seq or this is not a barrier
+        if (parse_seq_len == 0 || parse_seq[ind] != -1) {
+            int i = parse_seq_len ? parse_seq[ind] : ind;
+            struct ggml_tensor * node = gf->nodes[i];
+
+            // allocate parents (leafs)
             for (int j = 0; j < GGML_MAX_SRC; j++) {
                 struct ggml_tensor * parent = node->src[j];
                 if (parent == NULL) {
                     break;
                 }
-                hash_get(ht, parent)->n_children += 1;
+                allocate_node(galloc, parent);
             }
-        }
-    }
 
-    // allocate tensors
-    for (int g = 0; g < n_graphs; g++) {
-        struct ggml_cgraph * gf = graphs[g];
-        AT_PRINTF("####### graph %d/%d\n", g, n_graphs);
-        // graph inputs are allocated first to ensure that they are not overwritten by each other
-        if (inputs != NULL && inputs[g] != NULL) {
-            for (int i = 0; inputs[g][i] != NULL; i++) {
-                struct ggml_tensor * input = inputs[g][i];
-                AT_PRINTF("input: %s\n", input->name);
-                allocate_node(alloc, input);
+            // allocate node
+            allocate_node(galloc, node);
+
+            AT_PRINTF("exec: %s (%s) <= ", ggml_op_name(node->op), node->name);
+            for (int j = 0; j < GGML_MAX_SRC; j++) {
+                struct ggml_tensor * parent = node->src[j];
+                if (parent == NULL) {
+                    break;
+                }
+                AT_PRINTF("%s", parent->name);
+                if (j < GGML_MAX_SRC - 1 && node->src[j + 1] != NULL) {
+                    AT_PRINTF(", ");
+                }
             }
+            AT_PRINTF("\n");
         }
-        // if we have parse_seq then we allocate nodes following the list, and we only free nodes at barriers
-        int last_barrier_pos = 0;
-        int n_nodes = alloc->parse_seq_len ? alloc->parse_seq_len : gf->n_nodes;
 
-        for (int ind = 0; ind < n_nodes; ind++) {
-            // allocate a node if there is no parse_seq or this is not a barrier
-            if ((alloc->parse_seq_len==0) || alloc->parse_seq[ind] != -1) {
-                int i = alloc->parse_seq_len ? alloc->parse_seq[ind] : ind;
-                struct ggml_tensor * node = gf->nodes[i];
+        // update parents
+        // update immediately if there is no parse_seq
+        // update only at barriers if there is parse_seq
+        if ((parse_seq_len == 0) || parse_seq[ind] == -1) {
+            int update_start = parse_seq_len ? last_barrier_pos : ind;
+            int update_end   = parse_seq_len ? ind              : ind + 1;
+            for (int i = update_start; i < update_end; i++) {
+                int node_i = parse_seq_len ? parse_seq[i] : i;
+                struct ggml_tensor * node = gf->nodes[node_i];
 
-                // allocate parents (leafs)
                 for (int j = 0; j < GGML_MAX_SRC; j++) {
                     struct ggml_tensor * parent = node->src[j];
                     if (parent == NULL) {
                         break;
                     }
-                    allocate_node(alloc, parent);
-                }
+                    struct hash_node * p_hn = hash_get(galloc, parent);
+                    p_hn->n_children -= 1;
 
-                // allocate node
-                allocate_node(alloc, node);
+                    //AT_PRINTF("parent %s: %d children, %d views\n", parent->name, parent->n_children, parent->n_views);
 
-                AT_PRINTF("exec: %s (%s) <= ", ggml_op_name(node->op), node->name);
-                for (int j = 0; j < GGML_MAX_SRC; j++) {
-                    struct ggml_tensor * parent = node->src[j];
-                    if (parent == NULL) {
-                        break;
-                    }
-                    AT_PRINTF("%s", parent->name);
-                    if (j < GGML_MAX_SRC - 1 && node->src[j + 1] != NULL) {
-                        AT_PRINTF(", ");
-                    }
-                }
-                AT_PRINTF("\n");
-            }
-
-            // update parents
-            // update immediately if there is no parse_seq
-            // update only at barriers if there is parse_seq
-            if ((alloc->parse_seq_len == 0) || alloc->parse_seq[ind] == -1) {
-                int update_start = alloc->parse_seq_len ? last_barrier_pos : ind;
-                int update_end   = alloc->parse_seq_len ? ind              : ind + 1;
-                for (int i = update_start; i < update_end; i++) {
-                    int node_i = alloc->parse_seq_len ? alloc->parse_seq[i] : i;
-                    struct ggml_tensor * node = gf->nodes[node_i];
-
-                    for (int j = 0; j < GGML_MAX_SRC; j++) {
-                        struct ggml_tensor * parent = node->src[j];
-                        if (parent == NULL) {
-                            break;
-                        }
-                        struct hash_node * p_hn = hash_get(ht, parent);
-                        p_hn->n_children -= 1;
-
-                        //AT_PRINTF("parent %s: %d children, %d views\n", parent->name, parent->n_children, parent->n_views);
-
-                        if (p_hn->n_children == 0 && p_hn->n_views == 0) {
-                            if (ggml_is_view(parent)) {
-                                struct ggml_tensor * view_src = parent->view_src;
-                                struct hash_node * view_src_hn = hash_get(ht, view_src);
-                                view_src_hn->n_views -= 1;
-                                AT_PRINTF("view_src %s: %d children, %d views\n", view_src->name, view_src_hn->n_children, view_src_hn->n_views);
-                                if (view_src_hn->n_views == 0 && view_src_hn->n_children == 0 && view_src->data != node->data) {
-                                    ggml_allocr_free_tensor(alloc, view_src);
-                                }
-                            }
-                            else {
-                                if (parent->data != node->data) {
-                                    ggml_allocr_free_tensor(alloc, parent);
-                                }
+                    if (p_hn->n_children == 0 && p_hn->n_views == 0) {
+                        if (ggml_is_view(parent)) {
+                            struct ggml_tensor * view_src = parent->view_src;
+                            struct hash_node * view_src_hn = hash_get(galloc, view_src);
+                            view_src_hn->n_views -= 1;
+                            AT_PRINTF("view_src %s: %d children, %d views\n", view_src->name, view_src_hn->n_children, view_src_hn->n_views);
+                            if (view_src_hn->n_views == 0 && view_src_hn->n_children == 0) {
+                                free_node(galloc, view_src);
                             }
                         }
+                        else {
+                            free_node(galloc, parent);
+                        }
                     }
                 }
-                AT_PRINTF("\n");
-                if (alloc->parse_seq_len) {
-                    last_barrier_pos = ind + 1;
-                }
             }
-        }
-        // free graph outputs here that wouldn't be freed otherwise because they have no children
-        if (outputs != NULL && outputs[g] != NULL) {
-            for (int i = 0; outputs[g][i] != NULL; i++) {
-                struct ggml_tensor * output = outputs[g][i];
-                AT_PRINTF("output: %s\n", output->name);
-                ggml_allocr_free_tensor(alloc, output);
+            AT_PRINTF("\n");
+            if (parse_seq_len) {
+                last_barrier_pos = ind + 1;
             }
         }
     }
+}
 
-    return alloc->max_size;
+size_t ggml_gallocr_alloc_graph(ggml_gallocr_t galloc, ggml_tallocr_t talloc, struct ggml_cgraph * graph) {
+    size_t hash_size = graph->visited_hash_table.size;
+
+    // check if the hash table is initialized and large enough
+    if (galloc->hash_set.size < hash_size) {
+        if (galloc->hash_set.keys != NULL) {
+            free(galloc->hash_set.keys);
+        }
+        if (galloc->hash_values != NULL) {
+            free(galloc->hash_values);
+        }
+        galloc->hash_set.keys = malloc(sizeof(struct ggml_tensor *) * hash_size);
+        galloc->hash_set.size = hash_size;
+        galloc->hash_values = malloc(sizeof(struct hash_node) * hash_size);
+    }
+
+    // reset hash table
+    memset(galloc->hash_set.keys, 0, sizeof(struct ggml_tensor *) * hash_size);
+    memset(galloc->hash_values,   0, sizeof(struct hash_node) * hash_size);
+
+    galloc->talloc = talloc;
+    ggml_tallocr_alloc_graph_impl(galloc, graph);
+    galloc->talloc = NULL;
+
+    size_t max_size = ggml_tallocr_max_size(talloc);
+
+    return max_size;
+}
+
+void ggml_gallocr_alloc_graph_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, struct ggml_hash_set hash_set, ggml_tallocr_t * hash_node_alloct) {
+    const size_t hash_size = hash_set.size;
+
+    GGML_ASSERT(hash_size >= (size_t)(graph->n_nodes + graph->n_leafs));
+
+    galloc->talloc = NULL;
+
+    // alloc hash_values if needed
+    if (galloc->hash_values == NULL || galloc->hash_values_size < hash_size) {
+        free(galloc->hash_values);
+        galloc->hash_values      = malloc(sizeof(struct hash_node) * hash_size);
+        galloc->hash_values_size = hash_size;
+    }
+
+    // free hash_set.keys if needed
+    if (galloc->hash_set.keys != NULL) {
+        free(galloc->hash_set.keys);
+    }
+    galloc->hash_set = hash_set;
+
+    // reset hash values
+    memset(galloc->hash_values, 0, sizeof(struct hash_node) * hash_size);
+
+    galloc->hash_allocs = hash_node_alloct;
+
+    ggml_tallocr_alloc_graph_impl(galloc, graph);
+
+    // remove unowned resources
+    galloc->hash_set.keys = NULL;
+    galloc->hash_allocs = NULL;
+}
+
+// legacy API wrapper
+
+struct ggml_allocr {
+    ggml_tallocr_t talloc;
+    ggml_gallocr_t galloc;
+};
+
+static ggml_allocr_t ggml_allocr_new_impl(ggml_tallocr_t talloc) {
+    ggml_allocr_t alloc = (ggml_allocr_t)malloc(sizeof(struct ggml_allocr));
+    *alloc = (struct ggml_allocr) {
+        /*.talloc = */ talloc,
+        /*.galloc = */ ggml_gallocr_new(),
+    };
+    return alloc;
+}
+
+ggml_allocr_t ggml_allocr_new(void * data, size_t size, size_t alignment) {
+    return ggml_allocr_new_impl(ggml_tallocr_new(data, size, alignment));
+}
+
+ggml_allocr_t ggml_allocr_new_measure(size_t alignment) {
+    return ggml_allocr_new_impl(ggml_tallocr_new_measure(alignment));
+}
+
+ggml_allocr_t ggml_allocr_new_from_buffer(struct ggml_backend_buffer * buffer) {
+    return ggml_allocr_new_impl(ggml_tallocr_new_from_buffer(buffer));
+}
+
+ggml_allocr_t ggml_allocr_new_from_backend(struct ggml_backend * backend, size_t size) {
+    return ggml_allocr_new_impl(ggml_tallocr_new_from_backend(backend, size));
+}
+
+ggml_allocr_t ggml_allocr_new_measure_from_backend(struct ggml_backend * backend) {
+    return ggml_allocr_new_impl(ggml_tallocr_new_measure_from_backend(backend));
+}
+
+struct ggml_backend_buffer * ggml_allocr_get_buffer(ggml_allocr_t alloc) {
+    return ggml_tallocr_get_buffer(alloc->talloc);
+}
+
+void ggml_allocr_set_parse_seq(ggml_allocr_t alloc, const int * list, int n) {
+    ggml_gallocr_set_parse_seq(alloc->galloc, list, n);
+}
+
+void ggml_allocr_free(ggml_allocr_t alloc) {
+    ggml_gallocr_free(alloc->galloc);
+    ggml_tallocr_free(alloc->talloc);
+    free(alloc);
+}
+
+bool ggml_allocr_is_measure(ggml_allocr_t alloc) {
+    return ggml_tallocr_is_measure(alloc->talloc);
+}
+
+void ggml_allocr_reset(ggml_allocr_t alloc) {
+    ggml_tallocr_reset(alloc->talloc);
+}
+
+void ggml_allocr_alloc(ggml_allocr_t alloc, struct ggml_tensor * tensor) {
+    ggml_tallocr_alloc(alloc->talloc, tensor);
+}
+
+size_t ggml_allocr_max_size(ggml_allocr_t alloc) {
+    return ggml_tallocr_max_size(alloc->talloc);
 }
 
-size_t ggml_allocr_alloc_graph(struct ggml_allocr * alloc, struct ggml_cgraph * graph) {
-    return ggml_allocr_alloc_graph_tensors_n(alloc, &graph, 1, NULL, NULL);
+size_t ggml_allocr_alloc_graph(ggml_allocr_t alloc, struct ggml_cgraph * graph) {
+    return ggml_gallocr_alloc_graph(alloc->galloc, alloc->talloc, graph);
 }

ggml-alloc.h

@@ -6,20 +6,79 @@
 extern "C" {
 #endif
 
+struct ggml_backend;
+struct ggml_backend_buffer;
 
-GGML_API struct ggml_allocr * ggml_allocr_new(void * data, size_t size, size_t alignment);
-GGML_API struct ggml_allocr * ggml_allocr_new_measure(size_t alignment);
+//
+// Legacy API
+//
+
+typedef struct ggml_allocr * ggml_allocr_t;
+
+// initialize allocator for use with CPU backend only
+GGML_API ggml_allocr_t ggml_allocr_new(void * data, size_t size, size_t alignment);
+GGML_API ggml_allocr_t ggml_allocr_new_measure(size_t alignment);
+
+// initialize allocator for use with ggml-backend
+GGML_API ggml_allocr_t ggml_allocr_new_from_buffer(struct ggml_backend_buffer * buffer);
+GGML_API ggml_allocr_t ggml_allocr_new_from_backend(struct ggml_backend * backend, size_t size); // allocates an owned buffer
+GGML_API ggml_allocr_t ggml_allocr_new_measure_from_backend(struct ggml_backend * backend);
+
+GGML_API struct ggml_backend_buffer * ggml_allocr_get_buffer(ggml_allocr_t alloc);
 
 // tell the allocator to parse nodes following the order described in the list
 // you should call this if your graph are optimized to execute out-of-order
-GGML_API void   ggml_allocr_set_parse_seq(struct ggml_allocr * alloc, const int * list, int n);
+GGML_API void   ggml_allocr_set_parse_seq(ggml_allocr_t alloc, const int * list, int n);
+
+GGML_API void   ggml_allocr_free       (ggml_allocr_t alloc);
+GGML_API bool   ggml_allocr_is_measure (ggml_allocr_t alloc);
+GGML_API void   ggml_allocr_reset      (ggml_allocr_t alloc);
+GGML_API void   ggml_allocr_alloc      (ggml_allocr_t alloc, struct ggml_tensor * tensor);
+GGML_API size_t ggml_allocr_max_size   (ggml_allocr_t alloc);
+
+GGML_API size_t ggml_allocr_alloc_graph(ggml_allocr_t alloc, struct ggml_cgraph * graph);
+
+//
+// ggml-backend v2 API
+//
+
+// Seperate tensor and graph allocator objects
+// This is necessary for multi-backend allocation because the graph allocator needs to use multiple tensor allocators
+// The original API is kept as a wrapper around the new API
+
+// Tensor allocator
+typedef struct ggml_tallocr * ggml_tallocr_t;
+
+GGML_API ggml_tallocr_t ggml_tallocr_new(void * data, size_t size, size_t alignment);
+GGML_API ggml_tallocr_t ggml_tallocr_new_measure(size_t alignment);
+GGML_API ggml_tallocr_t ggml_tallocr_new_from_buffer(struct ggml_backend_buffer * buffer);
+GGML_API ggml_tallocr_t ggml_tallocr_new_from_backend(struct ggml_backend * backend, size_t size); // allocates an owned buffer
+GGML_API ggml_tallocr_t ggml_tallocr_new_measure_from_backend(struct ggml_backend * backend);
+
+GGML_API struct ggml_backend_buffer * ggml_tallocr_get_buffer(ggml_tallocr_t talloc);
+
+GGML_API void   ggml_tallocr_free       (ggml_tallocr_t talloc);
+GGML_API bool   ggml_tallocr_is_measure (ggml_tallocr_t talloc);
+GGML_API void   ggml_tallocr_reset      (ggml_tallocr_t talloc);
+GGML_API void   ggml_tallocr_alloc      (ggml_tallocr_t talloc, struct ggml_tensor * tensor);
+GGML_API size_t ggml_tallocr_max_size   (ggml_tallocr_t talloc);
+
+
+// Graph allocator
+typedef struct ggml_gallocr * ggml_gallocr_t;
+
+GGML_API ggml_gallocr_t ggml_gallocr_new(void);
+GGML_API void   ggml_gallocr_free(ggml_gallocr_t galloc);
 
-GGML_API void   ggml_allocr_free(struct ggml_allocr * alloc);
-GGML_API bool   ggml_allocr_is_measure(struct ggml_allocr * alloc);
-GGML_API void   ggml_allocr_reset(struct ggml_allocr * alloc);
-GGML_API void   ggml_allocr_alloc(struct ggml_allocr * alloc, struct ggml_tensor * tensor);
-GGML_API size_t ggml_allocr_alloc_graph(struct ggml_allocr * alloc, struct ggml_cgraph * graph);
+GGML_API void   ggml_gallocr_set_parse_seq(ggml_gallocr_t galloc, const int * list, int n);
+GGML_API size_t ggml_gallocr_alloc_graph(ggml_gallocr_t galloc, ggml_tallocr_t talloc, struct ggml_cgraph * graph);
 
+// Allocate tensors from the allocators given by the hash table
+GGML_API void   ggml_gallocr_alloc_graph_n(
+                    ggml_gallocr_t galloc,
+                    struct ggml_cgraph * graph,
+                    struct ggml_hash_set hash_set,
+                    ggml_tallocr_t * hash_node_talloc);
 
 #ifdef  __cplusplus
 }

ggml-backend-impl.h

@@ -0,0 +1,87 @@
+#pragma once
+
+// ggml-backend internal header
+
+#include "ggml-backend.h"
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+    //
+    // Backend buffer
+    //
+
+    typedef void * ggml_backend_buffer_context_t;
+
+    struct ggml_backend_buffer_i {
+        void   (*free_buffer)   (ggml_backend_buffer_t buffer);
+        void * (*get_base)      (ggml_backend_buffer_t buffer); // get base pointer
+        size_t (*get_alloc_size)(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); // pre-allocation callback
+        void   (*init_tensor)   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); // post-allocation callback
+        void   (*free_tensor)   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); // pre-free callback
+    };
+
+    struct ggml_backend_buffer {
+        struct ggml_backend_buffer_i iface;
+
+        ggml_backend_t                backend;
+        ggml_backend_buffer_context_t context;
+
+        size_t size;
+    };
+
+    GGML_API ggml_backend_buffer_t ggml_backend_buffer_init(
+            struct ggml_backend                  * backend,
+            struct ggml_backend_buffer_i           iface,
+                   ggml_backend_buffer_context_t   context,
+                   size_t                          size);
+
+    //
+    // Backend
+    //
+
+    typedef void * ggml_backend_context_t;
+
+    struct ggml_backend_i {
+        const char * (*get_name)(ggml_backend_t backend);
+
+        void (*free)(ggml_backend_t backend);
+
+        // buffer allocation
+        ggml_backend_buffer_t (*alloc_buffer)(ggml_backend_t backend, size_t size);
+
+        // get buffer alignment
+        size_t (*get_alignment)(ggml_backend_t backend);
+
+        // tensor data access
+        // these functions can be asynchronous, helper functions are provided for synchronous access that automatically call synchronize
+        void (*set_tensor_async)(ggml_backend_t backend,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
+        void (*get_tensor_async)(ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
+        void (*synchronize)     (ggml_backend_t backend);
+
+        // (optional) copy tensor between different backends, allow for single-copy tranfers
+        void (*cpy_tensor_from)(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst);
+        void (*cpy_tensor_to)  (ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst);
+
+        // compute graph with a plan
+        ggml_backend_graph_plan_t (*graph_plan_create) (ggml_backend_t backend, struct ggml_cgraph * cgraph);
+        void                      (*graph_plan_free)   (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
+        void                      (*graph_plan_compute)(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
+
+        // compute graph without a plan
+        void (*graph_compute)(ggml_backend_t backend, struct ggml_cgraph * cgraph);
+
+        // check if the backend supports an operation
+        bool (*supports_op)(ggml_backend_t backend, const struct ggml_tensor * op);
+    };
+
+    struct ggml_backend {
+        struct ggml_backend_i iface;
+
+        ggml_backend_context_t context;
+    };
+
+#ifdef  __cplusplus
+}
+#endif

ggml-backend.c

@@ -0,0 +1,950 @@
+#include "ggml-backend-impl.h"
+#include "ggml-alloc.h"
+#include "ggml-impl.h"
+
+#include <assert.h>
+#include <limits.h>
+#include <stdarg.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#define UNUSED GGML_UNUSED
+
+#define MAX(a, b) ((a) > (b) ? (a) : (b))
+
+// backend buffer
+
+ggml_backend_buffer_t ggml_backend_buffer_init(
+        struct ggml_backend                  * backend,
+        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));
+
+    GGML_ASSERT(iface.get_base != NULL);
+
+    (*buffer) = (struct ggml_backend_buffer) {
+        /* .interface = */ iface,
+        /* .backend   = */ backend,
+        /* .context   = */ context,
+        /* .size      = */ size,
+    };
+
+    return buffer;
+}
+
+void ggml_backend_buffer_free(ggml_backend_buffer_t buffer) {
+    if (buffer == NULL) {
+        return;
+    }
+
+    if (buffer->iface.free_buffer != NULL) {
+        buffer->iface.free_buffer(buffer);
+    }
+    free(buffer);
+}
+
+size_t ggml_backend_buffer_get_alignment(ggml_backend_buffer_t buffer) {
+    return ggml_backend_get_alignment(buffer->backend);
+}
+
+size_t ggml_backend_buffer_get_size(ggml_backend_buffer_t buffer) {
+    return buffer->size;
+}
+
+void * ggml_backend_buffer_get_base(ggml_backend_buffer_t buffer) {
+    void * base = buffer->iface.get_base(buffer);
+
+    GGML_ASSERT(base != NULL && "backend buffer base cannot be NULL");
+
+    return base;
+}
+
+size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
+    // get_alloc_size is optional, defaults to ggml_nbytes
+    if (buffer->iface.get_alloc_size) {
+        return buffer->iface.get_alloc_size(buffer, tensor);
+    }
+    return ggml_nbytes(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);
+    }
+}
+
+void ggml_backend_buffer_free_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
+    // free_tensor is optional
+    if (buffer->iface.free_tensor) {
+        buffer->iface.free_tensor(buffer, tensor);
+    }
+}
+
+// backend
+
+ggml_backend_t ggml_get_backend(const struct ggml_tensor * tensor) {
+    return tensor->buffer ? tensor->buffer->backend : NULL;
+}
+
+const char * ggml_backend_name(ggml_backend_t backend) {
+    if (backend == NULL) {
+        return "NULL";
+    }
+    return backend->iface.get_name(backend);
+}
+
+void ggml_backend_free(ggml_backend_t backend) {
+    if (backend == NULL) {
+        return;
+    }
+
+    backend->iface.free(backend);
+}
+
+ggml_backend_buffer_t ggml_backend_alloc_buffer(ggml_backend_t backend, size_t size) {
+    return backend->iface.alloc_buffer(backend, size);
+}
+
+size_t ggml_backend_get_alignment(ggml_backend_t backend) {
+    return backend->iface.get_alignment(backend);
+}
+
+void ggml_backend_tensor_set_async(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+    ggml_get_backend(tensor)->iface.set_tensor_async(ggml_get_backend(tensor), tensor, data, offset, size);
+}
+
+void ggml_backend_tensor_get_async(const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+    ggml_get_backend(tensor)->iface.get_tensor_async(ggml_get_backend(tensor), tensor, data, offset, size);
+}
+
+void ggml_backend_tensor_set(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+    ggml_backend_t backend = ggml_get_backend(tensor);
+
+    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
+    GGML_ASSERT(backend != NULL && "tensor backend not set");
+
+    backend->iface.set_tensor_async(backend, tensor, data, offset, size);
+    backend->iface.synchronize(backend);
+}
+
+void ggml_backend_tensor_get(const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+    ggml_backend_t backend = ggml_get_backend(tensor);
+
+    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
+    GGML_ASSERT(backend != NULL && "tensor backend not set");
+
+    backend->iface.get_tensor_async(backend, tensor, data, offset, size);
+    backend->iface.synchronize(backend);
+}
+
+void ggml_backend_synchronize(ggml_backend_t backend) {
+    backend->iface.synchronize(backend);
+}
+
+ggml_backend_graph_plan_t ggml_backend_graph_plan_create(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+    return backend->iface.graph_plan_create(backend, cgraph);
+}
+
+void ggml_backend_graph_plan_free(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
+    backend->iface.graph_plan_free(backend, plan);
+}
+
+void ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
+    backend->iface.graph_plan_compute(backend, plan);
+}
+
+void ggml_backend_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+    backend->iface.graph_compute(backend, cgraph);
+}
+
+bool ggml_backend_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
+    return backend->iface.supports_op(backend, op);
+}
+
+// backend copy
+
+static bool ggml_are_same_layout(const struct ggml_tensor * a, const struct ggml_tensor * b) {
+    if (a->type != b->type) {
+        return false;
+    }
+    for (int i = 0; i < GGML_MAX_DIMS; i++) {
+        if (a->ne[i] != b->ne[i]) {
+            return false;
+        }
+        if (a->nb[i] != b->nb[i]) {
+            return false;
+        }
+    }
+    return true;
+}
+
+void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst) {
+    //printf("src: %s ne: [%d %d %d %d] nb: [%d %d %d %d]\n", src->name, (int)src->ne[0], (int)src->ne[1], (int)src->ne[2], (int)src->ne[3], (int)src->nb[0], (int)src->nb[1], (int)src->nb[2], (int)src->nb[3]);
+    //printf("dst: %s ne: [%d %d %d %d] nb: [%d %d %d %d]\n", dst->name, (int)dst->ne[0], (int)dst->ne[1], (int)dst->ne[2], (int)dst->ne[3], (int)dst->nb[0], (int)dst->nb[1], (int)dst->nb[2], (int)dst->nb[3]);
+    GGML_ASSERT(ggml_are_same_layout(src, dst) && "cannot copy tensors with different layouts");
+
+    // fprintf(stderr, "cpy tensor %s from %s to %s (%lu bytes)\n", src->name, ggml_backend_name(src->backend), ggml_backend_name(dst->backend), ggml_nbytes(src));
+
+    if (src == dst) {
+        return;
+    }
+
+    // TODO: allow backends to support copy to/from same backend
+
+    if (ggml_get_backend(dst)->iface.cpy_tensor_from != NULL) {
+        ggml_get_backend(dst)->iface.cpy_tensor_from(ggml_get_backend(dst)->context, src, dst);
+    } else if (ggml_get_backend(src)->iface.cpy_tensor_to != NULL) {
+        ggml_get_backend(src)->iface.cpy_tensor_to(ggml_get_backend(src)->context, src, dst);
+    } else {
+        // shouldn't be hit when copying from/to CPU
+        #ifndef NDEBUG
+        fprintf(stderr, "ggml_backend_tensor_copy: neither cpy_tensor_from nor cpy_tensor_to are implemented for backends %s and %s, falling back to get/set\n", ggml_backend_name(src->buffer->backend), ggml_backend_name(dst->buffer->backend));
+        #endif
+        size_t nbytes = ggml_nbytes(src);
+        void * data = malloc(nbytes);
+        ggml_backend_tensor_get(src, data, 0, nbytes);
+        ggml_backend_tensor_set(dst, data, 0, nbytes);
+        free(data);
+    }
+}
+
+// backend CPU
+
+struct ggml_backend_cpu_context {
+    int n_threads;
+    void * work_data;
+    size_t work_size;
+};
+
+static const char * ggml_backend_cpu_name(ggml_backend_t backend) {
+    return "CPU";
+
+    UNUSED(backend);
+}
+
+static void ggml_backend_cpu_free(ggml_backend_t backend) {
+    struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
+    free(cpu_ctx->work_data);
+    free(cpu_ctx);
+    free(backend);
+}
+
+static void * ggml_backend_cpu_buffer_get_base(ggml_backend_buffer_t buffer) {
+    return (void *)buffer->context;
+}
+
+static void ggml_backend_cpu_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+    free(buffer->context);
+    UNUSED(buffer);
+}
+
+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,
+    /* .get_alloc_size = */ NULL, // defaults to ggml_nbytes
+    /* .init_tensor    = */ NULL, // no initialization required
+    /* .free_tensor    = */ NULL, // no cleanup required
+};
+
+// for buffers from ptr, free is not called
+static struct ggml_backend_buffer_i cpu_backend_buffer_i_from_ptr = {
+    /* .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,
+    /* .get_alloc_size = */ NULL, // defaults to ggml_nbytes
+    /* .init_tensor    = */ NULL,
+    /* .free_tensor    = */ NULL,
+};
+
+static const size_t TENSOR_ALIGNMENT = 64; // should be enough for AVX 512
+
+static ggml_backend_buffer_t ggml_backend_cpu_alloc_buffer(ggml_backend_t backend, size_t size) {
+    size += TENSOR_ALIGNMENT;   // malloc may return an address that is not aligned
+    void * data = malloc(size); // TODO: maybe use GGML_ALIGNED_MALLOC?
+
+    GGML_ASSERT(data != NULL && "failed to allocate buffer");
+
+    return ggml_backend_buffer_init(backend, cpu_backend_buffer_i, data, size);
+}
+
+static size_t ggml_backend_cpu_get_alignment(ggml_backend_t backend) {
+    return TENSOR_ALIGNMENT;
+    UNUSED(backend);
+}
+
+static void ggml_backend_cpu_set_tensor_async(ggml_backend_t backend, 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(backend);
+}
+
+static void ggml_backend_cpu_get_tensor_async(ggml_backend_t backend, 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(backend);
+}
+
+static void ggml_backend_cpu_synchronize(ggml_backend_t backend) {
+    UNUSED(backend);
+}
+
+static void ggml_backend_cpu_cpy_tensor_from(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst) {
+    ggml_backend_tensor_get(src, dst->data, 0, ggml_nbytes(src));
+
+    UNUSED(backend);
+}
+
+static void ggml_backend_cpu_cpy_tensor_to(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst) {
+    ggml_backend_tensor_set(dst, src->data, 0, ggml_nbytes(src));
+
+    UNUSED(backend);
+}
+
+struct ggml_backend_plan_cpu {
+    struct ggml_cplan cplan;
+    struct ggml_cgraph cgraph;
+};
+
+static ggml_backend_graph_plan_t ggml_backend_cpu_graph_plan_create(ggml_backend_t backend, 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;
+
+    if (cpu_plan->cplan.work_size > 0) {
+        cpu_plan->cplan.work_data = malloc(cpu_plan->cplan.work_size);
+    }
+
+    return cpu_plan;
+}
+
+static void ggml_backend_cpu_graph_plan_free(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
+    struct ggml_backend_plan_cpu * cpu_plan = (struct ggml_backend_plan_cpu *)plan;
+
+    free(cpu_plan->cplan.work_data);
+    free(cpu_plan);
+
+    UNUSED(backend);
+}
+
+static void 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;
+
+    ggml_graph_compute(&cpu_plan->cgraph, &cpu_plan->cplan);
+
+    UNUSED(backend);
+}
+
+static void 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) {
+        // TODO: may be faster to free and use malloc to avoid the copy
+        cpu_ctx->work_data = realloc(cpu_ctx->work_data, cplan.work_size);
+        cpu_ctx->work_size = cplan.work_size;
+    }
+
+    cplan.work_data = cpu_ctx->work_data;
+
+    ggml_graph_compute(cgraph, &cplan);
+}
+
+static bool ggml_backend_cpu_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
+    return true;
+    UNUSED(backend);
+    UNUSED(op);
+}
+
+static struct ggml_backend_i cpu_backend_i = {
+    /* .get_name            = */ ggml_backend_cpu_name,
+    /* .free                = */ ggml_backend_cpu_free,
+    /* .alloc_buffer        = */ ggml_backend_cpu_alloc_buffer,
+    /* .get_alignment       = */ ggml_backend_cpu_get_alignment,
+    /* .set_tensor_async    = */ ggml_backend_cpu_set_tensor_async,
+    /* .get_tensor_async    = */ ggml_backend_cpu_get_tensor_async,
+    /* .synchronize         = */ ggml_backend_cpu_synchronize,
+    /* .cpy_tensor_from     = */ ggml_backend_cpu_cpy_tensor_from,
+    /* .cpy_tensor_to       = */ ggml_backend_cpu_cpy_tensor_to,
+    /* .graph_plan_create   = */ ggml_backend_cpu_graph_plan_create,
+    /* .graph_plan_free     = */ ggml_backend_cpu_graph_plan_free,
+    /* .graph_plan_compute  = */ ggml_backend_cpu_graph_plan_compute,
+    /* .graph_compute       = */ ggml_backend_cpu_graph_compute,
+    /* .supports_op         = */ ggml_backend_cpu_supports_op,
+};
+
+ggml_backend_t ggml_backend_cpu_init(void) {
+    struct ggml_backend_cpu_context * ctx = malloc(sizeof(struct ggml_backend_cpu_context));
+
+    ctx->n_threads = GGML_DEFAULT_N_THREADS;
+    ctx->work_data = NULL;
+    ctx->work_size = 0;
+
+    ggml_backend_t cpu_backend = malloc(sizeof(struct ggml_backend));
+
+    *cpu_backend = (struct ggml_backend) {
+        /* .interface = */ cpu_backend_i,
+        /* .context   = */ ctx
+    };
+    return cpu_backend;
+}
+
+bool ggml_backend_is_cpu(ggml_backend_t backend) {
+    return backend->iface.get_name == ggml_backend_cpu_name;
+}
+
+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;
+}
+
+ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(ggml_backend_t backend_cpu, void * ptr, size_t size) {
+    return ggml_backend_buffer_init(backend_cpu, cpu_backend_buffer_i_from_ptr, ptr, size);
+}
+
+// scheduler
+
+#define GGML_MAX_BACKENDS 4
+#define GGML_MAX_SPLITS 256
+#define GGML_MAX_SPLIT_INPUTS 16
+
+struct ggml_backend_sched_split {
+    ggml_tallocr_t tallocr;
+    int i_start;
+    int i_end;
+    struct ggml_tensor * inputs[GGML_MAX_SPLIT_INPUTS];
+    int n_inputs;
+    struct ggml_cgraph * graph;
+};
+
+struct ggml_backend_sched {
+    int n_backends;
+    ggml_backend_t backends[GGML_MAX_BACKENDS];
+    ggml_tallocr_t  tallocs[GGML_MAX_BACKENDS];
+
+    ggml_gallocr_t galloc;
+
+    struct ggml_hash_set    hash_set;
+    ggml_tallocr_t *        node_talloc;                     // [hash_set.size]
+    struct ggml_tensor * (* node_copies)[GGML_MAX_BACKENDS]; // [hash_set.size][GGML_MAX_BACKENDS]
+
+    struct ggml_cgraph * graph;
+    struct ggml_backend_sched_split splits[GGML_MAX_SPLITS];
+    int n_splits;
+
+    struct ggml_context * ctx;
+
+    // align context_buffer to GGML_MEM_ALIGN
+    #ifdef _MSC_VER
+    __declspec(align(GGML_MEM_ALIGN))
+    #else
+    __attribute__((aligned(GGML_MEM_ALIGN)))
+    #endif
+    char context_buffer[GGML_MAX_SPLITS*GGML_MAX_SPLIT_INPUTS*sizeof(struct ggml_tensor) + GGML_MAX_SPLITS*sizeof(struct ggml_cgraph)];
+};
+
+#define hash_id(node) ggml_hash_find_or_insert(sched->hash_set, node)
+#define node_allocr(node) sched->node_talloc[hash_id(node)]
+
+static bool ggml_is_view_op(enum ggml_op op) {
+    return op == GGML_OP_VIEW || op == GGML_OP_RESHAPE || op == GGML_OP_PERMUTE || op == GGML_OP_TRANSPOSE;
+}
+
+// returns the priority of the backend, lower is better
+static int sched_backend_prio(ggml_backend_sched_t sched, ggml_backend_t backend) {
+    for (int i = 0; i < sched->n_backends; i++) {
+        if (sched->backends[i] == backend) {
+            return i;
+        }
+    }
+    return INT_MAX;
+}
+
+static int sched_allocr_prio(ggml_backend_sched_t sched, ggml_tallocr_t allocr) {
+    for (int i = 0; i < sched->n_backends; i++) {
+        if (sched->tallocs[i] == allocr) {
+            return i;
+        }
+    }
+    return INT_MAX;
+}
+
+// returns the backend that should be used for the node based on the current locations
+char causes[GGML_DEFAULT_GRAPH_SIZE*4 + GGML_MAX_SPLITS*GGML_MAX_SPLIT_INPUTS][128]; // debug, remove
+static ggml_backend_t sched_backend_from_cur(ggml_backend_sched_t sched, struct ggml_tensor * node) {
+    // if the dst tensor is already allocated in a buffer, we must assume that it is critical to keep it there
+    // ie. kv cache updates
+    // note that this doesn't allow fallback to CPU. need to add output tensors to the splits to copy the data back to the original backend.
+    // dst
+    ggml_backend_t cur_backend = ggml_get_backend(node);
+    if (cur_backend != NULL) {
+        sprintf(causes[hash_id(node)], "1.dst");
+        return cur_backend;
+    }
+
+    // view_src
+    if (node->view_src != NULL && ggml_get_backend(node->view_src) != NULL) {
+        sprintf(causes[hash_id(node)], "1.vsrc");
+        return ggml_get_backend(node->view_src);
+    }
+
+    // src
+    int cur_prio = INT_MAX;
+    size_t cur_size = 0;
+
+    for (int i = 0; i < GGML_MAX_SRC; i++) {
+        const struct ggml_tensor * src = node->src[i];
+        if (src == NULL) {
+            break;
+        }
+        ggml_backend_t src_backend = ggml_get_backend(src);
+        if (src_backend != NULL) {
+            int src_prio = sched_backend_prio(sched, src_backend);
+            size_t src_size = ggml_nbytes(src);
+            if (src_prio < cur_prio && src_size >= cur_size) {
+                cur_prio = src_prio;
+                cur_size = src_size;
+                cur_backend = src_backend;
+                sprintf(causes[hash_id(node)], "1.src%d", i);
+            }
+        }
+    }
+    return cur_backend;
+}
+
+static char * fmt_size(size_t size) {
+    static char buffer[128];
+    if (size >= 1024*1024) {
+        sprintf(buffer, "%zuM", size/1024/1024);
+    } else {
+        sprintf(buffer, "%zuK", size/1024);
+    }
+    return buffer;
+}
+
+static void sched_print_assignments(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
+    int cur_split = 0;
+    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 = ggml_tallocr_get_buffer(sched->splits[cur_split].tallocr)->backend;
+            fprintf(stderr, "\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, fmt_size(ggml_nbytes(sched->splits[cur_split].inputs[j])));
+            }
+            fprintf(stderr, "\n");
+            cur_split++;
+        }
+        struct ggml_tensor * node = graph->nodes[i];
+        if (ggml_is_view_op(node->op)) {
+            continue;
+        }
+        ggml_tallocr_t node_allocr = node_allocr(node);
+        ggml_backend_t node_backend = node_allocr ? ggml_tallocr_get_buffer(node_allocr)->backend : NULL;
+        fprintf(stderr, "node #%3d (%10.10s): %20.20s (%4.4s) [%4.4s %8.8s]:", i, ggml_op_name(node->op), node->name, fmt_size(ggml_nbytes(node)), node_allocr ? ggml_backend_name(node_backend) : "NULL", causes[hash_id(node)]);
+        for (int j = 0; j < GGML_MAX_SRC; j++) {
+            struct ggml_tensor * src = node->src[j];
+            if (src == NULL) {
+                break;
+            }
+            ggml_tallocr_t src_allocr = node_allocr(src);
+            ggml_backend_t src_backend = src_allocr ? ggml_tallocr_get_buffer(src_allocr)->backend : NULL;
+            fprintf(stderr, " %20.20s (%4.4s) [%4.4s %8.8s]", src->name, fmt_size(ggml_nbytes(src)), src_backend ? ggml_backend_name(src_backend) : "NULL", causes[hash_id(src)]);
+        }
+        fprintf(stderr, "\n");
+    }
+}
+
+// creates a copy of the tensor with the same memory layout
+static struct ggml_tensor * ggml_dup_tensor_layout(struct ggml_context * ctx, const struct ggml_tensor * tensor) {
+    struct ggml_tensor * dup = ggml_dup_tensor(ctx, tensor);
+    for (int i = 0; i < GGML_MAX_DIMS; i++) {
+        dup->nb[i] = tensor->nb[i];
+    }
+    return dup;
+}
+
+// assigns backends to ops and splits the graph into subgraphs that can be computed on the same backend
+// TODO: merge passes
+static void sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
+    // reset state
+    size_t hash_size = sched->hash_set.size;
+    memset(sched->hash_set.keys, 0, sizeof(sched->hash_set.keys[0]) * hash_size);
+    memset(sched->node_talloc,   0, sizeof(sched->node_talloc[0])   * hash_size);
+    memset(sched->node_copies,   0, sizeof(sched->node_copies[0])   * hash_size);
+    sched->n_splits = 0;
+
+    struct ggml_init_params params = {
+        /*.mem_size =   */ sizeof(sched->context_buffer),
+        /*.mem_buffer = */ sched->context_buffer,
+        /*.no_alloc =   */ true
+    };
+
+    if (sched->ctx != NULL) {
+        ggml_free(sched->ctx);
+    }
+
+    sched->ctx = ggml_init(params);
+
+    // pass 1: assign backends to ops with allocated inputs
+    for (int i = 0; i < graph->n_leafs; i++) {
+        struct ggml_tensor * leaf = graph->leafs[i];
+        if (node_allocr(leaf) != NULL) {
+            // do not overwrite user assignments
+            continue;
+        }
+        ggml_backend_t leaf_backend = ggml_get_backend(leaf);
+        if (leaf_backend == NULL && leaf->view_src != NULL) {
+            leaf_backend = ggml_get_backend(leaf->view_src);
+        }
+        if (leaf_backend != NULL) {
+            node_allocr(leaf) = ggml_backend_sched_get_tallocr(sched, leaf_backend);
+        }
+    }
+
+    for (int i = 0; i < graph->n_nodes; i++) {
+        struct ggml_tensor * node = graph->nodes[i];
+        if (node_allocr(node) != NULL) {
+            // do not overwrite user assignments
+            continue;
+        }
+        ggml_backend_t node_backend = sched_backend_from_cur(sched, node);
+        if (node_backend != NULL) {
+            node_allocr(node) = ggml_backend_sched_get_tallocr(sched, node_backend);
+        }
+    }
+    //printf("PASS 1 ASSIGNMENTS\n"); sched_print_assignments(sched, graph);
+
+    // pass 2: assign backends to ops from current assignments
+    // TODO:
+    //  - reuse sched_backend_from_cur
+    for (int i = 0; i < graph->n_nodes; i++) {
+        struct ggml_tensor * node = graph->nodes[i];
+        ggml_tallocr_t node_allocr = node_allocr(node);
+        if (node_allocr == NULL) {
+            int    cur_prio = INT_MAX;
+            size_t cur_size = 0;
+            for (int j = 0; j < GGML_MAX_SRC; j++) {
+                struct ggml_tensor * src = node->src[j];
+                if (src == NULL) {
+                    break;
+                }
+                ggml_tallocr_t src_allocr = node_allocr(src);
+                if (src_allocr != NULL) {
+                    int    src_prio = sched_allocr_prio(sched, src_allocr);
+                    size_t src_size = ggml_nbytes(src);
+                    if (src_prio < cur_prio && src_size >= cur_size) {
+                        cur_prio = src_prio;
+                        cur_size = src_size;
+                        node_allocr = src_allocr;
+                        sprintf(causes[hash_id(node)], "2.src%d", j);
+                    }
+                }
+            }
+            if (node_allocr != NULL) {
+                node_allocr(node) = node_allocr;
+            }
+        }
+    }
+    //printf("PASS 2 ASSIGNMENTS\n"); sched_print_assignments(sched, graph);
+
+    // pass 3: assign backends to remaining src from dst (should only be leafs)
+    for (int i = 0; i < graph->n_nodes; i++) {
+        struct ggml_tensor * node = graph->nodes[i];
+        ggml_tallocr_t node_allocr = node_allocr(node);
+        for (int j = 0; j < GGML_MAX_SRC; j++) {
+            struct ggml_tensor * src = node->src[j];
+            if (src == NULL) {
+                break;
+            }
+            ggml_tallocr_t src_allocr = node_allocr(src);
+            if (src_allocr == NULL) {
+                node_allocr(src) = node_allocr;
+            }
+        }
+    }
+    //printf("PASS 3 ASSIGNMENTS\n"); sched_print_assignments(sched, graph);
+
+    // pass 4: split graph, find tensors that need to be copied
+    // TODO:
+    //  - when switching from a less preferred backend to a more preferred backend, check if it is possible to move the switch to an earlier point for the same cost
+    // find first backend
+    int cur_split = 0;
+    for (int i = 0; i < graph->n_nodes; i++) {
+        struct ggml_tensor * node = graph->nodes[i];
+        if (node->view_src == NULL) {
+            sched->splits[0].tallocr = node_allocr(node);
+            break;
+        }
+    }
+    sched->splits[0].i_start = 0;
+    sched->splits[0].n_inputs = 0;
+    memset(sched->splits[0].inputs, 0, sizeof(sched->splits[0].inputs)); //HACK
+    ggml_tallocr_t cur_allocr = sched->splits[0].tallocr;
+    size_t cur_backend_id = sched_allocr_prio(sched, cur_allocr);
+    for (int i = 0; i < graph->n_nodes; i++) {
+        struct ggml_tensor * node = graph->nodes[i];
+
+        if (ggml_is_view_op(node->op)) {
+            continue;
+        }
+
+        ggml_tallocr_t node_allocr = node_allocr(node);
+
+        if (node_allocr != cur_allocr) {
+            sched->splits[cur_split].i_end = i;
+            cur_split++;
+            GGML_ASSERT(cur_split < GGML_MAX_SPLITS);
+            sched->splits[cur_split].tallocr = node_allocr;
+            sched->splits[cur_split].i_start = i;
+            sched->splits[cur_split].n_inputs = 0;
+            memset(sched->splits[cur_split].inputs, 0, sizeof(sched->splits[cur_split].inputs)); //HACK
+            cur_allocr = node_allocr;
+            cur_backend_id = sched_allocr_prio(sched, cur_allocr);
+        }
+
+        // find inputs that are not on the same backend
+        for (int j = 0; j < GGML_MAX_SRC; j++) {
+            struct ggml_tensor * src = node->src[j];
+            if (src == NULL) {
+                break;
+            }
+            ggml_tallocr_t src_allocr = node_allocr(src);
+            if (src_allocr != node_allocr) {
+                int n_inputs = sched->splits[cur_split].n_inputs++;
+                GGML_ASSERT(n_inputs < GGML_MAX_SPLIT_INPUTS);
+                sched->splits[cur_split].inputs[n_inputs] = (struct ggml_tensor *)src;
+
+                // create copies
+                size_t id = hash_id(src);
+                if (sched->node_copies[id][cur_backend_id] == NULL) {
+                    struct ggml_tensor * tensor_copy = ggml_dup_tensor_layout(sched->ctx, src);
+                    sched->node_copies[id][cur_backend_id] = tensor_copy;
+                    node_allocr(tensor_copy) = cur_allocr;
+                    ggml_backend_t backend = ggml_tallocr_get_buffer(cur_allocr)->backend;
+                    ggml_format_name(tensor_copy, "%s#%s", ggml_backend_name(backend), src->name);
+                }
+                node->src[j] = sched->node_copies[id][cur_backend_id];
+            }
+        }
+    }
+    sched->splits[cur_split].i_end = graph->n_nodes;
+    sched->n_splits = cur_split + 1;
+
+    //fprintf(stderr, "PASS 4 ASSIGNMENTS\n"); sched_print_assignments(sched, graph); fflush(stdout);
+
+#if 1
+    // sanity check: all sources should have the same backend as the node
+    for (int i = 0; i < graph->n_nodes; i++) {
+        struct ggml_tensor * node = graph->nodes[i];
+        ggml_tallocr_t node_allocr = node_allocr(node);
+        if (node_allocr == NULL) {
+            fprintf(stderr, "!!!!!!! %s has no backend\n", node->name);
+        }
+        for (int j = 0; j < GGML_MAX_SRC; j++) {
+            struct ggml_tensor * src = node->src[j];
+            if (src == NULL) {
+                break;
+            }
+            ggml_tallocr_t src_allocr = node_allocr(src);
+            if (src_allocr != node_allocr /* && src_backend != NULL */) { // ignore nulls for now
+                fprintf(stderr, "!!!! %s has backend %s, src %d (%s) has backend %s\n",
+                    node->name, node_allocr ? ggml_backend_name(ggml_tallocr_get_buffer(node_allocr)->backend) : "NULL",
+                    j, src->name, src_allocr ? ggml_backend_name(ggml_tallocr_get_buffer(src_allocr)->backend) : "NULL");
+            }
+        }
+    }
+#endif
+
+    // create copies of the graph for each split
+    // FIXME: avoid this copy, pass split inputs to ggml_gallocr_alloc_graph_n in some other way
+    struct ggml_cgraph * graph_copy = ggml_new_graph_custom(sched->ctx, graph->n_nodes + sched->n_splits*GGML_MAX_SPLIT_INPUTS, false);
+    for (int i = 0; i < sched->n_splits; i++) {
+        struct ggml_backend_sched_split * split = &sched->splits[i];
+        split->graph = ggml_graph_view(sched->ctx, graph, split->i_start, split->i_end);
+
+        // add inputs to the graph copy so that they are allocated by ggml-alloc at the start of the split
+        for (int j = 0; j < split->n_inputs; j++) {
+            struct ggml_tensor * input = split->inputs[j];
+            struct ggml_tensor * input_cpy = sched->node_copies[hash_id(input)][sched_allocr_prio(sched, split->tallocr)];
+            input_cpy->src[0] = input;
+            graph_copy->nodes[graph_copy->n_nodes++] = input_cpy;
+        }
+
+        for (int j = split->i_start; j < split->i_end; j++) {
+            graph_copy->nodes[graph_copy->n_nodes++] = graph->nodes[j];
+        }
+    }
+    sched->graph = graph_copy;
+}
+
+static void sched_alloc_splits(ggml_backend_sched_t sched) {
+    ggml_gallocr_alloc_graph_n(
+        sched->galloc,
+        sched->graph,
+        sched->hash_set,
+        sched->node_talloc);
+}
+
+static void sched_compute_splits(ggml_backend_sched_t sched) {
+    uint64_t copy_us[GGML_MAX_BACKENDS] = {0};
+    uint64_t compute_us[GGML_MAX_BACKENDS] = {0};
+
+    struct ggml_backend_sched_split * splits = sched->splits;
+
+    for (int i = 0; i < sched->n_splits; i++) {
+        struct ggml_backend_sched_split * split = &splits[i];
+        ggml_backend_t split_backend = ggml_tallocr_get_buffer(split->tallocr)->backend;
+        int split_backend_id = sched_backend_prio(sched, split_backend);
+
+        // copy the input tensors to the split backend
+        uint64_t copy_start_us = ggml_time_us();
+        for (int j = 0; j < split->n_inputs; j++) {
+            struct ggml_tensor * input_cpy = sched->node_copies[hash_id(split->inputs[j])][sched_backend_prio(sched, split_backend)];
+            if (split->inputs[j]->buffer == NULL) {
+                if (split->inputs[j]->view_src == NULL) {
+                    fprintf(stderr, "input %s has no buffer and no view_src\n", split->inputs[j]->name);
+                    exit(1);
+                }
+                struct ggml_tensor * view = split->inputs[j];
+                view->backend = view->view_src->backend;
+                view->buffer  = view->view_src->buffer;
+                view->data    = (char *)view->view_src->data + view->view_offs;
+                ggml_backend_buffer_init_tensor(ggml_backend_sched_get_buffer(sched, view->buffer->backend), view);
+            }
+            if (input_cpy->buffer == NULL) {
+                fprintf(stderr, "input_cpy %s has no buffer\n", input_cpy->name);
+                exit(1);
+            }
+            GGML_ASSERT(split->inputs[j]->buffer->backend != input_cpy->buffer->backend);
+            GGML_ASSERT(input_cpy->buffer->backend == split_backend);
+            ggml_backend_tensor_copy(split->inputs[j], input_cpy);
+        }
+        // ggml_backend_synchronize(split_backend);
+        int64_t copy_end_us = ggml_time_us();
+        copy_us[split_backend_id] += copy_end_us - copy_start_us;
+
+#if 0
+        char split_filename[GGML_MAX_NAME];
+        snprintf(split_filename, GGML_MAX_NAME, "split_%i_%s.dot", i, ggml_backend_name(split_backend));
+        ggml_graph_dump_dot(split->graph, NULL, split_filename);
+#endif
+
+        uint64_t compute_start_us = ggml_time_us();
+        ggml_backend_graph_compute(split_backend, split->graph);
+        // ggml_backend_synchronize(split_backend);
+        uint64_t compute_end_us = ggml_time_us();
+        compute_us[split_backend_id] += compute_end_us - compute_start_us;
+    }
+
+#if 0
+    // per-backend timings
+    fprintf(stderr, "sched_compute_splits times (%d splits):\n", sched->n_splits);
+    for (int i = 0; i < sched->n_backends; i++) {
+        if (copy_us[i] > 0 || compute_us[i] > 0) {
+            fprintf(stderr, "\t%5.5s: %lu us copy, %lu us compute\n", ggml_backend_name(sched->backends[i]), copy_us[i], compute_us[i]);
+        }
+    }
+#endif
+}
+
+static void sched_reset(ggml_backend_sched_t sched) {
+    for (int i = 0; i < sched->n_backends; i++) {
+        ggml_tallocr_reset(sched->tallocs[i]);
+    }
+}
+
+ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, int n_backends) {
+    GGML_ASSERT(n_backends <= GGML_MAX_BACKENDS);
+
+    struct ggml_backend_sched * sched = malloc(sizeof(struct ggml_backend_sched));
+    memset(sched, 0, sizeof(struct ggml_backend_sched));
+
+    fprintf(stderr, "ggml_backend_sched size: %lu KB\n", sizeof(struct ggml_backend_sched)/1024);
+
+    sched->n_backends = n_backends;
+    for (int i = 0; i < n_backends; i++) {
+        sched->backends[i] = backends[i];
+    }
+
+    sched->galloc = ggml_gallocr_new();
+
+    // init measure allocs for each backend
+    for (int i = 0; i < n_backends; i++) {
+        sched->tallocs[i] = ggml_tallocr_new_measure_from_backend(backends[i]);
+    }
+
+    return sched;
+}
+
+void ggml_backend_sched_free(ggml_backend_sched_t sched) {
+    if (sched == NULL) {
+        return;
+    }
+    for (int i = 0; i < sched->n_backends; i++) {
+        ggml_tallocr_free(sched->tallocs[i]);
+    }
+    ggml_gallocr_free(sched->galloc);
+    free(sched->hash_set.keys);
+    free(sched->node_talloc);
+    free(sched->node_copies);
+    free(sched);
+}
+
+void ggml_backend_sched_init_measure(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph) {
+    // initialize hash tables
+    size_t hash_size = measure_graph->visited_hash_table.size + GGML_MAX_SPLITS*GGML_MAX_SPLIT_INPUTS;
+    sched->hash_set.size = hash_size;
+    sched->hash_set.keys = malloc(sizeof(sched->hash_set.keys[0]) * hash_size);
+    sched->node_talloc   = malloc(sizeof(sched->node_talloc[0])   * hash_size);
+    sched->node_copies   = malloc(sizeof(sched->node_copies[0])   * hash_size);
+
+    sched_split_graph(sched, measure_graph);
+    sched_alloc_splits(sched);
+
+    // allocate buffers and reset allocators
+    for (int i = 0; i < sched->n_backends; i++) {
+        size_t size = ggml_tallocr_max_size(sched->tallocs[i]);
+        ggml_tallocr_free(sched->tallocs[i]);
+        sched->tallocs[i] = ggml_tallocr_new_from_backend(sched->backends[i], size);
+    }
+
+    sched_reset(sched);
+}
+
+void ggml_backend_sched_graph_compute(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
+    GGML_ASSERT(sched->hash_set.size >= graph->visited_hash_table.size + GGML_MAX_SPLITS*GGML_MAX_SPLIT_INPUTS);
+
+    sched_split_graph(sched, graph);
+    sched_alloc_splits(sched);
+    sched_compute_splits(sched);
+    sched_reset(sched);
+}
+
+ggml_tallocr_t ggml_backend_sched_get_tallocr(ggml_backend_sched_t sched, ggml_backend_t backend) {
+    int backend_index = sched_backend_prio(sched, backend);
+    return sched->tallocs[backend_index];
+}
+
+ggml_backend_buffer_t ggml_backend_sched_get_buffer(ggml_backend_sched_t sched, ggml_backend_t backend) {
+    int backend_index = sched_backend_prio(sched, backend);
+    return ggml_tallocr_get_buffer(sched->tallocs[backend_index]);
+}
+
+void ggml_backend_sched_set_node_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend) {
+    int backend_index = sched_backend_prio(sched, backend);
+    GGML_ASSERT(backend_index >= 0 && backend_index < sched->n_backends);
+    node_allocr(node) = sched->tallocs[backend_index];
+}

ggml-backend.h

@@ -0,0 +1,136 @@
+#pragma once
+
+#include "ggml.h"
+#include "ggml-alloc.h"
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+    //
+    // Backend buffer
+    //
+
+    struct ggml_backend_buffer;
+    typedef struct ggml_backend_buffer * ggml_backend_buffer_t;
+
+    // backend buffer functions
+    GGML_API void   ggml_backend_buffer_free          (ggml_backend_buffer_t buffer);
+    GGML_API size_t ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer);
+    GGML_API void * ggml_backend_buffer_get_base      (ggml_backend_buffer_t buffer);
+    GGML_API size_t ggml_backend_buffer_get_size      (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_init_tensor   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+    GGML_API void   ggml_backend_buffer_free_tensor   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+
+    //
+    // Backend
+    //
+
+    struct ggml_backend;
+    typedef struct ggml_backend * ggml_backend_t;
+    typedef void * ggml_backend_graph_plan_t;
+
+    GGML_API ggml_backend_t ggml_get_backend(const struct ggml_tensor * tensor);
+
+    GGML_API const char * ggml_backend_name(ggml_backend_t backend);
+    GGML_API void         ggml_backend_free(ggml_backend_t backend);
+
+    GGML_API ggml_backend_buffer_t ggml_backend_alloc_buffer(ggml_backend_t backend, size_t size);
+
+    GGML_API size_t ggml_backend_get_alignment(ggml_backend_t backend);
+
+    GGML_API void ggml_backend_tensor_set_async(      struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
+    GGML_API void ggml_backend_tensor_get_async(const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
+
+    GGML_API void ggml_backend_tensor_set(      struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
+    GGML_API void ggml_backend_tensor_get(const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
+
+    GGML_API void ggml_backend_synchronize(ggml_backend_t backend);
+
+    GGML_API ggml_backend_graph_plan_t ggml_backend_graph_plan_create (ggml_backend_t backend, struct ggml_cgraph * cgraph);
+
+    GGML_API void ggml_backend_graph_plan_free   (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
+    GGML_API void ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
+    GGML_API void ggml_backend_graph_compute     (ggml_backend_t backend, struct ggml_cgraph * cgraph);
+    GGML_API bool ggml_backend_supports_op       (ggml_backend_t backend, const struct ggml_tensor * op);
+
+    // tensor copy between different backends
+    GGML_API void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst);
+
+    //
+    // CPU backend
+    //
+
+    GGML_API ggml_backend_t ggml_backend_cpu_init(void);
+
+    GGML_API bool ggml_backend_is_cpu(ggml_backend_t backend);
+    GGML_API void ggml_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads);
+
+    // Create a backend buffer from an existing pointer
+    GGML_API ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(ggml_backend_t backend_cpu, void * ptr, size_t size);
+
+
+    //
+    // Backend scheduler
+    //
+
+    // The backend scheduler allows for multiple backends to be used together
+    // Handles compute buffer allocation, assignment of tensors to backends, and copying of tensors between backends
+    // The backends are selected based on:
+    // - the backend that supports the operation
+    // - the location of the pre-allocated tensors (e.g. the weights)
+    /*
+      Example usage:
+
+        sched = ggml_backend_sched_new({backend_gpu, backend_gpu2, backend_cpu}, num_backends);
+        // sched is initialized with measure allocators and cannot be used until allocated with a measure graph
+
+        // initialize buffers from a measure graph
+        measure_graph = build_graph(sched); // use the allocr to allocate inputs as needed
+
+        // in build_graph:
+        build_graph(...) {
+            // allocating tensors in a specific backend (optional, recommended: pre-allocate inputs in a different buffer)
+            alloc_cpu = ggml_backend_sched_get_allocr(sched, backend_cpu);
+            ggml_allocr_alloc(alloc_cpu, tensor);
+
+            // manually assigning nodes to a backend (optional, shouldn't be needed in most cases)
+            struct ggml_tensor * node = ggml_mul_mat(ctx, ...);
+            ggml_backend_sched_set_node_backend(sched, node, backend_gpu);
+        }
+
+        // allocate backend buffers from measure graph
+        ggml_backend_sched_init_measure(sched, measure_graph);
+
+        // the scheduler is now ready to compute graphs
+
+        // compute
+        graph = build_graph(sched);
+        ggml_backend_sched_graph_compute(sched, graph);
+    */
+
+    struct ggml_backend_sched;
+    typedef struct ggml_backend_sched * ggml_backend_sched_t;
+
+    // Initialize a backend scheduler
+    GGML_API ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, int n_backends);
+
+    GGML_API void ggml_backend_sched_free(ggml_backend_sched_t sched);
+
+    // Initialize backend buffers from a measure graph
+    GGML_API void ggml_backend_sched_init_measure(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph);
+
+    GGML_API ggml_tallocr_t        ggml_backend_sched_get_tallocr(ggml_backend_sched_t sched, ggml_backend_t backend);
+    GGML_API ggml_backend_buffer_t ggml_backend_sched_get_buffer (ggml_backend_sched_t sched, ggml_backend_t backend);
+
+    GGML_API void ggml_backend_sched_set_node_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend);
+
+    // Allocate a graph on the backend scheduler
+    GGML_API void ggml_backend_sched_graph_compute(
+            ggml_backend_sched_t sched,
+            struct ggml_cgraph * graph);
+
+#ifdef  __cplusplus
+}
+#endif

ggml-cuda.h

@@ -1,6 +1,7 @@
 #pragma once
 
 #include "ggml.h"
+#include "ggml-backend.h"
 
 #ifdef GGML_USE_HIPBLAS
 #define GGML_CUDA_NAME "ROCm"
@@ -31,6 +32,7 @@ GGML_API void   ggml_cuda_assign_buffers_force_inplace(struct ggml_tensor * tens
 
 GGML_API void   ggml_cuda_assign_buffers_no_alloc(struct ggml_tensor * tensor);
 GGML_API void   ggml_cuda_assign_scratch_offset(struct ggml_tensor * tensor, size_t offset);
+GGML_API void   ggml_cuda_copy_to_device(struct ggml_tensor * tensor);
 
 GGML_API void   ggml_cuda_set_main_device(int main_device);
 GGML_API void   ggml_cuda_set_mul_mat_q(bool mul_mat_q);
@@ -41,6 +43,9 @@ GGML_API bool   ggml_cuda_compute_forward(struct ggml_compute_params * params, s
 GGML_API int    ggml_cuda_get_device_count(void);
 GGML_API void   ggml_cuda_get_device_description(int device, char * description, size_t description_size);
 
+// backend API
+GGML_API ggml_backend_t ggml_backend_cuda_init(void); // TODO: take a list of devices to use
+
 #ifdef  __cplusplus
 }
 #endif

ggml-impl.h

@@ -0,0 +1,243 @@
+#pragma once
+
+#include "ggml.h"
+
+// GGML internal header
+
+#include <assert.h>
+#include <stddef.h>
+#include <stdbool.h>
+#include <string.h> // memcpy
+#include <math.h>   // fabsf
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// static_assert should be a #define, but if it's not,
+// fall back to the _Static_assert C11 keyword.
+// if C99 - static_assert is noop
+// ref: https://stackoverflow.com/a/53923785/4039976
+#ifndef static_assert
+#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201100L)
+#define static_assert(cond, msg) _Static_assert(cond, msg)
+#else
+#define static_assert(cond, msg) struct global_scope_noop_trick
+#endif
+#endif
+
+// __FMA__ and __F16C__ are not defined in MSVC, however they are implied with AVX2/AVX512
+#if defined(_MSC_VER) && (defined(__AVX2__) || defined(__AVX512F__))
+#ifndef __FMA__
+#define __FMA__
+#endif
+#ifndef __F16C__
+#define __F16C__
+#endif
+#ifndef __SSE3__
+#define __SSE3__
+#endif
+#endif
+
+// 16-bit float
+// on Arm, we use __fp16
+// on x86, we use uint16_t
+#if defined(__ARM_NEON) && !defined(_MSC_VER)
+
+// if YCM cannot find <arm_neon.h>, make a symbolic link to it, for example:
+//
+//   $ ln -sfn /Library/Developer/CommandLineTools/usr/lib/clang/13.1.6/include/arm_neon.h ./src/
+//
+#include <arm_neon.h>
+
+#define GGML_COMPUTE_FP16_TO_FP32(x) ((float) (x))
+#define GGML_COMPUTE_FP32_TO_FP16(x) (x)
+
+#define GGML_FP16_TO_FP32(x) ((float) (x))
+#define GGML_FP32_TO_FP16(x) (x)
+
+#else
+
+#ifdef __wasm_simd128__
+#include <wasm_simd128.h>
+#else
+#ifdef __POWER9_VECTOR__
+#include <altivec.h>
+#undef bool
+#define bool _Bool
+#else
+#if defined(_MSC_VER) || defined(__MINGW32__)
+#include <intrin.h>
+#else
+#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__) || defined(__SSE3__)
+#if !defined(__riscv)
+#include <immintrin.h>
+#endif
+#endif
+#endif
+#endif
+#endif
+
+#ifdef __riscv_v_intrinsic
+#include <riscv_vector.h>
+#endif
+
+#ifdef __F16C__
+
+#ifdef _MSC_VER
+#define GGML_COMPUTE_FP16_TO_FP32(x) _mm_cvtss_f32(_mm_cvtph_ps(_mm_cvtsi32_si128(x)))
+#define GGML_COMPUTE_FP32_TO_FP16(x) _mm_extract_epi16(_mm_cvtps_ph(_mm_set_ss(x), 0), 0)
+#else
+#define GGML_COMPUTE_FP16_TO_FP32(x) _cvtsh_ss(x)
+#define GGML_COMPUTE_FP32_TO_FP16(x) _cvtss_sh(x, 0)
+#endif
+
+#elif defined(__POWER9_VECTOR__)
+
+#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
+#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
+/* the inline asm below is about 12% faster than the lookup method */
+#define GGML_FP16_TO_FP32(x) GGML_COMPUTE_FP16_TO_FP32(x)
+#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
+
+static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
+    register float f;
+    register double d;
+    __asm__(
+        "mtfprd %0,%2\n"
+        "xscvhpdp %0,%0\n"
+        "frsp %1,%0\n" :
+        /* temp */ "=d"(d),
+        /* out */  "=f"(f):
+        /* in */   "r"(h));
+    return f;
+}
+
+static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
+    register double d;
+    register ggml_fp16_t r;
+    __asm__( /* xscvdphp can work on double or single precision */
+        "xscvdphp %0,%2\n"
+        "mffprd %1,%0\n" :
+        /* temp */ "=d"(d),
+        /* out */  "=r"(r):
+        /* in */   "f"(f));
+    return r;
+}
+
+#else
+
+// FP16 <-> FP32
+// ref: https://github.com/Maratyszcza/FP16
+
+static inline float fp32_from_bits(uint32_t w) {
+    union {
+        uint32_t as_bits;
+        float as_value;
+    } fp32;
+    fp32.as_bits = w;
+    return fp32.as_value;
+}
+
+static inline uint32_t fp32_to_bits(float f) {
+    union {
+        float as_value;
+        uint32_t as_bits;
+    } fp32;
+    fp32.as_value = f;
+    return fp32.as_bits;
+}
+
+static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
+    const uint32_t w = (uint32_t) h << 16;
+    const uint32_t sign = w & UINT32_C(0x80000000);
+    const uint32_t two_w = w + w;
+
+    const uint32_t exp_offset = UINT32_C(0xE0) << 23;
+#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)
+    const float exp_scale = 0x1.0p-112f;
+#else
+    const float exp_scale = fp32_from_bits(UINT32_C(0x7800000));
+#endif
+    const float normalized_value = fp32_from_bits((two_w >> 4) + exp_offset) * exp_scale;
+
+    const uint32_t magic_mask = UINT32_C(126) << 23;
+    const float magic_bias = 0.5f;
+    const float denormalized_value = fp32_from_bits((two_w >> 17) | magic_mask) - magic_bias;
+
+    const uint32_t denormalized_cutoff = UINT32_C(1) << 27;
+    const uint32_t result = sign |
+        (two_w < denormalized_cutoff ? fp32_to_bits(denormalized_value) : fp32_to_bits(normalized_value));
+    return fp32_from_bits(result);
+}
+
+static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
+#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)
+    const float scale_to_inf = 0x1.0p+112f;
+    const float scale_to_zero = 0x1.0p-110f;
+#else
+    const float scale_to_inf = fp32_from_bits(UINT32_C(0x77800000));
+    const float scale_to_zero = fp32_from_bits(UINT32_C(0x08800000));
+#endif
+    float base = (fabsf(f) * scale_to_inf) * scale_to_zero;
+
+    const uint32_t w = fp32_to_bits(f);
+    const uint32_t shl1_w = w + w;
+    const uint32_t sign = w & UINT32_C(0x80000000);
+    uint32_t bias = shl1_w & UINT32_C(0xFF000000);
+    if (bias < UINT32_C(0x71000000)) {
+        bias = UINT32_C(0x71000000);
+    }
+
+    base = fp32_from_bits((bias >> 1) + UINT32_C(0x07800000)) + base;
+    const uint32_t bits = fp32_to_bits(base);
+    const uint32_t exp_bits = (bits >> 13) & UINT32_C(0x00007C00);
+    const uint32_t mantissa_bits = bits & UINT32_C(0x00000FFF);
+    const uint32_t nonsign = exp_bits + mantissa_bits;
+    return (sign >> 16) | (shl1_w > UINT32_C(0xFF000000) ? UINT16_C(0x7E00) : nonsign);
+}
+
+#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
+#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
+
+#endif // __F16C__
+
+#endif // __ARM_NEON
+
+// precomputed f32 table for f16 (256 KB)
+// defined in ggml.c, initialized in ggml_init()
+extern float ggml_table_f32_f16[1 << 16];
+
+// On ARM NEON, it's quicker to directly convert x -> x instead of calling into ggml_lookup_fp16_to_fp32,
+// so we define GGML_FP16_TO_FP32 and GGML_FP32_TO_FP16 elsewhere for NEON.
+// This is also true for POWER9.
+#if !defined(GGML_FP16_TO_FP32) || !defined(GGML_FP32_TO_FP16)
+
+inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
+    uint16_t s;
+    memcpy(&s, &f, sizeof(uint16_t));
+    return ggml_table_f32_f16[s];
+}
+
+#define GGML_FP16_TO_FP32(x) ggml_lookup_fp16_to_fp32(x)
+#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
+
+#endif
+
+#define GGML_HASHTABLE_FULL ((size_t)-1)
+#define GGML_HASHTABLE_ALREADY_EXISTS ((size_t)-2)
+
+bool   ggml_hash_contains      (const struct ggml_hash_set hash_set, struct ggml_tensor * key);
+
+// returns GGML_HASHTABLE_FULL if table is full, otherwise the current index of the key or where it should be inserted
+size_t ggml_hash_find          (const struct ggml_hash_set hash_set, struct ggml_tensor * key);
+
+// returns GGML_HAHSHTABLE_ALREADY_EXISTS if key already exists, index otherwise, asserts if table is full
+size_t ggml_hash_insert        (      struct ggml_hash_set hash_set, struct ggml_tensor * key);
+
+// return index, asserts if table is full
+size_t ggml_hash_find_or_insert(      struct ggml_hash_set hash_set, struct ggml_tensor * key);
+
+#ifdef __cplusplus
+}
+#endif

ggml-metal.h

@@ -19,6 +19,9 @@
 
 #pragma once
 
+#include "ggml.h"
+#include "ggml-backend.h"
+
 #include <stddef.h>
 #include <stdbool.h>
 
@@ -33,8 +36,15 @@ struct ggml_cgraph;
 extern "C" {
 #endif
 
+//
+// internal API
+// temporary exposed to user-code
+//
+
 struct ggml_metal_context;
 
+void ggml_metal_log_set_callback(ggml_log_callback log_callback, void * user_data);
+
 // number of command buffers to use
 struct ggml_metal_context * ggml_metal_init(int n_cb);
 void ggml_metal_free(struct ggml_metal_context * ctx);
@@ -79,6 +89,17 @@ int * ggml_metal_get_concur_list(struct ggml_metal_context * ctx);
 // creates gf->n_threads command buffers in parallel
 void ggml_metal_graph_compute(struct ggml_metal_context * ctx, struct ggml_cgraph * gf);
 
+//
+// backend API
+// user-code should use only these functions
+//
+
+GGML_API ggml_backend_t ggml_backend_metal_init(void);
+
+GGML_API bool ggml_backend_is_metal(ggml_backend_t backend);
+
+GGML_API void ggml_backend_metal_set_n_cb(ggml_backend_t backend, int n_cb);
+
 #ifdef __cplusplus
 }
 #endif

ggml-metal.m

@@ -1,5 +1,6 @@
 #import "ggml-metal.h"
 
+#import "ggml-backend-impl.h"
 #import "ggml.h"
 
 #import <Foundation/Foundation.h>
@@ -11,16 +12,19 @@
 #define MIN(a, b) ((a) < (b) ? (a) : (b))
 #define MAX(a, b) ((a) > (b) ? (a) : (b))
 
-// TODO: temporary - reuse llama.cpp logging
 #ifdef GGML_METAL_NDEBUG
-#define metal_printf(...)
+#define GGML_METAL_LOG_INFO(...)
+#define GGML_METAL_LOG_WARN(...)
+#define GGML_METAL_LOG_ERROR(...)
 #else
-#define metal_printf(...) fprintf(stderr, __VA_ARGS__)
+#define GGML_METAL_LOG_INFO(...)  ggml_metal_log(GGML_LOG_LEVEL_INFO, __VA_ARGS__)
+#define GGML_METAL_LOG_WARN(...)  ggml_metal_log(GGML_LOG_LEVEL_WARN, __VA_ARGS__)
+#define GGML_METAL_LOG_ERROR(...) ggml_metal_log(GGML_LOG_LEVEL_ERROR, __VA_ARGS__)
 #endif
 
 #define UNUSED(x) (void)(x)
 
-#define GGML_MAX_CONCUR (2*GGML_MAX_NODES)
+#define GGML_MAX_CONCUR (2*GGML_DEFAULT_GRAPH_SIZE)
 
 struct ggml_metal_buffer {
     const char * name;
@@ -59,6 +63,7 @@ struct ggml_metal_context {
     GGML_METAL_DECL_KERNEL(mul);
     GGML_METAL_DECL_KERNEL(mul_row); // TODO: avoid this extra kernel, instead extend the "mul" kernel to support broadcast
     GGML_METAL_DECL_KERNEL(scale);
+    GGML_METAL_DECL_KERNEL(scale_4);
     GGML_METAL_DECL_KERNEL(silu);
     GGML_METAL_DECL_KERNEL(relu);
     GGML_METAL_DECL_KERNEL(gelu);
@@ -70,6 +75,8 @@ struct ggml_metal_context {
     GGML_METAL_DECL_KERNEL(get_rows_f16);
     GGML_METAL_DECL_KERNEL(get_rows_q4_0);
     GGML_METAL_DECL_KERNEL(get_rows_q4_1);
+    GGML_METAL_DECL_KERNEL(get_rows_q5_0);
+    GGML_METAL_DECL_KERNEL(get_rows_q5_1);
     GGML_METAL_DECL_KERNEL(get_rows_q8_0);
     GGML_METAL_DECL_KERNEL(get_rows_q2_K);
     GGML_METAL_DECL_KERNEL(get_rows_q3_K);
@@ -78,33 +85,40 @@ struct ggml_metal_context {
     GGML_METAL_DECL_KERNEL(get_rows_q6_K);
     GGML_METAL_DECL_KERNEL(rms_norm);
     GGML_METAL_DECL_KERNEL(norm);
-    GGML_METAL_DECL_KERNEL(mul_mat_f32_f32);
-    GGML_METAL_DECL_KERNEL(mul_mat_f16_f32);
-    GGML_METAL_DECL_KERNEL(mul_mat_f16_f32_1row);
-    GGML_METAL_DECL_KERNEL(mul_mat_f16_f32_l4);
-    GGML_METAL_DECL_KERNEL(mul_mat_q4_0_f32);
-    GGML_METAL_DECL_KERNEL(mul_mat_q4_1_f32);
-    GGML_METAL_DECL_KERNEL(mul_mat_q8_0_f32);
-    GGML_METAL_DECL_KERNEL(mul_mat_q2_K_f32);
-    GGML_METAL_DECL_KERNEL(mul_mat_q3_K_f32);
-    GGML_METAL_DECL_KERNEL(mul_mat_q4_K_f32);
-    GGML_METAL_DECL_KERNEL(mul_mat_q5_K_f32);
-    GGML_METAL_DECL_KERNEL(mul_mat_q6_K_f32);
+    GGML_METAL_DECL_KERNEL(mul_mv_f32_f32);
+    GGML_METAL_DECL_KERNEL(mul_mv_f16_f32);
+    GGML_METAL_DECL_KERNEL(mul_mv_f16_f32_1row);
+    GGML_METAL_DECL_KERNEL(mul_mv_f16_f32_l4);
+    GGML_METAL_DECL_KERNEL(mul_mv_q4_0_f32);
+    GGML_METAL_DECL_KERNEL(mul_mv_q4_1_f32);
+    GGML_METAL_DECL_KERNEL(mul_mv_q5_0_f32);
+    GGML_METAL_DECL_KERNEL(mul_mv_q5_1_f32);
+    GGML_METAL_DECL_KERNEL(mul_mv_q8_0_f32);
+    GGML_METAL_DECL_KERNEL(mul_mv_q2_K_f32);
+    GGML_METAL_DECL_KERNEL(mul_mv_q3_K_f32);
+    GGML_METAL_DECL_KERNEL(mul_mv_q4_K_f32);
+    GGML_METAL_DECL_KERNEL(mul_mv_q5_K_f32);
+    GGML_METAL_DECL_KERNEL(mul_mv_q6_K_f32);
     GGML_METAL_DECL_KERNEL(mul_mm_f32_f32);
     GGML_METAL_DECL_KERNEL(mul_mm_f16_f32);
     GGML_METAL_DECL_KERNEL(mul_mm_q4_0_f32);
     GGML_METAL_DECL_KERNEL(mul_mm_q4_1_f32);
+    GGML_METAL_DECL_KERNEL(mul_mm_q5_0_f32);
+    GGML_METAL_DECL_KERNEL(mul_mm_q5_1_f32);
     GGML_METAL_DECL_KERNEL(mul_mm_q8_0_f32);
     GGML_METAL_DECL_KERNEL(mul_mm_q2_K_f32);
     GGML_METAL_DECL_KERNEL(mul_mm_q3_K_f32);
     GGML_METAL_DECL_KERNEL(mul_mm_q4_K_f32);
     GGML_METAL_DECL_KERNEL(mul_mm_q5_K_f32);
     GGML_METAL_DECL_KERNEL(mul_mm_q6_K_f32);
-    GGML_METAL_DECL_KERNEL(rope);
+    GGML_METAL_DECL_KERNEL(rope_f32);
+    GGML_METAL_DECL_KERNEL(rope_f16);
     GGML_METAL_DECL_KERNEL(alibi_f32);
     GGML_METAL_DECL_KERNEL(cpy_f32_f16);
     GGML_METAL_DECL_KERNEL(cpy_f32_f32);
     GGML_METAL_DECL_KERNEL(cpy_f16_f16);
+    GGML_METAL_DECL_KERNEL(concat);
+    GGML_METAL_DECL_KERNEL(sqr);
 
 #undef GGML_METAL_DECL_KERNEL
 };
@@ -120,8 +134,37 @@ static NSString * const msl_library_source = @"see metal.metal";
 @implementation GGMLMetalClass
 @end
 
+ggml_log_callback ggml_metal_log_callback = NULL;
+void * ggml_metal_log_user_data = NULL;
+
+void ggml_metal_log_set_callback(ggml_log_callback log_callback, void * user_data) {
+    ggml_metal_log_callback  = log_callback;
+    ggml_metal_log_user_data = user_data;
+}
+
+static void ggml_metal_log(enum ggml_log_level level, const char* format, ...){
+    if (ggml_metal_log_callback != NULL) {
+        va_list args;
+        va_start(args, format);
+        char buffer[128];
+        int len = vsnprintf(buffer, 128, format, args);
+        if (len < 128) {
+            ggml_metal_log_callback(level, buffer, ggml_metal_log_user_data);
+        } else {
+            char* buffer2 = malloc(len+1);
+            vsnprintf(buffer2, len+1, format, args);
+            buffer2[len] = 0;
+            ggml_metal_log_callback(level, buffer2, ggml_metal_log_user_data);
+            free(buffer2);
+        }
+        va_end(args);
+    }
+}
+
+
+
 struct ggml_metal_context * ggml_metal_init(int n_cb) {
-    metal_printf("%s: allocating\n", __func__);
+    GGML_METAL_LOG_INFO("%s: allocating\n", __func__);
 
     id <MTLDevice> device;
     NSString * s;
@@ -131,14 +174,14 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
     NSArray * devices = MTLCopyAllDevices();
     for (device in devices) {
         s = [device name];
-        metal_printf("%s: found device: %s\n", __func__, [s UTF8String]);
+        GGML_METAL_LOG_INFO("%s: found device: %s\n", __func__, [s UTF8String]);
     }
 #endif
 
     // Pick and show default Metal device
     device = MTLCreateSystemDefaultDevice();
     s = [device name];
-    metal_printf("%s: picking default device: %s\n", __func__, [s UTF8String]);
+    GGML_METAL_LOG_INFO("%s: picking default device: %s\n", __func__, [s UTF8String]);
 
     // Configure context
     struct ggml_metal_context * ctx = malloc(sizeof(struct ggml_metal_context));
@@ -150,68 +193,69 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
 
     ctx->d_queue = dispatch_queue_create("ggml-metal", DISPATCH_QUEUE_CONCURRENT);
 
-#ifdef GGML_SWIFT
-    // load the default.metallib file
+    // load library
     {
-        NSError * error = nil;
-
-        NSBundle * bundle = [NSBundle bundleForClass:[GGMLMetalClass class]];
-        NSString * llamaBundlePath = [bundle pathForResource:@"llama_llama" ofType:@"bundle"];
-        NSBundle * llamaBundle = [NSBundle bundleWithPath:llamaBundlePath];
-        NSString * libPath = [llamaBundle pathForResource:@"default" ofType:@"metallib"];
-        NSURL * libURL = [NSURL fileURLWithPath:libPath];
-
-        // Load the metallib file into a Metal library
-        ctx->library = [ctx->device newLibraryWithURL:libURL error:&error];
-
-        if (error) {
-            metal_printf("%s: error: %s\n", __func__, [[error description] UTF8String]);
-            return NULL;
-        }
-    }
+        NSBundle * bundle = nil;
+#ifdef SWIFT_PACKAGE
+        bundle = SWIFTPM_MODULE_BUNDLE;
 #else
-    UNUSED(msl_library_source);
-
-    // read the source from "ggml-metal.metal" into a string and use newLibraryWithSource
-    {
+        bundle = [NSBundle bundleForClass:[GGMLMetalClass class]];
+#endif
         NSError * error = nil;
+        NSString * libPath = [bundle pathForResource:@"default" ofType:@"metallib"];
+        if (libPath != nil) {
+            NSURL * libURL = [NSURL fileURLWithPath:libPath];
+            GGML_METAL_LOG_INFO("%s: loading '%s'\n", __func__, [libPath UTF8String]);
+            ctx->library = [ctx->device newLibraryWithURL:libURL error:&error];
+        } else {
+            GGML_METAL_LOG_INFO("%s: default.metallib not found, loading from source\n", __func__);
+
+            NSString * sourcePath;
+            NSString * ggmlMetalPathResources = [[NSProcessInfo processInfo].environment objectForKey:@"GGML_METAL_PATH_RESOURCES"];
+            if (ggmlMetalPathResources) {
+                sourcePath = [ggmlMetalPathResources stringByAppendingPathComponent:@"ggml-metal.metal"];
+            } else {
+                sourcePath = [bundle pathForResource:@"ggml-metal" ofType:@"metal"];
+            }
+            if (sourcePath == nil) {
+                GGML_METAL_LOG_WARN("%s: error: could not use bundle path to find ggml-metal.metal, falling back to trying cwd\n", __func__);
+                sourcePath = @"ggml-metal.metal";
+            }
+            GGML_METAL_LOG_INFO("%s: loading '%s'\n", __func__, [sourcePath UTF8String]);
+            NSString * src = [NSString stringWithContentsOfFile:sourcePath encoding:NSUTF8StringEncoding error:&error];
+            if (error) {
+                GGML_METAL_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
+                return NULL;
+            }
 
-        //NSString * path = [[NSBundle mainBundle] pathForResource:@"../../examples/metal/metal" ofType:@"metal"];
-        NSBundle * bundle = [NSBundle bundleForClass:[GGMLMetalClass class]];
-        NSString * path   = [bundle pathForResource:@"ggml-metal" ofType:@"metal"];
-        metal_printf("%s: loading '%s'\n", __func__, [path UTF8String]);
-
-        NSString * src  = [NSString stringWithContentsOfFile:path encoding:NSUTF8StringEncoding error:&error];
-        if (error) {
-            metal_printf("%s: error: %s\n", __func__, [[error description] UTF8String]);
-            return NULL;
-        }
-
+            MTLCompileOptions* options = nil;
 #ifdef GGML_QKK_64
-        MTLCompileOptions* options = [MTLCompileOptions new];
-        options.preprocessorMacros = @{ @"QK_K" : @(64) };
-        ctx->library = [ctx->device newLibraryWithSource:src options:options error:&error];
-#else
-        ctx->library = [ctx->device newLibraryWithSource:src options:nil error:&error];
+            options = [MTLCompileOptions new];
+            options.preprocessorMacros = @{ @"QK_K" : @(64) };
 #endif
+            ctx->library = [ctx->device newLibraryWithSource:src options:options error:&error];
+        }
+
         if (error) {
-            metal_printf("%s: error: %s\n", __func__, [[error description] UTF8String]);
+            GGML_METAL_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
             return NULL;
         }
     }
-#endif
 
     // load kernels
     {
         NSError * error = nil;
+
+        /*
+        GGML_METAL_LOG_INFO("%s: loaded %-32s %16p | th_max = %4d | th_width = %4d\n", __func__, "kernel_"#name, (void *) ctx->pipeline_##name, \
+                (int) ctx->pipeline_##name.maxTotalThreadsPerThreadgroup, \
+                (int) ctx->pipeline_##name.threadExecutionWidth); \
+        */
 #define GGML_METAL_ADD_KERNEL(name) \
         ctx->function_##name = [ctx->library newFunctionWithName:@"kernel_"#name]; \
         ctx->pipeline_##name = [ctx->device newComputePipelineStateWithFunction:ctx->function_##name error:&error]; \
-        metal_printf("%s: loaded %-32s %16p | th_max = %4d | th_width = %4d\n", __func__, "kernel_"#name, (void *) ctx->pipeline_##name, \
-                (int) ctx->pipeline_##name.maxTotalThreadsPerThreadgroup, \
-                (int) ctx->pipeline_##name.threadExecutionWidth); \
         if (error) { \
-            metal_printf("%s: load pipeline error: %s\n", __func__, [[error description] UTF8String]); \
+            GGML_METAL_LOG_ERROR("%s: error: load pipeline error: %s\n", __func__, [[error description] UTF8String]); \
             return NULL; \
         }
 
@@ -220,6 +264,7 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(mul);
         GGML_METAL_ADD_KERNEL(mul_row);
         GGML_METAL_ADD_KERNEL(scale);
+        GGML_METAL_ADD_KERNEL(scale_4);
         GGML_METAL_ADD_KERNEL(silu);
         GGML_METAL_ADD_KERNEL(relu);
         GGML_METAL_ADD_KERNEL(gelu);
@@ -231,6 +276,8 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(get_rows_f16);
         GGML_METAL_ADD_KERNEL(get_rows_q4_0);
         GGML_METAL_ADD_KERNEL(get_rows_q4_1);
+        GGML_METAL_ADD_KERNEL(get_rows_q5_0);
+        GGML_METAL_ADD_KERNEL(get_rows_q5_1);
         GGML_METAL_ADD_KERNEL(get_rows_q8_0);
         GGML_METAL_ADD_KERNEL(get_rows_q2_K);
         GGML_METAL_ADD_KERNEL(get_rows_q3_K);
@@ -239,44 +286,66 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(get_rows_q6_K);
         GGML_METAL_ADD_KERNEL(rms_norm);
         GGML_METAL_ADD_KERNEL(norm);
-        GGML_METAL_ADD_KERNEL(mul_mat_f32_f32);
-        GGML_METAL_ADD_KERNEL(mul_mat_f16_f32);
-        GGML_METAL_ADD_KERNEL(mul_mat_f16_f32_1row);
-        GGML_METAL_ADD_KERNEL(mul_mat_f16_f32_l4);
-        GGML_METAL_ADD_KERNEL(mul_mat_q4_0_f32);
-        GGML_METAL_ADD_KERNEL(mul_mat_q4_1_f32);
-        GGML_METAL_ADD_KERNEL(mul_mat_q8_0_f32);
-        GGML_METAL_ADD_KERNEL(mul_mat_q2_K_f32);
-        GGML_METAL_ADD_KERNEL(mul_mat_q3_K_f32);
-        GGML_METAL_ADD_KERNEL(mul_mat_q4_K_f32);
-        GGML_METAL_ADD_KERNEL(mul_mat_q5_K_f32);
-        GGML_METAL_ADD_KERNEL(mul_mat_q6_K_f32);
-        GGML_METAL_ADD_KERNEL(mul_mm_f32_f32);
-        GGML_METAL_ADD_KERNEL(mul_mm_f16_f32);
-        GGML_METAL_ADD_KERNEL(mul_mm_q4_0_f32);
-        GGML_METAL_ADD_KERNEL(mul_mm_q8_0_f32);
-        GGML_METAL_ADD_KERNEL(mul_mm_q4_1_f32);
-        GGML_METAL_ADD_KERNEL(mul_mm_q2_K_f32);
-        GGML_METAL_ADD_KERNEL(mul_mm_q3_K_f32);
-        GGML_METAL_ADD_KERNEL(mul_mm_q4_K_f32);
-        GGML_METAL_ADD_KERNEL(mul_mm_q5_K_f32);
-        GGML_METAL_ADD_KERNEL(mul_mm_q6_K_f32);
-        GGML_METAL_ADD_KERNEL(rope);
+        GGML_METAL_ADD_KERNEL(mul_mv_f32_f32);
+        GGML_METAL_ADD_KERNEL(mul_mv_f16_f32);
+        GGML_METAL_ADD_KERNEL(mul_mv_f16_f32_1row);
+        GGML_METAL_ADD_KERNEL(mul_mv_f16_f32_l4);
+        GGML_METAL_ADD_KERNEL(mul_mv_q4_0_f32);
+        GGML_METAL_ADD_KERNEL(mul_mv_q4_1_f32);
+        GGML_METAL_ADD_KERNEL(mul_mv_q5_0_f32);
+        GGML_METAL_ADD_KERNEL(mul_mv_q5_1_f32);
+        GGML_METAL_ADD_KERNEL(mul_mv_q8_0_f32);
+        GGML_METAL_ADD_KERNEL(mul_mv_q2_K_f32);
+        GGML_METAL_ADD_KERNEL(mul_mv_q3_K_f32);
+        GGML_METAL_ADD_KERNEL(mul_mv_q4_K_f32);
+        GGML_METAL_ADD_KERNEL(mul_mv_q5_K_f32);
+        GGML_METAL_ADD_KERNEL(mul_mv_q6_K_f32);
+        if ([ctx->device supportsFamily:MTLGPUFamilyApple7]) {
+            GGML_METAL_ADD_KERNEL(mul_mm_f32_f32);
+            GGML_METAL_ADD_KERNEL(mul_mm_f16_f32);
+            GGML_METAL_ADD_KERNEL(mul_mm_q4_0_f32);
+            GGML_METAL_ADD_KERNEL(mul_mm_q4_1_f32);
+            GGML_METAL_ADD_KERNEL(mul_mm_q5_0_f32);
+            GGML_METAL_ADD_KERNEL(mul_mm_q5_1_f32);
+            GGML_METAL_ADD_KERNEL(mul_mm_q8_0_f32);
+            GGML_METAL_ADD_KERNEL(mul_mm_q2_K_f32);
+            GGML_METAL_ADD_KERNEL(mul_mm_q3_K_f32);
+            GGML_METAL_ADD_KERNEL(mul_mm_q4_K_f32);
+            GGML_METAL_ADD_KERNEL(mul_mm_q5_K_f32);
+            GGML_METAL_ADD_KERNEL(mul_mm_q6_K_f32);
+        }
+        GGML_METAL_ADD_KERNEL(rope_f32);
+        GGML_METAL_ADD_KERNEL(rope_f16);
         GGML_METAL_ADD_KERNEL(alibi_f32);
         GGML_METAL_ADD_KERNEL(cpy_f32_f16);
         GGML_METAL_ADD_KERNEL(cpy_f32_f32);
         GGML_METAL_ADD_KERNEL(cpy_f16_f16);
+        GGML_METAL_ADD_KERNEL(concat);
+        GGML_METAL_ADD_KERNEL(sqr);
 
 #undef GGML_METAL_ADD_KERNEL
     }
 
-    metal_printf("%s: hasUnifiedMemory              = %s\n",       __func__, ctx->device.hasUnifiedMemory ? "true" : "false");
 #if TARGET_OS_OSX
-    metal_printf("%s: recommendedMaxWorkingSetSize  = %8.2f MB\n", __func__, ctx->device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);
+    // print MTL GPU family:
+    GGML_METAL_LOG_INFO("%s: GPU name:   %s\n", __func__, [[ctx->device name] UTF8String]);
+
+    // determine max supported GPU family
+    // https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf
+    // https://developer.apple.com/metal/Metal-Feature-Set-Tables.pdf
+    for (int i = MTLGPUFamilyApple1 + 20; i >= MTLGPUFamilyApple1; --i) {
+        if ([ctx->device supportsFamily:i]) {
+            GGML_METAL_LOG_INFO("%s: GPU family: MTLGPUFamilyApple%d (%d)\n", __func__, i - MTLGPUFamilyApple1 + 1, i);
+            break;
+        }
+    }
+
+    GGML_METAL_LOG_INFO("%s: hasUnifiedMemory              = %s\n",       __func__, ctx->device.hasUnifiedMemory ? "true" : "false");
+    GGML_METAL_LOG_INFO("%s: recommendedMaxWorkingSetSize  = %8.2f MB\n", __func__, ctx->device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);
     if (ctx->device.maxTransferRate != 0) {
-        metal_printf("%s: maxTransferRate               = %8.2f MB/s\n", __func__, ctx->device.maxTransferRate / 1024.0 / 1024.0);
+        GGML_METAL_LOG_INFO("%s: maxTransferRate               = %8.2f MB/s\n", __func__, ctx->device.maxTransferRate / 1024.0 / 1024.0);
     } else {
-        metal_printf("%s: maxTransferRate               = built-in GPU\n", __func__);
+        GGML_METAL_LOG_INFO("%s: maxTransferRate               = built-in GPU\n", __func__);
     }
 #endif
 
@@ -284,7 +353,7 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
 }
 
 void ggml_metal_free(struct ggml_metal_context * ctx) {
-    metal_printf("%s: deallocating\n", __func__);
+    GGML_METAL_LOG_INFO("%s: deallocating\n", __func__);
 #define GGML_METAL_DEL_KERNEL(name) \
     [ctx->function_##name release]; \
     [ctx->pipeline_##name release];
@@ -294,6 +363,7 @@ void ggml_metal_free(struct ggml_metal_context * ctx) {
     GGML_METAL_DEL_KERNEL(mul);
     GGML_METAL_DEL_KERNEL(mul_row);
     GGML_METAL_DEL_KERNEL(scale);
+    GGML_METAL_DEL_KERNEL(scale_4);
     GGML_METAL_DEL_KERNEL(silu);
     GGML_METAL_DEL_KERNEL(relu);
     GGML_METAL_DEL_KERNEL(gelu);
@@ -305,6 +375,8 @@ void ggml_metal_free(struct ggml_metal_context * ctx) {
     GGML_METAL_DEL_KERNEL(get_rows_f16);
     GGML_METAL_DEL_KERNEL(get_rows_q4_0);
     GGML_METAL_DEL_KERNEL(get_rows_q4_1);
+    GGML_METAL_DEL_KERNEL(get_rows_q5_0);
+    GGML_METAL_DEL_KERNEL(get_rows_q5_1);
     GGML_METAL_DEL_KERNEL(get_rows_q8_0);
     GGML_METAL_DEL_KERNEL(get_rows_q2_K);
     GGML_METAL_DEL_KERNEL(get_rows_q3_K);
@@ -313,33 +385,42 @@ void ggml_metal_free(struct ggml_metal_context * ctx) {
     GGML_METAL_DEL_KERNEL(get_rows_q6_K);
     GGML_METAL_DEL_KERNEL(rms_norm);
     GGML_METAL_DEL_KERNEL(norm);
-    GGML_METAL_DEL_KERNEL(mul_mat_f32_f32);
-    GGML_METAL_DEL_KERNEL(mul_mat_f16_f32);
-    GGML_METAL_DEL_KERNEL(mul_mat_f16_f32_1row);
-    GGML_METAL_DEL_KERNEL(mul_mat_f16_f32_l4);
-    GGML_METAL_DEL_KERNEL(mul_mat_q4_0_f32);
-    GGML_METAL_DEL_KERNEL(mul_mat_q4_1_f32);
-    GGML_METAL_DEL_KERNEL(mul_mat_q8_0_f32);
-    GGML_METAL_DEL_KERNEL(mul_mat_q2_K_f32);
-    GGML_METAL_DEL_KERNEL(mul_mat_q3_K_f32);
-    GGML_METAL_DEL_KERNEL(mul_mat_q4_K_f32);
-    GGML_METAL_DEL_KERNEL(mul_mat_q5_K_f32);
-    GGML_METAL_DEL_KERNEL(mul_mat_q6_K_f32);
-    GGML_METAL_DEL_KERNEL(mul_mm_f32_f32);
-    GGML_METAL_DEL_KERNEL(mul_mm_f16_f32);
-    GGML_METAL_DEL_KERNEL(mul_mm_q4_0_f32);
-    GGML_METAL_DEL_KERNEL(mul_mm_q8_0_f32);
-    GGML_METAL_DEL_KERNEL(mul_mm_q4_1_f32);
-    GGML_METAL_DEL_KERNEL(mul_mm_q2_K_f32);
-    GGML_METAL_DEL_KERNEL(mul_mm_q3_K_f32);
-    GGML_METAL_DEL_KERNEL(mul_mm_q4_K_f32);
-    GGML_METAL_DEL_KERNEL(mul_mm_q5_K_f32);
-    GGML_METAL_DEL_KERNEL(mul_mm_q6_K_f32);
-    GGML_METAL_DEL_KERNEL(rope);
+    GGML_METAL_DEL_KERNEL(mul_mv_f32_f32);
+    GGML_METAL_DEL_KERNEL(mul_mv_f16_f32);
+    GGML_METAL_DEL_KERNEL(mul_mv_f16_f32_1row);
+    GGML_METAL_DEL_KERNEL(mul_mv_f16_f32_l4);
+    GGML_METAL_DEL_KERNEL(mul_mv_q4_0_f32);
+    GGML_METAL_DEL_KERNEL(mul_mv_q4_1_f32);
+    GGML_METAL_DEL_KERNEL(mul_mv_q5_0_f32);
+    GGML_METAL_DEL_KERNEL(mul_mv_q5_1_f32);
+    GGML_METAL_DEL_KERNEL(mul_mv_q8_0_f32);
+    GGML_METAL_DEL_KERNEL(mul_mv_q2_K_f32);
+    GGML_METAL_DEL_KERNEL(mul_mv_q3_K_f32);
+    GGML_METAL_DEL_KERNEL(mul_mv_q4_K_f32);
+    GGML_METAL_DEL_KERNEL(mul_mv_q5_K_f32);
+    GGML_METAL_DEL_KERNEL(mul_mv_q6_K_f32);
+    if ([ctx->device supportsFamily:MTLGPUFamilyApple7]) {
+        GGML_METAL_DEL_KERNEL(mul_mm_f32_f32);
+        GGML_METAL_DEL_KERNEL(mul_mm_f16_f32);
+        GGML_METAL_DEL_KERNEL(mul_mm_q4_0_f32);
+        GGML_METAL_DEL_KERNEL(mul_mm_q4_1_f32);
+        GGML_METAL_DEL_KERNEL(mul_mm_q5_0_f32);
+        GGML_METAL_DEL_KERNEL(mul_mm_q5_1_f32);
+        GGML_METAL_DEL_KERNEL(mul_mm_q8_0_f32);
+        GGML_METAL_DEL_KERNEL(mul_mm_q2_K_f32);
+        GGML_METAL_DEL_KERNEL(mul_mm_q3_K_f32);
+        GGML_METAL_DEL_KERNEL(mul_mm_q4_K_f32);
+        GGML_METAL_DEL_KERNEL(mul_mm_q5_K_f32);
+        GGML_METAL_DEL_KERNEL(mul_mm_q6_K_f32);
+    }
+    GGML_METAL_DEL_KERNEL(rope_f32);
+    GGML_METAL_DEL_KERNEL(rope_f16);
     GGML_METAL_DEL_KERNEL(alibi_f32);
     GGML_METAL_DEL_KERNEL(cpy_f32_f16);
     GGML_METAL_DEL_KERNEL(cpy_f32_f32);
     GGML_METAL_DEL_KERNEL(cpy_f16_f16);
+    GGML_METAL_DEL_KERNEL(concat);
+    GGML_METAL_DEL_KERNEL(sqr);
 
 #undef GGML_METAL_DEL_KERNEL
 
@@ -360,7 +441,7 @@ void * ggml_metal_host_malloc(size_t n) {
     void * data = NULL;
     const int result = posix_memalign((void **) &data, sysconf(_SC_PAGESIZE), n);
     if (result != 0) {
-        metal_printf("%s: error: posix_memalign failed\n", __func__);
+        GGML_METAL_LOG_ERROR("%s: error: posix_memalign failed\n", __func__);
         return NULL;
     }
 
@@ -388,7 +469,7 @@ int * ggml_metal_get_concur_list(struct ggml_metal_context * ctx) {
 // Metal buffer based on the host memory pointer
 //
 static id<MTLBuffer> ggml_metal_get_buffer(struct ggml_metal_context * ctx, struct ggml_tensor * t, size_t * offs) {
-    //metal_printf("%s: data tensor '%16s', offs_data = %8ld, offs_eval = %8ld, offs_cach = %8ld\n", __func__, t->name, offs_data, offs_eval, offs_cach);
+    //GGML_METAL_LOG_INFO("%s: data tensor '%16s', offs_data = %8ld, offs_eval = %8ld, offs_cach = %8ld\n", __func__, t->name, offs_data, offs_eval, offs_cach);
 
     const int64_t tsize = ggml_nbytes(t);
 
@@ -396,17 +477,17 @@ static id<MTLBuffer> ggml_metal_get_buffer(struct ggml_metal_context * ctx, stru
     for (int i = 0; i < ctx->n_buffers; ++i) {
         const int64_t ioffs = (int64_t) t->data - (int64_t) ctx->buffers[i].data;
 
-        //metal_printf("ioffs = %10ld, tsize = %10ld, sum = %10ld, ctx->buffers[%d].size = %10ld, name = %s\n", ioffs, tsize, ioffs + tsize, i, ctx->buffers[i].size, ctx->buffers[i].name);
+        //GGML_METAL_LOG_INFO("ioffs = %10ld, tsize = %10ld, sum = %10ld, ctx->buffers[%d].size = %10ld, name = %s\n", ioffs, tsize, ioffs + tsize, i, ctx->buffers[i].size, ctx->buffers[i].name);
         if (ioffs >= 0 && ioffs + tsize <= (int64_t) ctx->buffers[i].size) {
             *offs = (size_t) ioffs;
 
-            //metal_printf("%s: '%s' tensor '%16s', offs = %8ld\n", __func__, ctx->buffers[i].name, t->name, *offs);
+            //GGML_METAL_LOG_INFO("%s: '%s' tensor '%16s', offs = %8ld\n", __func__, ctx->buffers[i].name, t->name, *offs);
 
             return ctx->buffers[i].metal;
         }
     }
 
-    metal_printf("%s: error: buffer is nil\n", __func__);
+    GGML_METAL_LOG_ERROR("%s: error: buffer is nil\n", __func__);
 
     return nil;
 }
@@ -418,7 +499,7 @@ bool ggml_metal_add_buffer(
                          size_t   size,
                          size_t   max_size) {
     if (ctx->n_buffers >= GGML_METAL_MAX_BUFFERS) {
-        metal_printf("%s: too many buffers\n", __func__);
+        GGML_METAL_LOG_ERROR("%s: error: too many buffers\n", __func__);
         return false;
     }
 
@@ -428,7 +509,7 @@ bool ggml_metal_add_buffer(
             const int64_t ioffs = (int64_t) data - (int64_t) ctx->buffers[i].data;
 
             if (ioffs >= 0 && ioffs < (int64_t) ctx->buffers[i].size) {
-                metal_printf("%s: error: buffer '%s' overlaps with '%s'\n", __func__, name, ctx->buffers[i].name);
+                GGML_METAL_LOG_ERROR("%s: error: buffer '%s' overlaps with '%s'\n", __func__, name, ctx->buffers[i].name);
                 return false;
             }
         }
@@ -449,11 +530,11 @@ bool ggml_metal_add_buffer(
             ctx->buffers[ctx->n_buffers].metal = [ctx->device newBufferWithBytesNoCopy:data length:size_aligned options:MTLResourceStorageModeShared deallocator:nil];
 
             if (ctx->buffers[ctx->n_buffers].metal == nil) {
-                metal_printf("%s: failed to allocate '%-16s' buffer, size = %8.2f MB\n", __func__, name, size_aligned / 1024.0 / 1024.0);
+                GGML_METAL_LOG_ERROR("%s: error: failed to allocate '%-16s' buffer, size = %8.2f MB\n", __func__, name, size_aligned / 1024.0 / 1024.0);
                 return false;
             }
 
-            metal_printf("%s: allocated '%-16s' buffer, size = %8.2f MB", __func__, name, size_aligned / 1024.0 / 1024.0);
+            GGML_METAL_LOG_INFO("%s: allocated '%-16s' buffer, size = %8.2f MB", __func__, name, size_aligned / 1024.0 / 1024.0);
 
             ++ctx->n_buffers;
         } else {
@@ -473,13 +554,13 @@ bool ggml_metal_add_buffer(
                 ctx->buffers[ctx->n_buffers].metal = [ctx->device newBufferWithBytesNoCopy:(void *) ((uint8_t *) data + i) length:size_step_aligned options:MTLResourceStorageModeShared deallocator:nil];
 
                 if (ctx->buffers[ctx->n_buffers].metal == nil) {
-                    metal_printf("%s: failed to allocate '%-16s' buffer, size = %8.2f MB\n", __func__, name, size_step_aligned / 1024.0 / 1024.0);
+                    GGML_METAL_LOG_ERROR("%s: error: failed to allocate '%-16s' buffer, size = %8.2f MB\n", __func__, name, size_step_aligned / 1024.0 / 1024.0);
                     return false;
                 }
 
-                metal_printf("%s: allocated '%-16s' buffer, size = %8.2f MB, offs = %12ld", __func__, name, size_step_aligned / 1024.0 / 1024.0, i);
+                GGML_METAL_LOG_INFO("%s: allocated '%-16s' buffer, size = %8.2f MB, offs = %12ld", __func__, name, size_step_aligned / 1024.0 / 1024.0, i);
                 if (i + size_step < size) {
-                    metal_printf("\n");
+                    GGML_METAL_LOG_INFO("\n");
                 }
 
                 ++ctx->n_buffers;
@@ -487,17 +568,17 @@ bool ggml_metal_add_buffer(
         }
 
 #if TARGET_OS_OSX
-        metal_printf(", (%8.2f / %8.2f)",
+        GGML_METAL_LOG_INFO(", (%8.2f / %8.2f)",
                 ctx->device.currentAllocatedSize / 1024.0 / 1024.0,
                 ctx->device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);
 
         if (ctx->device.currentAllocatedSize > ctx->device.recommendedMaxWorkingSetSize) {
-            metal_printf(", warning: current allocated size is greater than the recommended max working set size\n");
+            GGML_METAL_LOG_WARN(", warning: current allocated size is greater than the recommended max working set size\n", __func__);
         } else {
-            metal_printf("\n");
+            GGML_METAL_LOG_INFO("\n");
         }
 #else
-        metal_printf(", (%8.2f)\n", ctx->device.currentAllocatedSize / 1024.0 / 1024.0);
+        GGML_METAL_LOG_INFO(", (%8.2f)\n", ctx->device.currentAllocatedSize / 1024.0 / 1024.0);
 #endif
     }
 
@@ -610,7 +691,7 @@ void ggml_metal_graph_find_concurrency(
     }
 
     if (ctx->concur_list_len > GGML_MAX_CONCUR) {
-        metal_printf("%s: too many elements for metal ctx->concur_list!\n", __func__);
+        GGML_METAL_LOG_WARN("%s: too many elements for metal ctx->concur_list!\n", __func__);
     }
 }
 
@@ -664,12 +745,26 @@ void ggml_metal_graph_compute(
                     continue;
                 }
 
-                //metal_printf("%s: encoding node %3d, op = %8s\n", __func__, i, ggml_op_name(gf->nodes[i]->op));
+                //GGML_METAL_LOG_INFO("%s: encoding node %3d, op = %8s\n", __func__, i, ggml_op_name(gf->nodes[i]->op));
 
                 struct ggml_tensor * src0 = gf->nodes[i]->src[0];
                 struct ggml_tensor * src1 = gf->nodes[i]->src[1];
                 struct ggml_tensor * dst  = gf->nodes[i];
 
+                switch (dst->op) {
+                    case GGML_OP_NONE:
+                    case GGML_OP_RESHAPE:
+                    case GGML_OP_VIEW:
+                    case GGML_OP_TRANSPOSE:
+                    case GGML_OP_PERMUTE:
+                        {
+                            // noop -> next node
+                        } continue;
+                    default:
+                        {
+                        } break;
+                }
+
                 const int64_t  ne00 = src0 ? src0->ne[0] : 0;
                 const int64_t  ne01 = src0 ? src0->ne[1] : 0;
                 const int64_t  ne02 = src0 ? src0->ne[2] : 0;
@@ -708,53 +803,117 @@ void ggml_metal_graph_compute(
                 id<MTLBuffer> id_src1 = src1 ? ggml_metal_get_buffer(ctx, src1, &offs_src1) : nil;
                 id<MTLBuffer> id_dst  = dst  ? ggml_metal_get_buffer(ctx, dst,  &offs_dst)  : nil;
 
-                //metal_printf("%s: op - %s\n", __func__, ggml_op_name(dst->op));
+                //GGML_METAL_LOG_INFO("%s: op - %s\n", __func__, ggml_op_name(dst->op));
                 //if (src0) {
-                //    metal_printf("%s: src0 - %4s [%5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(src0t), ne00, ne01, ne02,
+                //    GGML_METAL_LOG_INFO("%s: src0 - %4s [%5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(src0t), ne00, ne01, ne02,
                 //            ggml_is_contiguous(src0), src0->name);
                 //}
                 //if (src1) {
-                //    metal_printf("%s: src1 - %4s [%5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(src1t), ne10, ne11, ne12,
+                //    GGML_METAL_LOG_INFO("%s: src1 - %4s [%5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(src1t), ne10, ne11, ne12,
                 //            ggml_is_contiguous(src1), src1->name);
                 //}
                 //if (dst) {
-                //    metal_printf("%s: dst  - %4s [%5lld, %5lld, %5lld], 1, %s\n",  __func__, ggml_type_name(dstt),  ne0,  ne1,  ne2,
+                //    GGML_METAL_LOG_INFO("%s: dst  - %4s [%5lld, %5lld, %5lld], 1, %s\n",  __func__, ggml_type_name(dstt),  ne0,  ne1,  ne2,
                 //            dst->name);
                 //}
 
                 switch (dst->op) {
-                    case GGML_OP_NONE:
-                    case GGML_OP_RESHAPE:
-                    case GGML_OP_VIEW:
-                    case GGML_OP_TRANSPOSE:
-                    case GGML_OP_PERMUTE:
+                    case GGML_OP_CONCAT:
                         {
-                            // noop
+                            const int64_t nb = ne00;
+
+                            [encoder setComputePipelineState:ctx->pipeline_concat];
+                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                            [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
+                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
+                            [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
+                            [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4];
+                            [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:5];
+                            [encoder setBytes:&ne03 length:sizeof(ne03) atIndex:6];
+                            [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:7];
+                            [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:8];
+                            [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:9];
+                            [encoder setBytes:&nb03 length:sizeof(nb03) atIndex:10];
+                            [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:11];
+                            [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:12];
+                            [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:13];
+                            [encoder setBytes:&ne13 length:sizeof(ne13) atIndex:14];
+                            [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:15];
+                            [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:16];
+                            [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:17];
+                            [encoder setBytes:&nb13 length:sizeof(nb13) atIndex:18];
+                            [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:19];
+                            [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:20];
+                            [encoder setBytes:&ne2  length:sizeof(ne2)  atIndex:21];
+                            [encoder setBytes:&ne3  length:sizeof(ne3)  atIndex:22];
+                            [encoder setBytes:&nb0  length:sizeof(nb0)  atIndex:23];
+                            [encoder setBytes:&nb1  length:sizeof(nb1)  atIndex:24];
+                            [encoder setBytes:&nb2  length:sizeof(nb2)  atIndex:25];
+                            [encoder setBytes:&nb3  length:sizeof(nb3)  atIndex:26];
+                            [encoder setBytes:&nb   length:sizeof(nb)   atIndex:27];
+
+                            const int nth = MIN(1024, ne0);
+
+                            [encoder dispatchThreadgroups:MTLSizeMake(ne1, ne2, ne3) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
                         } break;
                     case GGML_OP_ADD:
                         {
                             GGML_ASSERT(ggml_is_contiguous(src0));
                             GGML_ASSERT(ggml_is_contiguous(src1));
 
-                            // utilize float4
-                            GGML_ASSERT(ne00 % 4 == 0);
-                            const int64_t nb = ne00/4;
+                            bool bcast_row = false;
 
-                            if (ggml_nelements(src1) == ne10) {
+                            int64_t nb = ne00;
+
+                            if (ggml_nelements(src1) == ne10 && ne00 % 4 == 0) {
                                 // src1 is a row
                                 GGML_ASSERT(ne11 == 1);
+
+                                nb = ne00 / 4;
                                 [encoder setComputePipelineState:ctx->pipeline_add_row];
+
+                                bcast_row = true;
                             } else {
                                 [encoder setComputePipelineState:ctx->pipeline_add];
                             }
                             [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                             [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
                             [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
-                            [encoder setBytes:&nb     length:sizeof(nb) atIndex:3];
-
-                            const int64_t n = ggml_nelements(dst)/4;
+                            [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
+                            [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4];
+                            [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:5];
+                            [encoder setBytes:&ne03 length:sizeof(ne03) atIndex:6];
+                            [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:7];
+                            [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:8];
+                            [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:9];
+                            [encoder setBytes:&nb03 length:sizeof(nb03) atIndex:10];
+                            [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:11];
+                            [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:12];
+                            [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:13];
+                            [encoder setBytes:&ne13 length:sizeof(ne13) atIndex:14];
+                            [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:15];
+                            [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:16];
+                            [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:17];
+                            [encoder setBytes:&nb13 length:sizeof(nb13) atIndex:18];
+                            [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:19];
+                            [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:20];
+                            [encoder setBytes:&ne2  length:sizeof(ne2)  atIndex:21];
+                            [encoder setBytes:&ne3  length:sizeof(ne3)  atIndex:22];
+                            [encoder setBytes:&nb0  length:sizeof(nb0)  atIndex:23];
+                            [encoder setBytes:&nb1  length:sizeof(nb1)  atIndex:24];
+                            [encoder setBytes:&nb2  length:sizeof(nb2)  atIndex:25];
+                            [encoder setBytes:&nb3  length:sizeof(nb3)  atIndex:26];
+                            [encoder setBytes:&nb   length:sizeof(nb)   atIndex:27];
+
+                            if (bcast_row) {
+                                const int64_t n = ggml_nelements(dst)/4;
+
+                                [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                            } else {
+                                const int nth = MIN(1024, ne0);
 
-                            [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                                [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+                            }
                         } break;
                     case GGML_OP_MUL:
                         {
@@ -787,13 +946,19 @@ void ggml_metal_graph_compute(
 
                             const float scale = *(const float *) src1->data;
 
-                            [encoder setComputePipelineState:ctx->pipeline_scale];
+                            int64_t n = ggml_nelements(dst);
+
+                            if (n % 4 == 0) {
+                                n /= 4;
+                                [encoder setComputePipelineState:ctx->pipeline_scale_4];
+                            } else {
+                                [encoder setComputePipelineState:ctx->pipeline_scale];
+                            }
+
                             [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                             [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
                             [encoder setBytes:&scale length:sizeof(scale) atIndex:2];
 
-                            const int64_t n = ggml_nelements(dst)/4;
-
                             [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                         } break;
                     case GGML_OP_UNARY:
@@ -804,9 +969,10 @@ void ggml_metal_graph_compute(
                                     [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                                     [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
 
-                                    const int64_t n = ggml_nelements(dst)/4;
+                                    const int64_t n = ggml_nelements(dst);
+                                    GGML_ASSERT(n % 4 == 0);
 
-                                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                                    [encoder dispatchThreadgroups:MTLSizeMake(n/4, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                                 } break;
                             case GGML_UNARY_OP_RELU:
                                 {
@@ -824,23 +990,39 @@ void ggml_metal_graph_compute(
                                     [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                                     [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
 
-                                    const int64_t n = ggml_nelements(dst)/4;
+                                    const int64_t n = ggml_nelements(dst);
+                                    GGML_ASSERT(n % 4 == 0);
 
-                                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                                    [encoder dispatchThreadgroups:MTLSizeMake(n/4, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                                 } break;
                             default:
                                 {
-                                    metal_printf("%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
+                                    GGML_METAL_LOG_WARN("%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
                                     GGML_ASSERT(false);
                                 }
                         } break;
+                    case GGML_OP_SQR:
+                        {
+                            GGML_ASSERT(ggml_is_contiguous(src0));
+
+                            [encoder setComputePipelineState:ctx->pipeline_sqr];
+                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                            [encoder setBuffer:id_dst  offset:offs_dst atIndex:1];
+
+                            const int64_t n = ggml_nelements(dst);
+                            [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                        } break;
                     case GGML_OP_SOFT_MAX:
                         {
-                            const int nth = 32;
+                            int nth = 32; // SIMD width
 
                             if (ne00%4 == 0) {
                                 [encoder setComputePipelineState:ctx->pipeline_soft_max_4];
                             } else {
+                                do {
+                                    nth *= 2;
+                                } while (nth <= ne00 && nth <= 1024);
+                                nth /= 2;
                                 [encoder setComputePipelineState:ctx->pipeline_soft_max];
                             }
                             [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
@@ -848,8 +1030,9 @@ void ggml_metal_graph_compute(
                             [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2];
                             [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3];
                             [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4];
+                            [encoder setThreadgroupMemoryLength:nth/32*sizeof(float) atIndex:0];
 
-                            [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+                            [encoder dispatchThreadgroups:MTLSizeMake(ne01*ne02*ne03, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
                         } break;
                     case GGML_OP_DIAG_MASK_INF:
                         {
@@ -875,26 +1058,53 @@ void ggml_metal_graph_compute(
                         } break;
                     case GGML_OP_MUL_MAT:
                         {
-                            // TODO: needs to be updated after PR: https://github.com/ggerganov/ggml/pull/224
-
                             GGML_ASSERT(ne00 == ne10);
-                            // GGML_ASSERT(ne02 == ne12); // Should be checked on individual data types until broadcast is implemented everywhere
-                            uint gqa = ne12/ne02;
                             GGML_ASSERT(ne03 == ne13);
 
+                            const uint gqa = ne12/ne02;
+
+                            // find the break-even point where the matrix-matrix kernel becomes more efficient compared
+                            // to the matrix-vector kernel
+                            int ne11_mm_min = 1;
+
+#if 0
+                            // the numbers below are measured on M2 Ultra for 7B and 13B models
+                            // these numbers do not translate to other devices or model sizes
+                            // TODO: need to find a better approach
+                            if ([ctx->device.name isEqualToString:@"Apple M2 Ultra"]) {
+                                switch (src0t) {
+                                    case GGML_TYPE_F16:  ne11_mm_min = 2;  break;
+                                    case GGML_TYPE_Q8_0: ne11_mm_min = 7;  break;
+                                    case GGML_TYPE_Q2_K: ne11_mm_min = 15; break;
+                                    case GGML_TYPE_Q3_K: ne11_mm_min = 7;  break;
+                                    case GGML_TYPE_Q4_0:
+                                    case GGML_TYPE_Q4_1: ne11_mm_min = 15; break;
+                                    case GGML_TYPE_Q4_K: ne11_mm_min = 11; break;
+                                    case GGML_TYPE_Q5_0:                          // not tested yet
+                                    case GGML_TYPE_Q5_1: ne11_mm_min = 13; break; // not tested yet
+                                    case GGML_TYPE_Q5_K: ne11_mm_min = 7;  break;
+                                    case GGML_TYPE_Q6_K: ne11_mm_min = 7;  break;
+                                    default:             ne11_mm_min = 1;  break;
+                                }
+                            }
+#endif
+
                             // for now the matrix-matrix multiplication kernel only works on A14+/M1+ SoCs
                             // AMD GPU and older A-chips will reuse matrix-vector multiplication kernel
-                            if (!ggml_is_transposed(src0) &&
+                            if ([ctx->device supportsFamily:MTLGPUFamilyApple7] &&
+                                !ggml_is_transposed(src0) &&
                                 !ggml_is_transposed(src1) &&
                                 src1t == GGML_TYPE_F32 &&
-                                [ctx->device supportsFamily:MTLGPUFamilyApple7] &&
-                                ne00%32 == 0 &&
-                                ne11 > 1) {
+                                ne00 % 32 == 0 && ne00 >= 64 &&
+                                ne11 > ne11_mm_min) {
+                                //printf("matrix: ne00 = %6d, ne01 = %6d, ne02 = %6d, ne11 = %6d, ne12 = %6d\n", ne00, ne01, ne02, ne11, ne12);
                                 switch (src0->type) {
                                     case GGML_TYPE_F32:  [encoder setComputePipelineState:ctx->pipeline_mul_mm_f32_f32];  break;
                                     case GGML_TYPE_F16:  [encoder setComputePipelineState:ctx->pipeline_mul_mm_f16_f32];  break;
                                     case GGML_TYPE_Q4_0: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_0_f32]; break;
                                     case GGML_TYPE_Q4_1: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_1_f32]; break;
+                                    case GGML_TYPE_Q5_0: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q5_0_f32]; break;
+                                    case GGML_TYPE_Q5_1: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q5_1_f32]; break;
                                     case GGML_TYPE_Q8_0: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q8_0_f32]; break;
                                     case GGML_TYPE_Q2_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q2_K_f32]; break;
                                     case GGML_TYPE_Q3_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q3_K_f32]; break;
@@ -918,17 +1128,18 @@ void ggml_metal_graph_compute(
                                 [encoder setBytes:&ne1     length:sizeof(ne1)  atIndex:12];
                                 [encoder setBytes:&gqa     length:sizeof(gqa)  atIndex:13];
                                 [encoder setThreadgroupMemoryLength:8192 atIndex:0];
-                                [encoder dispatchThreadgroups:MTLSizeMake( (ne11+31)/32, (ne01+63) / 64, ne12) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)];
+                                [encoder dispatchThreadgroups:MTLSizeMake( (ne11 + 31)/32, (ne01 + 63)/64, ne12) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)];
                             } else {
                                 int nth0 = 32;
                                 int nth1 = 1;
                                 int nrows = 1;
+                                //printf("vector: ne00 = %6d, ne01 = %6d, ne02 = %6d, ne11 = %6d, ne12 = %6d\n", ne00, ne01, ne02, ne11, ne12);
 
                                 // use custom matrix x vector kernel
                                 switch (src0t) {
                                     case GGML_TYPE_F32:
                                         {
-                                            [encoder setComputePipelineState:ctx->pipeline_mul_mat_f32_f32];
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_f32_f32];
                                             nrows = 4;
                                         } break;
                                     case GGML_TYPE_F16:
@@ -936,12 +1147,12 @@ void ggml_metal_graph_compute(
                                             nth0 = 32;
                                             nth1 = 1;
                                             if (ne11 * ne12 < 4) {
-                                                [encoder setComputePipelineState:ctx->pipeline_mul_mat_f16_f32_1row];
+                                                [encoder setComputePipelineState:ctx->pipeline_mul_mv_f16_f32_1row];
                                             } else if (ne00 >= 128 && ne01 >= 8 && ne00%4 == 0) {
-                                                [encoder setComputePipelineState:ctx->pipeline_mul_mat_f16_f32_l4];
+                                                [encoder setComputePipelineState:ctx->pipeline_mul_mv_f16_f32_l4];
                                                 nrows = ne11;
                                             } else {
-                                                [encoder setComputePipelineState:ctx->pipeline_mul_mat_f16_f32];
+                                                [encoder setComputePipelineState:ctx->pipeline_mul_mv_f16_f32];
                                                 nrows = 4;
                                             }
                                         } break;
@@ -952,7 +1163,7 @@ void ggml_metal_graph_compute(
 
                                             nth0 = 8;
                                             nth1 = 8;
-                                            [encoder setComputePipelineState:ctx->pipeline_mul_mat_q4_0_f32];
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_q4_0_f32];
                                         } break;
                                     case GGML_TYPE_Q4_1:
                                         {
@@ -961,7 +1172,25 @@ void ggml_metal_graph_compute(
 
                                             nth0 = 8;
                                             nth1 = 8;
-                                            [encoder setComputePipelineState:ctx->pipeline_mul_mat_q4_1_f32];
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_q4_1_f32];
+                                        } break;
+                                    case GGML_TYPE_Q5_0:
+                                        {
+                                            GGML_ASSERT(ne02 == 1);
+                                            GGML_ASSERT(ne12 == 1);
+
+                                            nth0 = 8;
+                                            nth1 = 8;
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_q5_0_f32];
+                                        } break;
+                                    case GGML_TYPE_Q5_1:
+                                        {
+                                            GGML_ASSERT(ne02 == 1);
+                                            GGML_ASSERT(ne12 == 1);
+
+                                            nth0 = 8;
+                                            nth1 = 8;
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_q5_1_f32];
                                         } break;
                                     case GGML_TYPE_Q8_0:
                                         {
@@ -970,7 +1199,7 @@ void ggml_metal_graph_compute(
 
                                             nth0 = 8;
                                             nth1 = 8;
-                                            [encoder setComputePipelineState:ctx->pipeline_mul_mat_q8_0_f32];
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_q8_0_f32];
                                         } break;
                                     case GGML_TYPE_Q2_K:
                                         {
@@ -979,7 +1208,7 @@ void ggml_metal_graph_compute(
 
                                             nth0 = 2;
                                             nth1 = 32;
-                                            [encoder setComputePipelineState:ctx->pipeline_mul_mat_q2_K_f32];
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_q2_K_f32];
                                         } break;
                                     case GGML_TYPE_Q3_K:
                                         {
@@ -988,7 +1217,7 @@ void ggml_metal_graph_compute(
 
                                             nth0 = 2;
                                             nth1 = 32;
-                                            [encoder setComputePipelineState:ctx->pipeline_mul_mat_q3_K_f32];
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_q3_K_f32];
                                         } break;
                                     case GGML_TYPE_Q4_K:
                                         {
@@ -997,7 +1226,7 @@ void ggml_metal_graph_compute(
 
                                             nth0 = 4; //1;
                                             nth1 = 8; //32;
-                                            [encoder setComputePipelineState:ctx->pipeline_mul_mat_q4_K_f32];
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_q4_K_f32];
                                         } break;
                                     case GGML_TYPE_Q5_K:
                                         {
@@ -1006,7 +1235,7 @@ void ggml_metal_graph_compute(
 
                                             nth0 = 2;
                                             nth1 = 32;
-                                            [encoder setComputePipelineState:ctx->pipeline_mul_mat_q5_K_f32];
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_q5_K_f32];
                                         } break;
                                     case GGML_TYPE_Q6_K:
                                         {
@@ -1015,11 +1244,11 @@ void ggml_metal_graph_compute(
 
                                             nth0 = 2;
                                             nth1 = 32;
-                                            [encoder setComputePipelineState:ctx->pipeline_mul_mat_q6_K_f32];
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_q6_K_f32];
                                         } break;
                                     default:
                                         {
-                                            metal_printf("Asserting on type %d\n",(int)src0t);
+                                            GGML_METAL_LOG_ERROR("Asserting on type %d\n", (int)src0t);
                                             GGML_ASSERT(false && "not implemented");
                                         }
                                 };
@@ -1043,8 +1272,9 @@ void ggml_metal_graph_compute(
                                 [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:16];
                                 [encoder setBytes:&gqa  length:sizeof(gqa)  atIndex:17];
 
-                                if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1 || src0t == GGML_TYPE_Q8_0 ||
-                                    src0t == GGML_TYPE_Q2_K) {// || src0t == GGML_TYPE_Q4_K) {
+                                if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1 ||
+                                    src0t == GGML_TYPE_Q5_0 || src0t == GGML_TYPE_Q5_1 || src0t == GGML_TYPE_Q8_0 ||
+                                    src0t == GGML_TYPE_Q2_K) { // || src0t == GGML_TYPE_Q4_K) {
                                     [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                                 }
                                 else if (src0t == GGML_TYPE_Q4_K) {
@@ -1075,6 +1305,8 @@ void ggml_metal_graph_compute(
                                 case GGML_TYPE_F16:  [encoder setComputePipelineState:ctx->pipeline_get_rows_f16];  break;
                                 case GGML_TYPE_Q4_0: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_0]; break;
                                 case GGML_TYPE_Q4_1: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_1]; break;
+                                case GGML_TYPE_Q5_0: [encoder setComputePipelineState:ctx->pipeline_get_rows_q5_0]; break;
+                                case GGML_TYPE_Q5_1: [encoder setComputePipelineState:ctx->pipeline_get_rows_q5_1]; break;
                                 case GGML_TYPE_Q8_0: [encoder setComputePipelineState:ctx->pipeline_get_rows_q8_0]; break;
                                 case GGML_TYPE_Q2_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q2_K]; break;
                                 case GGML_TYPE_Q3_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q3_K]; break;
@@ -1097,10 +1329,12 @@ void ggml_metal_graph_compute(
                         } break;
                     case GGML_OP_RMS_NORM:
                         {
+                            GGML_ASSERT(ne00 % 4 == 0);
+
                             float eps;
                             memcpy(&eps, dst->op_params, sizeof(float));
 
-                            const int nth = 512;
+                            const int nth = MIN(512, ne00);
 
                             [encoder setComputePipelineState:ctx->pipeline_rms_norm];
                             [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
@@ -1119,7 +1353,7 @@ void ggml_metal_graph_compute(
                             float eps;
                             memcpy(&eps, dst->op_params, sizeof(float));
 
-                            const int nth = 256;
+                            const int nth = MIN(256, ne00);
 
                             [encoder setComputePipelineState:ctx->pipeline_norm];
                             [encoder setBuffer:id_src0 offset:offs_src0        atIndex:0];
@@ -1137,17 +1371,16 @@ void ggml_metal_graph_compute(
                         {
                             GGML_ASSERT((src0t == GGML_TYPE_F32));
 
-                            const int n_past = ((int32_t *) dst->op_params)[0]; UNUSED(n_past);
+                            const int nth = MIN(1024, ne00);
+
+                            //const int n_past = ((int32_t *) dst->op_params)[0];
                             const int n_head = ((int32_t *) dst->op_params)[1];
                             float max_bias;
                             memcpy(&max_bias, (int32_t *) dst->op_params + 2, sizeof(float));
 
-                            if (__builtin_popcount(n_head) != 1) {
-                                GGML_ASSERT(false && "only power-of-two n_head implemented");
-                            }
-
                             const int n_heads_log2_floor = 1 << (int) floor(log2(n_head));
                             const float m0 = powf(2.0f, -(max_bias) / n_heads_log2_floor);
+                            const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_heads_log2_floor);
 
                             [encoder setComputePipelineState:ctx->pipeline_alibi_f32];
                             [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
@@ -1168,55 +1401,74 @@ void ggml_metal_graph_compute(
                             [encoder setBytes:&nb1  length:sizeof(uint64_t) atIndex:15];
                             [encoder setBytes:&nb2  length:sizeof(uint64_t) atIndex:16];
                             [encoder setBytes:&nb3  length:sizeof(uint64_t) atIndex:17];
-                            [encoder setBytes:&m0  length:sizeof(    float) atIndex:18];
-
-                            const int nth = 32;
+                            [encoder setBytes:&m0   length:sizeof(   float) atIndex:18];
+                            [encoder setBytes:&m1   length:sizeof(   float) atIndex:19];
+                            [encoder setBytes:&n_heads_log2_floor   length:sizeof(int) atIndex:20];
 
                             [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
                         } break;
                     case GGML_OP_ROPE:
                         {
-                            const int n_past = ((int32_t *) dst->op_params)[0];
-                            const int n_dims = ((int32_t *) dst->op_params)[1];
-                            const int mode   = ((int32_t *) dst->op_params)[2];
+                            GGML_ASSERT(ne10 == ne02);
 
-                            float freq_base;
-                            float freq_scale;
-                            memcpy(&freq_base,  (int32_t *) dst->op_params + 4, sizeof(float));
-                            memcpy(&freq_scale, (int32_t *) dst->op_params + 5, sizeof(float));
+                            const int nth = MIN(1024, ne00);
 
-                            [encoder setComputePipelineState:ctx->pipeline_rope];
-                            [encoder setBuffer:id_src0 offset:offs_src0        atIndex:0];
-                            [encoder setBuffer:id_dst  offset:offs_dst         atIndex:1];
-                            [encoder setBytes:&ne00    length:sizeof( int64_t) atIndex:2];
-                            [encoder setBytes:&ne01    length:sizeof( int64_t) atIndex:3];
-                            [encoder setBytes:&ne02    length:sizeof( int64_t) atIndex:4];
-                            [encoder setBytes:&ne03    length:sizeof( int64_t) atIndex:5];
-                            [encoder setBytes:&nb00    length:sizeof(uint64_t) atIndex:6];
-                            [encoder setBytes:&nb01    length:sizeof(uint64_t) atIndex:7];
-                            [encoder setBytes:&nb02    length:sizeof(uint64_t) atIndex:8];
-                            [encoder setBytes:&nb03    length:sizeof(uint64_t) atIndex:9];
-                            [encoder setBytes:&ne0     length:sizeof( int64_t) atIndex:10];
-                            [encoder setBytes:&ne1     length:sizeof( int64_t) atIndex:11];
-                            [encoder setBytes:&ne2     length:sizeof( int64_t) atIndex:12];
-                            [encoder setBytes:&ne3     length:sizeof( int64_t) atIndex:13];
-                            [encoder setBytes:&nb0     length:sizeof(uint64_t) atIndex:14];
-                            [encoder setBytes:&nb1     length:sizeof(uint64_t) atIndex:15];
-                            [encoder setBytes:&nb2     length:sizeof(uint64_t) atIndex:16];
-                            [encoder setBytes:&nb3     length:sizeof(uint64_t) atIndex:17];
-                            [encoder setBytes:&n_past  length:sizeof(     int) atIndex:18];
-                            [encoder setBytes:&n_dims  length:sizeof(     int) atIndex:19];
-                            [encoder setBytes:&mode    length:sizeof(     int) atIndex:20];
-                            [encoder setBytes:&freq_base  length:sizeof(float) atIndex:21];
-                            [encoder setBytes:&freq_scale length:sizeof(float) atIndex:22];
+                            const int n_past     = ((int32_t *) dst->op_params)[0];
+                            const int n_dims     = ((int32_t *) dst->op_params)[1];
+                            const int mode       = ((int32_t *) dst->op_params)[2];
+                            const int n_orig_ctx = ((int32_t *) dst->op_params)[3];
+
+                            float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
+                            memcpy(&freq_base,   (int32_t *) dst->op_params +  5, sizeof(float));
+                            memcpy(&freq_scale,  (int32_t *) dst->op_params +  6, sizeof(float));
+                            memcpy(&ext_factor,  (int32_t *) dst->op_params +  7, sizeof(float));
+                            memcpy(&attn_factor, (int32_t *) dst->op_params +  8, sizeof(float));
+                            memcpy(&beta_fast,   (int32_t *) dst->op_params +  9, sizeof(float));
+                            memcpy(&beta_slow,   (int32_t *) dst->op_params + 10, sizeof(float));
+
+                            switch (src0->type) {
+                                case GGML_TYPE_F32: [encoder setComputePipelineState:ctx->pipeline_rope_f32]; break;
+                                case GGML_TYPE_F16: [encoder setComputePipelineState:ctx->pipeline_rope_f16]; break;
+                                default: GGML_ASSERT(false);
+                            };
+
+                            [encoder setBuffer:id_src0     offset:offs_src0        atIndex:0];
+                            [encoder setBuffer:id_src1     offset:offs_src1        atIndex:1];
+                            [encoder setBuffer:id_dst      offset:offs_dst         atIndex:2];
+                            [encoder setBytes:&ne00        length:sizeof( int64_t) atIndex:3];
+                            [encoder setBytes:&ne01        length:sizeof( int64_t) atIndex:4];
+                            [encoder setBytes:&ne02        length:sizeof( int64_t) atIndex:5];
+                            [encoder setBytes:&ne03        length:sizeof( int64_t) atIndex:6];
+                            [encoder setBytes:&nb00        length:sizeof(uint64_t) atIndex:7];
+                            [encoder setBytes:&nb01        length:sizeof(uint64_t) atIndex:8];
+                            [encoder setBytes:&nb02        length:sizeof(uint64_t) atIndex:9];
+                            [encoder setBytes:&nb03        length:sizeof(uint64_t) atIndex:10];
+                            [encoder setBytes:&ne0         length:sizeof( int64_t) atIndex:11];
+                            [encoder setBytes:&ne1         length:sizeof( int64_t) atIndex:12];
+                            [encoder setBytes:&ne2         length:sizeof( int64_t) atIndex:13];
+                            [encoder setBytes:&ne3         length:sizeof( int64_t) atIndex:14];
+                            [encoder setBytes:&nb0         length:sizeof(uint64_t) atIndex:15];
+                            [encoder setBytes:&nb1         length:sizeof(uint64_t) atIndex:16];
+                            [encoder setBytes:&nb2         length:sizeof(uint64_t) atIndex:17];
+                            [encoder setBytes:&nb3         length:sizeof(uint64_t) atIndex:18];
+                            [encoder setBytes:&n_past      length:sizeof(     int) atIndex:19];
+                            [encoder setBytes:&n_dims      length:sizeof(     int) atIndex:20];
+                            [encoder setBytes:&mode        length:sizeof(     int) atIndex:21];
+                            [encoder setBytes:&n_orig_ctx  length:sizeof(     int) atIndex:22];
+                            [encoder setBytes:&freq_base   length:sizeof(   float) atIndex:23];
+                            [encoder setBytes:&freq_scale  length:sizeof(   float) atIndex:24];
+                            [encoder setBytes:&ext_factor  length:sizeof(   float) atIndex:25];
+                            [encoder setBytes:&attn_factor length:sizeof(   float) atIndex:26];
+                            [encoder setBytes:&beta_fast   length:sizeof(   float) atIndex:27];
+                            [encoder setBytes:&beta_slow   length:sizeof(   float) atIndex:28];
 
-                            [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(32, 1, 1)];
+                            [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
                         } break;
                     case GGML_OP_DUP:
                     case GGML_OP_CPY:
                     case GGML_OP_CONT:
                         {
-                            const int nth = 32;
+                            const int nth = MIN(1024, ne00);
 
                             switch (src0t) {
                                 case GGML_TYPE_F32:
@@ -1261,7 +1513,7 @@ void ggml_metal_graph_compute(
                         } break;
                     default:
                         {
-                            metal_printf("%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
+                            GGML_METAL_LOG_ERROR("%s: error: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
                             GGML_ASSERT(false);
                         }
                 }
@@ -1286,10 +1538,147 @@ void ggml_metal_graph_compute(
 
         MTLCommandBufferStatus status = (MTLCommandBufferStatus) [ctx->command_buffers[i] status];
         if (status != MTLCommandBufferStatusCompleted) {
-            metal_printf("%s: command buffer %d failed with status %lu\n", __func__, i, status);
+            GGML_METAL_LOG_INFO("%s: command buffer %d failed with status %lu\n", __func__, i, status);
             GGML_ASSERT(false);
         }
     }
 
     }
 }
+
+////////////////////////////////////////////////////////////////////////////////
+
+// backend interface
+
+static const char * ggml_backend_metal_name(ggml_backend_t backend) {
+    return "Metal";
+
+    UNUSED(backend);
+}
+
+static void ggml_backend_metal_free(ggml_backend_t backend) {
+    struct ggml_metal_context * ctx = (struct ggml_metal_context *)backend->context;
+    ggml_metal_free(ctx);
+    free(backend);
+}
+
+static void * ggml_backend_metal_buffer_get_base(ggml_backend_buffer_t buffer) {
+    return (void *)buffer->context;
+}
+
+static void ggml_backend_metal_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+    free(buffer->context);
+    UNUSED(buffer);
+}
+
+static struct ggml_backend_buffer_i metal_backend_buffer_i = {
+    /* .free_buffer    = */ ggml_backend_metal_buffer_free_buffer,
+    /* .get_base       = */ ggml_backend_metal_buffer_get_base,
+    /* .get_alloc_size = */ NULL, // defaults to ggml_nbytes
+    /* .init_tensor    = */ NULL, // no initialization required
+    /* .free_tensor    = */ NULL, // no cleanup required
+};
+
+static ggml_backend_buffer_t ggml_backend_metal_alloc_buffer(ggml_backend_t backend, size_t size) {
+    struct ggml_metal_context * ctx = (struct ggml_metal_context *)backend->context;
+
+    void * data = ggml_metal_host_malloc(size);
+
+    // TODO: set proper name of the buffers
+    ggml_metal_add_buffer(ctx, "backend", data, size, 0);
+
+    return ggml_backend_buffer_init(backend, metal_backend_buffer_i, data, size);
+}
+
+static size_t ggml_backend_metal_get_alignment(ggml_backend_t backend) {
+    return 32;
+    UNUSED(backend);
+}
+
+static void ggml_backend_metal_set_tensor_async(ggml_backend_t backend, 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(backend);
+}
+
+static void ggml_backend_metal_get_tensor_async(ggml_backend_t backend, 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(backend);
+}
+
+static void ggml_backend_metal_synchronize(ggml_backend_t backend) {
+    UNUSED(backend);
+}
+
+static void ggml_backend_metal_cpy_tensor_from(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst) {
+    ggml_backend_tensor_get(src, dst->data, 0, ggml_nbytes(src));
+
+    UNUSED(backend);
+}
+
+static void ggml_backend_metal_cpy_tensor_to(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst) {
+    ggml_backend_tensor_set_async(dst, src->data, 0, ggml_nbytes(src));
+
+    UNUSED(backend);
+}
+
+static void ggml_backend_metal_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+    struct ggml_metal_context * metal_ctx = (struct ggml_metal_context *)backend->context;
+
+    ggml_metal_graph_compute(metal_ctx, cgraph);
+}
+
+static bool ggml_backend_metal_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
+    return true;
+    UNUSED(backend);
+    UNUSED(op);
+}
+
+static struct ggml_backend_i metal_backend_i = {
+    /* .get_name            = */ ggml_backend_metal_name,
+    /* .free                = */ ggml_backend_metal_free,
+    /* .alloc_buffer        = */ ggml_backend_metal_alloc_buffer,
+    /* .get_alignment       = */ ggml_backend_metal_get_alignment,
+    /* .set_tensor_async    = */ ggml_backend_metal_set_tensor_async,
+    /* .get_tensor_async    = */ ggml_backend_metal_get_tensor_async,
+    /* .synchronize         = */ ggml_backend_metal_synchronize,
+    /* .cpy_tensor_from     = */ ggml_backend_metal_cpy_tensor_from,
+    /* .cpy_tensor_to       = */ ggml_backend_metal_cpy_tensor_to,
+    /* .graph_plan_create   = */ NULL, // the metal implementation does not require creating graph plans atm
+    /* .graph_plan_free     = */ NULL,
+    /* .graph_plan_compute  = */ NULL,
+    /* .graph_compute       = */ ggml_backend_metal_graph_compute,
+    /* .supports_op         = */ ggml_backend_metal_supports_op,
+};
+
+ggml_backend_t ggml_backend_metal_init(void) {
+    struct ggml_metal_context * ctx = malloc(sizeof(struct ggml_metal_context));
+
+    ctx = ggml_metal_init(GGML_DEFAULT_N_THREADS);
+
+    ggml_backend_t metal_backend = malloc(sizeof(struct ggml_backend));
+
+    *metal_backend = (struct ggml_backend) {
+        /* .interface = */ metal_backend_i,
+        /* .context   = */ ctx,
+    };
+
+    return metal_backend;
+}
+
+bool ggml_backend_is_metal(ggml_backend_t backend) {
+    return backend->iface.get_name == ggml_backend_metal_name;
+}
+
+void ggml_backend_metal_set_n_cb(ggml_backend_t backend, int n_cb) {
+    struct ggml_metal_context * ctx = (struct ggml_metal_context *)backend->context;
+
+    ggml_metal_set_n_cb(ctx, n_cb);
+}

ggml-metal.metal

@@ -13,23 +13,85 @@ typedef struct {
 
 #define QK4_1 32
 typedef struct {
-    half d;          // delta
-    half m;          // min
+    half d;                 // delta
+    half m;                 // min
     uint8_t qs[QK4_1 / 2];  // nibbles / quants
 } block_q4_1;
 
+#define QK5_0 32
+typedef struct {
+    half d;                // delta
+    uint8_t qh[4];         // 5-th bit of quants
+    uint8_t qs[QK5_0 / 2]; // nibbles / quants
+} block_q5_0;
+
+#define QK5_1 32
+typedef struct {
+    half d;                 // delta
+    half m;                 // min
+    uint8_t qh[4];          // 5-th bit of quants
+    uint8_t qs[QK5_1 / 2];  // nibbles / quants
+} block_q5_1;
+
 #define QK8_0 32
 typedef struct {
     half    d;         // delta
     int8_t  qs[QK8_0]; // quants
 } block_q8_0;
 
+// general-purpose kernel for addition of two tensors
+// pros: works for non-contiguous tensors, supports broadcast across dims 1, 2 and 3
+// cons: not very efficient
 kernel void kernel_add(
-        device const float4 * src0,
-        device const float4 * src1,
-        device       float4 * dst,
-        uint tpig[[thread_position_in_grid]]) {
-    dst[tpig] = src0[tpig] + src1[tpig];
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        constant  int64_t & ne00,
+        constant  int64_t & ne01,
+        constant  int64_t & ne02,
+        constant  int64_t & ne03,
+        constant  int64_t & nb00,
+        constant  int64_t & nb01,
+        constant  int64_t & nb02,
+        constant  int64_t & nb03,
+        constant  int64_t & ne10,
+        constant  int64_t & ne11,
+        constant  int64_t & ne12,
+        constant  int64_t & ne13,
+        constant  int64_t & nb10,
+        constant  int64_t & nb11,
+        constant  int64_t & nb12,
+        constant  int64_t & nb13,
+        constant  int64_t & ne0,
+        constant  int64_t & ne1,
+        constant  int64_t & ne2,
+        constant  int64_t & ne3,
+        constant  int64_t & nb0,
+        constant  int64_t & nb1,
+        constant  int64_t & nb2,
+        constant  int64_t & nb3,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    const int64_t i03 = tgpig.z;
+    const int64_t i02 = tgpig.y;
+    const int64_t i01 = tgpig.x;
+
+    const int64_t i13 = i03 % ne13;
+    const int64_t i12 = i02 % ne12;
+    const int64_t i11 = i01 % ne11;
+
+    device const char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01 + tpitg.x*nb00;
+    device const char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11 + tpitg.x*nb10;
+    device       char * dst_ptr  = dst  + i03*nb3  + i02*nb2  + i01*nb1  + tpitg.x*nb0;
+
+    for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
+        ((device float *)dst_ptr)[0] = ((device float *)src0_ptr)[0] + ((device float *)src1_ptr)[0];
+
+        src0_ptr += ntg.x*nb00;
+        src1_ptr += ntg.x*nb10;
+        dst_ptr  += ntg.x*nb0;
+    }
 }
 
 // assumption: src1 is a row
@@ -38,7 +100,7 @@ kernel void kernel_add_row(
         device const float4 * src0,
         device const float4 * src1,
         device       float4 * dst,
-        constant    int64_t & nb,
+        constant    int64_t & nb [[buffer(27)]],
         uint tpig[[thread_position_in_grid]]) {
     dst[tpig] = src0[tpig] + src1[tpig % nb];
 }
@@ -63,9 +125,17 @@ kernel void kernel_mul_row(
 }
 
 kernel void kernel_scale(
+        device const float * src0,
+        device       float * dst,
+        constant     float & scale,
+        uint tpig[[thread_position_in_grid]]) {
+    dst[tpig] = src0[tpig] * scale;
+}
+
+kernel void kernel_scale_4(
         device const float4 * src0,
         device       float4 * dst,
-        constant     float & scale,
+        constant     float  & scale,
         uint tpig[[thread_position_in_grid]]) {
     dst[tpig] = src0[tpig] * scale;
 }
@@ -85,6 +155,13 @@ kernel void kernel_relu(
     dst[tpig] = max(0.0f, src0[tpig]);
 }
 
+kernel void kernel_sqr(
+        device const float * src0,
+        device       float * dst,
+        uint tpig[[thread_position_in_grid]]) {
+    dst[tpig] = src0[tpig] * src0[tpig];
+}
+
 constant float GELU_COEF_A    = 0.044715f;
 constant float SQRT_2_OVER_PI = 0.79788456080286535587989211986876f;
 
@@ -107,36 +184,73 @@ kernel void kernel_soft_max(
         constant   int64_t & ne00,
         constant   int64_t & ne01,
         constant   int64_t & ne02,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]],
-        uint3   ntg[[threads_per_threadgroup]]) {
-    const int64_t i03 = tgpig[2];
-    const int64_t i02 = tgpig[1];
-    const int64_t i01 = tgpig[0];
+        threadgroup float  * buf [[threadgroup(0)]],
+        uint  tgpig[[threadgroup_position_in_grid]],
+        uint  tpitg[[thread_position_in_threadgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint    ntg[[threads_per_threadgroup]]) {
+    const int64_t i03 = (tgpig) / (ne02*ne01);
+    const int64_t i02 = (tgpig - i03*ne02*ne01) / ne01;
+    const int64_t i01 = (tgpig - i03*ne02*ne01 - i02*ne01);
 
     device const float * psrc0 = src0 + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
     device       float * pdst  = dst  + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
 
     // parallel max
-    float lmax = psrc0[tpitg[0]];
-    for (int i00 = tpitg[0] + ntg[0]; i00 < ne00; i00 += ntg[0]) {
+    float lmax = tpitg < ne00 ? psrc0[tpitg] : -INFINITY;
+
+    for (int i00 = tpitg + ntg; i00 < ne00; i00 += ntg) {
         lmax = MAX(lmax, psrc0[i00]);
     }
-    const float max = simd_max(lmax);
+
+    float max = simd_max(lmax);
+    if (tiisg == 0) {
+        buf[sgitg] = max;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    // broadcast, simd group number is ntg / 32
+    for (uint i = ntg / 32 / 2; i > 0; i /= 2) {
+       if (tpitg < i) {
+           buf[tpitg] = MAX(buf[tpitg], buf[tpitg + i]);
+       }
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    max = buf[0];
 
     // parallel sum
     float lsum = 0.0f;
-    for (int i00 = tpitg[0]; i00 < ne00; i00 += ntg[0]) {
+    for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
         const float exp_psrc0 = exp(psrc0[i00] - max);
         lsum += exp_psrc0;
         // Remember the result of exp here. exp is expensive, so we really do not
-        // whish to compute it twice.
+        // wish to compute it twice.
         pdst[i00] = exp_psrc0;
     }
 
-    const float sum = simd_sum(lsum);
+    float sum = simd_sum(lsum);
+    if (tiisg == 0) {
+        buf[sgitg] = sum;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    // broadcast, simd group number is ntg / 32
+    for (uint i = ntg / 32 / 2; i > 0; i /= 2) {
+       if (tpitg < i) {
+           buf[tpitg] += buf[tpitg + i];
+       }
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    sum = buf[0];
 
-    for (int i00 = tpitg[0]; i00 < ne00; i00 += ntg[0]) {
+    for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
         pdst[i00] /= sum;
     }
 }
@@ -147,37 +261,73 @@ kernel void kernel_soft_max_4(
         constant   int64_t & ne00,
         constant   int64_t & ne01,
         constant   int64_t & ne02,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]],
-        uint3   ntg[[threads_per_threadgroup]]) {
-    const int64_t i03 = tgpig[2];
-    const int64_t i02 = tgpig[1];
-    const int64_t i01 = tgpig[0];
+        threadgroup float  * buf [[threadgroup(0)]],
+        uint  tgpig[[threadgroup_position_in_grid]],
+        uint  tpitg[[thread_position_in_threadgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint    ntg[[threads_per_threadgroup]]) {
+    const int64_t i03 = (tgpig) / (ne02*ne01);
+    const int64_t i02 = (tgpig - i03*ne02*ne01) / ne01;
+    const int64_t i01 = (tgpig - i03*ne02*ne01 - i02*ne01);
 
     device const float4 * psrc4 = (device const float4 *)(src0 + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00);
     device       float4 * pdst4 = (device       float4 *)(dst  + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00);
 
     // parallel max
-    float4 lmax4 = psrc4[tpitg[0]];
-    for (int i00 = tpitg[0] + ntg[0]; i00 < ne00/4; i00 += ntg[0]) {
+    float4 lmax4 = tpitg < ne00/4 ? psrc4[tpitg] : -INFINITY;
+
+    for (int i00 = tpitg + ntg; i00 < ne00/4; i00 += ntg) {
         lmax4 = fmax(lmax4, psrc4[i00]);
     }
-    float lmax = MAX(MAX(lmax4[0], lmax4[1]), MAX(lmax4[2], lmax4[3]));
 
-    const float max = simd_max(lmax);
+    const float lmax = MAX(MAX(lmax4[0], lmax4[1]), MAX(lmax4[2], lmax4[3]));
+    float max = simd_max(lmax);
+    if (tiisg == 0) {
+        buf[sgitg] = max;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    // broadcast, simd group number is ntg / 32
+    for (uint i = ntg / 32 / 2; i > 0; i /= 2) {
+       if (tpitg < i) {
+           buf[tpitg] = MAX(buf[tpitg], buf[tpitg + i]);
+       }
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    max = buf[0];
 
     // parallel sum
     float4 lsum4 = 0.0f;
-    for (int i00 = tpitg[0]; i00 < ne00/4; i00 += ntg[0]) {
+    for (int i00 = tpitg; i00 < ne00/4; i00 += ntg) {
         const float4 exp_psrc4 = exp(psrc4[i00] - max);
         lsum4 += exp_psrc4;
         pdst4[i00] = exp_psrc4;
     }
-    float lsum = lsum4[0] + lsum4[1] + lsum4[2] + lsum4[3];
 
-    const float sum = simd_sum(lsum);
+    const float lsum = lsum4[0] + lsum4[1] + lsum4[2] + lsum4[3];
+    float sum = simd_sum(lsum);
+    if (tiisg == 0) {
+        buf[sgitg] = sum;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    // broadcast, simd group number is ntg / 32
+    for (uint i = ntg / 32 / 2; i > 0; i /= 2) {
+       if (tpitg < i) {
+           buf[tpitg] += buf[tpitg + i];
+       }
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    sum = buf[0];
 
-    for (int i00 = tpitg[0]; i00 < ne00/4; i00 += ntg[0]) {
+    for (int i00 = tpitg; i00 < ne00/4; i00 += ntg) {
         pdst4[i00] /= sum;
     }
 }
@@ -197,7 +347,7 @@ kernel void kernel_diag_mask_inf(
         dst[i02*ne01*ne00 + i01*ne00 + i00] = -INFINITY;
     } else {
         dst[i02*ne01*ne00 + i01*ne00 + i00] = src0[i02*ne01*ne00 + i01*ne00 + i00];
-     }
+    }
 }
 
 kernel void kernel_diag_mask_inf_8(
@@ -291,10 +441,11 @@ kernel void kernel_rms_norm(
         uint sgitg[[simdgroup_index_in_threadgroup]],
         uint tiisg[[thread_index_in_simdgroup]],
         uint   ntg[[threads_per_threadgroup]]) {
-    device const float4 * x = (device const float4 *) ((device const char *) src0 + tgpig*nb01);
-    device const float * x_scalar = (device const float *) x;
-    float4 sumf=0;
-    float all_sum=0;
+    device const float4 * x        = (device const float4 *) ((device const char *) src0 + tgpig*nb01);
+    device const float  * x_scalar = (device const float  *) x;
+
+    float4 sumf = 0;
+    float all_sum = 0;
 
     // parallel sum
     for (int i00 = tpitg; i00 < ne00/4; i00 += ntg) {
@@ -307,6 +458,7 @@ kernel void kernel_rms_norm(
     }
 
     threadgroup_barrier(mem_flags::mem_threadgroup);
+
     // broadcast, simd group number is ntg / 32
     for (uint i = ntg / 32 / 2; i > 0; i /= 2) {
        if (tpitg < i) {
@@ -314,7 +466,9 @@ kernel void kernel_rms_norm(
        }
     }
     if (tpitg == 0) {
-        for (int i = 4 * (ne00 / 4); i < ne00; i++) {sum[0] += x_scalar[i];}
+        for (int i = 4 * (ne00 / 4); i < ne00; i++) {
+            sum[0] += x_scalar[i];
+        }
         sum[0] /= ne00;
     }
 
@@ -329,7 +483,9 @@ kernel void kernel_rms_norm(
         y[i00] = x[i00] * scale;
     }
     if (tpitg == 0) {
-        for (int i00 = 4 * (ne00 / 4); i00 < ne00; i00++) {y_scalar[i00] = x_scalar[i00] * scale;}
+        for (int i00 = 4 * (ne00 / 4); i00 < ne00; i00++) {
+            y_scalar[i00] = x_scalar[i00] * scale;
+        }
     }
 }
 
@@ -339,8 +495,11 @@ kernel void kernel_rms_norm(
 // that corresponds to the missing bit shifts (1, 1/16, 1/256, 1/4096)
 inline float block_q_n_dot_y(device const block_q4_0 * qb_curr, float sumy, thread float * yl, int il) {
     float d = qb_curr->d;
+
     float2 acc = 0.f;
+
     device const uint16_t * qs = ((device const uint16_t *)qb_curr + 1 + il/2);
+
     for (int i = 0; i < 8; i+=2) {
         acc[0] += yl[i + 0] * (qs[i / 2] & 0x000F)
                 + yl[i + 1] * (qs[i / 2] & 0x0F00);
@@ -357,8 +516,11 @@ inline float block_q_n_dot_y(device const block_q4_0 * qb_curr, float sumy, thre
 inline float block_q_n_dot_y(device const block_q4_1 * qb_curr, float sumy, thread float * yl, int il) {
     float d = qb_curr->d;
     float m = qb_curr->m;
-    device const uint16_t * qs = ((device const uint16_t *)qb_curr + 2 + il/2);
+
     float2 acc = 0.f;
+
+    device const uint16_t * qs = ((device const uint16_t *)qb_curr + 2 + il/2);
+
     for (int i = 0; i < 8; i+=2) {
         acc[0] += yl[i + 0] * (qs[i / 2] & 0x000F)
                 + yl[i + 1] * (qs[i / 2] & 0x0F00);
@@ -368,9 +530,52 @@ inline float block_q_n_dot_y(device const block_q4_1 * qb_curr, float sumy, thre
     return d * (acc[0] + acc[1]) + sumy * m;
 }
 
+// function for calculate inner product between half a q5_0 block and 16 floats (yl), sumy is SUM(yl[i])
+// il indicates where the q5 quants begin (0 or QK5_0/4)
+// we assume that the yl's have been multiplied with the appropriate scale factor
+// that corresponds to the missing bit shifts (1, 1/16, 1/256, 1/4096)
+inline float block_q_n_dot_y(device const block_q5_0 * qb_curr, float sumy, thread float * yl, int il) {
+    float d = qb_curr->d;
+
+    float2 acc = 0.f;
+
+    device const uint16_t * qs =  ((device const uint16_t *)qb_curr + 3 + il/2);
+           const uint32_t   qh = *((device const uint32_t *)qb_curr->qh);
+
+    for (int i = 0; i < 8; i+=2) {
+        acc[0] += yl[i + 0] * ((qs[i / 2] & 0x000F) | ((qh >> (i+0+il        ) << 4 ) & 0x00010))
+                + yl[i + 1] * ((qs[i / 2] & 0x0F00) | ((qh >> (i+1+il        ) << 12) & 0x01000));
+        acc[1] += yl[i + 8] * ((qs[i / 2] & 0x00F0) | ((qh >> (i+0+il+QK5_0/2) << 8 ) & 0x00100))
+                + yl[i + 9] * ((qs[i / 2] & 0xF000) | ((qh >> (i+1+il+QK5_0/2) << 16) & 0x10000));
+    }
+    return d * (sumy * -16.f + acc[0] + acc[1]);
+}
+
+// function for calculate inner product between half a q5_1 block and 16 floats (yl), sumy is SUM(yl[i])
+// il indicates where the q5 quants begin (0 or QK5_1/4)
+// we assume that the yl's have been multiplied with the appropriate scale factor
+// that corresponds to the missing bit shifts (1, 1/16, 1/256, 1/4096)
+inline float block_q_n_dot_y(device const block_q5_1 * qb_curr, float sumy, thread float * yl, int il) {
+    float d = qb_curr->d;
+    float m = qb_curr->m;
+
+    float2 acc = 0.f;
+
+    device const uint16_t * qs =  ((device const uint16_t *)qb_curr + 4 + il/2);
+           const uint32_t   qh = *((device const uint32_t *)qb_curr->qh);
+
+    for (int i = 0; i < 8; i+=2) {
+        acc[0] += yl[i + 0] * ((qs[i / 2] & 0x000F) | ((qh >> (i+0+il        ) << 4 ) & 0x00010))
+                + yl[i + 1] * ((qs[i / 2] & 0x0F00) | ((qh >> (i+1+il        ) << 12) & 0x01000));
+        acc[1] += yl[i + 8] * ((qs[i / 2] & 0x00F0) | ((qh >> (i+0+il+QK5_0/2) << 8 ) & 0x00100))
+                + yl[i + 9] * ((qs[i / 2] & 0xF000) | ((qh >> (i+1+il+QK5_0/2) << 16) & 0x10000));
+    }
+    return d * (acc[0] + acc[1]) + sumy * m;
+}
+
 // putting them in the kernel cause a significant performance penalty
-#define N_DST 4 // each SIMD group works on 4 rows
-#define N_SIMDGROUP 2 // number of SIMD groups in a thread group
+#define N_DST 4        // each SIMD group works on 4 rows
+#define N_SIMDGROUP 2  // number of SIMD groups in a thread group
 #define N_SIMDWIDTH 32 // assuming SIMD group size is 32
 //Note: This is a template, but strictly speaking it only applies to
 //      quantizations where the block size is 32. It also does not
@@ -381,18 +586,23 @@ void mul_vec_q_n_f32(device const void * src0, device const float * src1, device
                     int64_t ne00, int64_t ne01, int64_t ne02, int64_t ne10, int64_t ne12, int64_t ne0, int64_t ne1, uint gqa,
                     uint3 tgpig, uint tiisg, uint sgitg) {
     const int nb = ne00/QK4_0;
+
     const int r0 = tgpig.x;
     const int r1 = tgpig.y;
     const int im = tgpig.z;
+
     const int first_row = (r0 * nsg + sgitg) * nr;
+
     const uint offset0 = first_row * nb + im/gqa*(nb*ne0);
+
     device const block_q_type * x = (device const block_q_type *) src0 + offset0;
     device const float        * y = (device const float        *) src1 + r1*ne10 + im*ne00*ne1;
-    float yl[16];       // src1 vector cache
-    float sumf[nr]={0.f};
 
-    const int ix = tiisg/2;
-    const int il = 8*(tiisg%2);
+    float yl[16]; // src1 vector cache
+    float sumf[nr] = {0.f};
+
+    const int ix = (tiisg/2);
+    const int il = (tiisg%2)*8;
 
     device const float * yb = y + ix * QK4_0 + il;
 
@@ -403,6 +613,7 @@ void mul_vec_q_n_f32(device const void * src0, device const float * src1, device
             sumy += yb[i] + yb[i+1];
             yl[i+0] = yb[i+ 0];
             yl[i+1] = yb[i+ 1]/256.f;
+
             sumy += yb[i+16] + yb[i+17];
             yl[i+8] = yb[i+16]/16.f;
             yl[i+9] = yb[i+17]/4096.f;
@@ -418,12 +629,12 @@ void mul_vec_q_n_f32(device const void * src0, device const float * src1, device
     for (int row = 0; row < nr; ++row) {
         const float tot = simd_sum(sumf[row]);
         if (tiisg == 0 && first_row + row < ne01) {
-            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = tot;
+            dst[im*ne0*ne1 + r1*ne0 + first_row + row] = tot;
         }
     }
 }
 
-kernel void kernel_mul_mat_q4_0_f32(
+kernel void kernel_mul_mv_q4_0_f32(
         device const  void * src0,
         device const float * src1,
         device       float * dst,
@@ -436,12 +647,12 @@ kernel void kernel_mul_mat_q4_0_f32(
         constant   int64_t & ne1[[buffer(16)]],
         constant   uint    & gqa[[buffer(17)]],
         uint3 tgpig[[threadgroup_position_in_grid]],
-        uint tiisg[[thread_index_in_simdgroup]],
-        uint sgitg[[simdgroup_index_in_threadgroup]]) {
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
     mul_vec_q_n_f32<block_q4_0, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(src0,src1,dst,ne00,ne01,ne02,ne10,ne12,ne0,ne1,gqa,tgpig,tiisg,sgitg);
 }
 
-kernel void kernel_mul_mat_q4_1_f32(
+kernel void kernel_mul_mv_q4_1_f32(
         device const  void * src0,
         device const float * src1,
         device       float * dst,
@@ -459,9 +670,46 @@ kernel void kernel_mul_mat_q4_1_f32(
      mul_vec_q_n_f32<block_q4_1, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(src0,src1,dst,ne00,ne01,ne02,ne10,ne12,ne0,ne1,gqa,tgpig,tiisg,sgitg);
 }
 
+kernel void kernel_mul_mv_q5_0_f32(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01[[buffer(4)]],
+        constant   int64_t & ne02[[buffer(5)]],
+        constant   int64_t & ne10[[buffer(9)]],
+        constant   int64_t & ne12[[buffer(11)]],
+        constant   int64_t & ne0[[buffer(15)]],
+        constant   int64_t & ne1[[buffer(16)]],
+        constant   uint    & gqa[[buffer(17)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+    mul_vec_q_n_f32<block_q5_0, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(src0,src1,dst,ne00,ne01,ne02,ne10,ne12,ne0,ne1,gqa,tgpig,tiisg,sgitg);
+}
+
+kernel void kernel_mul_mv_q5_1_f32(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01[[buffer(4)]],
+        constant   int64_t & ne02[[buffer(5)]],
+        constant   int64_t & ne10[[buffer(9)]],
+        constant   int64_t & ne12[[buffer(11)]],
+        constant   int64_t & ne0[[buffer(15)]],
+        constant   int64_t & ne1[[buffer(16)]],
+        constant   uint    & gqa[[buffer(17)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+    mul_vec_q_n_f32<block_q5_1, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(src0,src1,dst,ne00,ne01,ne02,ne10,ne12,ne0,ne1,gqa,tgpig,tiisg,sgitg);
+}
+
+
 #define NB_Q8_0 8
 
-kernel void kernel_mul_mat_q8_0_f32(
+kernel void kernel_mul_mv_q8_0_f32(
         device const  void * src0,
         device const float * src1,
         device       float * dst,
@@ -525,7 +773,7 @@ kernel void kernel_mul_mat_q8_0_f32(
 
 #define N_F32_F32 4
 
-kernel void kernel_mul_mat_f32_f32(
+kernel void kernel_mul_mv_f32_f32(
         device const  char * src0,
         device const  char * src1,
         device       float * dst,
@@ -596,7 +844,7 @@ kernel void kernel_mul_mat_f32_f32(
     }
 }
 
-kernel void kernel_mul_mat_f16_f32_1row(
+kernel void kernel_mul_mv_f16_f32_1row(
         device const  char * src0,
         device const  char * src1,
         device       float * dst,
@@ -615,7 +863,7 @@ kernel void kernel_mul_mat_f16_f32_1row(
         constant   int64_t & ne0,
         constant   int64_t & ne1,
         uint3 tgpig[[threadgroup_position_in_grid]],
-        uint tiisg[[thread_index_in_simdgroup]]) {
+        uint  tiisg[[thread_index_in_simdgroup]]) {
 
     const int64_t r0 = tgpig.x;
     const int64_t r1 = tgpig.y;
@@ -650,7 +898,7 @@ kernel void kernel_mul_mat_f16_f32_1row(
 
 #define N_F16_F32 4
 
-kernel void kernel_mul_mat_f16_f32(
+kernel void kernel_mul_mv_f16_f32(
         device const  char * src0,
         device const  char * src1,
         device       float * dst,
@@ -722,7 +970,7 @@ kernel void kernel_mul_mat_f16_f32(
 }
 
 // Assumes row size (ne00) is a multiple of 4
-kernel void kernel_mul_mat_f16_f32_l4(
+kernel void kernel_mul_mv_f16_f32_l4(
         device const  char * src0,
         device const  char * src1,
         device       float * dst,
@@ -783,7 +1031,9 @@ kernel void kernel_alibi_f32(
         constant  uint64_t & nb1,
         constant  uint64_t & nb2,
         constant  uint64_t & nb3,
-        constant      float & m0,
+        constant     float & m0,
+        constant     float & m1,
+        constant       int & n_heads_log2_floor,
         uint3 tgpig[[threadgroup_position_in_grid]],
         uint3 tpitg[[thread_position_in_threadgroup]],
         uint3   ntg[[threads_per_threadgroup]]) {
@@ -799,37 +1049,122 @@ kernel void kernel_alibi_f32(
     const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0);
 
     device float * dst_data = (device float *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
-    float m_k = pow(m0, i2 + 1);
+    float m_k;
+    if (i2 < n_heads_log2_floor) {
+        m_k = pow(m0, i2 + 1);
+    } else {
+        m_k = pow(m1, 2 * (i2 - n_heads_log2_floor) + 1);
+    }
     for (int64_t i00 = tpitg.x; i00 < ne00; i00 += ntg.x) {
         device const float * src = (device float *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
         dst_data[i00] = src[0] + m_k * (i00 - ne00 + 1);
     }
 }
 
+static float rope_yarn_ramp(const float low, const float high, const int i0) {
+    const float y = (i0 / 2 - low) / max(0.001f, high - low);
+    return 1.0f - min(1.0f, max(0.0f, y));
+}
+
+// YaRN algorithm based on LlamaYaRNScaledRotaryEmbedding.py from https://github.com/jquesnelle/yarn
+// MIT licensed. Copyright (c) 2023 Jeffrey Quesnelle and Bowen Peng.
+static void rope_yarn(
+    float theta_extrap, float freq_scale, float corr_dims[2], int64_t i0, float ext_factor, float mscale,
+    thread float * cos_theta, thread float * sin_theta
+) {
+    // Get n-d rotational scaling corrected for extrapolation
+    float theta_interp = freq_scale * theta_extrap;
+    float theta = theta_interp;
+    if (ext_factor != 0.0f) {
+        float ramp_mix = rope_yarn_ramp(corr_dims[0], corr_dims[1], i0) * ext_factor;
+        theta = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix;
+
+        // Get n-d magnitude scaling corrected for interpolation
+        mscale *= 1.0f + 0.1f * log(1.0f / freq_scale);
+    }
+    *cos_theta = cos(theta) * mscale;
+    *sin_theta = sin(theta) * mscale;
+}
+
+// Apparently solving `n_rot = 2pi * x * base^((2 * max_pos_emb) / n_dims)` for x, we get
+// `corr_fac(n_rot) = n_dims * log(max_pos_emb / (n_rot * 2pi)) / (2 * log(base))`
+static float rope_yarn_corr_factor(int n_dims, int n_orig_ctx, float n_rot, float base) {
+    return n_dims * log(n_orig_ctx / (n_rot * 2 * M_PI_F)) / (2 * log(base));
+}
+
+static void rope_yarn_corr_dims(
+    int n_dims, int n_orig_ctx, float freq_base, float beta_fast, float beta_slow, float dims[2]
+) {
+    // start and end correction dims
+    dims[0] = max(0.0f,         floor(rope_yarn_corr_factor(n_dims, n_orig_ctx, beta_fast, freq_base)));
+    dims[1] = min(n_dims - 1.0f, ceil(rope_yarn_corr_factor(n_dims, n_orig_ctx, beta_slow, freq_base)));
+}
+
+typedef void (rope_t)(
+        device const    void * src0,
+        device const int32_t * src1,
+        device         float * dst,
+        constant     int64_t & ne00,
+        constant     int64_t & ne01,
+        constant     int64_t & ne02,
+        constant     int64_t & ne03,
+        constant    uint64_t & nb00,
+        constant    uint64_t & nb01,
+        constant    uint64_t & nb02,
+        constant    uint64_t & nb03,
+        constant     int64_t & ne0,
+        constant     int64_t & ne1,
+        constant     int64_t & ne2,
+        constant     int64_t & ne3,
+        constant    uint64_t & nb0,
+        constant    uint64_t & nb1,
+        constant    uint64_t & nb2,
+        constant    uint64_t & nb3,
+        constant         int & n_past,
+        constant         int & n_dims,
+        constant         int & mode,
+        constant         int & n_orig_ctx,
+        constant       float & freq_base,
+        constant       float & freq_scale,
+        constant       float & ext_factor,
+        constant       float & attn_factor,
+        constant       float & beta_fast,
+        constant       float & beta_slow,
+        uint  tiitg[[thread_index_in_threadgroup]],
+        uint3 tptg[[threads_per_threadgroup]],
+        uint3 tgpig[[threadgroup_position_in_grid]]);
+
+template<typename T>
 kernel void kernel_rope(
-        device const  void * src0,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant   int64_t & ne03,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   int64_t & ne2,
-        constant   int64_t & ne3,
-        constant  uint64_t & nb0,
-        constant  uint64_t & nb1,
-        constant  uint64_t & nb2,
-        constant  uint64_t & nb3,
-        constant       int & n_past,
-        constant       int & n_dims,
-        constant       int & mode,
-        constant     float & freq_base,
-        constant     float & freq_scale,
+        device const    void * src0,
+        device const int32_t * src1,
+        device         float * dst,
+        constant     int64_t & ne00,
+        constant     int64_t & ne01,
+        constant     int64_t & ne02,
+        constant     int64_t & ne03,
+        constant    uint64_t & nb00,
+        constant    uint64_t & nb01,
+        constant    uint64_t & nb02,
+        constant    uint64_t & nb03,
+        constant     int64_t & ne0,
+        constant     int64_t & ne1,
+        constant     int64_t & ne2,
+        constant     int64_t & ne3,
+        constant    uint64_t & nb0,
+        constant    uint64_t & nb1,
+        constant    uint64_t & nb2,
+        constant    uint64_t & nb3,
+        constant         int & n_past,
+        constant         int & n_dims,
+        constant         int & mode,
+        constant         int & n_orig_ctx,
+        constant       float & freq_base,
+        constant       float & freq_scale,
+        constant       float & ext_factor,
+        constant       float & attn_factor,
+        constant       float & beta_fast,
+        constant       float & beta_slow,
         uint  tiitg[[thread_index_in_threadgroup]],
         uint3 tptg[[threads_per_threadgroup]],
         uint3 tgpig[[threadgroup_position_in_grid]]) {
@@ -839,23 +1174,28 @@ kernel void kernel_rope(
 
     const bool is_neox = mode & 2;
 
-    const int64_t p = ((mode & 1) == 0 ? n_past + i2 : i2);
+    float corr_dims[2];
+    rope_yarn_corr_dims(n_dims, n_orig_ctx, freq_base, beta_fast, beta_slow, corr_dims);
 
-    const float theta_0 = freq_scale * (float)p;
+    device const int32_t * pos = src1;
+
+    const int64_t p = pos[i2];
+
+    const float theta_0 = (float)p;
     const float inv_ndims = -1.f/n_dims;
 
     if (!is_neox) {
         for (int64_t i0 = 2*tiitg; i0 < ne0; i0 += 2*tptg.x) {
 
             const float theta = theta_0 * pow(freq_base, inv_ndims*i0);
-            const float cos_theta = cos(theta);
-            const float sin_theta = sin(theta);
+            float cos_theta, sin_theta;
+            rope_yarn(theta, freq_scale, corr_dims, i0, ext_factor, attn_factor, &cos_theta, &sin_theta);
 
-            device const float * const src = (device float *)((device char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
-            device       float * dst_data  = (device float *)((device char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+            device const T * const src = (device T *)((device char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+            device       T * dst_data  = (device T *)((device char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
 
-            const float x0 = src[0];
-            const float x1 = src[1];
+            const T x0 = src[0];
+            const T x1 = src[1];
 
             dst_data[0] = x0*cos_theta - x1*sin_theta;
             dst_data[1] = x0*sin_theta + x1*cos_theta;
@@ -864,14 +1204,17 @@ kernel void kernel_rope(
         for (int64_t ib = 0; ib < ne0/n_dims; ++ib) {
             for (int64_t ic = 2*tiitg; ic < n_dims; ic += 2*tptg.x) {
 
-                const float theta = theta_0 * pow(freq_base, inv_ndims*ic - ib);
-                const float cos_theta = cos(theta);
-                const float sin_theta = sin(theta);
+                // simplified from `(ib * n_dims + ic) * inv_ndims`
+                const float cur_rot = inv_ndims*ic - ib;
+
+                const float theta = theta_0 * pow(freq_base, cur_rot);
+                float cos_theta, sin_theta;
+                rope_yarn(theta, freq_scale, corr_dims, cur_rot, ext_factor, attn_factor, &cos_theta, &sin_theta);
 
                 const int64_t i0 = ib*n_dims + ic/2;
 
-                device const float * const src = (device float *)((device char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
-                device       float * dst_data  = (device float *)((device char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+                device const T * const src = (device T *)((device char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+                device       T * dst_data  = (device T *)((device char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
 
                 const float x0 = src[0];
                 const float x1 = src[n_dims/2];
@@ -883,6 +1226,9 @@ kernel void kernel_rope(
     }
 }
 
+template [[host_name("kernel_rope_f32")]] kernel rope_t kernel_rope<float>;
+template [[host_name("kernel_rope_f16")]] kernel rope_t kernel_rope<half>;
+
 kernel void kernel_cpy_f16_f16(
         device const half * src0,
         device       half * dst,
@@ -1008,6 +1354,62 @@ kernel void kernel_cpy_f32_f32(
     }
 }
 
+kernel void kernel_concat(
+    device const char * src0,
+    device const char * src1,
+    device       char * dst,
+    constant   int64_t & ne00,
+    constant   int64_t & ne01,
+    constant   int64_t & ne02,
+    constant   int64_t & ne03,
+    constant  uint64_t & nb00,
+    constant  uint64_t & nb01,
+    constant  uint64_t & nb02,
+    constant  uint64_t & nb03,
+    constant   int64_t & ne10,
+    constant   int64_t & ne11,
+    constant   int64_t & ne12,
+    constant   int64_t & ne13,
+    constant  uint64_t & nb10,
+    constant  uint64_t & nb11,
+    constant  uint64_t & nb12,
+    constant  uint64_t & nb13,
+    constant   int64_t & ne0,
+    constant   int64_t & ne1,
+    constant   int64_t & ne2,
+    constant   int64_t & ne3,
+    constant  uint64_t & nb0,
+    constant  uint64_t & nb1,
+    constant  uint64_t & nb2,
+    constant  uint64_t & nb3,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]]) {
+
+    const int64_t i03 = tgpig.z;
+    const int64_t i02 = tgpig.y;
+    const int64_t i01 = tgpig.x;
+
+    const int64_t i13 = i03 % ne13;
+    const int64_t i12 = i02 % ne12;
+    const int64_t i11 = i01 % ne11;
+
+    device const char * src0_ptr = src0 + i03 * nb03 + i02 * nb02 + i01 * nb01 + tpitg.x*nb00;
+    device const char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11 + tpitg.x*nb10;
+    device       char * dst_ptr  = dst  + i03*nb3  + i02*nb2  + i01*nb1  + tpitg.x*nb0;
+
+    for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
+        if (i02 < ne02) {
+            ((device float *)dst_ptr)[0] = ((device float *)src0_ptr)[0];
+            src0_ptr += ntg.x*nb00;
+        } else {
+            ((device float *)dst_ptr)[0] = ((device float *)src1_ptr)[0];
+            src1_ptr += ntg.x*nb10;
+        }
+        dst_ptr += ntg.x*nb0;
+    }
+}
+
 //============================================ k-quants ======================================================
 
 #ifndef QK_K
@@ -1100,7 +1502,7 @@ static inline uchar4 get_scale_min_k4(int j, device const uint8_t * q) {
 
 //====================================== dot products =========================
 
-kernel void kernel_mul_mat_q2_K_f32(
+kernel void kernel_mul_mv_q2_K_f32(
         device const  void * src0,
         device const float * src1,
         device       float * dst,
@@ -1244,7 +1646,7 @@ kernel void kernel_mul_mat_q2_K_f32(
 }
 
 #if QK_K == 256
-kernel void kernel_mul_mat_q3_K_f32(
+kernel void kernel_mul_mv_q3_K_f32(
         device const  void * src0,
         device const float * src1,
         device       float * dst,
@@ -1273,8 +1675,8 @@ kernel void kernel_mul_mat_q3_K_f32(
 
     float yl[32];
 
-    const uint16_t kmask1 = 0x3030;
-    const uint16_t kmask2 = 0x0f0f;
+    //const uint16_t kmask1 = 0x3030;
+    //const uint16_t kmask2 = 0x0f0f;
 
     const int tid = tiisg/4;
     const int ix  = tiisg%4;
@@ -1396,7 +1798,7 @@ kernel void kernel_mul_mat_q3_K_f32(
     }
 }
 #else
-kernel void kernel_mul_mat_q3_K_f32(
+kernel void kernel_mul_mv_q3_K_f32(
         device const  void * src0,
         device const float * src1,
         device       float * dst,
@@ -1467,7 +1869,7 @@ kernel void kernel_mul_mat_q3_K_f32(
 #endif
 
 #if QK_K == 256
-kernel void kernel_mul_mat_q4_K_f32(
+kernel void kernel_mul_mv_q4_K_f32(
         device const  void * src0,
         device const float * src1,
         device       float * dst,
@@ -1573,7 +1975,7 @@ kernel void kernel_mul_mat_q4_K_f32(
     }
 }
 #else
-kernel void kernel_mul_mat_q4_K_f32(
+kernel void kernel_mul_mv_q4_K_f32(
         device const  void * src0,
         device const float * src1,
         device       float * dst,
@@ -1662,7 +2064,7 @@ kernel void kernel_mul_mat_q4_K_f32(
 }
 #endif
 
-kernel void kernel_mul_mat_q5_K_f32(
+kernel void kernel_mul_mv_q5_K_f32(
         device const  void * src0,
         device const float * src1,
         device       float * dst,
@@ -1835,7 +2237,7 @@ kernel void kernel_mul_mat_q5_K_f32(
 
 }
 
-kernel void kernel_mul_mat_q6_K_f32(
+kernel void kernel_mul_mv_q6_K_f32(
         device const  void * src0,
         device const float * src1,
         device       float * dst,
@@ -1984,6 +2386,62 @@ void dequantize_q4_1(device const block_q4_1 *xb, short il, thread type4x4 & reg
     }
 }
 
+template <typename type4x4>
+void dequantize_q5_0(device const block_q5_0 *xb, short il, thread type4x4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 3);
+    const float d = xb->d;
+    const float md = -16.h * xb->d;
+    const ushort mask = il ? 0x00F0 : 0x000F;
+
+    const uint32_t qh = *((device const uint32_t *)xb->qh);
+
+    const int x_mv = il ? 4 : 0;
+
+    const int gh_mv = il ? 12 : 0;
+    const int gh_bk = il ?  0 : 4;
+
+    for (int i = 0; i < 8; i++) {
+        // extract the 5-th bits for x0 and x1
+        const uint8_t xh_0 = ((qh >> (gh_mv + 2*i  )) << gh_bk) & 0x10;
+        const uint8_t xh_1 = ((qh >> (gh_mv + 2*i+1)) << gh_bk) & 0x10;
+
+        // combine the 4-bits from qs with the 5th bit
+        const int32_t x0 = ((((qs[i]     ) & mask) >> x_mv) | xh_0);
+        const int32_t x1 = ((((qs[i] >> 8) & mask) >> x_mv) | xh_1);
+
+        reg[i/2][2*(i%2)+0] = d * x0 + md;
+        reg[i/2][2*(i%2)+1] = d * x1 + md;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q5_1(device const block_q5_1 *xb, short il, thread type4x4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 4);
+    const float d = xb->d;
+    const float m = xb->m;
+    const ushort mask = il ? 0x00F0 : 0x000F;
+
+    const uint32_t qh = *((device const uint32_t *)xb->qh);
+
+    const int x_mv = il ? 4 : 0;
+
+    const int gh_mv = il ? 12 : 0;
+    const int gh_bk = il ?  0 : 4;
+
+    for (int i = 0; i < 8; i++) {
+        // extract the 5-th bits for x0 and x1
+        const uint8_t xh_0 = ((qh >> (gh_mv + 2*i  )) << gh_bk) & 0x10;
+        const uint8_t xh_1 = ((qh >> (gh_mv + 2*i+1)) << gh_bk) & 0x10;
+
+        // combine the 4-bits from qs with the 5th bit
+        const int32_t x0 = ((((qs[i]     ) & mask) >> x_mv) | xh_0);
+        const int32_t x1 = ((((qs[i] >> 8) & mask) >> x_mv) | xh_1);
+
+        reg[i/2][2*(i%2)+0] = d * x0 + m;
+        reg[i/2][2*(i%2)+1] = d * x1 + m;
+    }
+}
+
 template <typename type4x4>
 void dequantize_q8_0(device const block_q8_0 *xb, short il, thread type4x4 & reg) {
     device const int8_t * qs = ((device const int8_t *)xb->qs);
@@ -2173,7 +2631,7 @@ kernel void kernel_get_rows(
 }
 
 #define BLOCK_SIZE_M 64 // 8 simdgroup matrices from matrix A
-#define BLOCK_SIZE_N 32 // 4 simdgroup matrices from matrix A
+#define BLOCK_SIZE_N 32 // 4 simdgroup matrices from matrix B
 #define BLOCK_SIZE_K 32
 #define THREAD_MAT_M 4 // each thread take 4 simdgroup matrices from matrix A
 #define THREAD_MAT_N 2 // each thread take 2 simdgroup matrices from matrix B
@@ -2210,9 +2668,11 @@ kernel void kernel_mul_mm(device const  uchar * src0,
     const uint r0 = tgpig.y;
     const uint r1 = tgpig.x;
     const uint im = tgpig.z;
+
     // if this block is of 64x32 shape or smaller
     short n_rows = (ne0 - r0 * BLOCK_SIZE_M < BLOCK_SIZE_M) ? (ne0 - r0 * BLOCK_SIZE_M) : BLOCK_SIZE_M;
     short n_cols = (ne1 - r1 * BLOCK_SIZE_N < BLOCK_SIZE_N) ? (ne1 - r1 * BLOCK_SIZE_N) : BLOCK_SIZE_N;
+
     // a thread shouldn't load data outside of the matrix
     short thread_row = ((short)tiitg/THREAD_PER_ROW) < n_rows ? ((short)tiitg/THREAD_PER_ROW) : n_rows - 1;
     short thread_col = ((short)tiitg/THREAD_PER_COL) < n_cols ? ((short)tiitg/THREAD_PER_COL) : n_cols - 1;
@@ -2236,26 +2696,30 @@ kernel void kernel_mul_mm(device const  uchar * src0,
         + nb10 * (BLOCK_SIZE_K / THREAD_PER_COL * (tiitg % THREAD_PER_COL)));
 
     for (int loop_k = 0; loop_k < ne00; loop_k += BLOCK_SIZE_K) {
-        //load data and store to threadgroup memory
+        // load data and store to threadgroup memory
         half4x4 temp_a;
         dequantize_func(x, il, temp_a);
         threadgroup_barrier(mem_flags::mem_threadgroup);
+
         #pragma unroll(16)
         for (int i = 0; i < 16; i++) {
             *(sa + SG_MAT_SIZE * ((tiitg / THREAD_PER_ROW / 8) \
-            + 16 * (tiitg % THREAD_PER_ROW) + 8 * (i / 8)) \
-            + (tiitg / THREAD_PER_ROW) % 8 + (i & 7) * 8) = temp_a[i/4][i%4];
+            +                     (tiitg % THREAD_PER_ROW) * 16 + (i / 8) * 8) \
+            +                     (tiitg / THREAD_PER_ROW) % 8  + (i & 7) * 8) = temp_a[i/4][i%4];
         }
-        *(threadgroup float2x4 *)(sb + (tiitg % THREAD_PER_COL) * 8 * 32 + 8 * (tiitg / THREAD_PER_COL)) \
-                = *((device float2x4 *)y);
+
+        *(threadgroup float2x4 *)(sb + (tiitg % THREAD_PER_COL) * 8 * 32 + 8 * (tiitg / THREAD_PER_COL)) = *((device float2x4 *)y);
+
         il = (il + 2 < nl) ? il + 2 : il % 2;
         x  = (il < 2) ? x + (2+nl-1)/nl : x;
         y += BLOCK_SIZE_K;
 
         threadgroup_barrier(mem_flags::mem_threadgroup);
-        //load matrices from threadgroup memory and conduct outer products
+
+        // load matrices from threadgroup memory and conduct outer products
         threadgroup half  * lsma = (sa + THREAD_MAT_M * SG_MAT_SIZE * (sgitg % 2));
         threadgroup float * lsmb = (sb + THREAD_MAT_N * SG_MAT_SIZE * (sgitg / 2));
+
         #pragma unroll(4)
         for (int ik = 0; ik < BLOCK_SIZE_K / 8; ik++) {
             #pragma unroll(4)
@@ -2270,6 +2734,7 @@ kernel void kernel_mul_mm(device const  uchar * src0,
 
             lsma += BLOCK_SIZE_M / SG_MAT_ROW * SG_MAT_SIZE;
             lsmb += BLOCK_SIZE_N / SG_MAT_ROW * SG_MAT_SIZE;
+
             #pragma unroll(8)
             for (int i = 0; i < 8; i++){
                 simdgroup_multiply_accumulate(c_res[i], mb[i/4], ma[i%4], c_res[i]);
@@ -2278,25 +2743,26 @@ kernel void kernel_mul_mm(device const  uchar * src0,
     }
 
     if ((r0 + 1) * BLOCK_SIZE_M <= ne0 && (r1 + 1) * BLOCK_SIZE_N <= ne1) {
-        device float *C = dst + BLOCK_SIZE_M * r0 + 32 * (sgitg&1) \
-                          + (BLOCK_SIZE_N * r1 + 16 * (sgitg>>1)) * ne0 + im*ne1*ne0;
+        device float * C = dst + (BLOCK_SIZE_M * r0 + 32 * (sgitg &  1)) \
+                               + (BLOCK_SIZE_N * r1 + 16 * (sgitg >> 1)) * ne0 + im*ne1*ne0;
         for (int i = 0; i < 8; i++) {
             simdgroup_store(c_res[i], C + 8 * (i%4) + 8 * ne0 * (i/4), ne0);
         }
     } else {
         // block is smaller than 64x32, we should avoid writing data outside of the matrix
         threadgroup_barrier(mem_flags::mem_threadgroup);
-        threadgroup float *temp_str = ((threadgroup float *)shared_memory) \
+        threadgroup float * temp_str = ((threadgroup float *)shared_memory) \
                                       + 32 * (sgitg&1) + (16 * (sgitg>>1)) * BLOCK_SIZE_M;
         for (int i = 0; i < 8; i++) {
             simdgroup_store(c_res[i], temp_str + 8 * (i%4) + 8 * BLOCK_SIZE_M * (i/4), BLOCK_SIZE_M);
         }
 
         threadgroup_barrier(mem_flags::mem_threadgroup);
-        device float *C = dst + BLOCK_SIZE_M * r0 + (BLOCK_SIZE_N * r1) * ne0 + im*ne1*ne0;
-        if (sgitg==0) {
+
+        device float * C = dst + (BLOCK_SIZE_M * r0) + (BLOCK_SIZE_N * r1) * ne0 + im*ne1*ne0;
+        if (sgitg == 0) {
             for (int i = 0; i < n_rows; i++) {
-                for (int j = tiitg; j< n_cols; j += BLOCK_SIZE_N) {
+                for (int j = tiitg; j < n_cols; j += BLOCK_SIZE_N) {
                     *(C + i + j * ne0) = *(temp_str + i + j * BLOCK_SIZE_M);
                 }
             }
@@ -2317,6 +2783,8 @@ template [[host_name("kernel_get_rows_f32")]]  kernel get_rows_t kernel_get_rows
 template [[host_name("kernel_get_rows_f16")]]  kernel get_rows_t kernel_get_rows<half4x4,    1, dequantize_f16>;
 template [[host_name("kernel_get_rows_q4_0")]] kernel get_rows_t kernel_get_rows<block_q4_0, 2, dequantize_q4_0>;
 template [[host_name("kernel_get_rows_q4_1")]] kernel get_rows_t kernel_get_rows<block_q4_1, 2, dequantize_q4_1>;
+template [[host_name("kernel_get_rows_q5_0")]] kernel get_rows_t kernel_get_rows<block_q5_0, 2, dequantize_q5_0>;
+template [[host_name("kernel_get_rows_q5_1")]] kernel get_rows_t kernel_get_rows<block_q5_1, 2, dequantize_q5_1>;
 template [[host_name("kernel_get_rows_q8_0")]] kernel get_rows_t kernel_get_rows<block_q8_0, 2, dequantize_q8_0>;
 template [[host_name("kernel_get_rows_q2_K")]] kernel get_rows_t kernel_get_rows<block_q2_K, QK_NL, dequantize_q2_K>;
 template [[host_name("kernel_get_rows_q3_K")]] kernel get_rows_t kernel_get_rows<block_q3_K, QK_NL, dequantize_q3_K>;
@@ -2345,6 +2813,8 @@ template [[host_name("kernel_mul_mm_f32_f32")]]  kernel mat_mm_t kernel_mul_mm<f
 template [[host_name("kernel_mul_mm_f16_f32")]]  kernel mat_mm_t kernel_mul_mm<half4x4,    1,     dequantize_f16>;
 template [[host_name("kernel_mul_mm_q4_0_f32")]] kernel mat_mm_t kernel_mul_mm<block_q4_0, 2,     dequantize_q4_0>;
 template [[host_name("kernel_mul_mm_q4_1_f32")]] kernel mat_mm_t kernel_mul_mm<block_q4_1, 2,     dequantize_q4_1>;
+template [[host_name("kernel_mul_mm_q5_0_f32")]] kernel mat_mm_t kernel_mul_mm<block_q5_0, 2,     dequantize_q5_0>;
+template [[host_name("kernel_mul_mm_q5_1_f32")]] kernel mat_mm_t kernel_mul_mm<block_q5_1, 2,     dequantize_q5_1>;
 template [[host_name("kernel_mul_mm_q8_0_f32")]] kernel mat_mm_t kernel_mul_mm<block_q8_0, 2,     dequantize_q8_0>;
 template [[host_name("kernel_mul_mm_q2_K_f32")]] kernel mat_mm_t kernel_mul_mm<block_q2_K, QK_NL, dequantize_q2_K>;
 template [[host_name("kernel_mul_mm_q3_K_f32")]] kernel mat_mm_t kernel_mul_mm<block_q3_K, QK_NL, dequantize_q3_K>;

ggml-opencl.cpp

@@ -19,7 +19,7 @@
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
-#define CL_DMMV_BLOCK_SIZE 32
+#define CL_DMMV_LOCAL_SIZE 32
 
 #ifndef K_QUANTS_PER_ITERATION
 #define K_QUANTS_PER_ITERATION 1
@@ -202,14 +202,14 @@ inline void get_scale_min_k4(int j, const __global uint8_t *q, uint8_t *d, uint8
 
 __kernel void dequantize_block_q2_K(__global const struct block_q2_K *x, __global float *yy)
 {
-    const int i = get_group_id(0);
+    const int i = get_group_id(0) + get_global_offset(0);
     const int tid = get_local_id(0);
     const int n = tid / 32;
     const int l = tid - 32 * n;
     const int is = 8 * n + l / 16;
 
     const uint8_t q = x[i].qs[32 * n + l];
-    __global float *y = yy + i * QK_K + 128 * n;
+    __global float *y = yy + get_group_id(0) * QK_K + 128 * n;
 
     const float dall = vload_half(0, &x[i].d);
     const float dmin = vload_half(0, &x[i].dmin);
@@ -223,7 +223,7 @@ __kernel void dequantize_block_q2_K(__global const struct block_q2_K *x, __globa
 __kernel void dequantize_block_q3_K(__global const struct block_q3_K *x, __global float *yy)
 {
     int r = get_local_id(0) / 4;
-    int i = get_group_id(0);
+    int i = get_group_id(0) + get_global_offset(0);
     int tid = r / 2;
     int is0 = r % 2;
     int l0 = 16 * is0 + 4 * (get_local_id(0) % 4);
@@ -241,7 +241,7 @@ __kernel void dequantize_block_q3_K(__global const struct block_q3_K *x, __globa
     float d_all = vload_half(0, &x[i].d);
     float dl = d_all * (us - 32);
 
-    __global float *y = yy + i * QK_K + 128 * n + 32 * j;
+    __global float *y = yy + get_group_id(0) * QK_K + 128 * n + 32 * j;
     const __global uint8_t *q = x[i].qs + 32 * n;
     const __global uint8_t *hm = x[i].hmask;
 
@@ -251,14 +251,14 @@ __kernel void dequantize_block_q3_K(__global const struct block_q3_K *x, __globa
 
 __kernel void dequantize_block_q4_K(__global const struct block_q4_K *x, __global float *yy)
 {
-    const int i = get_group_id(0);
+    const int i = get_group_id(0) + get_global_offset(0);
     const int tid = get_local_id(0);
     const int il = tid / 8;
     const int ir = tid % 8;
     const int is = 2 * il;
     const int n = 4;
 
-    __global float *y = yy + i * QK_K + 64 * il + n * ir;
+    __global float *y = yy + get_group_id(0) * QK_K + 64 * il + n * ir;
 
     const float dall = vload_half(0, &x[i].d);
     const float dmin = vload_half(0, &x[i].dmin);
@@ -281,13 +281,13 @@ __kernel void dequantize_block_q4_K(__global const struct block_q4_K *x, __globa
 
 __kernel void dequantize_block_q5_K(__global const struct block_q5_K *x, __global float *yy)
 {
-    const int i = get_group_id(0);
+    const int i = get_group_id(0) + get_global_offset(0);
     const int tid = get_local_id(0);
     const int il = tid / 16;
     const int ir = tid % 16;
     const int is = 2 * il;
 
-    __global float *y = yy + i * QK_K + 64 * il + 2 * ir;
+    __global float *y = yy + get_group_id(0) * QK_K + 64 * il + 2 * ir;
 
     const float dall = vload_half(0, &x[i].d);
     const float dmin = vload_half(0, &x[i].dmin);
@@ -313,13 +313,13 @@ __kernel void dequantize_block_q5_K(__global const struct block_q5_K *x, __globa
 
 __kernel void dequantize_block_q6_K(__global const struct block_q6_K *x, __global float *yy)
 {
-    const int i = get_group_id(0);
+    const int i = get_group_id(0) + get_global_offset(0);
     const int tid = get_local_id(0);
     const int ip = tid / 32;
     const int il = tid - 32 * ip;
     const int is = 8 * ip + il / 16;
 
-    __global float *y = yy + i * QK_K + 128 * ip + il;
+    __global float *y = yy + get_group_id(0) * QK_K + 128 * ip + il;
 
     const float d = vload_half(0, &x[i].d);
 
@@ -338,7 +338,7 @@ __kernel void dequantize_mul_mat_vec_q2_K(__global const struct block_q2_K * xx,
     const int row = get_group_id(0);
 
     const int num_blocks_per_row = ncols / QK_K;
-    const int ib0 = row*num_blocks_per_row;
+    const int ib0 = row*num_blocks_per_row + get_global_offset(0);
 
     __global const struct block_q2_K * x = xx + ib0;
 
@@ -413,7 +413,7 @@ __kernel void dequantize_mul_mat_vec_q3_K(__global const struct block_q3_K * xx,
     const int row = get_group_id(0);
 
     const int num_blocks_per_row = ncols / QK_K;
-    const int ib0 = row*num_blocks_per_row;
+    const int ib0 = row*num_blocks_per_row + get_global_offset(0);
 
     __global const struct block_q3_K * x = xx + ib0;
 
@@ -489,7 +489,7 @@ __kernel void dequantize_mul_mat_vec_q4_K(__global const struct block_q4_K * xx,
 
     const int row = get_group_id(0);
     const int num_blocks_per_row = ncols / QK_K;
-    const int ib0 = row*num_blocks_per_row;
+    const int ib0 = row*num_blocks_per_row + get_global_offset(0);
 
     const int tid = get_local_id(0)/K_QUANTS_PER_ITERATION;  // 0...15
     const int ix  = get_local_id(0)%K_QUANTS_PER_ITERATION;
@@ -562,7 +562,7 @@ __kernel void dequantize_mul_mat_vec_q5_K(__global const struct block_q5_K * xx,
 
     const int row = get_group_id(0);
     const int num_blocks_per_row = ncols / QK_K;
-    const int ib0 = row*num_blocks_per_row;
+    const int ib0 = row*num_blocks_per_row + get_global_offset(0);
 
     const int tid = get_local_id(0)/2;  // 0...15
     const int ix  = get_local_id(0)%2;
@@ -641,7 +641,7 @@ __kernel void dequantize_mul_mat_vec_q6_K(__global const struct block_q6_K * xx,
     const int row = get_group_id(0);
 
     const int num_blocks_per_row = ncols / QK_K;
-    const int ib0 = row*num_blocks_per_row;
+    const int ib0 = row*num_blocks_per_row + get_global_offset(0);
 
     __global const struct block_q6_K * x = xx + ib0;
 
@@ -730,7 +730,7 @@ __kernel void KERNEL_NAME(__global X_TYPE* x, __global float* y) {
     const uint qk = QUANT_K;
     const uint qr = QUANT_R;
 
-    const int ib = i/qk; // block index
+    const int ib = i/qk + get_global_offset(0); // block index
     const int iqs = (i%qk)/qr; // quant index
     const int iybs = i - i%qk; // y block start index
     const int y_offset = qr == 1 ? 1 : qk/2;
@@ -745,19 +745,21 @@ __kernel void KERNEL_NAME(__global X_TYPE* x, __global float* y) {
 
 std::string dequant_mul_mat_vec_template = MULTILINE_QUOTE(
 __kernel void KERNEL_NAME(__global X_TYPE* x, __local float* tmp, __global float* y, __global float* dst, const int ncols) {
-    const int block_size = get_local_size(0);
+    const int local_size = get_local_size(0);
     const int row = get_group_id(0);
     const int tid = get_local_id(0);
 
     const uint qk = QUANT_K;
     const uint qr = QUANT_R;
 
+    const int col_step = local_size * 2;
     const int y_offset = qr == 1 ? 1 : qk/2;
 
+    x += get_global_offset(0);
+
     tmp[tid] = 0;
 
-    for (int i = 0; i < ncols/block_size; i += 2) {
-        const int col = i*block_size + 2*tid;
+    for (int col = tid*2; col < ncols; col += col_step) {
         const int ib = (row*ncols + col)/qk; // block index
         const int iqs = (col%qk)/qr; // quant index
         const int iybs = col - col%qk; // y block start index
@@ -773,7 +775,7 @@ __kernel void KERNEL_NAME(__global X_TYPE* x, __local float* tmp, __global float
 
     // sum up partial sums and write back result
     barrier(CLK_LOCAL_MEM_FENCE);
-    for (int s=block_size/2; s>0; s>>=1) {
+    for (int s=local_size/2; s>0; s>>=1) {
         if (tid < s) {
             tmp[tid] += tmp[tid + s];
         }
@@ -847,7 +849,7 @@ std::array<std::string, 2> mul_str_values = {
     "mul_f32", "float"
 };
 
-std::string& replace(std::string& s, const std::string& from, const std::string& to) {
+static std::string& replace(std::string& s, const std::string& from, const std::string& to) {
     size_t pos = 0;
     while ((pos = s.find(from, pos)) != std::string::npos) {
          s.replace(pos, from.length(), to);
@@ -856,7 +858,7 @@ std::string& replace(std::string& s, const std::string& from, const std::string&
     return s;
 }
 
-std::string generate_kernels() {
+static std::string generate_kernels() {
     std::stringstream src;
     src << program_source << '\n';
     src << k_quants_source << '\n';
@@ -1349,30 +1351,42 @@ static cl_int ggml_cl_h2d_tensor_2d(cl_command_queue queue, cl_mem dst, size_t o
     const enum ggml_type type = src->type;
     const size_t ts = ggml_type_size(type);
     const size_t bs = ggml_blck_size(type);
+    const uint64_t row_size = ts*ne0/bs;
 
-    const void * x = (const void *) ((const char *) src->data + i2*nb2 + i3*nb3);
-    if (nb0 == ts && nb1 == ts*ne0/bs) {
-        err = clEnqueueWriteBuffer(queue, dst, CL_FALSE, offset, ne1*nb1, x, 0, NULL, ev);
-        return err;
+    const char * x = (const char *) src->data + i2*nb2 + i3*nb3;
+    if (nb0 == ts && nb1 == row_size) {
+        return clEnqueueWriteBuffer(queue, dst, CL_FALSE, offset, ne1*row_size, x, 0, NULL, ev);
     }
     if (nb0 == ts) {
         const size_t buffer_origin[3] = { offset, 0, 0 };
         const size_t host_origin[3] = { 0, 0, 0 };
-        const size_t region[3] = { ts*ne0/bs, ne1, 1 };
-        err = clEnqueueWriteBufferRect(queue, dst, CL_FALSE, buffer_origin, host_origin, region, ts*ne0/bs, 0, nb1, 0, x, 0, NULL, ev);
-        return err;
+        const size_t region[3] = { row_size, ne1, 1 };
+        return clEnqueueWriteBufferRect(queue, dst, CL_FALSE, buffer_origin, host_origin, region, row_size, 0, nb1, 0, x, 0, NULL, ev);
     }
+    std::vector<cl_event> events;
+    if (ev && ne1>1) events.reserve(ne1-1);
     for (uint64_t i1 = 0; i1 < ne1; i1++) {
         // pretend the row is a matrix with cols=1
-        const size_t buffer_origin[3] = { offset, i1, 0 };
+        const size_t buffer_origin[3] = { offset + i1*row_size, 0, 0 };
         const size_t host_origin[3] = { 0, 0, 0 };
-        const size_t region[3] = { ts/bs, ne0, 1 };
-        err = clEnqueueWriteBufferRect(queue, dst, CL_FALSE, buffer_origin, host_origin, region, 0, 0, nb0, 0, ((const char *)x) + i1*nb0, 0, NULL, ev);
+        const size_t region[3] = { ts, ne0/bs, 1 };
+        // if an event is requested, make the last write wait for all previous writes to complete
+        if (ev && i1) {
+            events.push_back(*ev);
+        }
+        cl_uint nevents = i1 == ne1-1 ? events.size() : 0U;
+        err = clEnqueueWriteBufferRect(queue, dst, CL_FALSE, buffer_origin, host_origin, region, ts, 0, nb0, 0, x + i1*nb1, nevents, nevents ? events.data() : nullptr, ev);
         if (err != CL_SUCCESS) {
-            break;
+            for (auto event : events) {
+                clReleaseEvent(event);
+            }
+            return err;
         }
     }
-    return err;
+    for (auto event : events) {
+        CL_CHECK(clReleaseEvent(event));
+    }
+    return CL_SUCCESS;
 }
 
 static void ggml_cl_mul_f32(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
@@ -1381,75 +1395,46 @@ static void ggml_cl_mul_f32(const ggml_tensor * src0, const ggml_tensor * src1,
     const int64_t ne01 = src0->ne[1];
     const int64_t ne02 = src0->ne[2];
     const int64_t ne03 = src0->ne[3];
-    const int64_t ne0 = ne00 * ne01 * ne02 * ne03;
     const int64_t ne10 = src1->ne[0];
     const int64_t ne11 = src1->ne[1];
     const int64_t ne12 = src1->ne[2];
     const int64_t ne13 = src1->ne[3];
-    const int64_t nb10 = src1->nb[0];
     const int nb2  = dst->nb[2];
     const int nb3  = dst->nb[3];
     size_t x_size;
     size_t d_size;
 
-    cl_mem d_X = ggml_cl_pool_malloc(ne0 * sizeof(float), &x_size); // src0
+    cl_mem d_X = ggml_cl_pool_malloc(ne00 * ne01 * sizeof(float), &x_size); // src0
     cl_mem d_Y = (cl_mem) src1->extra; // src1 is already on device, broadcasted.
-    cl_mem d_D = ggml_cl_pool_malloc(ne0 * sizeof(float), &d_size); // dst
+    cl_mem d_D = ggml_cl_pool_malloc(ne00 * ne01 * sizeof(float), &d_size); // dst
 
 
     for (int64_t i03 = 0; i03 < ne03; i03++) {
         for (int64_t i02 = 0; i02 < ne02; i02++) {
-            const int i0 = i03*ne02 + i02;
-
             cl_event ev;
 
             // copy src0 to device
-            CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_X, i0, src0, i03, i02, &ev));
-
-            if (nb10 == sizeof(float)) {
-                // Contiguous, avoid overhead from queueing many kernel runs
-                const int64_t i13 = i03%ne13;
-                const int64_t i12 = i02%ne12;
-                const int i1 = i13*ne12*ne11 + i12*ne11;
-
-                cl_int x_offset = 0;
-                cl_int y_offset = i1*ne10;
-                cl_int d_offset = 0;
-
-                size_t global = ne00 * ne01;
-                cl_int ky = ne10;
-                CL_CHECK(clSetKernelArg(mul_f32_cl, 0, sizeof(cl_mem), &d_X));
-                CL_CHECK(clSetKernelArg(mul_f32_cl, 1, sizeof(cl_int), &x_offset));
-                CL_CHECK(clSetKernelArg(mul_f32_cl, 2, sizeof(cl_mem), &d_Y));
-                CL_CHECK(clSetKernelArg(mul_f32_cl, 3, sizeof(cl_int), &y_offset));
-                CL_CHECK(clSetKernelArg(mul_f32_cl, 4, sizeof(cl_mem), &d_D));
-                CL_CHECK(clSetKernelArg(mul_f32_cl, 5, sizeof(cl_int), &d_offset));
-                CL_CHECK(clSetKernelArg(mul_f32_cl, 6, sizeof(cl_int), &ky));
-                CL_CHECK(clEnqueueNDRangeKernel(queue, mul_f32_cl, 1, NULL, &global, NULL, 1, &ev, NULL));
-            } else {
-                for (int64_t i01 = 0; i01 < ne01; i01++) {
-                    const int64_t i13 = i03%ne13;
-                    const int64_t i12 = i02%ne12;
-                    const int64_t i11 = i01%ne11;
-                    const int i1 = i13*ne12*ne11 + i12*ne11 + i11;
-
-                    cl_int x_offset = i01*ne00;
-                    cl_int y_offset = i1*ne10;
-                    cl_int d_offset = i01*ne00;
+            CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_X, 0, src0, i03, i02, &ev));
 
-                    // compute
-                    size_t global = ne00;
-                    cl_int ky = ne10;
-                    CL_CHECK(clSetKernelArg(mul_f32_cl, 0, sizeof(cl_mem), &d_X));
-                    CL_CHECK(clSetKernelArg(mul_f32_cl, 1, sizeof(cl_int), &x_offset));
-                    CL_CHECK(clSetKernelArg(mul_f32_cl, 2, sizeof(cl_mem), &d_Y));
-                    CL_CHECK(clSetKernelArg(mul_f32_cl, 3, sizeof(cl_int), &y_offset));
-                    CL_CHECK(clSetKernelArg(mul_f32_cl, 4, sizeof(cl_mem), &d_D));
-                    CL_CHECK(clSetKernelArg(mul_f32_cl, 5, sizeof(cl_int), &d_offset));
-                    CL_CHECK(clSetKernelArg(mul_f32_cl, 6, sizeof(cl_int), &ky));
-                    CL_CHECK(clEnqueueNDRangeKernel(queue, mul_f32_cl, 1, NULL, &global, NULL, 1, &ev, NULL));
-                }
-            }
+            const int64_t i13 = i03%ne13;
+            const int64_t i12 = i02%ne12;
+            const int i1 = i13*ne12*ne11 + i12*ne11;
+
+            cl_int x_offset = 0;
+            cl_int y_offset = i1*ne10;
+            cl_int d_offset = 0;
+
+            size_t global = ne00 * ne01;
+            cl_int ky = ne10 * ne11;
+
+            CL_CHECK(clSetKernelArg(mul_f32_cl, 0, sizeof(cl_mem), &d_X));
+            CL_CHECK(clSetKernelArg(mul_f32_cl, 1, sizeof(cl_int), &x_offset));
+            CL_CHECK(clSetKernelArg(mul_f32_cl, 2, sizeof(cl_mem), &d_Y));
+            CL_CHECK(clSetKernelArg(mul_f32_cl, 3, sizeof(cl_int), &y_offset));
+            CL_CHECK(clSetKernelArg(mul_f32_cl, 4, sizeof(cl_mem), &d_D));
+            CL_CHECK(clSetKernelArg(mul_f32_cl, 5, sizeof(cl_int), &d_offset));
+            CL_CHECK(clSetKernelArg(mul_f32_cl, 6, sizeof(cl_int), &ky));
+            CL_CHECK(clEnqueueNDRangeKernel(queue, mul_f32_cl, 1, NULL, &global, NULL, 1, &ev, NULL));
 
             CL_CHECK(clReleaseEvent(ev));
             CL_CHECK(clFinish(queue));
@@ -1476,10 +1461,15 @@ static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * sr
 
     const int64_t ne10 = src1->ne[0];
     const int64_t ne11 = src1->ne[1];
+    const int64_t ne12 = src1->ne[2];
+    const int64_t ne13 = src1->ne[3];
 
     const int nb2  = dst->nb[2];
     const int nb3  = dst->nb[3];
 
+    const int64_t r2 = ne12 / ne02;
+    const int64_t r3 = ne13 / ne03;
+
     const float alpha = 1.0f;
     const float beta = 0.0f;
     const int x_ne = ne01 * ne00;
@@ -1498,35 +1488,46 @@ static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * sr
     cl_mem d_Y = ggml_cl_pool_malloc(sizeof(float) * y_ne, &y_size);
     cl_mem d_D = ggml_cl_pool_malloc(sizeof(float) * d_ne, &d_size);
 
+    size_t x_offset = 0;
+
     for (int64_t i03 = 0; i03 < ne03; i03++) {
-        for (int64_t i02 = 0; i02 < ne02; i02++) {
-            // copy data to device
-            if (src0->backend != GGML_BACKEND_GPU) {
-                CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_X, 0, src0, i03, i02, NULL));
-            }
-            CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Y, 0, src1, i03, i02, NULL));
+        // TODO: copy src0 here when r3>1
+        for (int64_t i13 = i03 * r3, e13 = i13 + r3; i13 < e13; i13++) {
+            for (int64_t i02 = 0; i02 < ne02; i02++) {
+                if (src0->backend == GGML_BACKEND_GPU) {
+                    x_offset = (i03 * ne02 + i02) * x_ne;
+                } else {
+                    // copy src0 to device
+                    CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_X, 0, src0, i03, i02, NULL));
+                }
 
-            CL_CHECK(clFinish(queue));
+                for (int64_t i12 = i02 * r2, e12 = i12 + r2; i12 < e12; i12++) {
+                    // copy src1 to device
+                    CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Y, 0, src1, i13, i12, NULL));
 
-            // compute
-            cl_event ev_sgemm;
-            clblast::StatusCode status = clblast::Gemm<cl_float>(clblast::Layout::kColMajor,
-                                                       clblast::Transpose::kYes, clblast::Transpose::kNo,
-                                                       ne01, ne11, ne10,
-                                                       alpha,
-                                                       d_X, 0, ne00,
-                                                       d_Y, 0, ne10,
-                                                       beta,
-                                                       d_D, 0, ne01,
-                                                       &queue, &ev_sgemm);
-
-            if (status != clblast::StatusCode::kSuccess) {
-                GGML_ASSERT(false);
-            }
+                    CL_CHECK(clFinish(queue));
 
-            // copy dst to host
-            float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3);
-            CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(float) * d_ne, d, 1, &ev_sgemm, NULL));
+                    // compute
+                    cl_event ev_sgemm;
+                    clblast::StatusCode status = clblast::Gemm<cl_float>(clblast::Layout::kColMajor,
+                                                               clblast::Transpose::kYes, clblast::Transpose::kNo,
+                                                               ne01, ne11, ne10,
+                                                               alpha,
+                                                               d_X, x_offset, ne00,
+                                                               d_Y, 0, ne10,
+                                                               beta,
+                                                               d_D, 0, ne01,
+                                                               &queue, &ev_sgemm);
+
+                    if (status != clblast::StatusCode::kSuccess) {
+                        GGML_ASSERT(false);
+                    }
+
+                    // copy dst to host
+                    float * d = (float *) ((char *) dst->data + i12*nb2 + i13*nb3);
+                    CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(float) * d_ne, d, 1, &ev_sgemm, NULL));
+                }
+            }
         }
     }
 
@@ -1537,7 +1538,7 @@ static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * sr
     ggml_cl_pool_free(d_D, d_size);
 }
 
-static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, void * wdata, size_t /* wsize */) {
+static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, void * wdata, size_t wsize) {
     GGML_ASSERT(fp16_support);
 
     const int64_t ne00 = src0->ne[0];
@@ -1547,6 +1548,8 @@ static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * sr
 
     const int64_t ne10 = src1->ne[0];
     const int64_t ne11 = src1->ne[1];
+    const int64_t ne12 = src1->ne[2];
+    const int64_t ne13 = src1->ne[3];
 
     const int nb10 = src1->nb[0];
     const int nb11 = src1->nb[1];
@@ -1556,12 +1559,19 @@ static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * sr
     const int nb2  = dst->nb[2];
     const int nb3  = dst->nb[3];
 
+    const int64_t r2 = ne12 / ne02;
+    const int64_t r3 = ne13 / ne03;
+
     const ggml_fp16_t alpha = ggml_fp32_to_fp16(1.0f);
     const ggml_fp16_t beta = ggml_fp32_to_fp16(0.0f);
     const int x_ne = ne01 * ne00;
     const int y_ne = ne11 * ne10;
     const int d_ne = ne11 * ne01;
 
+    GGML_ASSERT(wsize >= sizeof(ggml_fp16_t) * y_ne);
+    GGML_ASSERT(wsize >= sizeof(ggml_fp16_t) * d_ne);
+    ggml_fp16_t * const tmp = (ggml_fp16_t *) wdata;
+
     size_t x_size;
     size_t y_size;
     size_t d_size;
@@ -1577,63 +1587,71 @@ static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * sr
     bool src1_cont_rows = nb10 == sizeof(float);
     bool src1_cont_cols = (size_t)nb11 == ne11*sizeof(float);
 
-    for (int64_t i03 = 0; i03 < ne03; i03++) {
-        for (int64_t i02 = 0; i02 < ne02; i02++) {
-            // copy src0 to device
-            if (src0->backend != GGML_BACKEND_GPU) {
-                CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_X, 0, src0, i03, i02, NULL));
-            }
+    size_t x_offset = 0;
 
-            // convert src1 to fp16
-            // TODO: use multiple threads
-            ggml_fp16_t * const tmp = (ggml_fp16_t *) wdata + (ne11 * ne10) * (i03 * ne02 + i02);
-            char * src1i = (char *) src1->data + i03*nb13 + i02*nb12;
-            if (src1_cont_rows) {
-                if (src1_cont_cols) {
-                    ggml_fp32_to_fp16_row((float *) src1i, tmp, ne10*ne11);
+    for (int64_t i03 = 0; i03 < ne03; i03++) {
+        // TODO: copy src0 here when r3>1
+        for (int64_t i13 = i03 * r3, e13 = i13 + r3; i13 < e13; i13++) {
+            for (int64_t i02 = 0; i02 < ne02; i02++) {
+                if (src0->backend == GGML_BACKEND_GPU) {
+                    x_offset = (i03 * ne02 + i02) * x_ne;
+                } else {
+                    // copy src0 to device
+                    CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_X, 0, src0, i03, i02, NULL));
                 }
-                else {
-                    for (int64_t i01 = 0; i01 < ne11; i01++) {
-                        ggml_fp32_to_fp16_row((float *) (src1i + i01*nb11), tmp + i01*ne10, ne10);
+
+                for (int64_t i12 = i02 * r2, e12 = i12 + r2; i12 < e12; i12++) {
+                    // convert src1 to fp16
+                    // TODO: use multiple threads
+                    char * src1i = (char *) src1->data + i13*nb13 + i12*nb12;
+                    if (src1_cont_rows) {
+                        if (src1_cont_cols) {
+                            ggml_fp32_to_fp16_row((float *) src1i, tmp, ne10*ne11);
+                        }
+                        else {
+                            for (int64_t i11 = 0; i11 < ne11; i11++) {
+                                ggml_fp32_to_fp16_row((float *) (src1i + i11*nb11), tmp + i11*ne10, ne10);
+                            }
+                        }
                     }
-                }
-            }
-            else {
-                for (int64_t i01 = 0; i01 < ne11; i01++) {
-                    for (int64_t i00 = 0; i00 < ne10; i00++) {
-                        // very slow due to no inlining
-                        tmp[i01*ne10 + i00] = ggml_fp32_to_fp16(*(float *) (src1i + i01*nb11 + i00*nb10));
+                    else {
+                        for (int64_t i11 = 0; i11 < ne11; i11++) {
+                            for (int64_t i10 = 0; i10 < ne10; i10++) {
+                                // very slow due to no inlining
+                                tmp[i11*ne10 + i10] = ggml_fp32_to_fp16(*(float *) (src1i + i11*nb11 + i10*nb10));
+                            }
+                        }
                     }
-                }
-            }
 
-            // copy src1 to device
-            CL_CHECK(clEnqueueWriteBuffer(queue, d_Y, false, 0, sizeof(ggml_fp16_t) * y_ne, tmp, 0, NULL, NULL));
+                    // copy src1 to device
+                    CL_CHECK(clEnqueueWriteBuffer(queue, d_Y, false, 0, sizeof(ggml_fp16_t) * y_ne, tmp, 0, NULL, NULL));
 
-            CL_CHECK(clFinish(queue));
+                    CL_CHECK(clFinish(queue));
 
-            // compute
-            cl_event ev_sgemm;
-            clblast::StatusCode status = clblast::Gemm<cl_half>(clblast::Layout::kColMajor,
-                                                       clblast::Transpose::kYes, clblast::Transpose::kNo,
-                                                       ne01, ne11, ne10,
-                                                       alpha,
-                                                       d_X, 0, ne00,
-                                                       d_Y, 0, ne10,
-                                                       beta,
-                                                       d_D, 0, ne01,
-                                                       &queue, &ev_sgemm);
-
-            if (status != clblast::StatusCode::kSuccess) {
-                GGML_ASSERT(false);
-            }
+                    // compute
+                    cl_event ev_sgemm;
+                    clblast::StatusCode status = clblast::Gemm<cl_half>(clblast::Layout::kColMajor,
+                                                               clblast::Transpose::kYes, clblast::Transpose::kNo,
+                                                               ne01, ne11, ne10,
+                                                               alpha,
+                                                               d_X, x_offset, ne00,
+                                                               d_Y, 0, ne10,
+                                                               beta,
+                                                               d_D, 0, ne01,
+                                                               &queue, &ev_sgemm);
+
+                    if (status != clblast::StatusCode::kSuccess) {
+                        GGML_ASSERT(false);
+                    }
 
-            // copy dst to host, then convert to float
-            CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(ggml_fp16_t) * d_ne, tmp, 1, &ev_sgemm, NULL));
+                    // copy dst to host, then convert to float
+                    CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(ggml_fp16_t) * d_ne, tmp, 1, &ev_sgemm, NULL));
 
-            float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3);
+                    float * d = (float *) ((char *) dst->data + i12*nb2 + i13*nb3);
 
-            ggml_fp16_to_fp32_row(tmp, d, d_ne);
+                    ggml_fp16_to_fp32_row(tmp, d, d_ne);
+                }
+            }
         }
     }
 
@@ -1652,18 +1670,24 @@ static void ggml_cl_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor *
 
     const int64_t ne10 = src1->ne[0];
     const int64_t ne11 = src1->ne[1];
+    const int64_t ne12 = src1->ne[2];
+    const int64_t ne13 = src1->ne[3];
 
     const int nb2  = dst->nb[2];
     const int nb3  = dst->nb[3];
     const ggml_type type = src0->type;
-    const bool mul_mat_vec = ne11 == 1;
+    const bool mul_mat_vec = ne11 == 1 && ne00%2 == 0;
+
+    const int64_t r2 = ne12 / ne02;
+    const int64_t r3 = ne13 / ne03;
 
     const float alpha = 1.0f;
     const float beta = 0.0f;
     const int x_ne = ne01 * ne00;
     const int y_ne = ne11 * ne10;
     const int d_ne = ne11 * ne01;
-    const size_t q_sz = ggml_type_size(type) * x_ne / ggml_blck_size(type);
+    const int x_bps = x_ne / ggml_blck_size(type); // blocks per 2D slice
+    const size_t q_sz = ggml_type_size(type) * x_bps;
 
     size_t x_size;
     size_t y_size;
@@ -1685,78 +1709,86 @@ static void ggml_cl_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor *
     GGML_ASSERT(to_fp32_cl != nullptr);
 
     const size_t global_denom = ggml_cl_global_denom(type);
-    const size_t local = ggml_cl_local_size(type);
+    const size_t local = mul_mat_vec ? CL_DMMV_LOCAL_SIZE : ggml_cl_local_size(type);
 
     size_t ev_idx = 0;
     std::vector<cl_event> events;
 
     for (int64_t i03 = 0; i03 < ne03; i03++) {
-        for (int64_t i02 = 0; i02 < ne02; i02++) {
-            // copy src0 to device if necessary
-            if (src0->backend == GGML_BACKEND_CPU) {
-                events.emplace_back();
-                CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Q, 0, src0, i03, i02, events.data() + ev_idx++));
-            } else if (src0->backend == GGML_BACKEND_GPU) {
-                d_Q = (cl_mem) src0->extra;
-            } else {
-                GGML_ASSERT(false);
-            }
-            if (mul_mat_vec) { // specialized dequantize_mul_mat_vec kernel
-                // copy src1 to device
-                events.emplace_back();
-                CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Y, 0, src1, i03, i02, events.data() + ev_idx++));
-
-                // compute
-                const size_t global = ne01 * CL_DMMV_BLOCK_SIZE;
-                const size_t local = CL_DMMV_BLOCK_SIZE;
-                const cl_int ncols = ne00;
-                events.emplace_back();
-                CL_CHECK(clSetKernelArg(*dmmv, 0, sizeof(cl_mem), &d_Q));
-                CL_CHECK(clSetKernelArg(*dmmv, 1, sizeof(float) * local, NULL));
-                CL_CHECK(clSetKernelArg(*dmmv, 2, sizeof(cl_mem), &d_Y));
-                CL_CHECK(clSetKernelArg(*dmmv, 3, sizeof(cl_mem), &d_D));
-                CL_CHECK(clSetKernelArg(*dmmv, 4, sizeof(cl_int), &ncols));
-                CL_CHECK(clEnqueueNDRangeKernel(queue, *dmmv, 1, NULL, &global, &local, events.size() - 1, events.data(), events.data() + ev_idx++));
-            } else { // general dequantization kernel + CLBlast matrix matrix multiplication
-                // convert src0 to fp32 on device
-                const size_t global = x_ne / global_denom;
-                CL_CHECK(clSetKernelArg(*to_fp32_cl, 0, sizeof(cl_mem), &d_Q));
-                CL_CHECK(clSetKernelArg(*to_fp32_cl, 1, sizeof(cl_mem), &d_X));
-                CL_CHECK(clEnqueueNDRangeKernel(queue, *to_fp32_cl, 1, NULL, &global, local > 0 ? &local : NULL, events.size(), !events.empty() ? events.data() : NULL, NULL));
-
-                // copy src1 to device
-                CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Y, 0, src1, i03, i02, NULL));
-
-                events.emplace_back();
-
-                // wait for conversion
-                CL_CHECK(clFinish(queue));
-
-                // compute
-                clblast::StatusCode status = clblast::Gemm<cl_float>(clblast::Layout::kColMajor,
-                                                           clblast::Transpose::kYes, clblast::Transpose::kNo,
-                                                           ne01, ne11, ne10,
-                                                           alpha,
-                                                           d_X, 0, ne00,
-                                                           d_Y, 0, ne10,
-                                                           beta,
-                                                           d_D, 0, ne01,
-                                                           &queue, events.data() + ev_idx++);
-
-                if (status != clblast::StatusCode::kSuccess) {
+        // TODO: copy and dequantize src0 here when r3>1
+        for (int64_t i13 = i03 * r3, e13 = i13 + r3; i13 < e13; i13++) {
+            for (int64_t i02 = 0; i02 < ne02; i02++) {
+                // copy src0 to device if necessary
+                if (src0->backend == GGML_BACKEND_CPU) {
+                    events.emplace_back();
+                    CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Q, 0, src0, i03, i02, events.data() + ev_idx++));
+                } else if (src0->backend == GGML_BACKEND_GPU) {
+                    d_Q = (cl_mem) src0->extra;
+                } else {
                     GGML_ASSERT(false);
                 }
-            }
 
-            // copy dst to host
-            float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3);
-            CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(float) * d_ne, d, 1, &events[events.size() - 1], NULL));
-            for (auto *event : events) {
-                clReleaseEvent(event);
-            }
+                if (!mul_mat_vec) {
+                    // convert src0 to fp32 on device
+                    const size_t global = x_ne / global_denom;
+                    const size_t offset = src0->backend == GGML_BACKEND_GPU ? (i03 * ne02 + i02) * x_bps : 0;
+                    CL_CHECK(clSetKernelArg(*to_fp32_cl, 0, sizeof(cl_mem), &d_Q));
+                    CL_CHECK(clSetKernelArg(*to_fp32_cl, 1, sizeof(cl_mem), &d_X));
+                    CL_CHECK(clEnqueueNDRangeKernel(queue, *to_fp32_cl, 1, &offset, &global, local > 0 ? &local : NULL, events.size(), !events.empty() ? events.data() : NULL, NULL));
+                }
+
+                for (int64_t i12 = i02 * r2, e12 = i12 + r2; i12 < e12; i12++) {
+                    if (mul_mat_vec) { // specialized dequantize_mul_mat_vec kernel
+                        // copy src1 to device
+                        events.emplace_back();
+                        CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Y, 0, src1, i13, i12, events.data() + ev_idx++));
+
+                        // compute
+                        const size_t global = ne01 * local;
+                        const size_t offset = src0->backend == GGML_BACKEND_GPU ? (i03 * ne02 + i02) * x_bps : 0;
+                        const cl_int ncols = ne00;
+                        events.emplace_back();
+                        CL_CHECK(clSetKernelArg(*dmmv, 0, sizeof(cl_mem), &d_Q));
+                        CL_CHECK(clSetKernelArg(*dmmv, 1, sizeof(float) * local, NULL));
+                        CL_CHECK(clSetKernelArg(*dmmv, 2, sizeof(cl_mem), &d_Y));
+                        CL_CHECK(clSetKernelArg(*dmmv, 3, sizeof(cl_mem), &d_D));
+                        CL_CHECK(clSetKernelArg(*dmmv, 4, sizeof(cl_int), &ncols));
+                        CL_CHECK(clEnqueueNDRangeKernel(queue, *dmmv, 1, &offset, &global, &local, events.size() - 1, events.data(), events.data() + ev_idx++));
+                    } else { // CLBlast matrix matrix multiplication
+                        // copy src1 to device
+                        CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Y, 0, src1, i13, i12, NULL));
+
+                        // wait for conversion
+                        CL_CHECK(clFinish(queue));
+
+                        // compute
+                        events.emplace_back();
+                        clblast::StatusCode status = clblast::Gemm<cl_float>(clblast::Layout::kColMajor,
+                                                                   clblast::Transpose::kYes, clblast::Transpose::kNo,
+                                                                   ne01, ne11, ne10,
+                                                                   alpha,
+                                                                   d_X, 0, ne00,
+                                                                   d_Y, 0, ne10,
+                                                                   beta,
+                                                                   d_D, 0, ne01,
+                                                                   &queue, events.data() + ev_idx++);
+
+                        if (status != clblast::StatusCode::kSuccess) {
+                            GGML_ASSERT(false);
+                        }
+                    }
 
-            ev_idx = 0;
-            events.clear();
+                    // copy dst to host
+                    float * d = (float *) ((char *) dst->data + i12*nb2 + i13*nb3);
+                    CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(float) * d_ne, d, 1, &events[events.size() - 1], NULL));
+                    for (auto *event : events) {
+                        clReleaseEvent(event);
+                    }
+
+                    ev_idx = 0;
+                    events.clear();
+                }
+            }
         }
     }
 
@@ -1788,7 +1820,7 @@ bool ggml_cl_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tens
     return false;
 }
 
-bool ggml_cl_mul_mat_use_f16(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * /* dst */) {
+static bool ggml_cl_mul_mat_use_f16(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * /* dst */) {
     // If device doesn't support FP16
     if (!fp16_support) {
         return false;
@@ -1831,8 +1863,8 @@ void ggml_cl_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor *
 }
 
 size_t ggml_cl_mul_mat_get_wsize(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) {
-    if (ggml_cl_mul_mat_use_f16(src0, src1, dst)) {
-        return ggml_nelements(src1) * sizeof(ggml_fp16_t);
+    if (src0->type == GGML_TYPE_F16 && ggml_cl_mul_mat_use_f16(src0, src1, dst)) {
+        return sizeof(ggml_fp16_t) * std::max(src1->ne[0] * src1->ne[1], dst->ne[0] * dst->ne[1]);
     }
     return 0;
 }
@@ -1844,17 +1876,19 @@ void ggml_cl_transform_tensor(void * data, ggml_tensor * tensor) {
     const int64_t ne3 = tensor->ne[3];
 
     const ggml_type type = tensor->type;
-    const size_t q_sz = ggml_type_size(type) * ne0 * ne1 * ne2 * ne3 / ggml_blck_size(type);
+    const size_t s_sz = ggml_type_size(type) * (size_t) (ne0 * ne1 / ggml_blck_size(type));
+    const size_t q_sz = s_sz * (size_t) (ne2 * ne3);
 
     size_t q_size;
     cl_mem dst = ggml_cl_pool_malloc(q_sz, &q_size);
 
     tensor->data = data;
     // copy tensor to device
+    size_t offset = 0;
     for (int64_t i3 = 0; i3 < ne3; i3++) {
         for (int64_t i2 = 0; i2 < ne2; i2++) {
-            int i = i3*ne2 + i2;
-            CL_CHECK(ggml_cl_h2d_tensor_2d(queue, dst, i*ne0*ne1, tensor, i3, i2, NULL));
+            CL_CHECK(ggml_cl_h2d_tensor_2d(queue, dst, offset, tensor, i3, i2, NULL));
+            offset += s_sz;
         }
     }
 

ggml-quants.h

@@ -0,0 +1,224 @@
+#pragma once
+
+#include "ggml-impl.h"
+
+// GGML internal header
+
+#include <stdint.h>
+#include <stddef.h>
+
+#define QK4_0 32
+typedef struct {
+    ggml_fp16_t d;          // delta
+    uint8_t qs[QK4_0 / 2];  // nibbles / quants
+} block_q4_0;
+static_assert(sizeof(block_q4_0) == sizeof(ggml_fp16_t) + QK4_0 / 2, "wrong q4_0 block size/padding");
+
+#define QK4_1 32
+typedef struct {
+    ggml_fp16_t d;          // delta
+    ggml_fp16_t m;          // min
+    uint8_t qs[QK4_1 / 2];  // nibbles / quants
+} block_q4_1;
+static_assert(sizeof(block_q4_1) == 2 * sizeof(ggml_fp16_t) + QK4_1 / 2, "wrong q4_1 block size/padding");
+
+#define QK5_0 32
+typedef struct {
+    ggml_fp16_t d;         // delta
+    uint8_t qh[4];         // 5-th bit of quants
+    uint8_t qs[QK5_0 / 2]; // nibbles / quants
+} block_q5_0;
+static_assert(sizeof(block_q5_0) == sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_0 / 2, "wrong q5_0 block size/padding");
+
+#define QK5_1 32
+typedef struct {
+    ggml_fp16_t d;         // delta
+    ggml_fp16_t m;         // min
+    uint8_t qh[4];         // 5-th bit of quants
+    uint8_t qs[QK5_1 / 2]; // nibbles / quants
+} block_q5_1;
+static_assert(sizeof(block_q5_1) == 2 * sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_1 / 2, "wrong q5_1 block size/padding");
+
+#define QK8_0 32
+typedef struct {
+    ggml_fp16_t d;         // delta
+    int8_t  qs[QK8_0];     // quants
+} block_q8_0;
+static_assert(sizeof(block_q8_0) == sizeof(ggml_fp16_t) + QK8_0, "wrong q8_0 block size/padding");
+
+#define QK8_1 32
+typedef struct {
+    float d;               // delta
+    float s;               // d * sum(qs[i])
+    int8_t  qs[QK8_1];     // quants
+} block_q8_1;
+static_assert(sizeof(block_q8_1) == 2*sizeof(float) + QK8_1, "wrong q8_1 block size/padding");
+
+//
+// Super-block quantization structures
+//
+
+// Super-block size
+#ifdef GGML_QKK_64
+#define QK_K 64
+#define K_SCALE_SIZE 4
+#else
+#define QK_K 256
+#define K_SCALE_SIZE 12
+#endif
+
+// 2-bit quantization
+// weight is represented as x = a * q + b
+// 16 blocks of 16 elements each
+// Effectively 2.5625 bits per weight
+typedef struct {
+    uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
+    uint8_t qs[QK_K/4];      // quants
+    ggml_fp16_t d;           // super-block scale for quantized scales
+    ggml_fp16_t dmin;        // super-block scale for quantized mins
+} block_q2_K;
+static_assert(sizeof(block_q2_K) == 2*sizeof(ggml_fp16_t) + QK_K/16 + QK_K/4, "wrong q2_K block size/padding");
+
+// 3-bit quantization
+// weight is represented as x = a * q
+// 16 blocks of 16 elements each
+// Effectively 3.4375 bits per weight
+#ifdef GGML_QKK_64
+typedef struct {
+    uint8_t hmask[QK_K/8];     // quants - high bit
+    uint8_t qs[QK_K/4];        // quants - low 2 bits
+    uint8_t scales[2];
+    ggml_fp16_t d;             // super-block scale
+} block_q3_K;
+static_assert(sizeof(block_q3_K) == sizeof(ggml_fp16_t) + QK_K / 4 + QK_K / 8 + 2, "wrong q3_K block size/padding");
+#else
+typedef struct {
+    uint8_t hmask[QK_K/8];     // quants - high bit
+    uint8_t qs[QK_K/4];        // quants - low 2 bits
+    uint8_t scales[12];        // scales, quantized with 6 bits
+    ggml_fp16_t d;             // super-block scale
+} block_q3_K;
+static_assert(sizeof(block_q3_K) == sizeof(ggml_fp16_t) + QK_K / 4 + QK_K / 8 + 12, "wrong q3_K block size/padding");
+#endif
+
+// 4-bit quantization
+// 8 blocks of 32 elements each
+// weight is represented as x = a * q + b
+// Effectively 4.5 bits per weight
+#ifdef GGML_QKK_64
+typedef struct {
+    ggml_fp16_t d[2];          // super-block scales/mins
+    uint8_t scales[2];         // 4-bit block scales/mins
+    uint8_t qs[QK_K/2];        // 4--bit quants
+} block_q4_K;
+static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_fp16_t) + QK_K/2 + 2, "wrong q4_K block size/padding");
+#else
+typedef struct {
+    ggml_fp16_t d;             // super-block scale for quantized scales
+    ggml_fp16_t dmin;          // super-block scale for quantized mins
+    uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
+    uint8_t qs[QK_K/2];        // 4--bit quants
+} block_q4_K;
+static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_fp16_t) + K_SCALE_SIZE + QK_K/2, "wrong q4_K block size/padding");
+#endif
+
+// 5-bit quantization
+// 8 blocks of 32 elements each
+// weight is represented as x = a * q + b
+// Effectively 5.5 bits per weight
+#ifdef GGML_QKK_64
+typedef struct {
+    ggml_fp16_t d;               // super-block scale
+    int8_t  scales[QK_K/16];     // 8-bit block scales
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+} block_q5_K;
+static_assert(sizeof(block_q5_K) == sizeof(ggml_fp16_t) + QK_K/2 + QK_K/8 + QK_K/16, "wrong q5_K block size/padding");
+#else
+typedef struct {
+    ggml_fp16_t d;               // super-block scale for quantized scales
+    ggml_fp16_t dmin;            // super-block scale for quantized mins
+    uint8_t scales[K_SCALE_SIZE];   // scales and mins, quantized with 6 bits
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+} block_q5_K;
+static_assert(sizeof(block_q5_K) == 2*sizeof(ggml_fp16_t) + K_SCALE_SIZE + QK_K/2 + QK_K/8, "wrong q5_K block size/padding");
+#endif
+
+// 6-bit quantization
+// weight is represented as x = a * q
+// 16 blocks of 16 elements each
+// Effectively 6.5625 bits per weight
+typedef struct {
+    uint8_t ql[QK_K/2];      // quants, lower 4 bits
+    uint8_t qh[QK_K/4];      // quants, upper 2 bits
+    int8_t  scales[QK_K/16]; // scales, quantized with 8 bits
+    ggml_fp16_t d;           // super-block scale
+} block_q6_K;
+static_assert(sizeof(block_q6_K) == sizeof(ggml_fp16_t) + QK_K / 16 + 3*QK_K/4, "wrong q6_K block size/padding");
+
+// This is only used for intermediate quantization and dot products
+typedef struct {
+    float   d;              // delta
+    int8_t  qs[QK_K];       // quants
+    int16_t bsums[QK_K/16]; // sum of quants in groups of 16
+} block_q8_K;
+static_assert(sizeof(block_q8_K) == sizeof(float) + QK_K + QK_K/16*sizeof(int16_t), "wrong q8_K block size/padding");
+
+
+// Quantization
+void quantize_row_q4_0_reference(const float * restrict x, block_q4_0 * restrict y, int k);
+void quantize_row_q4_1_reference(const float * restrict x, block_q4_1 * restrict y, int k);
+void quantize_row_q5_0_reference(const float * restrict x, block_q5_0 * restrict y, int k);
+void quantize_row_q5_1_reference(const float * restrict x, block_q5_1 * restrict y, int k);
+void quantize_row_q8_0_reference(const float * restrict x, block_q8_0 * restrict y, int k);
+void quantize_row_q8_1_reference(const float * restrict x, block_q8_1 * restrict y, int k);
+
+void quantize_row_q2_K_reference(const float * restrict x, block_q2_K * restrict y, int k);
+void quantize_row_q3_K_reference(const float * restrict x, block_q3_K * restrict y, int k);
+void quantize_row_q4_K_reference(const float * restrict x, block_q4_K * restrict y, int k);
+void quantize_row_q5_K_reference(const float * restrict x, block_q5_K * restrict y, int k);
+void quantize_row_q6_K_reference(const float * restrict x, block_q6_K * restrict y, int k);
+void quantize_row_q8_K_reference(const float * restrict x, block_q8_K * restrict y, int k);
+
+void quantize_row_q4_0(const float * restrict x, void * restrict y, int k);
+void quantize_row_q4_1(const float * restrict x, void * restrict y, int k);
+void quantize_row_q5_0(const float * restrict x, void * restrict y, int k);
+void quantize_row_q5_1(const float * restrict x, void * restrict y, int k);
+void quantize_row_q8_0(const float * restrict x, void * restrict y, int k);
+void quantize_row_q8_1(const float * restrict x, void * restrict y, int k);
+
+void quantize_row_q2_K(const float * restrict x, void * restrict y, int k);
+void quantize_row_q3_K(const float * restrict x, void * restrict y, int k);
+void quantize_row_q4_K(const float * restrict x, void * restrict y, int k);
+void quantize_row_q5_K(const float * restrict x, void * restrict y, int k);
+void quantize_row_q6_K(const float * restrict x, void * restrict y, int k);
+void quantize_row_q8_K(const float * restrict x, void * restrict y, int k);
+
+// Dequantization
+void dequantize_row_q4_0(const block_q4_0 * restrict x, float * restrict y, int k);
+void dequantize_row_q4_1(const block_q4_1 * restrict x, float * restrict y, int k);
+void dequantize_row_q5_0(const block_q5_0 * restrict x, float * restrict y, int k);
+void dequantize_row_q5_1(const block_q5_1 * restrict x, float * restrict y, int k);
+void dequantize_row_q8_0(const block_q8_0 * restrict x, float * restrict y, int k);
+//void dequantize_row_q8_1(const block_q8_1 * restrict x, float * restrict y, int k);
+
+void dequantize_row_q2_K(const block_q2_K * restrict x, float * restrict y, int k);
+void dequantize_row_q3_K(const block_q3_K * restrict x, float * restrict y, int k);
+void dequantize_row_q4_K(const block_q4_K * restrict x, float * restrict y, int k);
+void dequantize_row_q5_K(const block_q5_K * restrict x, float * restrict y, int k);
+void dequantize_row_q6_K(const block_q6_K * restrict x, float * restrict y, int k);
+void dequantize_row_q8_K(const block_q8_K * restrict x, float * restrict y, int k);
+
+// Dot product
+void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q5_0_q8_0(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+
+void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);

ggml.h

@@ -58,7 +58,8 @@
 //   {
 //       ...
 //
-//       struct ggml_cgraph gf = ggml_build_forward(f);
+//       struct ggml_cgraph * gf = ggml_new_graph(ctx);
+//       ggml_build_forward_expand(gf, f);
 //
 //       // set the input variable and parameter values
 //       ggml_set_f32(x, 2.0f);
@@ -213,15 +214,14 @@
 #define GGML_QNT_VERSION        2    // bump this on quantization format changes
 #define GGML_QNT_VERSION_FACTOR 1000 // do not change this
 
-#define GGML_MAX_DIMS          4
-#define GGML_MAX_NODES         4096
-#define GGML_MAX_PARAMS        256
-#define GGML_MAX_CONTEXTS      64
-#define GGML_MAX_SRC           6
-#define GGML_MAX_NAME          64
-#define GGML_MAX_OP_PARAMS     32
-#define GGML_DEFAULT_N_THREADS 4
-
+#define GGML_MAX_DIMS           4
+#define GGML_MAX_PARAMS         1024
+#define GGML_MAX_CONTEXTS       64
+#define GGML_MAX_SRC            6
+#define GGML_MAX_NAME           64
+#define GGML_MAX_OP_PARAMS      64
+#define GGML_DEFAULT_N_THREADS  4
+#define GGML_DEFAULT_GRAPH_SIZE 2048
 #if UINTPTR_MAX == 0xFFFFFFFF
     #define GGML_MEM_ALIGN 4
 #else
@@ -231,8 +231,9 @@
 #define GGML_EXIT_SUCCESS 0
 #define GGML_EXIT_ABORTED 1
 
-#define GGUF_MAGIC   0x46554747 // "GGUF"
-#define GGUF_VERSION 2
+#define GGUF_MAGIC "GGUF"
+
+#define GGUF_VERSION 3
 
 #define GGUF_DEFAULT_ALIGNMENT 32
 
@@ -244,10 +245,21 @@
     do { \
         if (!(x)) { \
             fprintf(stderr, "GGML_ASSERT: %s:%d: %s\n", __FILE__, __LINE__, #x); \
-            abort(); \
+            fflush(stderr); \
+            fflush(stdout); \
+            ggml_print_backtrace(); \
+            exit(1); \
         } \
     } while (0)
 
+#ifndef NDEBUG
+#define GGML_UNREACHABLE() GGML_ASSERT(!"statement should not be reached")
+#elif defined(__GNUC__)
+#define GGML_UNREACHABLE() __builtin_unreachable()
+#else
+#define GGML_UNREACHABLE() ((void) 0)
+#endif
+
 // used to copy the number of elements and stride in bytes of tensors into local variables.
 // main purpose is to reduce code duplication and improve readability.
 //
@@ -318,7 +330,7 @@ extern "C" {
         GGML_TYPE_COUNT,
     };
 
-    enum ggml_backend {
+    enum ggml_backend_type {
         GGML_BACKEND_CPU = 0,
         GGML_BACKEND_GPU = 10,
         GGML_BACKEND_GPU_SPLIT = 20,
@@ -392,7 +404,12 @@ extern "C" {
         GGML_OP_ALIBI,
         GGML_OP_CLAMP,
         GGML_OP_CONV_1D,
+        GGML_OP_CONV_1D_STAGE_0,  // internal
+        GGML_OP_CONV_1D_STAGE_1,  // internal
+        GGML_OP_CONV_TRANSPOSE_1D,
         GGML_OP_CONV_2D,
+        GGML_OP_CONV_2D_STAGE_0, // internal
+        GGML_OP_CONV_2D_STAGE_1, // internal
         GGML_OP_CONV_TRANSPOSE_2D,
         GGML_OP_POOL_1D,
         GGML_OP_POOL_2D,
@@ -437,6 +454,7 @@ extern "C" {
         GGML_UNARY_OP_GELU,
         GGML_UNARY_OP_GELU_QUICK,
         GGML_UNARY_OP_SILU,
+        GGML_UNARY_OP_LEAKY
     };
 
     enum ggml_object_type {
@@ -445,6 +463,12 @@ extern "C" {
         GGML_OBJECT_WORK_BUFFER
     };
 
+    enum ggml_log_level {
+        GGML_LOG_LEVEL_ERROR = 2,
+        GGML_LOG_LEVEL_WARN = 3,
+        GGML_LOG_LEVEL_INFO = 4
+    };
+
     // ggml object
     struct ggml_object {
         size_t offs;
@@ -461,14 +485,16 @@ extern "C" {
 
     // n-dimensional tensor
     struct ggml_tensor {
-        enum ggml_type    type;
-        enum ggml_backend backend;
+        enum ggml_type         type;
+        enum ggml_backend_type backend;
+
+        struct ggml_backend_buffer * buffer;
 
         int     n_dims;
         int64_t ne[GGML_MAX_DIMS]; // number of elements
         size_t  nb[GGML_MAX_DIMS]; // stride in bytes:
-                                   // nb[0] = sizeof(type)
-                                   // nb[1] = nb[0]   * ne[0] + padding
+                                   // nb[0] = ggml_type_size(type)
+                                   // nb[1] = nb[0]   * (ne[0] / ggml_blck_size(type)) + padding
                                    // nb[i] = nb[i-1] * ne[i-1]
 
         // compute data
@@ -496,7 +522,7 @@ extern "C" {
 
         void * extra; // extra things e.g. for ggml-cuda.cu
 
-        char padding[4];
+        char padding[12];
     };
 
     static const size_t GGML_TENSOR_SIZE = sizeof(struct ggml_tensor);
@@ -509,29 +535,35 @@ extern "C" {
 
         int n_threads;
 
-        // the `n_tasks` of nodes, 1:1 mapping to cgraph nodes
-        int n_tasks[GGML_MAX_NODES];
-
         // abort ggml_graph_compute when true
         bool (*abort_callback)(void * data);
         void * abort_callback_data;
     };
 
-    // next prime after GGML_MAX_NODES
-    // #define GGML_GRAPH_HASHTABLE_SIZE 4099
-    // next prime after GGML_MAX_NODES * 2 (nodes + leafs)
-    #define GGML_GRAPH_HASHTABLE_SIZE 8273
+    enum ggml_cgraph_eval_order {
+        GGML_CGRAPH_EVAL_ORDER_LEFT_TO_RIGHT = 0,
+        GGML_CGRAPH_EVAL_ORDER_RIGHT_TO_LEFT,
+        GGML_CGRAPH_EVAL_ORDER_COUNT
+    };
+
+    struct ggml_hash_set {
+        size_t size;
+        struct ggml_tensor ** keys;
+    };
 
     // computation graph
     struct ggml_cgraph {
+        int size;
         int n_nodes;
         int n_leafs;
 
-        struct ggml_tensor * nodes[GGML_MAX_NODES];
-        struct ggml_tensor * grads[GGML_MAX_NODES];
-        struct ggml_tensor * leafs[GGML_MAX_NODES];
+        struct ggml_tensor ** nodes;
+        struct ggml_tensor ** grads;
+        struct ggml_tensor ** leafs;
+
+        struct ggml_hash_set visited_hash_table;
 
-        void * visited_hash_table[GGML_GRAPH_HASHTABLE_SIZE];
+        enum ggml_cgraph_eval_order order;
 
         // performance
         int     perf_runs;
@@ -539,8 +571,6 @@ extern "C" {
         int64_t perf_time_us;
     };
 
-    static const size_t GGML_GRAPH_SIZE = sizeof(struct ggml_cgraph);
-
     // scratch buffer
     struct ggml_scratch {
         size_t offs;
@@ -585,6 +615,8 @@ extern "C" {
     GGML_API int64_t ggml_cycles(void);
     GGML_API int64_t ggml_cycles_per_ms(void);
 
+    GGML_API void    ggml_print_backtrace(void);
+
     GGML_API void    ggml_numa_init(void); // call once for better performance on NUMA systems
     GGML_API bool    ggml_is_numa(void); // true if init detected that system has >1 NUMA node
 
@@ -674,18 +706,30 @@ extern "C" {
     GGML_API struct ggml_tensor * ggml_dup_tensor (struct ggml_context * ctx, const struct ggml_tensor * src);
     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_next_tensor (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);
     GGML_API struct ggml_tensor * ggml_set_i32 (struct ggml_tensor * tensor, int32_t value);
     GGML_API struct ggml_tensor * ggml_set_f32 (struct ggml_tensor * tensor, float value);
 
+    // Converts a flat index into coordinates
+    GGML_API void    ggml_unravel_index(const struct ggml_tensor * tensor, int64_t i, int64_t * i0, int64_t * i1, int64_t * i2, int64_t * i3);
+
     GGML_API int32_t ggml_get_i32_1d(const struct ggml_tensor * tensor, int i);
     GGML_API void    ggml_set_i32_1d(const struct ggml_tensor * tensor, int i, int32_t value);
 
+    GGML_API int32_t ggml_get_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3);
+    GGML_API void    ggml_set_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3, int32_t value);
+
     GGML_API float   ggml_get_f32_1d(const struct ggml_tensor * tensor, int i);
     GGML_API void    ggml_set_f32_1d(const struct ggml_tensor * tensor, int i, float value);
 
+    GGML_API float   ggml_get_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3);
+    GGML_API void    ggml_set_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3, float value);
+
     GGML_API void *  ggml_get_data    (const struct ggml_tensor * tensor);
     GGML_API float * ggml_get_data_f32(const struct ggml_tensor * tensor);
 
@@ -719,6 +763,12 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
+    GGML_API struct ggml_tensor * ggml_add_cast(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            enum   ggml_type      type);
+
     GGML_API struct ggml_tensor * ggml_add1(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -828,6 +878,7 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
+    // sums repetitions in a into shape of b
     GGML_API struct ggml_tensor * ggml_repeat_back(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -892,6 +943,10 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    GGML_API struct ggml_tensor * ggml_leaky(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
     GGML_API struct ggml_tensor * ggml_relu_inplace(
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
@@ -970,9 +1025,9 @@ extern "C" {
             struct ggml_tensor  * b,
             float                 eps);
 
-    // A: n columns, m rows
-    // B: n columns, p rows  (i.e. we transpose it internally)
-    // result is m columns, p rows
+    // A: k columns, n rows => [ne03, ne02, n, k]
+    // B: k columns, m rows  (i.e. we transpose it internally) => [ne03 * x, ne02 * y, m, k]
+    // result is n columns, m rows => [ne03 * x, ne02 * y, m, n]
     GGML_API struct ggml_tensor * ggml_mul_mat(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -1049,7 +1104,6 @@ extern "C" {
             size_t                nb1,
             size_t                offset);
 
-
     // a -> b, return view(b)
     GGML_API struct ggml_tensor * ggml_cpy(
             struct ggml_context * ctx,
@@ -1072,6 +1126,33 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    // make contiguous, with new shape
+    GGML_API struct ggml_tensor * ggml_cont_1d(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int64_t               ne0);
+
+    GGML_API struct ggml_tensor * ggml_cont_2d(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int64_t               ne0,
+            int64_t               ne1);
+
+    GGML_API struct ggml_tensor * ggml_cont_3d(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int64_t               ne0,
+            int64_t               ne1,
+            int64_t               ne2);
+
+    GGML_API struct ggml_tensor * ggml_cont_4d(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int64_t               ne0,
+            int64_t               ne1,
+            int64_t               ne2,
+            int64_t               ne3);
+
     // return view(a), b specifies the new shape
     // TODO: when we start computing gradient, make a copy instead of view
     GGML_API struct ggml_tensor * ggml_reshape(
@@ -1219,14 +1300,15 @@ extern "C" {
             struct ggml_tensor  * b);
 
     // rotary position embedding
-    // if mode & 1 == 1, skip n_past elements
+    // if mode & 1 == 1, skip n_past elements (DEPRECATED)
     // if mode & 2 == 1, GPT-NeoX style
     // if mode & 4 == 1, ChatGLM style
-    // TODO: avoid creating a new tensor every time
+    //
+    // b is an int32 vector with size a->ne[2], it contains the positions
     GGML_API struct ggml_tensor * ggml_rope(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
-            int                   n_past,
+            struct ggml_tensor  * b,
             int                   n_dims,
             int                   mode,
             int                   n_ctx);
@@ -1235,7 +1317,7 @@ extern "C" {
     GGML_API struct ggml_tensor * ggml_rope_inplace(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
-            int                   n_past,
+            struct ggml_tensor  * b,
             int                   n_dims,
             int                   mode,
             int                   n_ctx);
@@ -1244,29 +1326,43 @@ extern "C" {
     GGML_API struct ggml_tensor * ggml_rope_custom(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
-            int                   n_past,
+            struct ggml_tensor  * b,
             int                   n_dims,
             int                   mode,
             int                   n_ctx,
+            int                   n_orig_ctx,
             float                 freq_base,
-            float                 freq_scale);
+            float                 freq_scale,
+            float                 ext_factor,
+            float                 attn_factor,
+            float                 beta_fast,
+            float                 beta_slow);
 
     // in-place, returns view(a)
     GGML_API struct ggml_tensor * ggml_rope_custom_inplace(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
-            int                   n_past,
+            struct ggml_tensor  * b,
             int                   n_dims,
             int                   mode,
             int                   n_ctx,
+            int                   n_orig_ctx,
             float                 freq_base,
-            float                 freq_scale);
+            float                 freq_scale,
+            float                 ext_factor,
+            float                 attn_factor,
+            float                 beta_fast,
+            float                 beta_slow);
+
+    // compute correction dims for YaRN RoPE scaling
+    void ggml_rope_yarn_corr_dims(
+        int n_dims, int n_orig_ctx, float freq_base, float beta_fast, float beta_slow, float dims[2]);
 
     // xPos RoPE, in-place, returns view(a)
     GGML_API struct ggml_tensor * ggml_rope_xpos_inplace(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
-            int                   n_past,
+            struct ggml_tensor  * b,
             int                   n_dims,
             float                 base,
             bool                  down);
@@ -1276,7 +1372,7 @@ extern "C" {
     GGML_API struct ggml_tensor * ggml_rope_back(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
-            int                   n_past,
+            struct ggml_tensor  * b,
             int                   n_dims,
             int                   mode,
             int                   n_ctx,
@@ -1287,7 +1383,7 @@ extern "C" {
 
     // alibi position embedding
     // in-place, returns view(a)
-    struct ggml_tensor * ggml_alibi(
+    GGML_API struct ggml_tensor * ggml_alibi(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             int                   n_past,
@@ -1296,7 +1392,7 @@ extern "C" {
 
     // clamp
     // in-place, returns view(a)
-    struct ggml_tensor * ggml_clamp(
+    GGML_API struct ggml_tensor * ggml_clamp(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             float                 min,
@@ -1319,6 +1415,14 @@ extern "C" {
             int                   s,
             int                   d);
 
+    GGML_API struct ggml_tensor * ggml_conv_transpose_1d(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            int                   s0,
+            int                   p0,
+            int                   d0);
+
     GGML_API struct ggml_tensor * ggml_conv_2d(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -1377,6 +1481,8 @@ extern "C" {
             int                   s0, // stride
             int                   p0); // padding
 
+    // the result will have 2*p0 padding for the first dimension
+    // and 2*p1 padding for the second dimension
     GGML_API struct ggml_tensor * ggml_pool_2d(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -1385,8 +1491,8 @@ extern "C" {
             int                   k1,
             int                   s0,
             int                   s1,
-            int                   p0,
-            int                   p1);
+            float                 p0,
+            float                 p1);
 
     // nearest interpolate
     // used in stable-diffusion
@@ -1627,19 +1733,22 @@ extern "C" {
     GGML_API void ggml_build_forward_expand (struct ggml_cgraph * cgraph, struct ggml_tensor * tensor);
     GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep);
 
-    GGML_API struct ggml_cgraph ggml_build_forward (struct ggml_tensor * tensor);
-    GGML_API struct ggml_cgraph ggml_build_backward(struct ggml_context * ctx, struct ggml_cgraph * gf, bool keep);
-
     // graph allocation in a context
-    GGML_API struct ggml_cgraph * ggml_new_graph        (struct ggml_context * ctx);
-    GGML_API struct ggml_cgraph * ggml_build_forward_ctx(struct ggml_context * ctx, struct ggml_tensor * tensor);
+    GGML_API struct ggml_cgraph * ggml_new_graph         (struct ggml_context * ctx); // size = GGML_DEFAULT_GRAPH_SIZE, grads = false
+    GGML_API struct ggml_cgraph * ggml_new_graph_custom  (struct ggml_context * ctx, size_t size, bool grads);
+    GGML_API struct ggml_cgraph * ggml_graph_dup         (struct ggml_context * ctx, struct ggml_cgraph * cgraph);
+    GGML_API struct ggml_cgraph * ggml_graph_view        (struct ggml_context * ctx, struct ggml_cgraph * cgraph, int i0, int i1);
+    GGML_API void                 ggml_graph_cpy         (struct ggml_cgraph * src, struct ggml_cgraph * dst);
+    GGML_API void                 ggml_graph_reset       (struct ggml_cgraph * cgraph);  // zero grads
+    GGML_API void                 ggml_graph_clear       (struct ggml_cgraph * cgraph);
+
     GGML_API size_t ggml_graph_overhead(void);
+    GGML_API size_t ggml_graph_overhead_custom(size_t size, bool grads);
 
     // ggml_graph_plan() has to be called before ggml_graph_compute()
     // when plan.work_size > 0, caller must allocate memory for plan.work_data
     GGML_API struct ggml_cplan ggml_graph_plan   (struct ggml_cgraph * cgraph, int n_threads /*= GGML_DEFAULT_N_THREADS*/);
-    GGML_API               int ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan);
-    GGML_API              void ggml_graph_reset  (struct ggml_cgraph * cgraph);
+    GGML_API int               ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan);
 
     // same as ggml_graph_compute() but the work data is allocated as a part of the context
     // note: the drawback of this API is that you must have ensured that the context has enough memory for the work data
@@ -1647,8 +1756,8 @@ extern "C" {
 
     GGML_API struct ggml_tensor * ggml_graph_get_tensor(struct ggml_cgraph * cgraph, const char * name);
 
-    GGML_API void               ggml_graph_export(const struct ggml_cgraph * cgraph, const char * fname);
-    GGML_API struct ggml_cgraph ggml_graph_import(const char * fname, struct ggml_context ** ctx_data, struct ggml_context ** ctx_eval);
+    GGML_API void                 ggml_graph_export(const struct ggml_cgraph * cgraph, const char * fname);
+    GGML_API struct ggml_cgraph * ggml_graph_import(const char * fname, struct ggml_context ** ctx_data, struct ggml_context ** ctx_eval);
 
     // print info and performance information for the graph
     GGML_API void ggml_graph_print(const struct ggml_cgraph * cgraph);
@@ -1656,6 +1765,16 @@ extern "C" {
     // dump the graph into a file using the dot format
     GGML_API void ggml_graph_dump_dot(const struct ggml_cgraph * gb, const struct ggml_cgraph * gf, const char * filename);
 
+    // build gradient checkpointing backward graph gb for gf using provided checkpoints
+    // gb_tmp will contain original backward graph with rewritten backward process nodes,
+    // but without the second forward pass nodes.
+    GGML_API void ggml_build_backward_gradient_checkpointing(
+            struct ggml_context   * ctx,
+            struct ggml_cgraph    * gf,
+            struct ggml_cgraph    * gb,
+            struct ggml_cgraph    * gb_tmp,
+            struct ggml_tensor  * * checkpoints,
+            int                     n_checkpoints);
     //
     // optimization
     //
@@ -1682,6 +1801,7 @@ extern "C" {
         GGML_OPT_NO_CONTEXT,
         GGML_OPT_INVALID_WOLFE,
         GGML_OPT_FAIL,
+        GGML_OPT_CANCEL,
 
         GGML_LINESEARCH_FAIL = -128,
         GGML_LINESEARCH_MINIMUM_STEP,
@@ -1690,7 +1810,8 @@ extern "C" {
         GGML_LINESEARCH_INVALID_PARAMETERS,
     };
 
-    typedef void (*ggml_opt_callback)(void * data, float * sched);
+    typedef void (*ggml_opt_callback)(void * data, int accum_step, float * sched, bool * cancel);
+    typedef void (*ggml_log_callback)(enum ggml_log_level level, const char * text, void * user_data);
 
     // optimization parameters
     //
@@ -1699,6 +1820,8 @@ extern "C" {
     struct ggml_opt_params {
         enum ggml_opt_type type;
 
+        size_t graph_size;
+
         int n_threads;
 
         // delta-based convergence test
@@ -1721,6 +1844,8 @@ extern "C" {
         bool print_forward_graph;
         bool print_backward_graph;
 
+        int n_gradient_accumulation;
+
         // ADAM parameters
         struct {
             int n_iter;
@@ -1766,6 +1891,7 @@ extern "C" {
         float loss_after;
 
         struct {
+            struct ggml_tensor * g;  // current gradient
             struct ggml_tensor * m;  // first moment
             struct ggml_tensor * v;  // second moment
             struct ggml_tensor * pf; // past function values
@@ -1829,12 +1955,19 @@ extern "C" {
     // quantization
     //
 
+    // TODO: these would probably get removed in favor of the more general ggml_quantize_chunk
     GGML_API size_t ggml_quantize_q4_0(const float * src, void * dst, int n, int k, int64_t * hist);
     GGML_API size_t ggml_quantize_q4_1(const float * src, void * dst, int n, int k, int64_t * hist);
     GGML_API size_t ggml_quantize_q5_0(const float * src, void * dst, int n, int k, int64_t * hist);
     GGML_API size_t ggml_quantize_q5_1(const float * src, void * dst, int n, int k, int64_t * hist);
     GGML_API size_t ggml_quantize_q8_0(const float * src, void * dst, int n, int k, int64_t * hist);
 
+    GGML_API size_t ggml_quantize_q2_K(const float * src, void * dst, int n, int k, int64_t * hist);
+    GGML_API size_t ggml_quantize_q3_K(const float * src, void * dst, int n, int k, int64_t * hist);
+    GGML_API size_t ggml_quantize_q4_K(const float * src, void * dst, int n, int k, int64_t * hist);
+    GGML_API size_t ggml_quantize_q5_K(const float * src, void * dst, int n, int k, int64_t * hist);
+    GGML_API size_t ggml_quantize_q6_K(const float * src, void * dst, int n, int k, int64_t * hist);
+
     GGML_API size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, int start, int n, int64_t * hist);
 
     //
@@ -1882,26 +2015,26 @@ extern "C" {
 
     GGML_API int          gguf_get_n_kv(const struct gguf_context * ctx);
     GGML_API int          gguf_find_key(const struct gguf_context * ctx, const char * key);
-    GGML_API const char * gguf_get_key (const struct gguf_context * ctx, int i);
-
-    GGML_API enum gguf_type gguf_get_kv_type (const struct gguf_context * ctx, int i);
-    GGML_API enum gguf_type gguf_get_arr_type(const struct gguf_context * ctx, int i);
-
-    // results are undefined if the wrong type is used for the key
-    GGML_API uint8_t      gguf_get_val_u8  (const struct gguf_context * ctx, int i);
-    GGML_API int8_t       gguf_get_val_i8  (const struct gguf_context * ctx, int i);
-    GGML_API uint16_t     gguf_get_val_u16 (const struct gguf_context * ctx, int i);
-    GGML_API int16_t      gguf_get_val_i16 (const struct gguf_context * ctx, int i);
-    GGML_API uint32_t     gguf_get_val_u32 (const struct gguf_context * ctx, int i);
-    GGML_API int32_t      gguf_get_val_i32 (const struct gguf_context * ctx, int i);
-    GGML_API float        gguf_get_val_f32 (const struct gguf_context * ctx, int i);
-    GGML_API uint64_t     gguf_get_val_u64 (const struct gguf_context * ctx, int i);
-    GGML_API int64_t      gguf_get_val_i64 (const struct gguf_context * ctx, int i);
-    GGML_API double       gguf_get_val_f64 (const struct gguf_context * ctx, int i);
-    GGML_API bool         gguf_get_val_bool(const struct gguf_context * ctx, int i);
-    GGML_API const char * gguf_get_val_str (const struct gguf_context * ctx, int i);
-    GGML_API int          gguf_get_arr_n   (const struct gguf_context * ctx, int i);
-    GGML_API const void * gguf_get_arr_data(const struct gguf_context * ctx, int i);
+    GGML_API const char * gguf_get_key (const struct gguf_context * ctx, int key_id);
+
+    GGML_API enum gguf_type gguf_get_kv_type (const struct gguf_context * ctx, int key_id);
+    GGML_API enum gguf_type gguf_get_arr_type(const struct gguf_context * ctx, int key_id);
+
+    // will abort if the wrong type is used for the key
+    GGML_API uint8_t      gguf_get_val_u8  (const struct gguf_context * ctx, int key_id);
+    GGML_API int8_t       gguf_get_val_i8  (const struct gguf_context * ctx, int key_id);
+    GGML_API uint16_t     gguf_get_val_u16 (const struct gguf_context * ctx, int key_id);
+    GGML_API int16_t      gguf_get_val_i16 (const struct gguf_context * ctx, int key_id);
+    GGML_API uint32_t     gguf_get_val_u32 (const struct gguf_context * ctx, int key_id);
+    GGML_API int32_t      gguf_get_val_i32 (const struct gguf_context * ctx, int key_id);
+    GGML_API float        gguf_get_val_f32 (const struct gguf_context * ctx, int key_id);
+    GGML_API uint64_t     gguf_get_val_u64 (const struct gguf_context * ctx, int key_id);
+    GGML_API int64_t      gguf_get_val_i64 (const struct gguf_context * ctx, int key_id);
+    GGML_API double       gguf_get_val_f64 (const struct gguf_context * ctx, int key_id);
+    GGML_API bool         gguf_get_val_bool(const struct gguf_context * ctx, int key_id);
+    GGML_API const char * gguf_get_val_str (const struct gguf_context * ctx, int key_id);
+    GGML_API int          gguf_get_arr_n   (const struct gguf_context * ctx, int key_id);
+    GGML_API const void * gguf_get_arr_data(const struct gguf_context * ctx, int key_id);
     GGML_API const char * gguf_get_arr_str (const struct gguf_context * ctx, int key_id, int i);
 
     GGML_API int    gguf_get_n_tensors    (const struct gguf_context * ctx);
@@ -2008,7 +2141,7 @@ extern "C" {
         enum ggml_type    vec_dot_type;
     } ggml_type_traits_t;
 
-    ggml_type_traits_t ggml_internal_get_type_traits(enum ggml_type type);
+    GGML_API ggml_type_traits_t ggml_internal_get_type_traits(enum ggml_type type);
 
 #ifdef  __cplusplus
 }

whisper.cpp

@@ -120,6 +120,7 @@ static void byteswap_tensor(ggml_tensor * tensor) {
 //#define WHISPER_USE_FLASH_ATTN
 //#define WHISPER_USE_FLASH_FF
 #define WHISPER_MAX_DECODERS 16
+#define WHISPER_MAX_NODES 4096
 
 //
 // ggml helpers
@@ -663,7 +664,7 @@ static void whisper_allocr_graph_init(struct whisper_allocr & allocr, std::funct
     auto & meta  = allocr.meta;
     auto & data  = allocr.data;
 
-    meta.resize(ggml_tensor_overhead()*GGML_MAX_NODES + ggml_graph_overhead());
+    meta.resize(ggml_tensor_overhead()*WHISPER_MAX_NODES + ggml_graph_overhead());
 
     alloc = ggml_allocr_new_measure(tensor_alignment);
 
@@ -1616,7 +1617,7 @@ static struct ggml_cgraph * whisper_build_graph_encoder(
 
     struct ggml_context * ctx0 = ggml_init(params);
 
-    ggml_cgraph * gf = ggml_new_graph(ctx0);
+    ggml_cgraph * gf = ggml_new_graph_custom(ctx0, WHISPER_MAX_NODES, false);
 
     ggml_allocr * alloc = wstate.alloc_encode.alloc;
 
@@ -2034,7 +2035,7 @@ static struct ggml_cgraph * whisper_build_graph_decoder(
 
     struct ggml_context * ctx0 = ggml_init(params);
 
-    ggml_cgraph * gf = ggml_new_graph(ctx0);
+    ggml_cgraph * gf = ggml_new_graph_custom(ctx0, WHISPER_MAX_NODES, false);
 
     ggml_allocr * alloc = wstate.alloc_decode.alloc;
 
@@ -3773,9 +3774,6 @@ struct whisper_full_params whisper_full_default_params(enum whisper_sampling_str
         /*.encoder_begin_callback           =*/ nullptr,
         /*.encoder_begin_callback_user_data =*/ nullptr,
 
-        /*.abort_callback           =*/ nullptr,
-        /*.abort_callback_user_data =*/ nullptr,
-
         /*.logits_filter_callback           =*/ nullptr,
         /*.logits_filter_callback_user_data =*/ nullptr,
     };
@@ -4535,7 +4533,7 @@ int whisper_full_with_state(
 
         // initial prompt
         if (!params.prompt_tokens && params.initial_prompt) {
-            prompt_tokens.resize(2048);
+            prompt_tokens.resize(1024);
             prompt_tokens.resize(whisper_tokenize(ctx, params.initial_prompt, prompt_tokens.data(), prompt_tokens.size()));
             params.prompt_tokens   = prompt_tokens.data();
             params.prompt_n_tokens = prompt_tokens.size();
@@ -5432,7 +5430,7 @@ WHISPER_API const char * whisper_bench_ggml_mul_mat_str(int n_threads) {
     // b: N*N*sizeof(float)
     // c: N*N*sizeof(float)
     // when F16 is used, there is an extra work buffer of size N*N*sizeof(float)
-    std::vector<uint8_t> buf(3llu*N_max*N_max*sizeof(float) + 3*ggml_tensor_overhead());
+    std::vector<uint8_t> buf(3llu*N_max*N_max*sizeof(float) + 3*ggml_tensor_overhead() + ggml_graph_overhead());
     std::vector<uint8_t> work;
 
     // put a bunch of random data in the buffer
@@ -5483,17 +5481,19 @@ WHISPER_API const char * whisper_bench_ggml_mul_mat_str(int n_threads) {
 
             struct ggml_tensor * c = ggml_mul_mat(ctx0, a, b);
 
-            struct ggml_cgraph gf = ggml_build_forward(c);
+            struct ggml_cgraph * gf = ggml_new_graph(ctx0);
+
+            ggml_build_forward_expand(gf, c);
 
             double tsum = 0.0;
 
             // heat-up
-            ggml_graph_compute_helper(work, &gf, n_threads, nullptr , nullptr);
+            ggml_graph_compute_helper(work, gf, n_threads, nullptr, nullptr);
 
             for (int i = 0; i < n_max; ++i) {
                 const int64_t t0 = ggml_time_us();
 
-                ggml_graph_compute_helper(work, &gf, n_threads, nullptr, nullptr);
+                ggml_graph_compute_helper(work, gf, n_threads, nullptr, nullptr);
 
                 const int64_t t1 = ggml_time_us();