extern/whisper.cpp
1
0
Fork 1
mirror of https://github.com/ggerganov/whisper.cpp.git synced 2025-08-14 16:38:44 +02:00
Code Issues Packages Projects Releases Wiki Activity

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

Browse Source 
* 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 <ggerganov@gmail.com>

* 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 <codesoda@users.noreply.github.com>
This commit is contained in:
Georgi Gerganov
2023-11-03 21:35:05 +02:00
committed by GitHub
parent 8a2bee6717
commit f96e1c5b78
38 changed files with 30875 additions and 7843 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
bindings/ruby/ext/extconf.rb
View File

@ -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')} .")

87
bindings/ruby/ext/ggml-backend-impl.h Normal file
View File

@ -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

950
bindings/ruby/ext/ggml-backend.c Normal file
View File

@ -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];
}

136
bindings/ruby/ext/ggml-backend.h Normal file
View File

@ -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

249
bindings/ruby/ext/ggml-impl.h Normal file
View File

@ -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

7282
bindings/ruby/ext/ggml-quants.c Normal file
View File

File diff suppressed because it is too large Load Diff

224
bindings/ruby/ext/ggml-quants.h Normal file
View File

@ -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);