sync : ggml (backend v2) (#3912)
* sync : ggml (backend v2) (wip) * sync : migrate examples and llama.cpp to dynamic graphs (wip) * sync : update tests + fix max op params to 64 ggml-ci * sync : ggml-cuda ggml-ci * llama : fix save/load state context size ggml-ci * sync : try to fix build on tvOS * sync : pass custom graph sizes in training examples * sync : update graph copies to new ggml API * sync : update sync-ggml.sh with new files * scripts : fix header in sync script * train : fix context size calculations * llama : increase inference graph size up to 4096 nodes * train : allocate grads for backward graphs * train : allocate grads for gb_tmp
This commit is contained in:
Georgi Gerganov
GitHub
parent
bb50a792ec
commit
4760e7cc0b
22 changed files with 1994 additions and 864 deletions
591
ggml-backend.c
591
ggml-backend.c
|
|
@ -1,7 +1,9 @@
|
|||
#include "ggml-backend.h"
|
||||
#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>
|
||||
|
|
@ -33,6 +35,10 @@ ggml_backend_buffer_t ggml_backend_buffer_init(
|
|||
}
|
||||
|
||||
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);
|
||||
}
|
||||
|
|
@ -43,15 +49,20 @@ size_t ggml_backend_buffer_get_alignment(ggml_backend_buffer_t buffer) {
|
|||
return ggml_backend_get_alignment(buffer->backend);
|
||||
}
|
||||
|
||||
void * ggml_backend_buffer_get_base(ggml_backend_buffer_t buffer) {
|
||||
return buffer->iface.get_base(buffer);
|
||||
}
|
||||
|
||||
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);
|
||||
}
|
||||
|
|
@ -59,12 +70,14 @@ size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct g
|
|||
}
|
||||
|
||||
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);
|
||||
}
|
||||
|
|
@ -73,14 +86,21 @@ void ggml_backend_buffer_free_tensor(ggml_backend_buffer_t buffer, struct ggml_t
|
|||
// backend
|
||||
|
||||
ggml_backend_t ggml_get_backend(const struct ggml_tensor * tensor) {
|
||||
return tensor->buffer->backend;
|
||||
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);
|
||||
}
|
||||
|
||||
|
|
@ -101,13 +121,23 @@ void ggml_backend_tensor_get_async(const struct ggml_tensor * tensor, void * dat
|
|||
}
|
||||
|
||||
void ggml_backend_tensor_set(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);
|
||||
ggml_get_backend(tensor)->iface.synchronize(ggml_get_backend(tensor));
|
||||
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_get_backend(tensor)->iface.get_tensor_async(ggml_get_backend(tensor), tensor, data, offset, size);
|
||||
ggml_get_backend(tensor)->iface.synchronize(ggml_get_backend(tensor));
|
||||
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) {
|
||||
|
|
@ -156,7 +186,7 @@ void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst
|
|||
//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");
|
||||
|
||||
// printf("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));
|
||||
// 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;
|
||||
|
|
@ -234,6 +264,8 @@ static ggml_backend_buffer_t ggml_backend_cpu_alloc_buffer(ggml_backend_t backen
|
|||
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);
|
||||
}
|
||||
|
||||
|
|
@ -271,8 +303,7 @@ static void ggml_backend_cpu_cpy_tensor_from(ggml_backend_t backend, struct ggml
|
|||
}
|
||||
|
||||
static void ggml_backend_cpu_cpy_tensor_to(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst) {
|
||||
// for a backend such as CUDA that can queue async calls, it is ok to do this asynchronously, but it may not be the case for other backends
|
||||
ggml_backend_tensor_set_async(dst, src->data, 0, ggml_nbytes(src));
|
||||
ggml_backend_tensor_set(dst, src->data, 0, ggml_nbytes(src));
|
||||
|
||||
UNUSED(backend);
|
||||
}
|
||||
|
|
@ -383,3 +414,537 @@ void ggml_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int 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];
|
||||
}
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue