vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

use async copy and compute to improve multi-gpu performance

Browse source 
ggml-ci
This commit is contained in:
slaren 2024-01-12 00:50:00 +01:00
parent c4867196b4
commit 23c14ef53e
6 changed files with 201 additions and 167 deletions
Download patch file Download diff file Expand all files Collapse all files

5
ggml-alloc.c
View file

@ -102,8 +102,6 @@ void ggml_tallocr_alloc(ggml_tallocr_t alloc, struct ggml_tensor * tensor) {
} }
} }
AT_PRINTF("block %d\n", best_fit_block);
if (best_fit_block == -1) { if (best_fit_block == -1) {
// the last block is our last resort // the last block is our last resort
struct free_block * block = &alloc->free_blocks[alloc->n_free_blocks - 1]; struct free_block * block = &alloc->free_blocks[alloc->n_free_blocks - 1];
@ -117,6 +115,7 @@ void ggml_tallocr_alloc(ggml_tallocr_t alloc, struct ggml_tensor * tensor) {
return; return;
} }
} }
struct free_block * block = &alloc->free_blocks[best_fit_block]; struct free_block * block = &alloc->free_blocks[best_fit_block];
void * addr = block->addr; void * addr = block->addr;
block->addr = (char*)block->addr + size; block->addr = (char*)block->addr + size;
@ -129,6 +128,8 @@ void ggml_tallocr_alloc(ggml_tallocr_t alloc, struct ggml_tensor * tensor) {
} }
} }
AT_PRINTF("block %d, addr %p\n", best_fit_block, addr);
tensor->data = addr; tensor->data = addr;
tensor->buffer = alloc->buffer; tensor->buffer = alloc->buffer;
if (!alloc->measure) { if (!alloc->measure) {

33
ggml-backend-impl.h
View file

@ -35,17 +35,15 @@ extern "C" {
typedef void * ggml_backend_buffer_context_t; typedef void * ggml_backend_buffer_context_t;
struct ggml_backend_buffer_i { struct ggml_backend_buffer_i {
const char * (*get_name) (ggml_backend_buffer_t buffer); const char * (*get_name) (ggml_backend_buffer_t buffer);
void (*free_buffer) (ggml_backend_buffer_t buffer); void (*free_buffer)(ggml_backend_buffer_t buffer);
void * (*get_base) (ggml_backend_buffer_t buffer); void * (*get_base) (ggml_backend_buffer_t buffer);
void (*init_tensor) (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); void (*init_tensor)(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
void (*set_tensor) (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size); void (*set_tensor) (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
void (*get_tensor) (ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size); void (*get_tensor) (ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size);
// (optional) copy tensor between different buffer-type, allow for single-copy tranfers bool (*cpy_tensor) (ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst); // dst is in the buffer, src may be in any buffer
void (*cpy_tensor_from)(ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst); void (*clear) (ggml_backend_buffer_t buffer, uint8_t value);
void (*cpy_tensor_to) (ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst); void (*reset) (ggml_backend_buffer_t buffer); // reset any internal state due to tensor initialization, such as tensor extras
void (*clear) (ggml_backend_buffer_t buffer, uint8_t value);
void (*reset) (ggml_backend_buffer_t buffer); // reset any internal state due to tensor initialization, such as tensor extras
}; };
struct ggml_backend_buffer { struct ggml_backend_buffer {
@ -62,6 +60,8 @@ extern "C" {
ggml_backend_buffer_context_t context, ggml_backend_buffer_context_t context,
size_t size); size_t size);
// do not use directly, use ggml_backend_tensor_copy instead
bool ggml_backend_buffer_copy_tensor(const struct ggml_tensor * src, struct ggml_tensor * dst);
// //
// Backend // Backend
@ -77,14 +77,12 @@ extern "C" {
// buffer allocation // buffer allocation
ggml_backend_buffer_type_t (*get_default_buffer_type)(ggml_backend_t backend); ggml_backend_buffer_type_t (*get_default_buffer_type)(ggml_backend_t backend);
// (optional) asynchroneous tensor data access // (optional) asynchronous tensor data access
void (*set_tensor_async)(ggml_backend_t backend, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size); 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 (*get_tensor_async)(ggml_backend_t backend, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size);
bool (*cpy_tensor_async)(ggml_backend_t backend, const struct ggml_tensor * src, struct ggml_tensor * dst);
// (optional) asynchroneous tensor copy // (optional) complete all pending operations
void (*cpy_tensor_from_async)(ggml_backend_t backend, const struct ggml_tensor * src, struct ggml_tensor * dst);
void (*cpy_tensor_to_async) (ggml_backend_t backend, const struct ggml_tensor * src, struct ggml_tensor * dst);
void (*synchronize)(ggml_backend_t backend); void (*synchronize)(ggml_backend_t backend);
// compute graph with a plan // compute graph with a plan
@ -92,7 +90,7 @@ extern "C" {
void (*graph_plan_free) (ggml_backend_t backend, ggml_backend_graph_plan_t plan); 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); void (*graph_plan_compute)(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
// compute graph without a plan // compute graph without a plan (async)
bool (*graph_compute)(ggml_backend_t backend, struct ggml_cgraph * cgraph); bool (*graph_compute)(ggml_backend_t backend, struct ggml_cgraph * cgraph);
// check if the backend supports an operation // check if the backend supports an operation
@ -105,7 +103,6 @@ extern "C" {
ggml_backend_context_t context; ggml_backend_context_t context;
}; };
// //
// Backend registry // Backend registry
// //

255
ggml-backend.c
View file

@ -132,6 +132,14 @@ void ggml_backend_buffer_reset(ggml_backend_buffer_t buffer) {
} }
} }
bool ggml_backend_buffer_copy_tensor(const struct ggml_tensor * src, struct ggml_tensor * dst) {
ggml_backend_buffer_t dst_buf = dst->view_src ? dst->view_src->buffer : dst->buffer;
if (dst_buf->iface.cpy_tensor) {
return src->buffer->iface.cpy_tensor(dst_buf, src, dst);
}
return false;
}
// backend // backend
const char * ggml_backend_name(ggml_backend_t backend) { const char * ggml_backend_name(ggml_backend_t backend) {
@ -165,30 +173,42 @@ void ggml_backend_tensor_set_async(ggml_backend_t backend, struct ggml_tensor *
GGML_ASSERT(tensor->data != NULL && "tensor not allocated"); GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor write out of bounds"); GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor write out of bounds");
backend->iface.set_tensor_async(backend, tensor, data, offset, size); if (backend->iface.set_tensor_async == NULL) {
ggml_backend_tensor_set(tensor, data, offset, size);
} else {
backend->iface.set_tensor_async(backend, tensor, data, offset, size);
}
} }
void ggml_backend_tensor_get_async(ggml_backend_t backend, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) { void ggml_backend_tensor_get_async(ggml_backend_t backend, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
GGML_ASSERT(tensor->data != NULL && "tensor not allocated"); GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor read out of bounds"); GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor read out of bounds");
backend->iface.get_tensor_async(backend, tensor, data, offset, size); if (backend->iface.get_tensor_async == NULL) {
ggml_backend_tensor_get(tensor, data, offset, size);
} else {
backend->iface.get_tensor_async(backend, tensor, data, offset, size);
}
} }
void ggml_backend_tensor_set(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) { void ggml_backend_tensor_set(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
GGML_ASSERT(tensor->data != NULL && "tensor not allocated"); GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
GGML_ASSERT(tensor->buffer != NULL && "tensor buffer not set"); GGML_ASSERT(buf != NULL && "tensor buffer not set");
GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor write out of bounds"); GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor write out of bounds");
tensor->buffer->iface.set_tensor(tensor->buffer, tensor, data, offset, size); tensor->buffer->iface.set_tensor(buf, tensor, data, offset, size);
} }
void ggml_backend_tensor_get(const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) { void ggml_backend_tensor_get(const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
GGML_ASSERT(tensor->data != NULL && "tensor not allocated"); GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
GGML_ASSERT(tensor->buffer != NULL && "tensor buffer not set"); GGML_ASSERT(tensor->buffer != NULL && "tensor buffer not set");
GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor read out of bounds"); GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor read out of bounds");
tensor->buffer->iface.get_tensor(tensor->buffer, tensor, data, offset, size); tensor->buffer->iface.get_tensor(buf, tensor, data, offset, size);
} }
void ggml_backend_synchronize(ggml_backend_t backend) { void ggml_backend_synchronize(ggml_backend_t backend) {
@ -209,19 +229,10 @@ void ggml_backend_graph_plan_free(ggml_backend_t backend, ggml_backend_graph_pla
void ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) { void ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
backend->iface.graph_plan_compute(backend, plan); backend->iface.graph_plan_compute(backend, plan);
// TODO: optional sync
ggml_backend_synchronize(backend);
} }
bool ggml_backend_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) { bool ggml_backend_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
if (!backend->iface.graph_compute(backend, cgraph)) { return backend->iface.graph_compute(backend, cgraph);
return false;
}
// TODO: optional sync
ggml_backend_synchronize(backend);
return true;
} }
bool ggml_backend_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) { bool ggml_backend_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
@ -246,28 +257,20 @@ static bool ggml_are_same_layout(const struct ggml_tensor * a, const struct ggml
} }
void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst) { 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"); 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) { if (src == dst) {
return; return;
} }
// TODO: allow backends to support copy to/from same backend if (ggml_backend_buffer_is_host(src->buffer)) {
ggml_backend_tensor_set(dst, src->data, 0, ggml_nbytes(src));
if (dst->buffer->iface.cpy_tensor_from != NULL) { } else if (ggml_backend_buffer_is_host(dst->buffer)) {
dst->buffer->iface.cpy_tensor_from(dst->buffer, src, dst); ggml_backend_tensor_get(src, dst->data, 0, ggml_nbytes(src));
} else if (src->buffer->iface.cpy_tensor_to != NULL) { } else if (!ggml_backend_buffer_copy_tensor(src, dst)) {
src->buffer->iface.cpy_tensor_to(src->buffer, src, dst); #ifndef NDEBUG
} else { fprintf(stderr, "%s: warning: slow copy from %s to %s\n", __func__, ggml_backend_buffer_name(src->buffer), ggml_backend_buffer_name(dst->buffer));
// shouldn't be hit when copying from/to CPU #endif
#ifndef NDEBUG
fprintf(stderr, "ggml_backend_tensor_copy: neither cpy_tensor_from nor cpy_tensor_to "
"are implemented for %s and %s, falling back to get/set\n", src->name, dst->name);
#endif
size_t nbytes = ggml_nbytes(src); size_t nbytes = ggml_nbytes(src);
void * data = malloc(nbytes); void * data = malloc(nbytes);
ggml_backend_tensor_get(src, data, 0, nbytes); ggml_backend_tensor_get(src, data, 0, nbytes);
@ -276,6 +279,31 @@ void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst
} }
} }
void ggml_backend_tensor_copy_async(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst) {
GGML_ASSERT(ggml_are_same_layout(src, dst) && "cannot copy tensors with different layouts");
if (src == dst) {
return;
}
if (ggml_backend_buft_supports_backend(src->buffer->buft, backend) && ggml_backend_buft_supports_backend(dst->buffer->buft, backend)) {
if (backend->iface.cpy_tensor_async != NULL) {
if (backend->iface.cpy_tensor_async(backend, src, dst)) {
return;
}
}
}
size_t nbytes = ggml_nbytes(src);
if (ggml_backend_buffer_is_host(src->buffer)) {
ggml_backend_tensor_set_async(backend, dst, src->data, 0, nbytes);
}
else {
ggml_backend_tensor_copy(src, dst);
}
}
// backend registry // backend registry
#define GGML_MAX_BACKENDS_REG 16 #define GGML_MAX_BACKENDS_REG 16
@ -437,14 +465,12 @@ static void ggml_backend_cpu_buffer_get_tensor(ggml_backend_buffer_t buffer, con
GGML_UNUSED(buffer); GGML_UNUSED(buffer);
} }
static void ggml_backend_cpu_buffer_cpy_tensor_from(ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst) { static bool ggml_backend_cpu_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst) {
ggml_backend_tensor_get(src, dst->data, 0, ggml_nbytes(src)); if (ggml_backend_buffer_is_host(src->buffer)) {
memcpy(dst->data, src->data, ggml_nbytes(src));
GGML_UNUSED(buffer); return true;
} }
return false;
static void ggml_backend_cpu_buffer_cpy_tensor_to(ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst) {
ggml_backend_tensor_set(dst, src->data, 0, ggml_nbytes(src));
GGML_UNUSED(buffer); GGML_UNUSED(buffer);
} }
@ -460,8 +486,7 @@ static struct ggml_backend_buffer_i cpu_backend_buffer_i = {
/* .init_tensor = */ NULL, // no initialization required /* .init_tensor = */ NULL, // no initialization required
/* .set_tensor = */ ggml_backend_cpu_buffer_set_tensor, /* .set_tensor = */ ggml_backend_cpu_buffer_set_tensor,
/* .get_tensor = */ ggml_backend_cpu_buffer_get_tensor, /* .get_tensor = */ ggml_backend_cpu_buffer_get_tensor,
/* .cpy_tensor_from = */ ggml_backend_cpu_buffer_cpy_tensor_from, /* .cpy_tensor = */ ggml_backend_cpu_buffer_cpy_tensor,
/* .cpy_tensor_to = */ ggml_backend_cpu_buffer_cpy_tensor_to,
/* .clear = */ ggml_backend_cpu_buffer_clear, /* .clear = */ ggml_backend_cpu_buffer_clear,
/* .reset = */ NULL, /* .reset = */ NULL,
}; };
@ -474,8 +499,7 @@ static struct ggml_backend_buffer_i cpu_backend_buffer_i_from_ptr = {
/* .init_tensor = */ NULL, // no initialization required /* .init_tensor = */ NULL, // no initialization required
/* .set_tensor = */ ggml_backend_cpu_buffer_set_tensor, /* .set_tensor = */ ggml_backend_cpu_buffer_set_tensor,
/* .get_tensor = */ ggml_backend_cpu_buffer_get_tensor, /* .get_tensor = */ ggml_backend_cpu_buffer_get_tensor,
/* .cpy_tensor_from = */ ggml_backend_cpu_buffer_cpy_tensor_from, /* .cpy_tensor = */ ggml_backend_cpu_buffer_cpy_tensor,
/* .cpy_tensor_to = */ ggml_backend_cpu_buffer_cpy_tensor_to,
/* .clear = */ ggml_backend_cpu_buffer_clear, /* .clear = */ ggml_backend_cpu_buffer_clear,
/* .reset = */ NULL, /* .reset = */ NULL,
}; };
@ -683,8 +707,7 @@ static struct ggml_backend_i cpu_backend_i = {
/* .get_default_buffer_type = */ ggml_backend_cpu_get_default_buffer_type, /* .get_default_buffer_type = */ ggml_backend_cpu_get_default_buffer_type,
/* .set_tensor_async = */ NULL, /* .set_tensor_async = */ NULL,
/* .get_tensor_async = */ NULL, /* .get_tensor_async = */ NULL,
/* .cpy_tensor_from_async = */ NULL, /* .cpy_tensor_async = */ NULL,
/* .cpy_tensor_to_async = */ NULL,
/* .synchronize = */ NULL, /* .synchronize = */ NULL,
/* .graph_plan_create = */ ggml_backend_cpu_graph_plan_create, /* .graph_plan_create = */ ggml_backend_cpu_graph_plan_create,
/* .graph_plan_free = */ ggml_backend_cpu_graph_plan_free, /* .graph_plan_free = */ ggml_backend_cpu_graph_plan_free,
@ -748,12 +771,6 @@ struct ggml_backend_sched_split {
struct ggml_cgraph graph; struct ggml_cgraph graph;
}; };
// TODO: group all the hash values into a single struct for clarity
//struct sched_hash_value {
// ggml_tallocr_t tallocr;
// struct ggml_tensor * copies[GGML_MAX_BACKENDS];
//};
struct ggml_backend_sched { struct ggml_backend_sched {
int n_backends; int n_backends;
ggml_backend_t backends[GGML_MAX_BACKENDS]; ggml_backend_t backends[GGML_MAX_BACKENDS];
@ -810,14 +827,22 @@ static int sched_allocr_prio(ggml_backend_sched_t sched, ggml_tallocr_t allocr)
return INT_MAX; return INT_MAX;
} }
static ggml_backend_t get_buffer_backend(ggml_backend_sched_t sched, ggml_backend_buffer_t buffer) { static ggml_tallocr_t sched_allocr_from_buffer(ggml_backend_sched_t sched, ggml_backend_buffer_t buffer) {
if (buffer == NULL) { if (buffer == NULL) {
return NULL; return NULL;
} }
// check if this is already allocate in a allocr buffer (from user manual allocations)
for (int i = 0; i < sched->n_backends; i++) {
if (ggml_tallocr_get_buffer(sched->tallocs[i]) == buffer) {
return sched->tallocs[i];
}
}
// find highest prio backend that supports the buffer type // find highest prio backend that supports the buffer type
for (int i = 0; i < sched->n_backends; i++) { for (int i = 0; i < sched->n_backends; i++) {
if (ggml_backend_buft_supports_backend(buffer->buft, sched->backends[i])) { if (ggml_backend_buft_supports_backend(buffer->buft, sched->backends[i])) {
return sched->backends[i]; return sched->tallocs[i];
} }
} }
GGML_ASSERT(false && "tensor buffer type not supported by any backend"); GGML_ASSERT(false && "tensor buffer type not supported by any backend");
@ -827,7 +852,6 @@ static ggml_backend_t get_allocr_backend(ggml_backend_sched_t sched, ggml_talloc
if (allocr == NULL) { if (allocr == NULL) {
return NULL; return NULL;
} }
// find highest prio backend that supports the buffer type
for (int i = 0; i < sched->n_backends; i++) { for (int i = 0; i < sched->n_backends; i++) {
if (sched->tallocs[i] == allocr) { if (sched->tallocs[i] == allocr) {
return sched->backends[i]; return sched->backends[i];
@ -837,7 +861,7 @@ static ggml_backend_t get_allocr_backend(ggml_backend_sched_t sched, ggml_talloc
} }
#if 0 #if 0
static char causes[GGML_DEFAULT_GRAPH_SIZE*16 + GGML_MAX_SPLITS*GGML_MAX_SPLIT_INPUTS][128]; // debug, remove static char causes[GGML_DEFAULT_GRAPH_SIZE*16 + GGML_MAX_SPLITS*GGML_MAX_SPLIT_INPUTS][128]; // debug only
#define SET_CAUSE(node, ...) sprintf(causes[hash_id(node)], __VA_ARGS__) #define SET_CAUSE(node, ...) sprintf(causes[hash_id(node)], __VA_ARGS__)
#define GET_CAUSE(node) causes[hash_id(node)] #define GET_CAUSE(node) causes[hash_id(node)]
#else #else
@ -846,49 +870,37 @@ static char causes[GGML_DEFAULT_GRAPH_SIZE*16 + GGML_MAX_SPLITS*GGML_MAX_SPLIT_I
#endif #endif
// returns the backend that should be used for the node based on the current locations // returns the backend that should be used for the node based on the current locations
static ggml_backend_t sched_backend_from_cur(ggml_backend_sched_t sched, struct ggml_tensor * node) { static ggml_tallocr_t sched_allocr_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 // assign pre-allocated nodes to their backend
// 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 // dst
ggml_backend_t cur_backend = get_buffer_backend(sched, node->buffer); ggml_tallocr_t cur_allocr = sched_allocr_from_buffer(sched, node->buffer);
if (cur_backend != NULL) { if (cur_allocr != NULL) {
SET_CAUSE(node, "1.dst"); SET_CAUSE(node, "1.dst");
return cur_backend; return cur_allocr;
} }
// view_src // view_src
if (node->view_src != NULL && get_buffer_backend(sched, node->view_src->buffer) != NULL) { if (node->view_src != NULL) {
SET_CAUSE(node, "1.vsrc"); cur_allocr = sched_allocr_from_buffer(sched, node->view_src->buffer);
return get_buffer_backend(sched, node->view_src->buffer); if (cur_allocr != NULL) {
SET_CAUSE(node, "1.vsrc");
return cur_allocr;
}
} }
// assign nodes that use weights to the backend of the weights
// src
size_t cur_size = 0;
for (int i = 0; i < GGML_MAX_SRC; i++) { for (int i = 0; i < GGML_MAX_SRC; i++) {
const struct ggml_tensor * src = node->src[i]; const struct ggml_tensor * src = node->src[i];
if (src == NULL) { if (src == NULL) {
break; break;
} }
ggml_backend_t src_backend = get_buffer_backend(sched, src->buffer);
if (src->buffer != NULL && src->buffer->usage == GGML_BACKEND_BUFFER_USAGE_WEIGHTS) { if (src->buffer != NULL && src->buffer->usage == GGML_BACKEND_BUFFER_USAGE_WEIGHTS) {
// operations with weights are always on the same backend as the weights ggml_tallocr_t src_allocr = sched_allocr_from_buffer(sched, src->buffer);
cur_backend = src_backend; // operations with weights are always run on the same backend as the weights
SET_CAUSE(node, "1.wgt%d", i); SET_CAUSE(node, "1.wgt%d", i);
break; return src_allocr;
}
size_t src_size = ggml_nbytes(src);
if (src_size >= cur_size) {
cur_size = src_size;
cur_backend = src_backend;
SET_CAUSE(node, "1.src%d", i);
} }
} }
return cur_backend;
return NULL;
} }
static char * fmt_size(size_t size) { static char * fmt_size(size_t size) {
@ -953,7 +965,6 @@ static struct ggml_tensor * ggml_dup_tensor_layout(struct ggml_context * ctx, co
//#define DEBUG_PASS4 //#define DEBUG_PASS4
// assigns backends to ops and splits the graph into subgraphs that can be computed on the same backend // 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) { static void sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
// reset splits // reset splits
sched->n_splits = 0; sched->n_splits = 0;
@ -972,20 +983,14 @@ static void sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgraph * g
GGML_ASSERT(false); GGML_ASSERT(false);
} }
// pass 1: assign backends to ops with allocated inputs // pass 1: assign backends to ops with pre-allocated inputs
for (int i = 0; i < graph->n_leafs; i++) { for (int i = 0; i < graph->n_leafs; i++) {
struct ggml_tensor * leaf = graph->leafs[i]; struct ggml_tensor * leaf = graph->leafs[i];
if (node_allocr(leaf) != NULL) { if (node_allocr(leaf) != NULL) {
// do not overwrite user assignments // do not overwrite user assignments
continue; continue;
} }
ggml_backend_t leaf_backend = get_buffer_backend(sched, leaf->buffer); node_allocr(leaf) = sched_allocr_from_cur(sched, leaf);
if (leaf_backend == NULL && leaf->view_src != NULL) {
leaf_backend = get_buffer_backend(sched, leaf->view_src->buffer);
}
if (leaf_backend != NULL) {
node_allocr(leaf) = ggml_backend_sched_get_tallocr(sched, leaf_backend);
}
} }
for (int i = 0; i < graph->n_nodes; i++) { for (int i = 0; i < graph->n_nodes; i++) {
@ -994,18 +999,24 @@ static void sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgraph * g
// do not overwrite user assignments // do not overwrite user assignments
continue; continue;
} }
ggml_backend_t node_backend = sched_backend_from_cur(sched, node); node_allocr(node) = sched_allocr_from_cur(sched, node);
if (node_backend != NULL) { // src
node_allocr(node) = ggml_backend_sched_get_tallocr(sched, node_backend); for (int j = 0; j < GGML_MAX_SRC; j++) {
struct ggml_tensor * src = node->src[j];
if (src == NULL) {
break;
}
if (node_allocr(src) == NULL) {
node_allocr(src) = sched_allocr_from_cur(sched, src);
}
} }
} }
#ifdef DEBUG_PASS1 #ifdef DEBUG_PASS1
fprintf(stderr, "PASS 1 ASSIGNMENTS\n"); sched_print_assignments(sched, graph); fprintf(stderr, "PASS 1 ASSIGNMENTS\n"); sched_print_assignments(sched, graph);
#endif #endif
// pass 2: assign backends to ops from current assignments // pass 2: expand current backend assignments
// start from the end and assign the same backend to previous ops // assign the same backend to adjacent nodes
// expand gpu backends (i.e. non last prio) up and down, ignoring cpu (the lowest priority backend) // expand gpu backends (i.e. non last prio) up and down, ignoring cpu (the lowest priority backend)
// thus, cpu will never be used unless weights are on cpu, or there are no gpu ops between cpu ops // thus, cpu will never be used unless weights are on cpu, or there are no gpu ops between cpu ops
@ -1027,7 +1038,7 @@ static void sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgraph * g
} }
} else { } else {
node_allocr(node) = cur_allocr; node_allocr(node) = cur_allocr;
SET_CAUSE(node, "2.cur"); SET_CAUSE(node, "2.1");
} }
} }
} }
@ -1050,7 +1061,7 @@ static void sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgraph * g
} }
} else { } else {
node_allocr(node) = cur_allocr; node_allocr(node) = cur_allocr;
SET_CAUSE(node, "2.cur"); SET_CAUSE(node, "2.2");
} }
} }
} }
@ -1068,7 +1079,7 @@ static void sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgraph * g
cur_allocr = node_allocr; cur_allocr = node_allocr;
} else { } else {
node_allocr(node) = cur_allocr; node_allocr(node) = cur_allocr;
SET_CAUSE(node, "2.cur"); SET_CAUSE(node, "2.3");
} }
} }
} }
@ -1080,7 +1091,7 @@ static void sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgraph * g
for (int i = 0; i < graph->n_nodes; i++) { for (int i = 0; i < graph->n_nodes; i++) {
struct ggml_tensor * node = graph->nodes[i]; struct ggml_tensor * node = graph->nodes[i];
ggml_tallocr_t cur_allocr = node_allocr(node); ggml_tallocr_t cur_allocr = node_allocr(node);
if (ggml_is_view_op(node->op) && cur_allocr == NULL) { if (node->view_src != NULL && cur_allocr == NULL) {
cur_allocr = node_allocr(node) = node_allocr(node->view_src); cur_allocr = node_allocr(node) = node_allocr(node->view_src);
SET_CAUSE(node, "3.vsrc"); SET_CAUSE(node, "3.vsrc");
} }
@ -1094,8 +1105,10 @@ static void sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgraph * g
if (src->view_src != NULL) { if (src->view_src != NULL) {
// views are always on the same backend as the source // views are always on the same backend as the source
node_allocr(src) = node_allocr(src->view_src); node_allocr(src) = node_allocr(src->view_src);
SET_CAUSE(src, "3.vsrc");
} else { } else {
node_allocr(src) = cur_allocr; node_allocr(src) = cur_allocr;
SET_CAUSE(src, "3.cur");
} }
} }
} }
@ -1136,7 +1149,6 @@ static void sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgraph * g
sched->splits[cur_split].tallocr = node_allocr; sched->splits[cur_split].tallocr = node_allocr;
sched->splits[cur_split].i_start = i; sched->splits[cur_split].i_start = i;
sched->splits[cur_split].n_inputs = 0; 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_allocr = node_allocr;
cur_backend_id = sched_allocr_prio(sched, cur_allocr); cur_backend_id = sched_allocr_prio(sched, cur_allocr);
} }
@ -1148,6 +1160,7 @@ static void sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgraph * g
break; break;
} }
ggml_tallocr_t src_allocr = node_allocr(src); ggml_tallocr_t src_allocr = node_allocr(src);
GGML_ASSERT(src_allocr != NULL); // all inputs should be assigned by now
if (src_allocr != node_allocr) { if (src_allocr != node_allocr) {
// check if the input is already in the split // check if the input is already in the split
bool found = false; bool found = false;
@ -1162,17 +1175,19 @@ static void sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgraph * g
int n_inputs = sched->splits[cur_split].n_inputs++; int n_inputs = sched->splits[cur_split].n_inputs++;
//printf("split %d input %d: %s (%s)\n", cur_split, n_inputs, src->name, ggml_backend_name(get_allocr_backend(sched, src_allocr))); //printf("split %d input %d: %s (%s)\n", cur_split, n_inputs, src->name, ggml_backend_name(get_allocr_backend(sched, src_allocr)));
GGML_ASSERT(n_inputs < GGML_MAX_SPLIT_INPUTS); GGML_ASSERT(n_inputs < GGML_MAX_SPLIT_INPUTS);
sched->splits[cur_split].inputs[n_inputs] = (struct ggml_tensor *)src; sched->splits[cur_split].inputs[n_inputs] = src;
} }
// create a copy of the input in the split's backend // create a copy of the input in the split's backend
size_t id = hash_id(src); size_t id = hash_id(src);
if (sched->node_copies[id][cur_backend_id] == NULL) { if (sched->node_copies[id][cur_backend_id] == NULL) {
ggml_backend_t backend = get_allocr_backend(sched, cur_allocr);
struct ggml_tensor * tensor_copy = ggml_dup_tensor_layout(sched->ctx, src); struct ggml_tensor * tensor_copy = ggml_dup_tensor_layout(sched->ctx, src);
ggml_format_name(tensor_copy, "%s#%s", ggml_backend_name(backend), src->name);
sched->node_copies[id][cur_backend_id] = tensor_copy; sched->node_copies[id][cur_backend_id] = tensor_copy;
node_allocr(tensor_copy) = cur_allocr; node_allocr(tensor_copy) = cur_allocr;
ggml_backend_t backend = get_allocr_backend(sched, cur_allocr); SET_CAUSE(tensor_copy, "4.cpy");
ggml_format_name(tensor_copy, "%s#%s", ggml_backend_name(backend), src->name);
} }
node->src[j] = sched->node_copies[id][cur_backend_id]; node->src[j] = sched->node_copies[id][cur_backend_id];
} }
@ -1231,6 +1246,7 @@ static void sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgraph * g
struct ggml_tensor * input = split->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)]; struct ggml_tensor * input_cpy = sched->node_copies[hash_id(input)][sched_allocr_prio(sched, split->tallocr)];
// add a dependency to the input source so that it is not freed before the copy is done // add a dependency to the input source so that it is not freed before the copy is done
GGML_ASSERT(input_cpy->src[0] == NULL || input_cpy->src[0] == input);
input_cpy->src[0] = input; input_cpy->src[0] = input;
graph_copy->nodes[graph_copy->n_nodes++] = input_cpy; graph_copy->nodes[graph_copy->n_nodes++] = input_cpy;
} }
@ -1265,25 +1281,16 @@ static void sched_compute_splits(ggml_backend_sched_t sched) {
uint64_t copy_start_us = ggml_time_us(); uint64_t copy_start_us = ggml_time_us();
for (int j = 0; j < split->n_inputs; j++) { for (int j = 0; j < split->n_inputs; j++) {
struct ggml_tensor * input = split->inputs[j]; struct ggml_tensor * input = split->inputs[j];
struct ggml_tensor * input_cpy = sched->node_copies[hash_id(input)][sched_backend_prio(sched, split_backend)]; struct ggml_tensor * input_cpy = sched->node_copies[hash_id(input)][split_backend_id];
if (input->buffer == NULL) {
GGML_ASSERT(false); GGML_ASSERT(input->buffer != NULL);
if (input->view_src == NULL) { GGML_ASSERT(input_cpy->buffer != NULL);
fprintf(stderr, "input %s has no buffer and no view_src\n", input->name);
GGML_ASSERT(false);
}
// FIXME: may need to use the sched buffer instead
ggml_backend_view_init(input->view_src->buffer, input);
}
if (input_cpy->buffer == NULL) {
fprintf(stderr, "input_cpy %s has no buffer\n", input_cpy->name);
GGML_ASSERT(false);
}
// TODO: avoid this copy if it was already copied in a previous split, and the input didn't change // TODO: avoid this copy if it was already copied in a previous split, and the input didn't change
// this is important to avoid copying constants such as KQ_mask and inp_pos multiple times // this is important to avoid copying constants such as KQ_mask and inp_pos multiple times
ggml_backend_tensor_copy(input, input_cpy); ggml_backend_tensor_copy_async(split_backend, input, input_cpy);
} }
// ggml_backend_synchronize(split_backend); //ggml_backend_synchronize(split_backend); // necessary to measure copy time
int64_t copy_end_us = ggml_time_us(); int64_t copy_end_us = ggml_time_us();
copy_us[split_backend_id] += copy_end_us - copy_start_us; copy_us[split_backend_id] += copy_end_us - copy_start_us;
@ -1295,7 +1302,7 @@ static void sched_compute_splits(ggml_backend_sched_t sched) {
uint64_t compute_start_us = ggml_time_us(); uint64_t compute_start_us = ggml_time_us();
ggml_backend_graph_compute(split_backend, &split->graph); ggml_backend_graph_compute(split_backend, &split->graph);
// ggml_backend_synchronize(split_backend); //ggml_backend_synchronize(split_backend); // necessary to measure compute time
uint64_t compute_end_us = ggml_time_us(); uint64_t compute_end_us = ggml_time_us();
compute_us[split_backend_id] += compute_end_us - compute_start_us; compute_us[split_backend_id] += compute_end_us - compute_start_us;
} }

1
ggml-backend.h
View file

@ -43,7 +43,6 @@ extern "C" {
GGML_API ggml_backend_buffer_type_t ggml_backend_buffer_get_type (ggml_backend_buffer_t buffer); GGML_API ggml_backend_buffer_type_t ggml_backend_buffer_get_type (ggml_backend_buffer_t buffer);
GGML_API void ggml_backend_buffer_reset (ggml_backend_buffer_t buffer); GGML_API void ggml_backend_buffer_reset (ggml_backend_buffer_t buffer);
// //
// Backend // Backend
// //

54
ggml-cuda.cu
View file

@ -9899,6 +9899,10 @@ static const char * ggml_backend_cuda_buffer_get_name(ggml_backend_buffer_t buff
return ctx->name.c_str(); return ctx->name.c_str();
} }
static bool ggml_backend_buffer_is_cuda(ggml_backend_buffer_t buffer) {
return buffer->iface.get_name == ggml_backend_cuda_buffer_get_name;
}
static void ggml_backend_cuda_buffer_free_buffer(ggml_backend_buffer_t buffer) { static void ggml_backend_cuda_buffer_free_buffer(ggml_backend_buffer_t buffer) {
ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context; ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
CUDA_CHECK(cudaFree(ctx->dev_ptr)); CUDA_CHECK(cudaFree(ctx->dev_ptr));
@ -9961,6 +9965,24 @@ static void ggml_backend_cuda_buffer_get_tensor(ggml_backend_buffer_t buffer, co
ggml_cuda_set_device(ctx->device); ggml_cuda_set_device(ctx->device);
CUDA_CHECK(cudaDeviceSynchronize()); CUDA_CHECK(cudaDeviceSynchronize());
CUDA_CHECK(cudaMemcpy(data, (const char *)tensor->data + offset, size, cudaMemcpyDeviceToHost)); CUDA_CHECK(cudaMemcpy(data, (const char *)tensor->data + offset, size, cudaMemcpyDeviceToHost));
CUDA_CHECK(cudaDeviceSynchronize());
}
static bool ggml_backend_cuda_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * src, ggml_tensor * dst) {
if (ggml_backend_buffer_is_cuda(src->buffer)) {
ggml_backend_cuda_buffer_context * src_ctx = (ggml_backend_cuda_buffer_context *)src->buffer->context;
ggml_backend_cuda_buffer_context * dst_ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
ggml_cuda_set_device(src_ctx->device);
CUDA_CHECK(cudaDeviceSynchronize());
ggml_cuda_set_device(dst_ctx->device);
CUDA_CHECK(cudaDeviceSynchronize());
CUDA_CHECK(cudaMemcpy((char *)dst->data, (const char *)src->data, ggml_nbytes(src), cudaMemcpyDeviceToDevice));
CUDA_CHECK(cudaDeviceSynchronize());
return true;
}
return false;
} }
static void ggml_backend_cuda_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) { static void ggml_backend_cuda_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
@ -9979,8 +10001,7 @@ static ggml_backend_buffer_i ggml_backend_cuda_buffer_interface = {
/* .init_tensor = */ ggml_backend_cuda_buffer_init_tensor, /* .init_tensor = */ ggml_backend_cuda_buffer_init_tensor,
/* .set_tensor = */ ggml_backend_cuda_buffer_set_tensor, /* .set_tensor = */ ggml_backend_cuda_buffer_set_tensor,
/* .get_tensor = */ ggml_backend_cuda_buffer_get_tensor, /* .get_tensor = */ ggml_backend_cuda_buffer_get_tensor,
/* .cpy_tensor_from = */ NULL, /* .cpy_tensor = */ ggml_backend_cuda_buffer_cpy_tensor,
/* .cpy_tensor_to = */ NULL,
/* .clear = */ ggml_backend_cuda_buffer_clear, /* .clear = */ ggml_backend_cuda_buffer_clear,
/* .reset = */ NULL, /* .reset = */ NULL,
}; };
@ -10114,6 +10135,11 @@ static const char * ggml_backend_cuda_split_buffer_get_name(ggml_backend_buffer_
UNUSED(buffer); UNUSED(buffer);
} }
// unused at the moment
//static bool ggml_backend_buffer_is_cuda_split(ggml_backend_buffer_t buffer) {
// return buffer->iface.get_name == ggml_backend_cuda_split_buffer_get_name;
//}
static void ggml_backend_cuda_split_buffer_free_buffer(ggml_backend_buffer_t buffer) { static void ggml_backend_cuda_split_buffer_free_buffer(ggml_backend_buffer_t buffer) {
ggml_backend_cuda_split_buffer_context * ctx = (ggml_backend_cuda_split_buffer_context *)buffer->context; ggml_backend_cuda_split_buffer_context * ctx = (ggml_backend_cuda_split_buffer_context *)buffer->context;
delete ctx; delete ctx;
@ -10256,8 +10282,7 @@ static struct ggml_backend_buffer_i ggml_backend_cuda_split_buffer_interface = {
/* .init_tensor = */ ggml_backend_cuda_split_buffer_init_tensor, /* .init_tensor = */ ggml_backend_cuda_split_buffer_init_tensor,
/* .set_tensor = */ ggml_backend_cuda_split_buffer_set_tensor, /* .set_tensor = */ ggml_backend_cuda_split_buffer_set_tensor,
/* .get_tensor = */ ggml_backend_cuda_split_buffer_get_tensor, /* .get_tensor = */ ggml_backend_cuda_split_buffer_get_tensor,
/* .cpy_tensor_from = */ NULL, /* .cpy_tensor = */ NULL,
/* .cpy_tensor_to = */ NULL,
/* .clear = */ ggml_backend_cuda_split_buffer_clear, /* .clear = */ ggml_backend_cuda_split_buffer_clear,
/* .reset = */ NULL, /* .reset = */ NULL,
}; };
@ -10457,6 +10482,17 @@ static void ggml_backend_cuda_get_tensor_async(ggml_backend_t backend, const ggm
CUDA_CHECK(cudaMemcpyAsync(data, (const char *)tensor->data + offset, size, cudaMemcpyDeviceToHost, g_cudaStreams[cuda_ctx->device][0])); CUDA_CHECK(cudaMemcpyAsync(data, (const char *)tensor->data + offset, size, cudaMemcpyDeviceToHost, g_cudaStreams[cuda_ctx->device][0]));
} }
static bool ggml_backend_cuda_cpy_tensor_async(ggml_backend_t backend, const ggml_tensor * src, ggml_tensor * dst) {
ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
if (dst->buffer->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device) && ggml_backend_buffer_is_cuda(src->buffer)) {
CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(dst), cudaMemcpyDeviceToDevice, g_cudaStreams[cuda_ctx->device][0]));
return true;
}
return false;
}
static void ggml_backend_cuda_synchronize(ggml_backend_t backend) { static void ggml_backend_cuda_synchronize(ggml_backend_t backend) {
ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context; ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
@ -10476,8 +10512,9 @@ static bool ggml_backend_cuda_graph_compute(ggml_backend_t backend, ggml_cgraph
for (int i = 0; i < cgraph->n_nodes; i++) { for (int i = 0; i < cgraph->n_nodes; i++) {
ggml_tensor * node = cgraph->nodes[i]; ggml_tensor * node = cgraph->nodes[i];
if (node->op == GGML_OP_RESHAPE || node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE) if (node->op == GGML_OP_RESHAPE || node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE || node->op == GGML_OP_NONE) {
continue; continue;
}
#ifndef NDEBUG #ifndef NDEBUG
assert(node->backend == GGML_BACKEND_GPU || node->backend == GGML_BACKEND_GPU_SPLIT); assert(node->backend == GGML_BACKEND_GPU || node->backend == GGML_BACKEND_GPU_SPLIT);
@ -10487,7 +10524,7 @@ static bool ggml_backend_cuda_graph_compute(ggml_backend_t backend, ggml_cgraph
for (int j = 0; j < GGML_MAX_SRC; j++) { for (int j = 0; j < GGML_MAX_SRC; j++) {
if (node->src[j] != nullptr) { if (node->src[j] != nullptr) {
assert(node->src[j]->backend == GGML_BACKEND_GPU || node->src[j]->backend == GGML_BACKEND_GPU_SPLIT); assert(node->src[j]->backend == GGML_BACKEND_GPU || node->src[j]->backend == GGML_BACKEND_GPU_SPLIT);
//assert(node->src[j]->buffer->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device)); assert(node->src[j]->buffer->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device));
assert(node->src[j]->extra != nullptr); assert(node->src[j]->extra != nullptr);
} }
} }
@ -10500,8 +10537,6 @@ static bool ggml_backend_cuda_graph_compute(ggml_backend_t backend, ggml_cgraph
GGML_ASSERT(ok); GGML_ASSERT(ok);
} }
UNUSED(backend);
return true; return true;
} }
@ -10622,8 +10657,7 @@ static ggml_backend_i ggml_backend_cuda_interface = {
/* .get_default_buffer_type = */ ggml_backend_cuda_get_default_buffer_type, /* .get_default_buffer_type = */ ggml_backend_cuda_get_default_buffer_type,
/* .set_tensor_async = */ ggml_backend_cuda_set_tensor_async, /* .set_tensor_async = */ ggml_backend_cuda_set_tensor_async,
/* .get_tensor_async = */ ggml_backend_cuda_get_tensor_async, /* .get_tensor_async = */ ggml_backend_cuda_get_tensor_async,
/* .cpy_tensor_from_async = */ NULL, /* .cpy_tensor_async = */ ggml_backend_cuda_cpy_tensor_async,
/* .cpy_tensor_to_async = */ NULL,
/* .synchronize = */ ggml_backend_cuda_synchronize, /* .synchronize = */ ggml_backend_cuda_synchronize,
/* .graph_plan_create = */ NULL, /* .graph_plan_create = */ NULL,
/* .graph_plan_free = */ NULL, /* .graph_plan_free = */ NULL,

20
ggml-metal.m
View file

@ -2525,14 +2525,12 @@ static void ggml_backend_metal_buffer_get_tensor(ggml_backend_buffer_t buffer, c
UNUSED(buffer); UNUSED(buffer);
} }
static void ggml_backend_metal_buffer_cpy_tensor_from(ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst) { static bool ggml_backend_metal_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst) {
ggml_backend_tensor_get(src, dst->data, 0, ggml_nbytes(src)); if (ggml_backend_buffer_is_host(src->buffer)) {
memcpy(dst->data, src->data, ggml_nbytes(src));
UNUSED(buffer); return true;
} }
return false;
static void ggml_backend_metal_buffer_cpy_tensor_to(ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst) {
ggml_backend_tensor_set(dst, src->data, 0, ggml_nbytes(src));
UNUSED(buffer); UNUSED(buffer);
} }
@ -2550,8 +2548,7 @@ static struct ggml_backend_buffer_i ggml_backend_metal_buffer_i = {
/* .init_tensor = */ NULL, /* .init_tensor = */ NULL,
/* .set_tensor = */ ggml_backend_metal_buffer_set_tensor, /* .set_tensor = */ ggml_backend_metal_buffer_set_tensor,
/* .get_tensor = */ ggml_backend_metal_buffer_get_tensor, /* .get_tensor = */ ggml_backend_metal_buffer_get_tensor,
/* .cpy_tensor_from = */ ggml_backend_metal_buffer_cpy_tensor_from, /* .cpy_tensor = */ ggml_backend_metal_buffer_cpy_tensor,
/* .cpy_tensor_to = */ ggml_backend_metal_buffer_cpy_tensor_to,
/* .clear = */ ggml_backend_metal_buffer_clear, /* .clear = */ ggml_backend_metal_buffer_clear,
/* .reset = */ NULL, /* .reset = */ NULL,
}; };
@ -2774,8 +2771,7 @@ static struct ggml_backend_i ggml_backend_metal_i = {
/* .get_default_buffer_type = */ ggml_backend_metal_get_default_buffer_type, /* .get_default_buffer_type = */ ggml_backend_metal_get_default_buffer_type,
/* .set_tensor_async = */ NULL, /* .set_tensor_async = */ NULL,
/* .get_tensor_async = */ NULL, /* .get_tensor_async = */ NULL,
/* .cpy_tensor_from_async = */ NULL, /* .cpy_tensor_async = */ NULL,
/* .cpy_tensor_to_async = */ NULL,
/* .synchronize = */ NULL, /* .synchronize = */ NULL,
/* .graph_plan_create = */ NULL, /* .graph_plan_create = */ NULL,
/* .graph_plan_free = */ NULL, /* .graph_plan_free = */ NULL,