ggml : remove src0 and src1 from ggml_tensor and rename opt to src (#2178)

* Add ggml changes

* Update train-text-from-scratch for change

* mpi : adapt to new ggml_tensor->src

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
This commit is contained in:
Spencer Sutton 2023-07-11 12:31:10 -04:00 committed by GitHub
parent c9c74b4e3f
commit 5bf2a27718
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
6 changed files with 371 additions and 421 deletions

View file

@ -1354,17 +1354,9 @@ struct ggml_tensor * expand(struct ggml_cgraph * g, struct ggml_tensor * t) {
}
}
if (t->src0) {
expand(g, t->src0);
}
if (t->src1) {
expand(g, t->src1);
}
for (int i = 0; i < GGML_MAX_OPT; ++i) {
if (t->opt[i]) {
expand(g, t->opt[i]);
for (int i = 0; i < GGML_MAX_SRC; ++i) {
if (t->src[i]) {
expand(g, t->src[i]);
}
}

View file

@ -3200,36 +3200,36 @@ void ggml_cuda_assign_buffers_impl(struct ggml_tensor * tensor, bool scratch, bo
}
// recursively assign CUDA buffers until a compute tensor is found
if (tensor->src0 != nullptr && tensor->src0->backend == GGML_BACKEND_CPU) {
const ggml_op src0_op = tensor->src0->op;
if (tensor->src[0] != nullptr && tensor->src[0]->backend == GGML_BACKEND_CPU) {
const ggml_op src0_op = tensor->src[0]->op;
if (src0_op == GGML_OP_RESHAPE || src0_op == GGML_OP_TRANSPOSE || src0_op == GGML_OP_VIEW) {
ggml_cuda_assign_buffers_impl(tensor->src0, scratch, force_inplace);
ggml_cuda_assign_buffers_impl(tensor->src[0], scratch, force_inplace);
}
}
if (tensor->op == GGML_OP_CPY && tensor->src1->backend == GGML_BACKEND_CPU) {
ggml_cuda_assign_buffers_impl(tensor->src1, scratch, force_inplace);
if (tensor->op == GGML_OP_CPY && tensor->src[1]->backend == GGML_BACKEND_CPU) {
ggml_cuda_assign_buffers_impl(tensor->src[1], scratch, force_inplace);
}
tensor->backend = GGML_BACKEND_GPU;
struct ggml_tensor_extra_gpu * extra = new ggml_tensor_extra_gpu;
memset(extra, 0, sizeof(*extra));
const bool inplace = (tensor->src0 != nullptr && tensor->src0->data == tensor->data) ||
const bool inplace = (tensor->src[0] != nullptr && tensor->src[0]->data == tensor->data) ||
tensor->op == GGML_OP_VIEW ||
force_inplace;
const size_t size = ggml_nbytes(tensor);
CUDA_CHECK(cudaSetDevice(g_main_device));
if (inplace && (tensor->src0->backend == GGML_BACKEND_GPU || tensor->src0->backend == GGML_BACKEND_GPU_SPLIT)) {
struct ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu * ) tensor->src0->extra;
if (inplace && (tensor->src[0]->backend == GGML_BACKEND_GPU || tensor->src[0]->backend == GGML_BACKEND_GPU_SPLIT)) {
struct ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu * ) tensor->src[0]->extra;
char * src0_ddc = (char *) src0_extra->data_device[g_main_device];
size_t offset = 0;
if (tensor->op == GGML_OP_VIEW) {
memcpy(&offset, tensor->opt[0]->data, sizeof(size_t));
memcpy(&offset, tensor->src[2]->data, sizeof(size_t));
}
extra->data_device[g_main_device] = src0_ddc + offset;
} else if (tensor->op == GGML_OP_CPY) {
struct ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu * ) tensor->src1->extra;
struct ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu * ) tensor->src[1]->extra;
void * src1_ddv = src1_extra->data_device[g_main_device];
extra->data_device[g_main_device] = src1_ddv;
} else if (scratch) {
@ -3300,8 +3300,8 @@ void ggml_cuda_free_scratch() {
bool ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor){
ggml_cuda_func_t func;
const bool any_on_device = tensor->backend == GGML_BACKEND_GPU
|| (tensor->src0 != nullptr && (tensor->src0->backend == GGML_BACKEND_GPU || tensor->src0->backend == GGML_BACKEND_GPU_SPLIT))
|| (tensor->src1 != nullptr && tensor->src1->backend == GGML_BACKEND_GPU);
|| (tensor->src[0] != nullptr && (tensor->src[0]->backend == GGML_BACKEND_GPU || tensor->src[0]->backend == GGML_BACKEND_GPU_SPLIT))
|| (tensor->src[1] != nullptr && tensor->src[1]->backend == GGML_BACKEND_GPU);
switch (tensor->op) {
case GGML_OP_ADD:
@ -3329,7 +3329,7 @@ bool ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_
func = ggml_cuda_rms_norm;
break;
case GGML_OP_MUL_MAT:
if (!any_on_device && !ggml_cuda_can_mul_mat(tensor->src0, tensor->src1, tensor)) {
if (!any_on_device && !ggml_cuda_can_mul_mat(tensor->src[0], tensor->src[1], tensor)) {
return false;
}
func = ggml_cuda_mul_mat;
@ -3383,6 +3383,6 @@ bool ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_
if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
return true;
}
func(tensor->src0, tensor->src1, tensor);
func(tensor->src[0], tensor->src[1], tensor);
return true;
}

View file

@ -393,8 +393,8 @@ void ggml_metal_graph_compute(
for (int i = node_start; i < node_end; ++i) {
metal_printf("%s: encoding node %3d, op = %8s\n", __func__, i, ggml_op_name(gf->nodes[i]->op));
struct ggml_tensor * src0 = gf->nodes[i]->src0;
struct ggml_tensor * src1 = gf->nodes[i]->src1;
struct ggml_tensor * src0 = gf->nodes[i]->src[0];
struct ggml_tensor * src1 = gf->nodes[i]->src[1];
struct ggml_tensor * dst = gf->nodes[i];
const int64_t ne00 = src0 ? src0->ne[0] : 0;

View file

@ -175,11 +175,11 @@ void ggml_mpi_graph_compute_pre(
// attach the input data to all nodes that need it
// TODO: not great - should be able to do this without modifying the compute graph (see next TODO below)
for (int i = idx_l0; i < idx_l1; i++) {
if (gf->nodes[i]->src0 == gf->nodes[idx_l0]) {
gf->nodes[i]->src0 = inp0;
if (gf->nodes[i]->src[0] == gf->nodes[idx_l0]) {
gf->nodes[i]->src[0] = inp0;
}
if (gf->nodes[i]->src1 == gf->nodes[idx_l0]) {
gf->nodes[i]->src1 = inp0;
if (gf->nodes[i]->src[1] == gf->nodes[idx_l0]) {
gf->nodes[i]->src[1] = inp0;
}
}

728
ggml.c

File diff suppressed because it is too large Load diff

10
ggml.h
View file

@ -132,10 +132,10 @@
// {
// struct ggml_tensor * a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, 2, 3);
//
// // a[1, 2] = 1.0f;
// // a[2, 1] = 1.0f;
// *(float *) ((char *) a->data + 2*a->nb[1] + 1*a->nb[0]) = 1.0f;
//
// // a[2, 0] = 2.0f;
// // a[0, 2] = 2.0f;
// *(float *) ((char *) a->data + 0*a->nb[1] + 2*a->nb[0]) = 2.0f;
//
// ...
@ -197,7 +197,7 @@
#define GGML_MAX_NODES 4096
#define GGML_MAX_PARAMS 256
#define GGML_MAX_CONTEXTS 64
#define GGML_MAX_OPT 4
#define GGML_MAX_SRC 6
#define GGML_MAX_NAME 48
#define GGML_DEFAULT_N_THREADS 4
@ -414,9 +414,7 @@ extern "C" {
bool is_param;
struct ggml_tensor * grad;
struct ggml_tensor * src0;
struct ggml_tensor * src1;
struct ggml_tensor * opt[GGML_MAX_OPT];
struct ggml_tensor * src[GGML_MAX_SRC];
// performance
int perf_runs;