root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-24 10:24:35 +00:00
Code Issues Actions 5 Packages Projects Releases Wiki Activity

ggml : reuse ggml_get_n_tasks() in ggml_graph_plan() (#4308)

Browse Source 
* ggml : fix soft max out-of-bounds access

ggml-ci

* ggml : reuse ggml_get_n_tasks() in ggml_graph_plan()

ggml-ci
This commit is contained in:
Georgi Gerganov 2023-12-03 15:56:35 +02:00 committed by GitHub
parent adf3de4f69
commit fbbc42827b
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 2 additions and 21 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

23
ggml.c
View File

@ -15879,18 +15879,16 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) {
// thread scheduling for the different operations + work buffer size estimation
for (int i = 0; i < cgraph->n_nodes; i++) {
int n_tasks = 1;
struct ggml_tensor * node = cgraph->nodes[i];
const int n_tasks = ggml_get_n_tasks(node, n_threads);
size_t cur = 0;
switch (node->op) {
case GGML_OP_CPY:
case GGML_OP_DUP:
{
n_tasks = n_threads;
if (ggml_is_quantized(node->type)) {
cur = ggml_type_size(GGML_TYPE_F32) * node->ne[0] * n_tasks;
}
@ -15898,16 +15896,12 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) {
case GGML_OP_ADD:
case GGML_OP_ADD1:
{
n_tasks = n_threads;
if (ggml_is_quantized(node->src[0]->type)) {
cur = ggml_type_size(GGML_TYPE_F32) * node->src[0]->ne[0] * n_tasks;
}
} break;
case GGML_OP_ACC:
{
n_tasks = n_threads;
if (ggml_is_quantized(node->src[0]->type)) {
cur = ggml_type_size(GGML_TYPE_F32) * node->src[1]->ne[0] * n_tasks;
}
@ -15935,16 +15929,12 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) {
} break;
case GGML_OP_OUT_PROD:
{
n_tasks = n_threads;
if (ggml_is_quantized(node->src[0]->type)) {
cur = ggml_type_size(GGML_TYPE_F32) * node->src[0]->ne[0] * n_tasks;
}
} break;
case GGML_OP_SOFT_MAX:
{
n_tasks = MIN(MIN(4, n_threads), ggml_nrows(node->src[0]));
cur = ggml_type_size(GGML_TYPE_F32) * node->ne[0] * n_tasks;
} break;
case GGML_OP_CONV_TRANSPOSE_1D:
@ -15974,7 +15964,6 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) {
} break;
case GGML_OP_IM2COL:
{
n_tasks = n_threads;
} break;
case GGML_OP_CONV_TRANSPOSE_2D:
{
@ -15992,8 +15981,6 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) {
} break;
case GGML_OP_FLASH_ATTN:
{
n_tasks = n_threads;
const int64_t ne11 = ggml_up(node->src[1]->ne[1], GGML_SOFT_MAX_UNROLL);
if (node->src[1]->type == GGML_TYPE_F32) {
@ -16006,8 +15993,6 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) {
} break;
case GGML_OP_FLASH_FF:
{
n_tasks = n_threads;
if (node->src[1]->type == GGML_TYPE_F32) {
cur = sizeof(float)*node->src[1]->ne[1]*n_tasks; // TODO: this can become (n_tasks-1)
cur += sizeof(float)*node->src[1]->ne[1]*n_tasks; // this is overestimated by x2
@ -16018,8 +16003,6 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) {
} break;
case GGML_OP_FLASH_ATTN_BACK:
{
n_tasks = n_threads;
const int64_t D = node->src[0]->ne[0];
const int64_t ne11 = ggml_up(node->src[1]->ne[1], GGML_SOFT_MAX_UNROLL);
const int64_t mxDn = MAX(D, ne11) * 2; // *2 because of S and SM in ggml_compute_forward_flash_attn_back
@ -16034,8 +16017,6 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) {
case GGML_OP_CROSS_ENTROPY_LOSS:
{
n_tasks = n_threads;
cur = ggml_type_size(node->type)*(n_tasks + node->src[0]->ne[0]*n_tasks);
} break;
case GGML_OP_COUNT: