Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

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

Merged
merged 2 commits into from
Dec 3, 2023
Merged

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

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
bfb98ad
ggml : fix soft max out-of-bounds access
ggerganov Dec 3, 2023
45331b6
ggml : reuse ggml_get_n_tasks() in ggml_graph_plan()
ggerganov Dec 3, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
28 changes: 6 additions & 22 deletions ggml.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -15629,7 +15629,6 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
} break;
case GGML_OP_DIAG_MASK_ZERO:
case GGML_OP_DIAG_MASK_INF:
case GGML_OP_SOFT_MAX:
case GGML_OP_SOFT_MAX_BACK:
case GGML_OP_ROPE:
case GGML_OP_ROPE_BACK:
Expand All @@ -15645,6 +15644,10 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
{
n_tasks = 1; //TODO
} break;
case GGML_OP_SOFT_MAX:
{
n_tasks = MIN(MIN(4, n_threads), ggml_nrows(node->src[0]));
} break;
case GGML_OP_CONV_TRANSPOSE_1D:
{
n_tasks = n_threads;
Expand Down Expand Up @@ -15876,35 +15879,29 @@ 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;
}
} break;
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;
}
Expand Down Expand Up @@ -15932,16 +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:
Expand Down Expand Up @@ -15971,7 +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:
{
Expand All @@ -15989,8 +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) {
Expand All @@ -16003,8 +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
Expand All @@ -16015,8 +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
Expand All @@ -16031,8 +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:
Expand Down
Loading