artin/llama2.c
1
0
Fork 0
mirror of https://github.com/trholding/llama2.c.git synced 2026-02-06 11:26:53 +00:00
Code Issues Packages Projects Releases Wiki Activity

Update runq.c

Browse Source 
runq - Undo #pragma omp parallel sections for matmuls for now as there is no real benefit with low number of cores
This commit is contained in:
Vulcan 2024-07-20 19:20:30 +05:30
parent 725faaa608
commit 16e223fbca
1 changed files with 0 additions and 9 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

9
runq.c
View File

@ -636,18 +636,9 @@ float* forward(Transformer* transformer, int token, int pos) {
// qkv matmuls for this position
quantize(&s->xq, s->xb, dim);
// L2E Addition
#pragma omp parallel sections
{
#pragma omp section
matmul(s->q, &s->xq, w->wq + l, dim, dim);
#pragma omp section
matmul(s->k, &s->xq, w->wk + l, dim, kv_dim);
#pragma omp section
matmul(s->v, &s->xq, w->wv + l, dim, kv_dim);
}
// END L2E Addition
// RoPE relative positional encoding: complex-valued rotate q and k in each head
for (int i = 0; i < dim; i+=2) {