mirror of
https://github.com/ggerganov/whisper.cpp.git
synced 2025-04-29 13:44:39 +02:00
cuda : fix data race in soft max (llama/5853)
This commit is contained in:
parent
bd26876267
commit
93a84a143b
@ -6947,6 +6947,7 @@ static __global__ void soft_max_f32(const float * x, const float * mask, const f
|
|||||||
// find the sum of exps in the block
|
// find the sum of exps in the block
|
||||||
tmp = warp_reduce_sum(tmp);
|
tmp = warp_reduce_sum(tmp);
|
||||||
if (block_size > WARP_SIZE) {
|
if (block_size > WARP_SIZE) {
|
||||||
|
__syncthreads();
|
||||||
if (warp_id == 0) {
|
if (warp_id == 0) {
|
||||||
buf_iw[lane_id] = 0.0f;
|
buf_iw[lane_id] = 0.0f;
|
||||||
}
|
}
|
||||||
|
Loading…
Reference in New Issue
Block a user