Skip to content

Commit 8957dc8

Browse files
digantdesaifacebook-github-bot
authored andcommitted
Update to new version with new 16x4 kleidi kernels (#6101)
Summary: * This xnnpack has kleidi v0.3.0 * Doesn't yet enable Kleidi yet by default `-DEXECUTORCH_XNNPACK_ENABLE_KLEIDI=OFF` * Minor fix for model load stat with warmup which was reporting 0.0 Pull Request resolved: #6101 Reviewed By: mergennachin Differential Revision: D64220150 Pulled By: digantdesai fbshipit-source-id: 98c8c4e079ee09dfd869970ec1abb9dccae29deb
1 parent cd2d2b4 commit 8957dc8

File tree

3 files changed

+12
-5
lines changed

3 files changed

+12
-5
lines changed
Submodule XNNPACK updated 487 files

extension/llm/runner/stats.h

Lines changed: 10 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -59,9 +59,16 @@ struct ET_EXPERIMENTAL Stats {
5959
aggregate_sampling_timer_start_timestamp = 0;
6060
}
6161

62-
void reset() {
63-
model_load_start_ms = 0;
64-
model_load_end_ms = 0;
62+
void reset(bool all_stats = false) {
63+
// Not resetting model_load_start_ms and model_load_end_ms because reset is
64+
// typically called after warmup and before running the actual run.
65+
// However, we don't load the model again during the actual run after
66+
// warmup. So, we don't want to reset these timestamps unless we are
67+
// resetting everything.
68+
if (all_stats) {
69+
model_load_start_ms = 0;
70+
model_load_end_ms = 0;
71+
}
6572
inference_start_ms = 0;
6673
prompt_eval_end_ms = 0;
6774
first_token_ms = 0;

0 commit comments

Comments
 (0)