-
Notifications
You must be signed in to change notification settings - Fork 684
backends/cuda: use async malloc/free #14976
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.
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
Conversation
Stack from ghstack (oldest at bottom): |
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/14976
Note: Links to docs will display an error until the docs builds have been completed. ❌ 4 New FailuresAs of commit a0e9faa with merge base 9764269 ( NEW FAILURES - The following jobs have failed:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
test-qnn-wheel-packages-linux is broken on main. test-multimodal-linux (gemma3-4b) looks like it's a bit flaky on main and segfaulting, so the segfault here is not blocking considering that we have more specific tests for this PR that are passing and this PR should be unrelated. merging. |
Let me see if I can fix the gemma3 issue |
Found device synchronize in aoti_torch_delete_tensor_object via Linux perf. This change appears to significantly improve self-reported latency from voxtral_runner as found in https://github.com/pytorch/executorch/blob/main/.github/workflows/cuda.yml#L111-L172:
Baseline:
Run latency (ms):
audio_encoder: 575.797
token_embedding: 14.571
text_decoder: 3095.356
With this PR:
Run latency (ms):
audio_encoder: 175.807
token_embedding: 8.799
text_decoder: 344.367