[inductor] make mask_rcnn inference work in max-autotune mode #123008
Conversation
[ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/123008
Note: Links to docs will display an error until the docs builds have been completed. ✅ You can merge normally! (1 Unrelated Failure)As of commit 06f4713 with merge base 57a9a64 ( BROKEN TRUNK - The following job failed but were present on the merge base:👉 Rebase onto the `viable/strict` branch to avoid these failures
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
shunting314
changed the title
|
@pytorchbot merge |
Merge failedReason: This PR needs a If not, please add the To add a label, you can comment to pytorchbot, for example For more information, see Details for Dev Infra teamRaised by workflow job |
|
@pytorchbot merge |
Merge startedYour change will be merged once all checks pass (ETA 0-4 Hours). Learn more about merging in the wiki. Questions? Feedback? Please reach out to the PyTorch DevX Team |
…h#123008) inference for vision_maskrcnn model fail when max-autotune is enabled. Repro: ``` TORCHINDUCTOR_MAX_AUTOTUNE=1 time python benchmarks/dynamo/torchbench.py --accuracy --inference --bfloat16 --backend inductor --only vision_maskrcnn ``` It turns out that MA code receives empty input tensor for convolution and some places in MA related code does not handle this corner case properly. This PR enhance that and now the accuracy test above can pass. Regarding why the input tensor is empty, I think it's probably due to no objects are detected in the input images (random data?). Pull Request resolved: pytorch#123008 Approved by: https://github.com/jansel
| @@ -360,6 +360,7 @@ def channels_last_conv(): | |||
| and not transposed | |||
| and is_zeros(output_padding) | |||
| and groups == 1 | |||
| and sympy_product(x.get_size()) > 0 | |||
There was a problem hiding this comment.
This needs to account for the case when x has unbacked sizes
Stack from ghstack (oldest at bottom):
inference for vision_maskrcnn model fail when max-autotune is enabled.
Repro:
It turns out that MA code receives empty input tensor for convolution and some places in MA related code does not handle this corner case properly. This PR enhance that and now the accuracy test above can pass.
Regarding why the input tensor is empty, I think it's probably due to no objects are detected in the input images (random data?).
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @peterbell10 @ipiszy @yf225 @chenyang78 @kadeng @muchulee8 @aakhundov @ColinPeppler @amjames @desertfire @chauhang