Issue Station
Sign In
Tools hereToolbox

pytorch - 💡(How to fix) Fix [Inductor] functorch_dp_cifar10 accuracy regression in dynamic_inductor_torchbench [1 participants]

Official PRs (…)
ON THIS PAGE

Recommended Tools

×6

Utilities matched from this issue’s tags and category — try them while you read without losing context.

GitHub issue graph ai analysis

Paste a GitHub issue URL. We fetch that issue, discover linked issues from bodies/comments/timeline, collect linked pull requests, and produce a structured English report.

The report is written in English Markdown for sharing and archival.

Helpful · Quick feedback

Loading…
GitHub stats
pytorch/pytorch#180740•Fetched 2026-04-19 15:03:54
View on GitHub
Comments
0
Participants
1
Timeline
51
Reactions
0
Author
NikhilAPatelNikhilAPatel
Participants
NikhilAPatelNikhilAPatel
Timeline (top)
mentioned ×22subscribed ×22labeled ×7

functorch_dp_cifar10 fails accuracy checks in the dynamic_inductor_torchbench periodic CI job (CUDA, A10G):

functorch_dp_cifar10    FAIL:     accuracy=fail_accuracy, expected=pass

The error is a numerical accuracy regression in batch norm parameters:

Accuracy failed for key name layer2.0.bn2.bias
RMSE (res-fp64): 0.00606, (ref-fp64): 0.00145, shape=torch.Size([128])

Error Message

The error is a numerical accuracy regression in batch norm parameters:

Root Cause

functorch_dp_cifar10 fails accuracy checks in the dynamic_inductor_torchbench periodic CI job (CUDA, A10G):

functorch_dp_cifar10    FAIL:     accuracy=fail_accuracy, expected=pass

The error is a numerical accuracy regression in batch norm parameters:

Accuracy failed for key name layer2.0.bn2.bias
RMSE (res-fp64): 0.00606, (ref-fp64): 0.00145, shape=torch.Size([128])

Code Example

functorch_dp_cifar10    FAIL:     accuracy=fail_accuracy, expected=pass

---

Accuracy failed for key name layer2.0.bn2.bias
RMSE (res-fp64): 0.00606, (ref-fp64): 0.00145, shape=torch.Size([128])
RAW_BUFFERClick to expand / collapse

Description

functorch_dp_cifar10 fails accuracy checks in the dynamic_inductor_torchbench periodic CI job (CUDA, A10G):

functorch_dp_cifar10    FAIL:     accuracy=fail_accuracy, expected=pass

The error is a numerical accuracy regression in batch norm parameters:

Accuracy failed for key name layer2.0.bn2.bias
RMSE (res-fp64): 0.00606, (ref-fp64): 0.00145, shape=torch.Size([128])

Details

  • Only affects dynamic_inductor_torchbench (dynamic shapes + inductor), not static shapes
  • Consistent across at least 4 consecutive periodic CI runs (April 17-18)
  • Inference pass succeeds; only training accuracy fails
  • The model compiles successfully (1 graph, 69 ops, 0 graph breaks)

Expected CSV entry

benchmarks/dynamo/ci_expected_accuracy/dynamic_inductor_torchbench_inference.csv: functorch_dp_cifar10,pass,0

CI Impact

Blocks periodic-dynamo-benchmarks-test / test (dynamic_inductor_torchbench, 1, 2, linux.g5.4xlarge.nvidia.gpu).

cc @ezyang @penguinwu @bobrenjc93 @aditvenk @laithsakka @voznesenskym @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @jataylo

extent analysis

TL;DR

Investigate and address the numerical accuracy regression in batch norm parameters for the functorch_dp_cifar10 model in the dynamic_inductor_torchbench CI job.

Guidance

  • Review the differences in batch norm parameter calculations between the dynamic and static shape configurations to identify the source of the numerical accuracy regression.
  • Verify that the issue is specific to the training pass and not the inference pass, as indicated by the successful inference pass.
  • Check the model compilation process to ensure that the graph and ops are correctly generated, given that the model compiles successfully with 1 graph, 69 ops, and 0 graph breaks.
  • Investigate the benchmarks/dynamo/ci_expected_accuracy/dynamic_inductor_torchbench_inference.csv file to understand the expected accuracy benchmarks for the functorch_dp_cifar10 model.

Example

No specific code example can be provided without further details on the implementation.

Notes

The issue seems to be specific to the dynamic shape configuration with inductor, and the numerical accuracy regression is limited to the batch norm parameters. Further investigation is needed to determine the root cause of the regression.

Recommendation

Apply a workaround to adjust the batch norm parameter calculations or the accuracy checks to mitigate the numerical accuracy regression, as the root cause of the issue is not immediately clear.

Vote matrix · Quick signals

Works
Did the solution work? Tap to confirm.
Easy Fix
Was it a quick fix?
Time Saver
Did it save you time?
Blocking
Was it severely blocking?
Common Issue
Are others likely hitting this too?
Flaky / Intermittent
Is it intermittent?
Verified / Reproducible
Can you reproduce it reliably?
Loading…
#database connection#vector store#embedding generation#cache error#pipeline error

Still need to ship something?

×6

Another batch ranked right after the header list — different links, same matching logic.

Back to top recommendations

TRENDING

pytorch - 💡(How to fix) Fix [Inductor] functorch_dp_cifar10 accuracy regression in dynamic_inductor_torchbench [1 participants]