🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/161412

• 📄 Preview Python docs built from this PR
• 📄 Preview C++ docs built from this PR
• ❓ Need help or want to give feedback on the CI? Visit the bot commands wiki

Note: Links to docs will display an error until the docs builds have been completed.

✅ No Failures

As of commit 1398d84 with merge base c321111 ():
💚 Looks good so far! There are no failures yet. 💚

This comment was automatically generated by Dr. CI and updates every 15 minutes.

The committers listed above are authorized under a signed CLA.

• ✅ login: mansiag05 / name: Mansi Agarwal (1398d84, c9a84ae, 5041fdb)

Hello all, thanks for the reviews and clarifications.
I wanted to follow up with an example to clarify the edge case we are discussing.

Examples:

Valid case:

input = torch.randn(10)        # [C] -> Internally [1,C]
target = torch.tensor(3)       # scalar, class 3 is the correct label.
F.nll_loss(input, target)      # works

• Conceptually, nll_loss expects input.shape = [N, C], target.shape = [N].
• When input is 1D ([C]), PyTorch internally treats it as a batch of size 1 ([1, C]). This means the target should be scalar (or [1]) to match the single batch sample.

Invalid case:

# 1D input and 1D target of same length
input1 = torch.randn(10)                           # shape [10] -> internally [1,10]
target1 = torch.randint(0, 10, (10,), dtype=torch.long)  # shape [10]
F.nll_loss(input1, target1)

• Here PyTorch currently does not raise an error, even though logically this is invalid: a batch of size 1 cannot have 10 labels.
• This seems to be a quirk in the current shape-checking logic: the implementation treats the 1D input as batch size 1 internally but does not fully enforce that the target must be scalar. So it accidentally accepts [10] as a target, even though it doesn’t align with the “batch size =1” interpretation.

Next Step

Given this, what would be the recommended approach for the PR?

1. Update the documentation to clearly explain that for 1D input, only scalar targets are valid?
2. Or additionally introduce a deprecation warning / stricter error check for the invalid 1D input + 1D target case?

Happy to adjust the PR based on guidance.

neither. We could make the 1d-1d shape check correct, but we won't limit target to 1d, and we won'd deprecate.

I've updated the check as per the comment.

Also, one more observation, when both the input and target tensors are 1D, the NLL loss function only uses the value at index 0 of the target tensor. The remaining values in the target tensor are ignored.

Right, that's why we need better shape checks (to make sure target has 1 element in this case)

I've updated the shape check. Could you please review.

@pytorchbot merge

Merge started

Your 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

Merge failed

Reason: 26 mandatory check(s) failed. The first few are:

• pull / linux-jammy-cuda12.8-cudnn9-py3.10-clang12 / build
• pull / linux-jammy-py3.10-clang12 / test (default, 2, 5, linux.4xlarge)
• pull / linux-jammy-py3.10-clang12 / test (default, 4, 5, linux.4xlarge)
• pull / linux-jammy-py3.10-clang12 / test (default, 3, 5, linux.4xlarge)
• pull / linux-jammy-py3.10-clang12 / test (crossref, 1, 2, linux.2xlarge)

Dig deeper by viewing the failures on hud

@pytorchbot rebase -b main

@pytorchbot started a rebase job onto refs/remotes/origin/main. Check the current status here

Successfully rebased fix-issue-157420 onto refs/remotes/origin/main, please pull locally before adding more changes (for example, via git checkout fix-issue-157420 && git pull --rebase)

@pytorchbot merge

Merge started

Your 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

PyTorch used to allow 1d size 0 target with 1d size 0 input, is it intentional that this PR blocks this? It's a corner case, just wondering. I guess part of it is that there is no 0 element 0 dim tensor, but that's just how dimensions work.