[aw471]

Summary

Record that a nonempty input column produces at least one unique category. This lets torch.compile resolve subsequent empty-vocabulary checks without changing category values or codes.

Description of the change

Encountered during raw-label preparation for compiled KumoTabular fitting. This localized reproduction isolates the same empty-vocabulary check used by AlignCategories (CPU, PyTorch 2.14):

import torch
from sdm import CategoricalTensor
from sdm.tensor.categorical import _iter_categories

def prepare(labels):
    categorical = CategoricalTensor.from_tensor(labels)
    for category in _iter_categories(categorical):
        if category.numel() == 0:
            return category
    return categorical.code

torch.compile(prepare, backend="inductor", fullgraph=True)(
    torch.tensor([[0], [1], [0]])
)

torch._dynamo.exc.UserError:
Could not guard on data-dependent expression Eq(u0, 0)
Caused by: if category.numel() == 0

unique() already compiles, but its learned output size can appear to be zero. Four added lines use torch._check to tell the compiler each vocabulary is nonempty when the input is nonempty. Empty rows and columns retain their existing behavior. This fixes this bound only; other KumoTabular fitting blockers remain.

Validation

• Reproducer fails before and passes after the fix with Inductor and fullgraph=True.
• Six strict CPU cases passed: regular, all-NaN, empty rows, empty columns, batched and strided inputs; one graph and zero breaks each.
• Existing categorical checks: 84 passed, 24 skipped; applicable pre-commit checks passed.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 82.91%. Comparing base (ec0203b) to head (b58d1bd).
✅ All tests successful. No failed tests found.

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #1260   +/-   ##
=======================================
  Coverage   82.91%   82.91%           
=======================================
  Files         173      173           
  Lines       11191    11194    +3     
=======================================
+ Hits         9279     9282    +3     
  Misses       1912     1912           

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

• Source mode: Auto
• Repository: NVIDIA/structured-data-models
• Branch: compile-tabular-align-value-routing
• File: .coderabbit.yaml
• Commit: b58d1bd
• Layers: Repository YAML
• Review profile: QUIET
• Plan: Enterprise
• Run ID: 3e3f8dd0-5819-4382-b628-f8cbe891e587

📥 Commits

Reviewing files that changed from the base of the PR and between 567e527 and b58d1bd.

📒 Files selected for processing: 1

• sdm/tensor/categorical.py

Included review availability: This review used your included allowance. Your plan provides up to 12 included reviews per hour; 8 remain after this review.

📝 Summary

Summary by CodeRabbit

• Bug Fixes
  • Categorical tensor conversion now checks that each generated category vector contains at least one category when the input has elements. Empty inputs remain supported and are not subject to this check. This validation helps identify invalid empty category vectors during conversion.

CategoricalTensor.from_tensor checks that generated category vectors are nonempty when the input tensor has at least one element. Empty inputs bypass the check.

Changes: Categorical tensor validation

Layer / File(s)	Summary
Check generated category vectors sdm/tensor/categorical.py	When the input tensor has elements, from_tensor checks that each generated category vector contains at least one value. Empty inputs skip this check.

Priority: ➖ Normal

Estimated code review effort: 2 (Simple) | ~5 minutes

Merge Risk: ⚪ Minimal · up to b58d1

No actionable merge-blocking risk was found in this change.

Pre-merge checks: 5

Check name	Status	Explanation
Title check		The title clearly and concisely describes the main change: preserving nonempty categorical vocabulary bounds during compilation.
Description check		The description directly explains the compilation issue, the torch._check fix, preserved empty-input behavior, and validation results.
Docstring Coverage		Docstring coverage is 100.00% which is sufficient. The required threshold is 80.00%. Docstring coverage is scoped to functions touched by this diff. Analyzed 1 functions across 1 files.
Linked Issues check		Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check		Check skipped because no linked issues were found for this pull request.

✨ Finishing Touches

📝 Generate docstrings

• Commit to this branch
• Create a new PR

🧪 Generate unit tests (beta)

• Commit to this branch
• Create a new PR

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

_{Comment @coderabbitai help to get the list of available commands.}