[BugFix] Exclude sparse padding from categorical entropy

[AHMETHAKANBEZIR1]

Fixes #4512

Summary

Sparse MaskedCategorical.entropy() skipped masking entirely. Padded indices therefore produce NaN with neg_inf=-inf, or contribute to entropy when neg_inf is finite. A valid sparse action with probability zero has the same 0 * -inf problem.

Apply the existing finite-minimum masking and entropy renormalization to sparse padding and non-finite logits as well as dense invalid actions. Sampling, log_prob, mode and distribution construction are unchanged. This is independent of #4509 and #4511.

Tests in the existing distribution test file compare public entropy and gradients with compact torch.distributions.Categorical references. They cover both masked distribution classes, two padding sentinels, finite/infinite masking values, float32/float64, two batch dimensions and mixed padded/unpadded rows. Separate zero-probability tests include dense controls.

Validation

• Current main 8f92a50: 18 new cases fail, 2 dense controls pass.
• Final Windows / Python 3.12.14 / PyTorch 2.10 CPU, MaskedCategorical and MaskedOneHotCategorical selection: 88 passed, 917 deselected.
• Both classes pass repeated torch.compile(backend="eager", fullgraph=True) entropy and gradient comparisons against the independent reference.
• Full configured pre-commit run --all-files passes; git diff --check passes. The Windows symlink placeholder for AGENTS.md was normalized by the EOF hook during the initial run and restored afterward; it is not part of this change.
• TorchRL Python sources are loaded from this checkout, using TensorDict 0.14.2's official wheel. The optional TorchRL C++ extension is unbuilt and emits its existing warning.
• No CUDA/MPS/Inductor/C++ build, full test suite or full documentation build was run. No public API/documentation change or performance claim.

Issue #4512 was assigned to AHMETHAKANBEZIR1 through the repository's /assign workflow before implementation. Prepared with Codex assistance; the reproductions, independent reference/gradient checks, compile checks and lint checks were run locally, and Codex is recorded as co-author.

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/rl/4513

• 📄 Preview Python docs built from this PR

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

⚠️ 16 Awaiting Approval

As of commit 85de661 with merge base 8f92a50 ():

AWAITING APPROVAL - The following workflows need approval before CI can run:

• Build Aarch64 Linux Wheels (gh)
• Build Linux Wheels (gh)
• Build M1 Wheels (gh)
• Build Windows Wheels (gh)
• Continuous Benchmark (PR) (gh)
• Examples Tests on Linux (gh)
• Generate documentation (gh)
• Habitat Tests on Linux (gh)
• Libs Tests on Linux (gh)
• Lint (gh)
• LLM Tests on Linux (gh)
• Push Binary Nightly (gh)
• SOTA Tests on Linux (gh)
• Tutorials Tests on Linux (gh)
• Unit-tests on Linux (gh)
• Validate Test Partitioning (gh)

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