[chaofengw-nv]

Background

Fixes #1779. FP16 PersonaPlex decoder builds round RMSNorm's 1e-8 epsilon to zero before promoting it to FP32. Zero inputs then produce NaNs; a small-input regression differs from Native by 0.29296875.

Exit Criteria

• Preserve the epsilon used by Native rms_norm_f32 through constant creation and variance addition.
• Keep zero inputs finite and match Native normalization for small and ordinary FP16 inputs, with identical assertions and tolerance for the original and fixed controls.

Implementation

Store the epsilon constant in FP32 in the standard and tensor-parallel builders. The family's dual-profile builder already does this. Model checkpoints, requests, gates, references and runtime source remain unchanged.

Add a one-layer standard-decoder regression. Identity value/output projections and one unmasked attention key expose normalization through the existing post-attention debug output. The expected value uses the independent Native FP32-normalization equation followed by FP16 output/residual rounding.

Change categories

• Model or runtime behavior
• Public API
• ABI
• Bundle or artifact format
• Dependencies
• Documentation only
• CI or developer tooling

Validation

Commands and Results

• GB300: PYTHONPATH=core/builder:apps/benchmark:. python -m pytest -c pyproject.toml -q -rA -p no:cacheprovider families/personaplex/tests/test_fp16_decoder_norm_epsilon.py: 3 passed for zero, 1e-4, and 0.1 inputs.
• Original-builder negative control, using the identical test and atol=5e-4: 2 failed, 1 passed. Zero input produces NaNs; all 64 small-input elements differ by 0.29296875. Ordinary input passes.
• These commands used isolated original/fixed family copies. Import-path guards verified the loaded builders, and their hashes matched the corresponding source files; the test file was identical in both runs.
• CPU equation control: epsilon 1e-8 gives 0.70703125 for a constant 1e-4 FP16 input; FP16-stored epsilon is zero and gives 1.0.
• ruff check --config ruff.toml on the three changed files, python -m tools.model_ci validate, python tools/test_impact.py --validate, and git diff --check: passed.

Hardware, Environment, and Revisions

• NVIDIA GB300, ARM64, CUDA 13.0, TensorRT 11.2.0.113, PyTorch 2.12.0+cu130.
• Head 39486fac00d8b435955c7be04b9b48445d92a678; base 9643800b3b5a951382c87ae1ebfbbeb24b1a8090. The tested fixed production files and test are byte-identical to this head.
• Native contract: the pinned PersonaPlex source's rms_norm_f32 uses FP32 variance with epsilon 1e-8.

Not Run / Remaining Gaps

Full PersonaPlex audio-parity acceptance for this epsilon change, multi-GPU tensor-parallel execution, other platforms, and protected exact-head premerge have not been qualified. The component result does not establish that this alone resolves the remaining audio-parity failure.

Contributor Self-Review

• I have completed a self-review of this change.

Reviewed the three-file diff, family ownership, FP32 constant propagation, independent reference, fixed tolerance, original-source negative control, DCO and validation limits.

Notes For Future Readers

• Rebuild engines to apply the corrected constant. Bundle interfaces and runtime loading are unchanged.
• The FP32 promotion must happen before storing the epsilon; casting an already rounded zero cannot recover it.
• Other PersonaPlex pipeline work is tracked in fix(personaplex): align FP16 inference with Native #1751. Its full-model validation is separate from this numerical regression.

Risk level

• Low
• Medium
• High

The change preserves one family-owned scalar in its intended computation dtype. Existing FP32 builds already use this value.

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

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

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

[chaofengw-nv]

Superseded by #1751. This change and its regression coverage are preserved in commit bf45b2e of the consolidated PR. Original issue links and acceptance criteria remain tracked there.