[chaofengw-nv]

Background

Fixes #1796. PersonaPlex's Half residual path produces a value consistent with combining an unrounded GEMM accumulator and the skip input; the exact fused-kernel mechanism has not been isolated. For a projection of 1 + 2^-12 and skip input -1, Native first rounds the projection to Half and returns zero; the current decoder returns 2^-12.

Exit Criteria

• Preserve linear-output rounding before the residual sum through an actual decoder layer's normal output.
• Retain Half weights, hidden states and KV caches and the existing FP32 graph behavior.
• Change no benchmark data, reference baseline or accuracy gate.

Implementation

A family-local residual helper casts the already rounded branch outputs to FP32, adds them and restores the residual dtype. This is value-equivalent to the Half addition while preserving the producer's rounding boundary. Standard, dual-profile and tensor-parallel decoder residuals use the helper. No new kernel, tactic constraint, platform branch or dependency is introduced; affected engines require rebuilding.

Change categories

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

Validation

Commands and Results

• python -m pytest -q families/personaplex/tests/test_residual_precision.py: identical original/fixed test, separate processes and three fresh engines per precision. Original: 3 numerical FP16 failures and 3 FP32 passes, zero errors or skips. Fixed: all 6 cases pass, zero errors or skips. The test SHA-256 is identical on both sides (e6616ed9d0d6d92015b2874cc5544e62afea8ffa92bae5c6b33739cbd7739916).
• A separate dense-GEMM boundary probe at widths 1024 and 4096: original returns 2^-12 on all three builds per width; FP32 residual-sum candidate returns zero on all three per width, with zero setup failures. An earlier multi-engine probe crashed and is excluded from these completed results.
• Recorded Native depth-layer data (16 codebooks x 6 layers at step 28): recomputing each attention and MLP residual from its rounded Half branch outputs matches the recorded Native tensors with zero differing bits across all 96 layers (CPU check of captured data).
• ruff check and python -m py_compile on all five changed files: passed.
• PYTHONPATH=core/builder:apps/benchmark:. python -m tools.model_ci validate, PYTHONPATH=core/builder:apps/benchmark:. python tools/test_impact.py --validate, and git diff --check: passed with CUDA hidden.

Hardware, Environment, and Revisions

Base 9a8b545f5e8c22c653c8029c95a999dc070b4f3f; tested head 30faad4b1a58d1b6acb57bc6ebf5ca28a3c06401. GB300, Linux ARM64, CUDA 13, TensorRT 11.2.0.113 and PyTorch 2.12.0+cu130. The GPU A/B verifies source and test hashes. Synthetic weights and inputs need no downloaded checkpoint or dataset. Native residual order follows the pinned official source.

Not Run / Remaining Gaps

The regression directly covers the attention-side residual in standard decode. The MLP-side residual, parallel residual, dynamic prefill and TP call sites were reviewed but are not directly exercised by that regression. A fresh original-20 run of the published integration stack plus this exact production patch remains 18/20 against both frozen and refreshed Native FP16: clip 11 is 6.89 dB and clip 18 is 3.78 dB under the unchanged 3 dB gate. All 20 output WAVs are bit-identical to the same stack without this patch, so this PR earns no whole-model accuracy credit in that build. Task P50 is Native 2201.01 ms and TRTMC 1492.99 ms (20 requests each); output sample counts and rates match on all requests. These are one-build measurements. Three-build model acceptance, BF16, distributed tensor-parallel execution, other hardware and exact-head premerge remain unqualified. This PR claims component correctness only.

Contributor Self-Review

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

Reviewed every standard, dynamic and TP residual call site and the new normal-output regression. Changes remain within PersonaPlex.

Notes For Future Readers

The important boundary is the linear result rounded to the model dtype before the residual addition. Promoting the entire core to FP32 changes the requested model precision; this helper retains model-dtype inputs and outputs. Supported call sites provide equally typed branch outputs; mixed-dtype helper calls are outside the supported contract. Independent input, RoPE, SwiGLU and profile fixes remain separate PRs.

Risk level

• Low
• Medium
• High

Residual arithmetic changes across the family. Component evidence does not establish whole-model parity or performance.

[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 50849eb of the consolidated PR. Original issue links and acceptance criteria remain tracked there.