[chaofengw-nv]

Background

Refs #1786. The primary PersonaPlex profile declares precision: fp16 but selects fp32_layers: [0, 1]. Those
selectors promote both autoregressive components, the temporal transformer (selector 0) and the depth transformer
(selector 1), to FP32, while the formal AIPerf benchmark's Native baseline runs that core in FP16. (The manifest's E2E testcase
reference_precision is bf16 and is unchanged.) The requested product configuration is
a true FP16 core: Half temporal and depth weights, hidden states and KV caches, with the Mimi codec in FP32 as before.

Exit Criteria

• The authoritative manifest families/personaplex/tests/manifests/personaplex-7b.json resolves to an FP16 build with
  no FP32 component selectors.
• The resulting bundle's temporal and depth engines use Half KV caches.
• Non-goals:
  • This PR does not claim the model passes the audio-parity benchmark.
  • It changes no gate, threshold, reference_precision or Native baseline.
  • It leaves the reference-only apps/aiperf_qual/config/models/personaplex-7b.yaml, which holds no precision
    settings, unchanged.

Implementation

• fp32_layers: [0, 1] becomes fp32_layers: [] in the authoritative manifest. No code changes.

Change categories

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

Validation

Commands and Results

• Catalog resolution (CPU only, CUDA hidden): isolated trees of main 9a8b545 and of this change resolved
  personaplex-7b through trtmc_benchmark.catalog.ManifestCatalog and generated the build command with
  trtmc_benchmark.builder._build_command.
  • main: precision fp16, fp32_layers [0, 1], command arguments --fp32-layer 0 --fp32-layer 1.
  • This change: precision fp16, fp32_layers [], no --fp32-layer arguments.
• Engine dtypes on GB300 (diagnostic builds, not this head): bundles built with the same request (fp32_layers []) on
  development trees that also carried pending accuracy changes. The serialized plans were inspected directly.
  • The temporal engine.plan and all 16 depth.N.plan engines use cache_k_0 HALF.
  • The Mimi encoder and decoder plans are FP32.

Hardware, Environment, and Revisions

GB300 (SBSA), CUDA 13.0, TensorRT 11.2.0.113. Checkpoint nvidia/personaplex-7b-v1 revision
fdaf4090a61cb315c138a1faee287ffd6c716309. Tested head 05b41f087ad7a41f00e4b21494428468af1007fd.

Not Run / Remaining Gaps

• The whole model does not pass yet. True-FP16 builds still fail the original audio-parity clip 7 against the frozen
  Native FP16 baseline (5.58 dB under the unchanged 3 dB gate). This holds even with the pending component fixes fix(personaplex): round the FP16 depth input like the reference #1789,
  fix(personaplex): preserve reference RoPE precision #1791 and fix(personaplex): round the SwiGLU gate product once #1794 and the FP32 RMSNorm epsilon applied. The remaining temporal-path differences are under investigation.
• Clip 11 differs because the frozen Native baseline runs its FP32 Mimi convolutions with cuDNN TF32. Whether to change
  that reference policy is a pending maintainer decision, and this PR does not change it.
• No bundle was built from exactly this head, and no full original-20 run with this profile is claimed. No CI results
  yet for this head.

Contributor Self-Review

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

Notes For Future Readers

• Selector meaning in this family: 0 = temporal, 1 = depth, 2 and 3 = Mimi components, 4 and above = temporal blocks.
  The FP16 profile now requests none of them.
• Bundles built from the earlier profile have FP32 temporal and depth engines and should be rebuilt.

Risk level

• Low
• Medium
• High

Medium: the default FP16 PersonaPlex build changes from an FP32 core to a Half core, which is the requested precision but
not yet audio-parity qualified.

[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 40288a4 of the consolidated PR. Original issue links and acceptance criteria remain tracked there.