Repository navigation
[Accuracy] PersonaPlex FP16 profile promotes the entire autoregressive core to FP32 #1786
Description
Description
The primary PersonaPlex profile declares FP16 but promotes both temporal and depth components with fp32_layers: [0, 1]. The autoregressive core therefore runs in FP32 while its benchmark Native baseline uses FP16. A comparison that changes both the implementation and core precision cannot establish FP16 conversion accuracy.
Reproduction and Evidence
Affected source: main at 9643800b3b5a951382c87ae1ebfbbeb24b1a8090; checkpoint nvidia/personaplex-7b-v1, revision fdaf4090a61cb315c138a1faee287ffd6c716309.
The profile's component selectors promote both core engines. Explicitly removing the selectors produces FP16 hidden states and KV caches and a substantially smaller bundle. This true FP16 path still fails the original audio-parity case 7; therefore changing the precision declaration alone does not fix the model.
The mixed-precision FP32-core candidate's case 7 passes against a separate Native FP32 diagnostic at 0.02 dB and fails against the canonical Native FP16 baseline at 5.58 dB under the same unchanged 3 dB gate. These are diagnostic controls, not a passing FP16 qualification.
A separate codec control explains the FP32 diagnostic's remaining case 11: cuDNN TF32 changes 8 of 888 conditioning codes; Native CUDA with TF32 disabled, CPU FP32, and the TensorRT encoder agree on all 888. This does not establish a fix for the true FP16 core and does not authorize replacing the historical Native FP16 baseline.
Hardware: NVIDIA GB300, ARM64, CUDA 13, TensorRT 11.2.0.113, PyTorch 2.12.0+cu130.
Required Outcome
The maintainer explicitly requires a true FP16 temporal/depth core aligned with the existing Native FP16 baseline. Remove the full-core promotions only as part of a verified FP16 fix. Keep the original 20 requests, reference checkpoint, and all quality gates. Validate actual core/cache dtypes and complete three fresh independent builds before declaring acceptance.
The alternative FP32-profile change is not being submitted. Its four queued baseline/profile jobs were held before execution, and the numerical/profile changes were restored. Previous FP32 controls remain diagnostics. Related component fixes are #1751, #1780, and #1781; avoiding discarded Native decoding is independently tracked in #1783.
Activity
[-][Accuracy] PersonaPlex benchmark references use mismatched numerical precision[/-][+][Accuracy] PersonaPlex FP16 profile promotes the entire autoregressive core to FP32[/+]fix(personaplex): build the FP16 profile with a Half core
fix(personaplex): build the FP16 profile with a Half core
fix(personaplex): align FP16 inference with Native (#1751)
chaofengw-nv commentedon Oct 21, 2026
The maintainer has explicitly accepted the current 50-question VoiceBench pilot when TRTMC aggregate accuracy matches Native. This supersedes the historical original-20 waveform qualification as the acceptance scope for this issue; the old results are retained as diagnostics.
The configuration defect is fixed by merged #1751. On merged main
59fe695bdd5cd328bfc276a534cf762beeede787(including benchmark #1809):Limits retained explicitly:
Closing as completed for the FP32-promotion defect and the revised maintainer-approved acceptance scope. This is not a claim of full-waveform parity or full-455-question acceptance.