[zhangzhenghao]

Summary

• Support and document all five published CLAP2 sizes: Large (960), Medium (640), Small (512), XS (320), and XXS (224), in F32/F16.
• Infer per-size source identity/revision instead of hardcoding Large; retain explicit revision overrides and avoid assigning Large provenance to custom geometries. The native implementation was already dimension-generic, so no inference kernel changes are needed.
• Audit every published tensor name/shape offline; exercise all five RoPE geometries including XXS head56, and run official asset-specific numerical oracles for reduced head56 and asymmetric head64 towers.

Verification

• Fresh Linux system-ICU static build: 36/36 default CTests.
• Fresh Linux system-ICU shared build: 35/35 default CTests.
• Optional official CLAP2 tests: 7/7, including all three tiny profiles, each F32/F16, with tokenizer/padding and crop-sensitive >30-second audio coverage.
• All five published checkpoints: 20/20 native comparisons (five sizes × F32/F16 × text/audio), against pinned official open_clip and timm 1.0.30. Six diverse texts and three waveform lengths per storage mode.
• Real XXS F16 CLI text/audio smoke: both return unit-normalized 224-dimensional vectors.
• Independent staged-code review and scoped follow-up reviews.

Published checkpoint maximum absolute embedding errors

Size	Text F32	Audio F32	Text F16	Audio F16
large	8.9407e-08	1.01961e-05	5.96046e-08	1.01961e-05
medium	7.45058e-08	1.20811e-05	8.9407e-08	1.20923e-05
small	8.9407e-08	7.33929e-06	1.19209e-07	7.36583e-06
xs	1.49012e-07	1.34697e-05	1.49012e-07	1.35042e-05
xxs	9.68575e-08	1.67266e-05	1.19209e-07	1.668e-05

Published-checkpoint tolerance remains 8e-5. The synthetic asymmetric-head64 multiple pure-tone case alone has an explicit 2e-4 waveform tolerance: F32 MKL FFT residuals near the −100 dB clamp differ from native double FFT. Independent boundary injection gave <3e-7 tower error with identical official mel, and Torch double FFT reproduced native-like differences. Text, other waveform cases and published thresholds are unchanged; negative probes confirmed their strict checks still reject altered outputs. No production FFT or inference behavior was changed.

Independent frontend/tower boundary evidence

For the unchanged asymmetric-head64 multiple input and weights:

Boundary experiment	Maximum absolute embedding error
Native complete waveform path	1.0035932e-4
Official Torch mel → native patchify/audio tower	2.83122e-7
Native power spectrum + Torch mel filters	1.0034442e-4
Torch F32 power spectrum + native mel filters	3.27826e-7

The worst near-floor mel bin differed by 0.28366 dB (native −99.24464, Torch −99.52831). Exact Torch Hann with native double FFT still showed the discrepancy; official double FFT reproduced native-like results. This evidence was independently reviewed using retained raw mel artifacts. It is not presented as a production inference change.

CLAP2 block quantization is not added by this change. Cross-platform CI status is reported separately from the local/published checks above.

Remote CI

Commit 8f0e96680c61c96edc9c2a5493fe88368c0af1f0, run 42484177700: 9/9 checks passed.

• Native build/tests/install: Linux x64/ARM64, macOS ARM64, Windows x64/ARM64.
• Python/Node package tests: Linux x64/ARM64, macOS ARM64, Windows x64. Windows ARM64 intentionally skips these binding jobs in the existing workflow.
• Rust: Linux, macOS, Windows.
• clang-format.