History

Fabio c36510c21e Some checks failed Linux Build / Linux cpu (push) Has been cancelled Details Linux Build / Linux cpu-clang (push) Has been cancelled Details Linux Build / Linux vulkan (push) Has been cancelled Details mac-build / macOS CPU (push) Has been cancelled Details Nix Build / Nix (cpu) (push) Has been cancelled Details Nix Build / Nix (vulkan) (push) Has been cancelled Details windows-build / Windows CPU (push) Has been cancelled Details Initial import: audio.cpp with Orange Pi documentation		2026-11-04 19:05:22 +01:00
..
.gitignore	Initial import: audio.cpp with Orange Pi documentation	2026-11-04 19:05:22 +01:00
compare_warmbench_results.py	Initial import: audio.cpp with Orange Pi documentation	2026-11-04 19:05:22 +01:00
higgs_audio_tts_cuda_bench_cases.json	Initial import: audio.cpp with Orange Pi documentation	2026-11-04 19:05:22 +01:00
higgs_audio_tts_cuda_cache_transition_cases.json	Initial import: audio.cpp with Orange Pi documentation	2026-11-04 19:05:22 +01:00
higgs_audio_tts_cuda_mixed_cases.json	Initial import: audio.cpp with Orange Pi documentation	2026-11-04 19:05:22 +01:00
higgs_audio_tts_cuda_perf_cases.json	Initial import: audio.cpp with Orange Pi documentation	2026-11-04 19:05:22 +01:00
higgs_audio_tts_python_warm_bench.py	Initial import: audio.cpp with Orange Pi documentation	2026-11-04 19:05:22 +01:00
higgs_audio_tts_sampler_logits.bin	Initial import: audio.cpp with Orange Pi documentation	2026-11-04 19:05:22 +01:00
higgs_audio_tts_warm_bench.cpp	Initial import: audio.cpp with Orange Pi documentation	2026-11-04 19:05:22 +01:00
higgs_audio_tts_warm_bench_cases.json	Initial import: audio.cpp with Orange Pi documentation	2026-11-04 19:05:22 +01:00
prefill_memory_bench.py	Initial import: audio.cpp with Orange Pi documentation	2026-11-04 19:05:22 +01:00
README.md	Initial import: audio.cpp with Orange Pi documentation	2026-11-04 19:05:22 +01:00
run_cuda_performance.ps1	Initial import: audio.cpp with Orange Pi documentation	2026-11-04 19:05:22 +01:00
test_sampler.cpp	Initial import: audio.cpp with Orange Pi documentation	2026-11-04 19:05:22 +01:00

README.md

Higgs Audio v3 TTS tests

The focused framework unit test validates that packed QKV/gate-up projections match the separate projections, suffix causal masks are correct, F16 KV writes preserve their values, and the decode graph exposes the intended CUDA paths: grouped FlashAttention, packed SwiGLU, direct KV updates, and the ROPE -> VIEW -> SET_ROWS fusion pattern.

cmake --build build/windows-cuda-release --config Release --target decoder_packed_projection_test higgs_audio_tts_warm_bench -j 8
ctest --test-dir build/windows-cuda-release -C Release -R decoder_packed_projection_test --output-on-failure

Run the fixed-seed, five-request CUDA benchmark and save every generated WAV:

tests/higgs_audio_tts/run_cuda_performance.ps1 `
  -Model ../models/higgs-audio-v3-tts-4b_Q8/higgs-audio-v3-tts-4b_Q8.gguf `
  -Label candidate

Compare a candidate run with a prior result directory request by request:

tests/higgs_audio_tts/run_cuda_performance.ps1 `
  -Model ../models/higgs-audio-v3-tts-4b_Q8/higgs-audio-v3-tts-4b_Q8.gguf `
  -Label candidate `
  -Baseline tests/higgs_audio_tts/results/baseline

The comparison reports frame counts, wall time, RTF, speedup, waveform cosine, and 80-band log-mel cosine. Result WAVs, logs, and JSON reports are written below tests/higgs_audio_tts/results/, which is intentionally ignored by Git. The comparison helper requires Python 3 with NumPy.

Add -RequireSameFrames when comparing paths that are expected to be deterministic and numerically identical. Sampled or mixed-precision paths still report their frame drift and similarity metrics without hiding the results.