Run wheel builds on pull requests

[andimarafioti]

Adds a pull_request trigger for the CUDA wheel workflow while keeping workflow_dispatch inputs for manual builds. The PR path also defines default qwentts.cpp ref, CUDA architectures, and wheel build tag because dispatch inputs are not available on pull_request events.

[andimarafioti]

Added generated wheel audio samples in samples/pr-wheel-audio/ for the runnable wheels on the local GB10 machine.

Settings: Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice, BF16, speaker aiden, English, seed 42, stream chunk 8 / 12.5s, left context 2.0s.

Greedy parity prompt: This is a short deterministic backend comparison.

Backend	Greedy sample	Duration
CPU	cpu_greedy_20f_customvoice_bf16.wav	1.60s
CUDA 12.8	cu128_greedy_20f_customvoice_bf16.wav	1.60s
CUDA 13.0	cu130_greedy_20f_customvoice_bf16.wav	1.60s

Sampled prompt: Hello from the qwentts cpp Python wheel. This sample compares the same Qwen three custom voice settings across backends.

Backend	Sampled sample	Duration
CPU	cpu_sampled_seed42_customvoice_bf16.wav	3.68s
CUDA 12.8	cu128_sampled_seed42_customvoice_bf16.wav	6.40s
CUDA 13.0	cu130_sampled_seed42_customvoice_bf16.wav	6.40s

manifest.json in the same folder has the measured generation metadata.

[andimarafioti]

CPU audio issue corroborated with STT and fixed in this branch.

What failed before, using the CPU wheel built from eda8b59092b8d7b7142177c599b08909b87f63b4:

File	nano-parakeet transcript
cpu_greedy_20f_customvoice_bf16.wav	D
cpu_sampled_seed42_customvoice_bf16.wav	What uh

The CPU failure tracks to the newer native backend text-projection revision. A local CPU wheel built from a62fde62e64bf81a49af6e6c07ac5817df93f19e produces intelligible speech on the same aarch64 machine, so CPU wheel builds are now pinned to that native ref while CUDA wheels stay on eda8b59092b8d7b7142177c599b08909b87f63b4.

Final STT validation with nano-parakeet / nvidia/parakeet-tdt-0.6b-v3:

File	nano-parakeet transcript
cpu_greedy_20f_customvoice_bf16.wav	This is a short deterministic.
cpu_sampled_seed42_customvoice_bf16.wav	Hello from the Quentz CPP Python wheel. This sample compares the same Quen 3 custom
cu128_greedy_20f_customvoice_bf16.wav	This is a short deterministic.
cu128_sampled_seed42_customvoice_bf16.wav	Hello from the Quent's TP by Python wheel. This sample compares the same Quen 3 custom voice setting.
cu130_greedy_20f_customvoice_bf16.wav	This is a short determination.
cu130_sampled_seed42_customvoice_bf16.wav	Hello from the Quint GPP Python wheel. This sample compares the same Quen 3 custom voice settings.

Updated files are in samples/pr-wheel-audio/, and manifest.json now includes the STT transcripts.

[andimarafioti]

Replaced the CPU-only old-ref workaround with a native patch applied to the current qwentts.cpp ref for all wheel builds.

What changed:

• CPU and CUDA now both default to eda8b59092b8d7b7142177c599b08909b87f63b4.
• CI applies patches/qwentts-cpu-host-prompt-projection.patch after checkout.
• CUDA keeps the backend prompt projection path.
• CPU uses host F32 prompt text projection and host ref-code embedding sum when no GPU backend is present.

Local validation on the aarch64 GB10 machine:

• Before the fix, the current-ref CPU wheel transcribed as D and What uh with nano-parakeet.
• After the fix, the current-ref CPU wheel transcribed as:
  • This is a short determination.
  • Hello from the Quince TPP Python wheel. This sample compares the same Quen 3 custom voice set.
• Final STT pass succeeded for all six checked-in WAVs: CPU, CUDA 12.8, and CUDA 13.0, greedy and sampled.

Local checks:

• python3 -m py_compile ...
• manifest/WAV metadata validator
• patch apply check against a clean eda8b59 checkout
• uv run --with pytest --with numpy --with huggingface-hub pytest -q -> 1 passed, 1 skipped

[andimarafioti]

Uploaded reset-before-alloc CPU validation artifacts here:

• Timings and STT: https://raw.githubusercontent.com/andimarafioti/qwentts.cpp/cpu-reset-audio-artifacts/timings.md
• Greedy WAV: https://raw.githubusercontent.com/andimarafioti/qwentts.cpp/cpu-reset-audio-artifacts/samples/reset_before_greedy_20f_customvoice_bf16.wav
• Sampled WAV: https://raw.githubusercontent.com/andimarafioti/qwentts.cpp/cpu-reset-audio-artifacts/samples/reset_before_sampled_seed42_customvoice_bf16.wav

Summary, CPU BF16 CustomVoice aiden, qwentts.cpp eda8b59 plus scheduler reset before talker graph allocation:

file	mode	audio	first yield	native total	STT
reset_before_greedy_20f_customvoice_bf16.wav	greedy, 20 frames	1.60s	3059ms	7833ms	This is a short deterministic.
reset_before_sampled_seed42_customvoice_bf16.wav	sampled seed 42, 80 frames	6.40s	5160ms	24610ms	Hello from the QNTTS CPP Python wheel. This sample compares the same QN three.

The reset-only candidate keeps the batched prompt projection path. The prompt embeddings are identical to the failing CPU run, but the first talker hidden/logits and first-frame codes match the good split-special path after resetting the scheduler before ggml_backend_sched_alloc_graph.

[andimarafioti]

Added all-options audio/timing artifacts:

• Summary table: https://raw.githubusercontent.com/andimarafioti/qwentts.cpp/cpu-reset-audio-artifacts/all-options/timings.md
• Full manifest: https://raw.githubusercontent.com/andimarafioti/qwentts.cpp/cpu-reset-audio-artifacts/all-options/manifest.json
• WAV folder: https://github.com/andimarafioti/qwentts.cpp/tree/cpu-reset-audio-artifacts/all-options/samples

Covered rows:

option	source	greedy first yield / total	sampled first yield / total	STT
cpu_reset_before_local	local native CPU, reset-before-alloc candidate	2880ms / 5893ms	4172ms / 28631ms	pass
cuda13_reset_before_local	local native CUDA 13.0, reset-before-alloc candidate	692ms / 1102ms	259ms / 2713ms	pass
cu128_wheel	installed CUDA 12.8 wheel	887ms / 1318ms	281ms / 2770ms	pass
cu130_wheel	installed CUDA 13.0 wheel	779ms / 1199ms	266ms / 2704ms	pass

Notes:

• Downloads are not included; all runs use local GGUF files.
• Each option loaded the model once, generated the greedy clip, then generated the sampled clip.
• The CUDA 12.8 wheel needed the venv-provided CUDA 12 runtime path in LD_LIBRARY_PATH on this machine.
• I did not include the stale cpu_wheel venv as a passing row because it still contains the old failing CPU backend path. The CPU validation row is the reset-before-alloc native candidate.

[andimarafioti]

Pushed an update replacing the broad host-projection patch with the minimal scheduler-reset patch.

Latest package commit: 2263ea5 (Use scheduler reset patch for wheel builds)

What changed:

• removed patches/qwentts-cpu-host-prompt-projection.patch
• added patches/qwentts-reset-scheduler-before-talker-alloc.patch
• CI applies the scheduler-reset patch for CPU and CUDA builds
• checked-in CPU sample metadata now points at the reset-before validation path

I also opened the native upstream PR here:
ServeurpersoCom/qwentts.cpp#4

Once that lands upstream, this package can drop the local patch and just build from the fixed qwentts.cpp ref.

[andimarafioti]

Upstream qwentts.cpp PR #4 was merged, so I pushed another package update.

Latest package commit: 98dbb7d (Build from merged qwentts.cpp scheduler fix)

What changed:

• wheel CI now defaults to qwentts.cpp f79b23a56702584d51e3a66509b881cd99a95bba
• removed the package-side scheduler patch entirely
• removed patch application from both CPU and CUDA jobs
• updated README and sample metadata to note that the fix is upstream

I also verified locally that qwentts.cpp f79b23a builds as CPU with no patch and produces valid speech on the short greedy STT check.

A fresh six-job wheel build is running from 98dbb7d:
https://github.com/andimarafioti/qwentts-cpp-python/actions/runs/27198934466