fix(qwen3-tts-cpp): hold qwentts.cpp at 35ebe537, upstream master hangs in synthesis

[localai-bot]

Problem

tests-qwen3-tts-cpp has been red on master since 2026-07-31. The suite loads every component successfully and then stops:

Will run 27 of 27 specs
••••••••••••••••[qwen3-tts-cpp] Loading talker=... codec=...
[Talker] Loaded: 28 layers, hidden 1024, ...
[CodePredictor] Loaded: 5 layers, ...
[Quantizer] Loaded: 16 codebooks ...
[Transformer] Loaded: 8 layers, ...
panic: test timed out after 20m0s

TTS() never returns from the native call:

goroutine 74 [syscall, 19 minutes]:
github.com/ebitengine/purego.RegisterFunc.func4
github.com/mudler/LocalAI/backend/go/qwen3-tts-cpp.(*Qwen3TtsCpp).TTS
    backend/go/qwen3-tts-cpp/goqwen3ttscpp.go:154
github.com/mudler/LocalAI/backend/go/qwen3-tts-cpp.init.func2.4
    backend/go/qwen3-tts-cpp/e2e_test.go:90

A job that normally takes about 5 minutes runs into the 20 minute Go test timeout instead.

Not a flake. Reproduced on master (run 30689716806) and again on an explicit re-run of the same job.

Which bump did it

Bisected across this cycle's eight qwentts.cpp bumps using each PR's own tests-qwen3-tts-cpp check:

Bump PR	Pin	tests-qwen3-tts-cpp
#10832	7bb91886	pass, 4m38s
#10850	9e11ce41	pass, 4m58s
#10902	95b4840a	pass, 5m9s
#10964	e93292be	pass, 5m15s
#11006	ba4c7f78	pass, 5m33s
#11039	82cd05b9	pass, 4m58s
#11127	35ebe537	pass, 5m1s
#11241	abab6b3	fail, 1h58m26s

#11241 was merged with this check already red, which is how the hang reached master.

Upstream cause

35ebe537..abab6b3 is three upstream commits:

• 26dd8adb predictor: unroll the frame into one cgraph and sample in standard ops
• 9a9a4252 tests: exercise the fused decode in every cossim harness
• abab6b3b tests: refresh the clone grid logs

Only 26dd8adb is functional; the other two are test-only. Rewriting the predictor's sampling into standard ops is consistent with a generation loop that never reaches its stop condition, which is exactly what an unbounded native call looks like from the Go side.

Fix

Hold the pin at 35ebe537, the last known-good commit.

The bump_deps entry is commented out rather than left in place: it tracks upstream master, so the nightly job would put the pin straight back onto the hang. Both the Makefile and the workflow carry a comment pointing at the other, so the hold is discoverable rather than mysterious, and restoring it is uncommenting four lines once upstream is fixed.

Why this blocks the release

This is not only a CI problem. TTS() hanging means a real qwen3-tts-cpp synthesis request never returns, so the backend wedges for users on the current master. Found during the pre-release CI sweep for v4.8.0.

Follow-up (not in this PR)

Report 26dd8adb upstream to ServeurpersoCom/qwentts.cpp with this bisection and stack trace, then restore the bump once it is fixed.

[localai-bot]

Reported upstream: ServeurpersoCom/qwentts.cpp#19

CI has now confirmed the hold works: tests-qwen3-tts-cpp passes in 5m15s on this PR, against 1h58m (timeout) on the abab6b3 pin. That matches the historical baseline for this job exactly, so 35ebe537 is confirmed good rather than merely older.

Restoring the bump is: revert the pin, and uncomment the four bump_deps lines.

[localai-bot]

Correcting the mechanism described in the commit message, now that the upstream fix is understood.

This commit (and issue #19 as originally filed) says TTS() never returns. That is imprecise. It is a runaway, not a deadlock, and the CI log says so directly if you read past the panic:

[Pipeline] Generated 2048 frames (slot 0)
[Pipeline] Generation done : 2048 frames
panic: test timed out after 20m0s

Generation runs to the 2048-frame cap instead of stopping at EOS around frame 17-28, so each synthesis does roughly 100x the work and the suite blows its 20 minute budget. The call does terminate, just far too late to be useful.

Root cause upstream, from ServeurpersoCom/qwentts.cpp#20: 26dd8ad moved the sub-talker draw into a graph tail that crosses the cdf in descending probability order, while the host sampler it replaced walks the cdf in vocabulary order. Both are valid multinomial samplers with the same marginal, but they map a given philox u to different tokens. Every stochastic acoustic code changed, the talker went off-manifold and stopped emitting codec_eos. Greedy was unaffected, which is why the commit's own greedy parity check passed.

None of this changes what this PR does or why. Holding the pin at 35ebe537 is still the right call for the release, and tests-qwen3-tts-cpp passing here in 5m15s is unchanged evidence. Once #20 lands upstream, restoring the bump is reverting the pin and uncommenting the four bump_deps lines.