* fix(indextts): accept the config name upstream ships, and keep long text alive
Two independent defects, both reported on a working IndexTTS 2.5 install.
Install always failed. IndexTeam/IndexTTS-2.5 ships the model config as
config.yaml — at the pinned revision d0aa86e7 and at HEAD; config_v2_5.yaml
exists in no upstream revision. VoiceStudio demanded that name, so
_weights_floor_ok never found it and the install died claiming 'the download
was likely interrupted' when the download had been perfect. The only way
through was to hand-rename the file. Both names are accepted now, in the
installer and on the load path, so installs created with the workaround keep
working without a reinstall.
Long text was killed at 60s. infer() is one blocking upstream call that puts
nothing on the wire, and IndexTTS was the only sidecar still on the 60s
recv_timeout_s class default while pockettts and omnivoice-subprocess had both
raised theirs. Raising the default alone does not fix it — which is why the
reporter's RECV_TIMEOUT_S=3600 edit didn't help: progress frames are also what
report activity to the GPU pool's execution clock (#1367), so a silent sidecar
still trips the outer generate budget. The sidecar now heartbeats every 5s
while infer() runs (and during the cold model construction), _send takes a
lock so the beat thread can't interleave framing, and the deadline rises to
900s via OMNIVOICE_INDEXTTS_RECV_TIMEOUT_S.
test_indextts25_health_requires_25_config_name asserted the bug — that a
checkout holding only config.yaml is unhealthy — so it is rewritten to the
corrected contract, including that a genuinely truncated download is still
caught.
Fixes#1611
* test(indextts): follow the installed config name in the sidecar loader tests
Two more tests encoded the config_v2_5.yaml assumption, both asserting
cfg_path against a directory where no config existed at all — so they were
pinning the literal name rather than the resolution. They now lay down a real
checkpoints/ tree and assert the resolved path, including that a checkout
carrying the pre-fix hand-renamed config still resolves.
Caught by the full suite; the targeted runs during development did not reach
tests/backend/services/.
* test(indextts): event-driven heartbeat tests, real interleaving proof, precedence pin
Review round on #1619 — all four findings taken.
- The docs line naming 0.5.1 is version-neutral now ('Earlier installs') —
version labels are the owner's call.
- The heartbeat tests waited on wall-clock sleeps; they now block on a
per-write Event with a bounded deadline, so scheduler load can't flake them.
- The _send test asserted the lock EXISTS — a tautology. It now drives four
concurrent writers through a stream that yields between every byte and
asserts every frame decodes; verified fail-before by removing the lock
(torn frame) and pass-after.
- The precedence test deleted config.yaml before creating the renamed one, so
reversed precedence still passed. Both files now coexist for the assertion;
verified fail-before by reversing _CFG_NAMES.
* fix(openai-compat): reuse cached engine instances in _resolve_engine
The direct engine-ID path in /v1/audio/speech constructed a fresh
backend per request (return cls()). For SubprocessBackend engines that
meant: a new sidecar process, a full torch import and an engine model
reload on EVERY request (measured ~28s floor per pockettts request on
an M3 Pro), plus another atexit hook registration each time — exactly
what get_engine_instance_for()'s docstring warns against.
Route the explicit-ID path through the same cached-singleton seam the
active-engine path already uses. Unknown/unavailable IDs keep their
400s; tts-1/tts-1-hd and the OmniVoiceBackend special case are
unchanged.
* fix(openai-compat): unload the outgoing engine on explicit-ID switches
Review follow-up (Greptile/CodeRabbit on #1614): caching instances without
a switch rule would let each distinct explicit engine ID stay resident,
accumulating sidecars / multi-GB in-process models. Mirror
get_active_tts_backend's MM2-01 switch rule: a different explicit ID
(omnivoice included, which resolves to the active engine) unloads the
outgoing instance first, best-effort.
* fix(openai-compat): evict via the shared single-engine-resident seam, not a router-local cache
The explicit-ID unload cache (13c14e2c) kept its own instance ref keyed by
model id. The shared engine cache is deliberately keyed by CLASS (registry
rebinds, idle sweeps and engine_memory eviction all mutate it), so the
router's id-keyed ref could go stale and keep serving an instance the
lifecycle system no longer tracked — caught by
test_openai_speech_toggle_off_sends_raw_text in full-suite order, and it
also introduced a novel unload path that ignored the
OMNIVOICE_SINGLE_ENGINE_RESIDENT opt-out.
Drop the router-local cache entirely: _resolve_engine returns the shared
cached singleton (get_engine_instance_for), and create_speech calls
evict_other_tts_engines(backend.id) before warming the engine — the exact
seam /generate uses. That covers every transition (explicit id → explicit
id, explicit id → tts-1/omnivoice aliases), honors the policy opt-out, and
leaves no per-router state to drift. Regression pinned at the route level in
test_speech_request_evicts_other_resident_engines.
* chore(changelog): trim the #1614 entry to the one-liner limit
415 chars against the 400 the style test allows — CI would have failed on it.
---------
Co-authored-by: debpalash <4178343+debpalash@users.noreply.github.com>
* fix(setup): tolerate reserved memory in the RAM preflight, add OMNIVOICE_RAM_PREFLIGHT=0 escape hatch (#1618)
An "8 GB" machine reports ~7.8 GB usable (firmware/iGPU/kernel
reservations), so comparing OS-reported RAM against the marketing-size
8 GB threshold hard-blocked exactly the boundary hardware the minimum is
meant to admit — with no way past the wizard. Both thresholds are now
compared with a 7% reserved-memory allowance, and
OMNIVOICE_RAM_PREFLIGHT=0 downgrades a genuine fail to a warning for
users who accept the OOM risk (same opt-out shape as
OMNIVOICE_ASR_VRAM_PREFLIGHT).
Regression tests: backend/tests/test_ram_preflight_1618.py.
Docs: troubleshooting §1c.
* review: hermetic preflight stubs in tests; correct the escape-hatch doc
Greptile P1: the Settings panel can't set OMNIVOICE_RAM_PREFLIGHT (and the
blocker appears before setup completes anyway) — the doc now points at
PowerShell / shell env only.
CodeRabbit: stub _network_check and media_tools.summary so each RAM
assertion stays fast and offline (26s -> 6s locally).
* fix(watermark): run AudioSeal eagerly instead of through torch.compile
AudioSeal vendors moshi's @torch_compile_lazy on SEANetEncoder.forward, so
the first embed of a session — not the model load, which #1576's prefetch
already warms — called torch.compile and dropped into Inductor's C++ codegen.
On a macOS arm64 deployment that compile raised CppCompileError on 10/10
takes: the embed fail-opened and the audio shipped UNMARKED, an EU AI Act
Art. 50(2) provenance gap, after burning 30-40s on the first take and 5-8s on
each later one.
The compile is pure cost even where it succeeds. Measured on an M3 (5s of
24kHz audio, three consecutive embeds): compiled 9.70/0.26/0.23s vs eager
0.30/0.28/0.27s — a ~10s first-embed tax to save ~0.03s afterwards, on CPU
work already bounded by the 30s chunk loop. Both embed and detect now run
inside audioseal's own no_compile() switch, restored on the way out (it is a
process global, and other models are entitled to compile).
Verified end-to-end: first embed 9.70s -> 0.26s, watermark still round-trips
at confidence 1.0 with the OmniVoice message intact.
Fixes#1615
* fix(watermark): collapse the eager-guard globals into one lock-guarded state
CodeQL flagged _eager_saved's module-level initializer as dead, and it was
right: depth 0->1 always writes the field before depth 1->0 reads it, so the
None at import was never observed. Depth and saved-value are only meaningful
together and only under _eager_lock, so they become one dict rather than two
module scalars — which also drops the global statement.
Also splits three semicolon-joined statements in the regression test (Ruff
E702, CodeRabbit).
Mutation re-checked after the refactor: a naive no_compile() body still fails
with 'compile was handed back mid-embed'.
* fix(dictation): refresh accessibility blocker
* docs(changelog): note accessibility refresh
* test(dictation): assert the native widget hide on accessibility grant
The recheck regression asserted only that the Accessibility pill text left
the DOM, so it still passed with hideWidgetWindow() removed and the native
capsule stranded on screen. Hold one stable getCurrentWindow().hide spy and
assert it after the poll (fails before the fix, passes after).
Palash Debnath