Skip to content

docs(tts): document Kokoro ANE A12 memory and routing limitations - #1009

Merged
Alex-Wengg merged 1 commit into
FluidInference:mainfrom
mvanhorn:fix/979-kokoro-ane-a12-device-limitations
Oct 11, 2026
Merged

Alex-Wengg merged 1 commit into
FluidInference:mainfrom
mvanhorn:fix/979-kokoro-ane-a12-device-limitations

Conversation

@mvanhorn

Copy link
Copy Markdown
Contributor

Why is this change needed?

The diff documents the legacy A12 observations, clarifies that neither a minimum supported chip nor a universal Vocoder memory requirement has been established, and separates the Noise stack-guard fault from the existing OS issues. It includes the maintainer's proposed CPU-only Vocoder override while retaining the iOS 18 defaults, explicitly identifying it as an unverified diagnostic experiment. It also limits these observations to the legacy runtime and explains why simulator results do not validate physical-device compatibility. Refs #979.

The Kokoro ANE documentation did not explain the legacy runtime's hardware limits or distinguish successful model loading from working synthesis. On the reported iPhone XR configuration, loading Vocoder with CPU+ANE killed the process; moving Vocoder to CPU allowed loading but synthesis later crashed in Noise. The documentation omitted the A12 memory-limit report and the distinction between forced aneTailCpu routing and the iOS 18 defaults, which request CPU+GPU for Noise and Tail. It also did not distinguish the reported BNNS worker-stack exhaustion from the existing input-overread and segfault cases.

Refs #979

The diff documents the legacy A12 observations, clarifies that neither a
minimum supported chip nor a universal Vocoder memory requirement has
been established, and separates the Noise stack-guard fault from the
existing OS issues. It includes the maintainer's proposed CPU-only
Vocoder override while retaining the iOS 18 defaults, explicitly
identifying it as an unverified diagnostic experiment. It also limits
these observations to the legacy runtime and explains why simulator
results do not validate physical-device compatibility. Fixes FluidInference#979.

Refs FluidInference#979
@Alex-Wengg
Alex-Wengg merged commit e560c64 into FluidInference:main Oct 11, 2026
12 of 13 checks passed
@mvanhorn

Copy link
Copy Markdown
Contributor Author

@Alex-Wengg thanks for taking the Kokoro ANE A12 notes, hope they save the next person some digging.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants