OmniVoice TTS: boot/long-reply OOM fixes + precomputed clone prompts - fix: pre-computed VoiceClonePrompt workflow (scripts/prepare_omnivoice_prompt.py) avoids the ~3.2 GiB DAC-encode transient at boot; handler --omnivoice_voice_clone_prompt loads the … #17

Merged
troed merged 6 commits from devel into main 2026-09-08 11:34:55 +02:00
Owner
No description provided.
Encoding --omnivoice_ref_audio at startup needs a ~3.2 GiB transient
audio-tokenizer activation (independent of quantization) that OOMs a
12 GB GPU alongside other resident models. Add a script to pre-compute
the VoiceClonePrompt offline (~18 KB) for --omnivoice_voice_clone_prompt,
and warn when the boot-time encode path is used.
fix: offload idle OmniVoice encoder modules to reclaim VRAM for long replies
All checks were successful
CI / Sanity check (ubuntu-latest) (pull_request) Successful in 4m31s
3793f8a637
troed merged commit 5b2353d8a4 into main 2026-09-08 11:34:55 +02:00
Sign in to join this conversation.
No reviewers
No labels
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
starfleet/computer!17
No description provided.