The Unmute workflow starts on a real host - #328
Merged
Conversation
Brought up on a two-GPU node. Three things the upstream compose assumed were not true there: the speech services' start script logs into Hugging Face unconditionally and dies on an empty token, though the weights and stock voices are public, so the login is now optional; GPU reservations by CDI name failed on a stale device spec, so services are pinned by NVIDIA runtime device index, with separate inputs for the STT and TTS cards so the two can be split across GPUs; and the backend reaches a Studio on the host through host.docker.internal. The compiled moshi-server lives in a named volume so only the first start builds it. Verified: STT healthy on a 6 GB card at about 2.5 GB.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Brought up on a two-GPU node (one busy card, one free 6 GB card). Three things the upstream compose assumed were not true there:
uvx hf auth login --token $HUGGING_FACE_HUB_TOKENunconditionally and dies on an empty token, although the Kyutai STT/TTS weights and stock voices are public (gated: false). A Studio start script mounted over the image's makes the login optional./dev/dri/card2missing). Services are pinned by NVIDIA-runtime device index instead, with separatestt_gpu/tts_gpuinputs so the two can be split across cards (STT ~2.5 GB, TTS ~5.3 GB).host.docker.internal(host-gateway).The compiled
moshi-serverlives in a named volume so only the first start builds it (about 90 s here with a warm crate cache). Verified: STT healthy on the 6 GB card at 2,516 MiB. Workflow-only.