The wrapper's `optimized` backend (torch.compile + CUDA graphs +
real-time streaming) reads its model registry from a YAML config:
default path is ~/qwen3-tts/config.yaml inside the container, which
doesn't exist. Without TTS_CONFIG set, the backend boots with an
empty registry and every synthesis request fails with
"Unknown model key: '<name>'. Available: []".
The repo ships /app/config.yaml with all 4 model variants defined.
Pointing TTS_CONFIG at it lets the optimized backend load cleanly.
This is a prerequisite for benching the optimized backend properly
— it's the path to the upstream's claimed 97 ms streaming TTFB. The
default `official` backend uses naive HF transformers autoregressive
generation that pegged GPU at only 27% utilization and gave us 8-12 s
TTFB on bench (no recompile theory needed — same phrase repeated 4x
plateaued at 8.5 s, ruling out shape-specific recompilation).
Three fixes from the first deploy attempt on irv-ml1:
- build.target=production. Upstream Dockerfile is multistage; the last
stage `cpu-base` was selected by default, producing a CPU-only image
with no flash-attn and `torch ... whl/cpu`.
- user: "0:0". Upstream image declares USER appuser but writes runtime
state under /root (mode 0700). appuser cannot traverse /root, so
/v1/voices 500s on PermissionError. Run as root to sidestep.
- QWEN3_TTS_MODEL=Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice. The bare
`1.7B` id we had isn't a real HF identifier; upstream publishes
-CustomVoice / -Base variants of each size. Use -CustomVoice so
`voice="clone:<name>"` works.
Tag bumped to v2 to keep the v1 cpu image distinguishable in the local
registry.
After: all 5 verify steps pass, GPU synthesis ~5s for 3-4s of audio,
three contrasting English `instructions` produce three distinct
hashes — emotion steering actually works (unlike CosyVoice's English
path).
Alibaba's open-weight TTS (Apache 2.0, Jan 2026), deployed via
groxaxo/Qwen3-TTS-Openai-Fastapi wrapper. Built locally from a
pinned git SHA via docker buildx's git context — no source
vendored. 1.7B flagship model by default; 0.6B available via
QWEN3_TTS_MODEL env override.
Why we need a second TTS stack: cosyvoice 3 emits Chinese phonemes
for English content per upstream FunAudioLLM/CosyVoice#1790
(unfixed). Qwen3-TTS is from the same Alibaba team but with
English first-class in the checkpoint — 10 languages, 97 ms
streaming TTFB, instruction-driven emotion. Coexists with cosyvoice
on irv-ml1 (port 8191; cosyvoice keeps 8190).
Voice cloning shape DIFFERS from cosyvoice: profile-based, not
voice-id. Profiles live under voice_library/profiles/<name>/ and
are referenced as voice="clone:<name>".
Path layout: /worktank/qwen3-tts/{cache,voices}/, with cache excluded
from restic (regenerable from HF Hub) and voices included (cloned
profiles need original reference audio to recreate).
playbooks/deploy-qwen3-tts.yaml: 10 steps + 5 verify, idempotent;
the wait step polls /health for up to ~10 min to absorb first-run
model download.
Stack only — restic profile update for /worktank/qwen3-tts/voices/
to follow when this is empirically validated against the GLaDOS
voice (the "did Qwen inherit the Chinese-bias bug?" question).
vh