Two fixes from the failed first deploy on irv-ml1:
1. CPU/GPU variant. Kokoro's GPU image needs CUDA >= 12.9; irv-ml1's
driver 570.124.06 caps at 12.8 so the gpu variant fails with
"nvidia-container-cli: requirement error: unsatisfied condition:
cuda>=12.9". Make the variant a knob:
KOKORO_VARIANT=cpu (default — works anywhere)
KOKORO_VARIANT=gpu (after driver bump)
KOKORO_USE_GPU=false|true (matches the variant)
Kokoro is tiny (82M params) so CPU is workable: TTFA ~1s vs ~300ms
on GPU. Acceptable while the driver bump gets scheduled. compose.yaml
no longer hard-codes `runtime: nvidia` — relies on the daemon's
default-runtime + NVIDIA_VISIBLE_DEVICES gating, same as how the
wrapper's USE_GPU flag selects the inference path inside the
container. Toggling between variants is now a `.env` edit + restart.
2. Tighter pull-log filter. --quiet on `docker compose pull` only
suppresses the pull command's stdout; the docker daemon still
emits per-layer extraction events on stderr ("ffbfd7a09415
Extracting 64.06MB" repeated dozens of times per layer). Drop those
too via grep on the SHA-prefixed pattern. set -o pipefail keeps a
real pull failure visible.
For existing deployments: removing /opt/docker/compose/kokoro/.env
on the host and rerunning the playbook re-seeds with the new schema.
Three TTS additions to round out coverage on irv-ml1, each filling a
distinct niche the existing slate doesn't own.
Final coverage matrix (all on irv-ml1):
Kokoro — low-latency English, fixed voice library, ~300ms TTFA
Chatterbox Turbo — low-latency English w/ voice cloning + paralinguistic tags
IndexTTS-2 — English voice cloning + emotion vector / text control
Qwen3-TTS-1.7B-Base — high-quality English voice cloning
CosyVoice 3 — multilingual (Chinese-leaning)
VibeVoice 1.5B — long-form / multi-speaker dialogue
stacks/kokoro:
- port 8193, GPU device 0 (3090)
- pulls ghcr.io/remsky/kokoro-fastapi-gpu:v0.2.4-master (no Dockerfile,
no first-run model download — models baked in)
- 60+ built-in voices, OpenAI-compat with stream=true over chunked HTTP
- Apache-2.0 weights + code, ~1 GB VRAM
stacks/vibevoice:
- port 8194, GPU device 1 (A6000 — for 7B headroom)
- builds groxaxo/VibeVoice-FastAPI1 (more current fork of ncoder-ai)
pinned to 7614c469a145
- default model microsoft/VibeVoice-1.5B (~7 GB bf16 VRAM); env var
swap to rsxdalv/VibeVoice-Large (7B) or FabioSarracino/VibeVoice-Large-Q8
- multi-speaker dialogue via /v1/vibevoice/generate with Speaker N: format
- long-form niche only — not low-latency
stacks/chatterbox:
- port 8196, GPU device 0 (3090)
- builds devnen/Chatterbox-TTS-Server (most active Turbo-supporting wrapper)
- default model ResembleAI/chatterbox-turbo (~2.5 GB fp16, ~75ms latency)
- paralinguistic tags inline ([laugh] [whisper] etc) — different shape
from IndexTTS-2's emotion vector; fills the speed+cloning niche
Kokoro/IndexTTS don't cover together
- mandatory PerTh watermark on outputs (Resemble policy)
Three matching playbooks under playbooks/deploy-{kokoro,vibevoice,
chatterbox}.yaml. All idempotent, creates-/when-gated.
Cold-deploy disk on /worktank/: ~7 GB Kokoro + ~19 GB VibeVoice 1.5B
+ ~12 GB Chatterbox = ~38 GB total. VRAM concurrent: ~10-11 GB across
both GPUs.
Skipped from the original four-stack proposal: VibeVoice Realtime
(overlaps Kokoro's niche; Kokoro wins on latency, license, and not
needing a build).
vh