vh
14a0004a47
The Zonos TTS engine that zonos-gateway fronts (irv-ml1 3090, feeds asset-engine + gateway-chat) ran as a bare native process with its real invocation existing ONLY in the running process argv — the committed harness/zonos_server.sh on irv-ml1 was STALE (said A6000/:1919, no perf flags; live is 3090/:1920 with cuda-graph/num-pages/ max-running-requests/memory-ratio). Captured the corrected canonical invocation + tunables + the containerization plan here so the config survives a process death. Engine = stock Zyphra/Zonos2 @ 194c0a3 (no custom PFI server code); torch 2.9.1+cu128; 15 GB HF weights. Next: containerize in-place on the 3090 (operator: keep off the A6000, it OOMs under ComfyUI). Not yet built — this commit is the config capture only.
zonos-engine — ZONOS2 native TTS engine (:1920, irv-ml1 3090)
The production Zonos TTS engine that zonos-gateway (stacks/zonos, :8890)
fronts via ZONOS_URL=http://host.docker.internal:1920, feeding asset-engine +
gateway-chat. This is the ENGINE; stacks/zonos/ is the OpenAI-compatible ADAPTER.
Status (2026-07-16): NATIVE, config was live-only — now captured here
The engine currently runs as a bare native process on irv-ml1 (launched detached, reparented to init — no systemd, won't survive a reboot). This dir captures its real config so it isn't lost, ahead of containerizing it (below).
⚠️ The old ~/tts-audition/harness/zonos_server.sh on irv-ml1 is STALE — it
says A6000 / port 1919 / no perf flags, but the live process runs on the
3090 / port 1920 with the perf flags below. zonos2-server.sh here is the
CORRECTED, canonical invocation (verified against the live process's argv).
The engine
- Upstream (stock):
github.com/Zyphra/Zonos2.git@ commit194c0a3, cloned at~/tts-audition/models/zonos2on irv-ml1 (uv-managed.venv). No custom PFI server code — thezonos2package ships its own server (python -m zonos2). The only "custom" part is the launch flags + this stack. - Stack: torch 2.9.1+cu128 (CUDA 12.8); heavy compiled-CUDA deps
(flashinfer / sgl_kernel / cutlass-dsl / apache-tvm-ffi / pynini).
uv.lockin the repo pins the whole env deterministically. - GPU: 3090 (irv-ml1 device 0,
GPU-12c15b3c-…). Kept on the 3090 by operator decision 2026-07-16 — the A6000 already OOMs under ComfyUI load, so it can't host Zonos too. Footprint ~17 GB (the 3090 runs Zonos + chatterbox-fast + parakeet, ~1.4 GB free). - Model weights:
Zyphra/ZONOS2, ~15 GB HF cache at~/.cache/huggingface/hub/models--Zyphra--ZONOS2— MOUNT this into the container, do NOT bake it. - Endpoint:
:1920(/tts/generate), reused across prompts; emotion directions auto-load from the repo'semotion_directions/.
Live invocation (source of truth)
uv run python -m zonos2 \
--model-path Zyphra/ZONOS2 \
--host 0.0.0.0 --port 1920 \
--tts-default-voices-dir ./default_voices/ \
--cuda-graph-max-bs 1 \
--num-pages 16384 \
--max-running-requests 2 \
--memory-ratio 0.3
Containerization plan (pending build)
Target shape = a self-contained image (fleet-standard, like stacks/zonos +
comfyui): FROM a CUDA 12.8 base → uv sync against the committed uv.lock (env
baked, reproducible) → mount the 15 GB HF weights + default_voices/ → pin the 3090
(NVIDIA_VISIBLE_DEVICES=0) → restart: unless-stopped → CMD = the invocation above.
Cutover = stop the native process (frees ~17 GB on the 3090) → docker compose up -d
(re-allocates ~17 GB, same footprint) → repoint zonos-gateway ZONOS_URL at the
container (or keep the :1920 host-port publish). Brief prod-TTS blip at cutover.
Build risk = getting the compiled-CUDA deps to install in the base image (mostly
prebuilt wheels + the uv.lock, so likely a couple of build iterations).