Placed on GPU 3 first because it was the empty card. That was the wrong read:
the seat is ~800 MiB, under 1% of a 96 GB card, so the question was never "where
does it fit" but "whose headroom is cheapest to spend".
vLLM sizes its KV cache as a fraction of TOTAL VRAM, not free VRAM. A resident
tenant on an otherwise-clean card therefore does not cost its own megabytes — it
costs the profiling margin of whatever full-size seat lands there later, and
flash-next needs 93 GiB of 96. A 96 GB card at 2 MiB can still take that; the
same card at 922 MiB is one where the next big seat needs its utilization
hand-trimmed, which this repo's flash-next history shows is both thin and silent
when it goes wrong.
Committed utilization per card is the number that governs, not free bytes:
GPU 0 0.40 + 0.48 = 0.88 ~13 GB spare <- moved here
GPU 1 0.52+0.24+0.10+0.055+0.03+0.03 = 0.975 ~4.3 GB
GPU 2 0.96 ~1.8 GB
GPU 3 - kept empty as reserve
GPU 3 is back to 2 MiB / 97,247 MiB free and is now documented as a deliberate
reserve rather than a spare.
Post-move n=5 on the same clip: 0.68 / 0.54 / 0.54 / 0.52 / 0.53 s, median 0.54 s
against 0.50 s on GPU 3. The spreads overlap at this sample size and no difference
is claimed; the GPU 3 figure was taken on an idle card and is now noted as a best
case, since the seat shares GPU 0 with the hot serving path. Silence control and
the gateway round-trip both re-verified after the move.
Also records both Parakeet endpoints (FV v3 on Blackwell, IRV v2 on a 3090) and
the four confounds that make them not an A/B pair, sent to tts-dev for the bench.
Retargets the existing sherpa-onnx stack from irv-ml1 to fv-ml1's utility card
and puts it behind the gateway. GPU 3 was the only card with room: 0/1/2 carry
the vLLM seats at 84-95.5 GB of 96.
Changes:
- compose: pin GPU via `device_ids: ["3"]` (the dead on-host stub used
`count: all`, which would have handed a 0.6B ASR seat all four cards);
join traefik-net; port 8300; homepage href to the live FV address.
- .env.example: default to the v3 int8 model (25 European languages, 464 MiB)
rather than English-only v2; models to /tank/parakeet/models.
- app.py: warm the recognizer at startup before uvicorn accepts traffic.
The warmup is not an optimisation. ONNX Runtime's CUDA EP compiles and autotunes
lazily on the FIRST DECODE, and on sm_120 that measured 45.7s cold (reproduced at
45.1s on a second container) against ~0.50s warm. A 45s first request is
indistinguishable from a hang and LiteLLM's default timeout abandons it long
before it returns. Decoding 1s of silence at load moves the cost inside the
healthcheck's 300s start_period; first real request after restart is now 0.65s.
Verification, because "provider=cuda" in the log is only an echo of the env var:
ORT falls back to CPU silently and still returns correct text, so the service
being up and the transcript being right establishes nothing. The discriminator is
a process on GPU 3 (922 MiB), confirmed. Controls both directions — a known TTS
sentence transcribes near-exactly (positive), 3s of digital silence returns
empty (null). Warm throughput 0.50s median on an 8.52s clip, n=5, spread
0.47-0.65s, single-stream, one clip: a smoke measurement with its harness
stated, not a benchmark.
Gateway aliases `ext-stt` (engine-neutral, mirrors ext-tts) and `whisper-1`
(OpenAI-compatible drop-in) registered via POST /model/new, i.e. LiteLLM's
Postgres store where the ext-tts family already lives — no gateway restart, and
config.yaml is consequently not a complete picture of what the gateway serves.
Both verified end to end.
The aliases use a raw IP deliberately: ana-docker resolves no .internal names at
all (resolv.conf points at 1.1.1.1), and LiteLLM only reaches irv-ml1 through a
hand-pinned extra_hosts entry. A second hosts entry would mean recreating the
container and bouncing the gateway for every consumer.
Also records the svos_miranda plugin validation pass and its structural findings,
and notes that the irv-ml1 parakeet is still running — there are two now, and
retiring the old one is the operator's call.
The Shadowfita FastAPI wrapper hit two unfixed upstream bugs on the
first real /transcribe call — chunker return-shape mismatch (open
issue #16) and a `torchaudio.tensor` that doesn't exist (open #10).
Rather than babysit someone else's half-tested code, switched to
sherpa-onnx with the prebuilt int8 Parakeet-TDT tarball from k2-fsa,
and wrote our own ~60-line FastAPI wrapper.
Moving parts now owned in-tree:
Dockerfile CUDA 12.8 + cuDNN 9 runtime base, installs
sherpa-onnx==1.12.39+cuda12.cudnn9 + fastapi +
soundfile + libasound2 (sherpa-onnx links to ALSA
at load time even when we never touch a mic).
app.py OfflineRecognizer.from_transducer() once at startup;
/transcribe and /v1/audio/transcriptions both accept
multipart uploads and return {"text": ...}.
entrypoint.sh Idempotent model download to /models on first run
(~400 MB int8 tarball), then exec uvicorn.
Smoke test: 0.wav (bundled in the tarball, The House of the Seven
Gables excerpt) transcribes cleanly in ~1.2s on GPU.
PARAKEET_MODEL_URL in .env lets you swap to the v3 (25-language)
tarball without touching any other files. Wipe *.onnx + tokens.txt
from the models dir and the entrypoint re-downloads.
Two new speech stacks on irv-ml1, both on the /worktank/<stack>/
pattern, no tnet (irv-ml1 is local-endpoints-only for now).
parakeet — ASR via Shadowfita/parakeet-tdt-0.6b-v2-fastapi:
- docker buildx git context pinned to SHA 31c5652; no source
vendored. Rebuild on SHA bump.
- GPU-capable FastAPI + Silero VAD + WS streaming.
- API: POST /transcribe, WS /ws/transcribe, GET /healthz. Not the
literal OpenAI `/v1/audio/transcriptions` path — note in README.
- HF cache at /worktank/parakeet/models/ (excluded from restic).
- Build ~158s first time; steady-state start ~40s.
cosyvoice — TTS via neosun/cosyvoice:v1.3.2 shipping
Fun-CosyVoice3-0.5B-2512 (CosyVoice 3, chosen over v2 for the
expanded 5,000-hour instruction-following data covering emotions,
speed, tones, dialects, accents, role-playing; ~150ms streaming
TTFB matches v2). API: /v1/audio/speech (OpenAI drop-in),
/v1/voices/create (cloning), /health.
- Host port 8190 (container 8188; host 8188 already taken by comfyui).
- /worktank/cosyvoice/{voices,input,output}/; voices include in
restic (precious — reproducing a clone needs the original ref
audio), input+output excluded (scratch).
- Model weights (~2-3 GB) live inside image layer; re-download on
tag bump, persist across `compose up -d`.
Both healthy on first deploy.
vh