* docs: add AGENTS.md with verified Tesla T4 (16GB) inference notes
Documents two things found while verifying inference on a real T4:
1. Cold-cache first /v1/audio/speech call can hit the 300s
OMNIVOICE_GENERATE_TIMEOUT_S because the checkpoint download happens
inside that budget — workaround via existing POST /models/install or
raising the timeout, no code change needed.
2. The OpenAI-compatible endpoint silently ignores num_step/guidance_scale
(schema doesn't declare them) — use native /generate for those.
Also documents the T4 acceleration checklist (dtype/attention/int8/CUDA
graphs) and measured VRAM (peak 2.05GB). No code changes.
* fix(docs): make /models/install workaround command actually executable
Addresses Greptile review: the instruction omitted the required
repo_id body field (InstallModelRequest rejects an empty body).
* fix(docs): correct port in /models/install example (3900, not 8000)
The app serves on port 3900 (confirmed: /health returns 200 there,
connection refused on 8000). Verified the exact corrected curl command
returns 200 {"status":"install_started",...}.
* move T4 notes to docs/hardware-notes-tesla-t4.md — AGENTS.md is the auto-loaded agent-instructions filename
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---------
Co-authored-by: mergetest <test@local>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
3.8 KiB
Verified Tesla T4 (16GB) inference notes
Measured on a real NVIDIA Tesla T4 (16GB, Turing/sm_75), driver 550.163.01 (CUDA 12.8), torch
2.8.0+cu128, transformers 5.3.0, Python 3.11.15 (uv-managed). Engine under test: the default
omnivoice TTS backend (OMNIVOICE_TTS_BACKEND=omnivoice).
Cold-cache first call can time out at 300s
The first generate() call lazily downloads the ~2.3GB k2-fsa/OmniVoice checkpoint, and that
download happens inside the OMNIVOICE_GENERATE_TIMEOUT_S budget (default 300s). On a fresh
install, the very first POST /v1/audio/speech can fail like this even though the GPU isn't
actually short on memory:
ERROR [omnivoice.openai_compat] OpenAI TTS failed: OpenAI TTS generate exceeded 300s and was
abandoned — the backend is running, but the job was too heavy for the available compute.
... most often the GPU is VRAM-starved ...
VRAM sampling during the failure showed a flat ~2GB with 0% GPU utilization for the whole 300s — consistent with waiting on a download, not compute. Once the checkpoint is cached, the identical request succeeds in ~1s (reproduced 5x: 1.574s / 1.034s / 1.065s / 0.995s / 0.911s).
Workaround (no code change needed, both already exist):
- For headless/API-only setups, pre-fetch the checkpoint before your first real TTS request:
(
curl -X POST http://localhost:3900/models/install \ -H "Content-Type: application/json" \ -d '{"repo_id": "k2-fsa/OmniVoice"}'repo_idis required —InstallModelRequestinbackend/api/schemas.pyrejects a bare/empty body — and must match one of the entries inKNOWN_MODELS, e.g. the default engine'sk2-fsa/OmniVoice.) Progress streams over the existing/setup/download-streamSSE feed. - Or raise
OMNIVOICE_GENERATE_TIMEOUT_Sfor the first request.
OpenAI-compatible endpoint doesn't expose num_step / guidance_scale
POST /v1/audio/speech's request schema doesn't declare num_step or guidance_scale fields —
sending them in the JSON body returns 200 OK but they're silently discarded (pydantic's default
extra=ignore behavior). The native multipart POST /generate endpoint does expose both as
explicit form fields, so use that endpoint if you need to control them.
Separately: the app's own default for num_step is 16 — half of the model's documented default of
32 (see docs/generation-parameters.md, "Use 16 for faster inference"). Not a bug, just not stated
that the app already runs the "fast" preset unless you override it via /generate.
T4 acceleration checklist
| Option | Status |
|---|---|
| dtype | torch.float16 hardcoded for the omnivoice engine (model_manager.py) — correct for Turing (no bf16 tensor cores this generation). No env var override for this engine specifically (ASR engines have ASR_COMPUTE_TYPE; dots_tts/indextts have their own precision vars; omnivoice doesn't). |
| Attention | sdpa, selected automatically since flash_attn isn't installed (_supports_flash_attn_2=True is declared but the package itself is absent) — safe on T4. |
| int8 | No int8 path for this engine (ASR's CTranslate2 int8 and sherpa-onnx's int8 ONNX models are separate/unrelated). |
| CUDA Graphs | No direct API usage in the app. Reachable indirectly via torch.compile(mode="reduce-overhead"), which the app attempts by default on this GPU (T4/sm_75 isn't in the framework's compile-exclusion list, unlike newer/Blackwell GPUs). The numbers above were measured with TORCH_COMPILE_DISABLE=1 for a clean eager baseline. |
| torch.compile | Attempted by default on T4 (see above) — not evaluated further here. |
VRAM
Peak measured: 2487 MiB (nvidia-smi) / 2.050 GB (torch.cuda.max_memory_allocated()) for the
default omnivoice engine — comfortably fits even the README's stated "minimum" (4GB) tier.