DEDUPED Hot AI repos, turned into use cases Get the Monday 5

How to choose a latency or throughput profile for load with Nari Qwen3 TTS

Switch between ttfa, balanced, and throughput profiles to match a latency target or a request volume.

Short answer: Install the environment with uv sync --frozen --extra codec --extra cuda --extra serving; then start the server with a profile: uv run --frozen nari-qwen3-tts-server --profile ttfa for latency, or --profile throughput for aggregate load; check readiness with curl --fail http://127.0.0.1:8000/ready before sending traffic. All steps.

Who it is forA platform engineer sizing a speech service that must meet a first-audio latency target or handle heavy batch load on one H100.

The problemOne fixed setting cannot favor both fast first audio for interactive users and high aggregate throughput for batch jobs.

The outcomeA chosen profile and a repeatable server start command that matches the workload.

What you need first

  • A Linux x86_64 host with an NVIDIA H100 GPU
  • NVIDIA Container Toolkit and a driver compatible with CUDA 13.0 (for Docker)
  • Docker, or uv with Python and the system packages build-essential, libsndfile1, and sox
  • A Hugging Face token (HF_TOKEN) if the model files need to be downloaded

Steps

  1. Install the environment with uv sync --frozen --extra codec --extra cuda --extra serving.
  2. Start the server with a profile: uv run --frozen nari-qwen3-tts-server --profile ttfa for latency, or --profile throughput for aggregate load. (inferred, check the README)
  3. Check readiness with curl --fail http://127.0.0.1:8000/ready before sending traffic.
  4. Record the time to first audio and the request rate for each profile, then keep the profile that meets the target.

Prompt for your agent

In the nari-labs/nari-qwen3-tts repo, start the server with the <profile> profile (ttfa, balanced, or throughput) and send a fixed set of test requests to `/v1/audio/speech`. Expected output: a short table of time to first audio and completed requests for each profile tested. Verify that each run started only after `/ready` returned success.

Caveat. The README gives the profile purposes but no numeric targets for each one, so any measured figures must come from the user's own runs on their hardware.

Use this in your AI agent

Ask your agent

Using the Deduped MCP, get the setup steps for nari-labs/nari-qwen3-tts and "Choose a latency or throughput profile for load"

Add Deduped to Claude Code once

claude mcp add --transport http deduped https://dedupedai.com/mcp

Cursor, ChatGPT, others

Share: X LinkedIn Hacker News Email Copy link

Third-party repos we don't own. AI-written use cases can be wrong. Disclaimer