Add Qwen3-TTS benchmark scripts

[linyueqian]

Summary

• Add benchmark scripts comparing vLLM-Omni streaming serving vs HuggingFace Transformers offline inference for Qwen3-TTS
• Measures TTFP, E2E latency, RTF, and throughput across configurable concurrency levels
• Includes orchestration script (run_benchmark.sh) with env vars for GPU, model, batch size, and concurrency

Files

• bench_tts_serve.py - Streaming serving benchmark client
• bench_tts_hf.py - HuggingFace offline baseline
• run_benchmark.sh - Orchestration script
• plot_results.py - Comparison plotting utility
• configs/qwen3_tts_bs1.yaml - batch_size=1 stage config
• configs/qwen3_tts_bs4.yaml - batch_size=4 stage config

Usage

# Full benchmark (vllm-omni + HF)
bash run_benchmark.sh

# vllm-omni only, 1.7B model
MODEL=Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice bash run_benchmark.sh --async-only

# Custom GPU and concurrency
GPU_DEVICE=1 CONCURRENCY="1 4" bash run_benchmark.sh

Test plan

• Verified on H200 with 0.6B model, bs1 config, concurrency 1 and 4
• Reviewer runs on their own hardware to validate

Related: #938

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 0b573d1ab9

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".