[skip CI][Docs] Add Qwen3-Omni and Qwen3-TTS performance blog and figures

[Shirley125]

PLEASE FILL IN THE PR DESCRIPTION HERE ENSURING ALL CHECKLIST ITEMS (AT THE BOTTOM) HAVE BEEN CONSIDERED.

Purpose

Add Qwen3-Omni and Qwen3-TTS performance blog and figures

Test Plan

Test Result

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan. Please provide the test scripts & test commands. Please state the reasons if your codes don't require additional test scripts. For test file guidelines, please check the test style doc
• The test results. Please paste the results comparison before and after, or the e2e results.
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model. Please run mkdocs serve to sync the documentation editions to ./docs.
• (Optional) Release notes update. If your change is user-facing, please update the release notes draft.

BEFORE SUBMITTING, PLEASE READ https://github.com/vllm-project/vllm-omni/blob/main/CONTRIBUTING.md (anything written below this line will be removed by GitHub Actions)

[Gaohan123]

@hsliuustc0106 @amy-why-3459 @R2-Y @linyueqian PTAL

[wtomin]

Wouldn't it be better to use https://user-images.githubusercontent.com/xxx/xxxx/xxx.png rather than to upload these pictures to repository?

[Shirley125]

The total size of all PNGsis only about 2–3 MB, which is negligible for the repository. Keeping the images together with the blog content in the same version ensures consistency.

[linyueqian]

Updated TTS benchmark results with latest vLLM v0.18.0 / vllm-omni v0.18.0rc2 data (H200).

Key changes:

• All 4 benchmark phases now pass with 0 failures (CUDA graph was previously broken)
• CUDA Graph section updated: now shows ~1.8x RTF improvement (was previously "negligible impact")
• Added "Why E2EL/RTF are higher with async chunk" subsection explaining the TTFP vs throughput tradeoff
• Updated YouTube demo link and port numbers

Headline numbers (concurrency 1):

• TTFP: 64ms (async chunk) / 733ms (CUDA graph)
• RTF: 0.124 (CUDA graph) / 0.160 (async chunk)
• vs HF Transformers: 242x faster TTFP, 16.5x faster RTF

@Sy0307 @JuanPZuluaga - could you take a look at these results and see if they align with what you're seeing?

[fake0fan]

@yinpeiqi pls also check if the corresponding descriptions and results are consistent with the paper.

[JuanPZuluaga]

@linyueqian ran a benchmark locally with the latest main:

• Qwen3-TTS Benchmark Results
• Hardware: RTX 6000 Ada (48GB) vs PR's H200
  -Model: Qwen3-TTS-12Hz-1.7B-CustomVoice
• Config: bs16 (max_num_seqs=16), async_chunk enabled, CUDA Graph on (max inflight 16)

Results on RTX 6000 Ada

Metric | Concurrency 1 | Concurrency 4 | Concurrency 10
-- | -- | -- | --
TTFP (ms) | 44.9 | 102.0 | 217.3
E2EL (ms) | 1033.5 | 1289.6 | 1782.2
RTF | 0.183 | 0.239 | 0.310
Throughput (audio-s/s) | 5.48 | 16.67 | 31.16

Results from PR (H200)

Metric | Concurrency 1 | Concurrency 4 | Concurrency 10
-- | -- | -- | --
TTFP (ms) | 64 | 119 | 425
E2EL (ms) | 941 | ~987 | 1767
RTF | 0.160 | — | 0.314

(i'm seeing a bit fast TTFP)

[linyueqian]

@linyueqian ran a benchmark locally with the latest main:

• Qwen3-TTS Benchmark Results
• Hardware: RTX 6000 Ada (48GB) vs PR's H200
  -Model: Qwen3-TTS-12Hz-1.7B-CustomVoice
• Config: bs16 (max_num_seqs=16), async_chunk enabled, CUDA Graph on (max inflight 16)

Results on RTX 6000 Ada

Metric | Concurrency 1 | Concurrency 4 | Concurrency 10
-- | -- | -- | --
TTFP (ms) | 44.9 | 102.0 | 217.3
E2EL (ms) | 1033.5 | 1289.6 | 1782.2
RTF | 0.183 | 0.239 | 0.310
Throughput (audio-s/s) | 5.48 | 16.67 | 31.16

Results from PR (H200)

Metric | Concurrency 1 | Concurrency 4 | Concurrency 10
-- | -- | -- | --
TTFP (ms) | 64 | 119 | 425
E2EL (ms) | 941 | ~987 | 1767
RTF | 0.160 | — | 0.314

(i'm seeing a bit fast TTFP)

Thank you very much!