Nextstep online e2e

[Joshna-Medisetty]

Purpose

Add an online serving E2E test for NextStep-1.1 text-to-image using OmniServer and OpenAIClientHandler.send_diffusion_request, consistent with other online diffusion tests in the repo. Supports broader L4 diffusion online coverage under RFC #1832.

What’s included

• Model: NextStep-1.1 (stepfun-ai/NextStep-1.1).
• Server config: TP=2, NextStep11Pipeline (single lightweight case; no Cache-DiT / Ulysses / FP8 stack).
• Pytest marks: advanced_model, diffusion, L4, distributed_cuda(2).

Test plan

• Ran the test successfully on Intel XPU using the same server flags (local validation; the committed test stays CUDA L4–marked for CI).
• Expect 2× L4 to be sufficient in CI.

Test Results (Intel XPU / 2× L4)

Config	Resolution	Steps	Latency	Peak Memory reserved	Peak memory allocated
TP=2	256x256	2	25.7 s	~20.2 GB	~16.1 GB
TP=2	256x256	20	57.3 s	~20.2 GB	~16.1 GB
TP=2	512x512	15	191.5 s	~20.2 GB	~16.6 GB
TP=2	512x512	20	230.0 s	~20.2 GB	~16.6 GB
TP=2	512x512	28	286.0 s	~20.2 GB	~16.6 GB

Sample Output

256×256 @ 2 steps

256×256 @ 20 steps

512×512 @ 15 steps

512×512 @ 20 steps

512×512 @ 28 steps

[Joshna-Medisetty]

@xuechendi @fhfuih

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: b65437cd72

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".

[chatgpt-codex-connector]

Exercise an actual VAE optimization path for NextStep

For stepfun-ai/NextStep-1.1, these flags do not change the server behavior today: vllm_omni/diffusion/models/nextstep_1_1/modeling_flux_vae.py implements AutoencoderKL.decode() as a plain decoder call and the class never defines use_slicing, while vllm_omni/diffusion/registry.py only applies vae_use_slicing when that attribute exists. In other words, this parametrization cannot catch regressions in VAE slicing or tiling for NextStep and only ends up validating TP=2, which gives the nightly L4 expansion suite false coverage for the advertised VAE paths.

Useful? React with 👍 / 👎.

[hsliuustc0106]

Test Evidence Request

Thank you for adding NextStep-1.1 online serving coverage! Before merging, could you please provide some baseline test evidence to help us understand the model's characteristics and validate the CI configuration?

Requested Evidence

1. Latency

• Time to generate a single image with the test config (TP=2, 256x256, 2 steps)
• If available, also include latency for a more realistic config (e.g., 512x512, 20-30 steps)

2. Accuracy / Correctness

• Sample output image(s) from a test run, OR
• Validation that send_diffusion_request returns a valid image with expected dimensions

3. Memory

• Peak VRAM usage with TP=2 on L4 (or your XPU equivalent)
• This helps validate that 2× L4 is sufficient (per your PR description) vs. needing TP=4 on 4× L4

How to Provide

You can add this directly in the PR description or as a comment. Example format:

### Test Results (Intel XPU / 2× L4)

| Config | Resolution | Steps | Latency | Peak Memory |
|--------|------------|-------|---------|-------------|
| TP=2   | 256x256    | 2     | X.XX s  | XX GB       |
| TP=2   | 512x512    | 28    | XX.XX s | XX GB       |

Sample output: [attach image or describe validation]

This information helps us:

• Validate the CI hardware allocation is appropriate
• Establish a baseline for future regression detection
• Document expected behavior for this model

---

Note: The docs/readthedocs.org failure appears transient (this PR doesn't modify documentation). It should pass on re-run.

[congw729]

LGTM

[xuechendi]

test only on CUDA?

[Joshna-Medisetty]

i tested it locally on xpu but removed it for pr to stay consistent with existing diffusion E2E tests, which currently target CUDA L4 . Would you prefer that I also include XPU coverage or should we keep it aligned with CUDA- only tests for now?

[Joshna-Medisetty]

@hsliuustc0106 Thanks for the checklist. I’ve updated the PR description with test results and sample outputs.

For memory, we see about ~20.2 GB peak reserved per GPU (PyTorch worker log on Intel XPU, TP=2) and about ~16–16.6 GB allocated depending on resolution; reserved vs allocated is called out in the description.
With ~20 GB per device, 2× L4 (24 GB each) should be enough for this setup without TP=4, noting CUDA/L4 may differ a bit from these XPU numbers.

I also changed the E2E config to 512×512 @ 20 steps so the test matches a more realistic setting.

[Gaohan123]

Please fix CI error. Thanks

[Gaohan123]

Are you ready? If not, I can remove the label. Then we can save some resources

[Joshna-Medisetty]

@Gaohan123 PR is ready and passes tests on L4.
I’m currently seeing a failure in the Qwen image diffusion perf test, which does not include the NextStep test introduced in this PR, so it appears unrelated to these changes.

[Joshna-Medisetty]

Hi @Gaohan123, PR is ready for merge. Could please add the label. Thank you.

[congw729]

Hi @Gaohan123, PR is ready for merge. Could please add the label. Thank you.

Hi, I just added the nightly label for L4 level tests to double check.

[Joshna-Medisetty]

The tests it fails are unrelated. May be need a force merge

[fhfuih]

Seems like #2435. It's OK to not merge main if there is no conflict.

Thanks for your patience and continuous tracking of this PR. It has really been a while. But since you happen to have done another merge, let's wait one more time (after that issue is resolved -> you update the branch) 🙏