[Test] Add L4 diffusion feature test for GLM-Image

[herotai214]

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

Purpose

This PR adds L4 test that covers the baseline & 2 of the supported features Tensor-Parallel, HSDP in both t2i & i2i cases for GLM-Image
Referred to #2167. Since it only covers i2i case, didn't effectively test the features, and the codes are quite different now, decided to raise this separate PR.

Tests are picked up by nightly: pytest -sv test_glm_image_expansion.py -m "full_model" --run-level "full_model".
The tests are intended to be picked up by the L4 nightly diffusion pipeline described in RFC #1832.

Test Plan

Model

• zai-org/GLM-Image

Changes:

• .buildkite/test-nightly.yml to trigger test
• tests/e2e/online_serving/test_glm_image_expansion.py (the new test file)

Tests added:

After referring to [RFC]: Continuous Diffusion Model Acceleration Support #1217:

• A normal baseline execution
• Tensor-Parallel=2
• HSDP=2

Above 3 tests loop through both t2i and i2i cases -> 6 tests in total.
All tests use 2 cards.

Remark:

• By default, a baseline GLM-Image instance has Stage-0 AR on device 0 and Stage-1 DiT on device 1 -> Total 2 cards.
• For TP and HSDP, only Stage-1 DiT part gets parallel.
  • To occupy the least cards as possible, we put 1 part of the Stage-1 to device 0 too -> Total 2 cards.
• Skip CFG parallel test since it is not working properly now, referring to [Bug]: GLM-Image CFG Parallel is not working #3382.
• This test only works correctly after the CLI bugfix by [Bugfix] Fix for multi-stage CLI overrides, which cause features like HSDP not working for multi-stage models #3384 is merged.
• (latest update: 8 May, 2026)

Test Result

pytest -sv tests/e2e/online_serving/test_glm_image_expansion.py -m "full_model" --run-level "full_model"
(With #3384 fix!)

================= 6 passed, 22 warnings in 1418.49s (0:23:38) ==================

pytest -sv tests/e2e/online_serving/test_glm_image_expansion.py -m "full_model" --collect-only

<Dir vllm-omni>
  <Package tests>
    <Package e2e>
      <Package online_serving>
        <Module test_glm_image_expansion.py>
          Comprehensive tests of diffusion features that are available in online serving mode
          and are supported by the following models:
          - zai-org/GLM-Image: (supports t2i & i2i)
          Coverage:
              For both t2i & i2i cases:
              - Tensor-Parallel
              - HSDP
          
          assert_diffusion_response validates successful generation and the expected
          1024x1024 resolution.
          <Function test_glm_image[baseline_t2i]>
            Test GLM-Image in both T2I and I2I modes across all parallel configurations.
          <Function test_glm_image[tensor_parallel_2_t2i]>
            Test GLM-Image in both T2I and I2I modes across all parallel configurations.
          <Function test_glm_image[hsdp_2_t2i]>
            Test GLM-Image in both T2I and I2I modes across all parallel configurations.
          <Function test_glm_image[baseline_i2i]>
            Test GLM-Image in both T2I and I2I modes across all parallel configurations.
          <Function test_glm_image[tensor_parallel_2_i2i]>
            Test GLM-Image in both T2I and I2I modes across all parallel configurations.
          <Function test_glm_image[hsdp_2_i2i]>
            Test GLM-Image in both T2I and I2I modes across all parallel configurations.

---

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)

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: a80dae5abc

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".

[herotai214]

@yenuo26 PTAL 🙏 not sure if I edit the .buildkite/test-nightly.yml correctly

[hsliuustc0106]

is this going to be an indepdent test pipeline?

[hsliuustc0106]

@JaredforReal

[hsliuustc0106]

LGTM — clean test PR with good use of shared helpers and parametrization. No substantive issues.

[hsliuustc0106]

COMMENT

This PR adds comprehensive L4 diffusion feature tests for GLM-Image, covering baseline, Tensor-Parallel, and HSDP configurations in both T2I and I2I modes.

Notes:

1. Good test structure - 6 tests total (3 configurations × 2 modes) with proper parametrization.

2. The Buildkite step is well-configured:

   • Appropriate timeout (120 minutes for 6 tests averaging ~3.8 minutes each)
   • Correct resource allocation (2 H100 GPUs)
   • Proper markers (full_model, diffusion, H100)

3. Stage overrides are correctly configured:

   • TP: Stage-0 stays on device 0, Stage-1 uses devices 0,1
   • HSDP: Only Stage-1 is sharded with devices 0,1 (runtime avoids sharding Stage-0)
   • Baseline: Stage-0 on device 0, Stage-1 on device 1

4. Test evidence is solid:

   • All 6 tests passed in 1418.49s (~23:38 total)
   • Tests validate both successful generation and expected 1024x1024 resolution

5. The PR correctly notes dependency on PR [Bugfix] Fix for multi-stage CLI overrides, which cause features like HSDP not working for multi-stage models #3384 (multi-stage CLI overrides bugfix).

No blocking issues found. This is a valuable addition to the test suite that will help catch regressions in GLM-Image's parallel features.

[yenuo26]

LGTM

[hsliuustc0106]

Hi @herotai214, friendly reminder — this PR hasn't had any activity (commits or reviews) in the past 7 days. 🕐

Could you please provide an update?

• If you're still working on it, that's great — just let us know.
• If you're blocked on something, feel free to ask for help.
• If this PR is no longer being pursued, please consider closing it so we can keep the review queue manageable.

Thanks for your contribution! 🙏

[herotai214]

Hi @herotai214, friendly reminder — this PR hasn't had any activity (commits or reviews) in the past 7 days. 🕐

Could you please provide an update?

• If you're still working on it, that's great — just let us know.
• If you're blocked on something, feel free to ask for help.
• If this PR is no longer being pursued, please consider closing it so we can keep the review queue manageable.

Thanks for your contribution! 🙏

Thanks for reminding; I was working for Hunyuan for a period previously, and I was aware that the multi-stage CLI issue solution seems to be uncertain; Therefore, I put this PR on hold.

Now I recognize the #3483 solves to CLI issue and have been merged today;

I'll verify asap if it helps this L4 test (support TP, HSDP properly through CLI), or otherwise I may need to push a revised version after removing those unsupported features (through CLI).

[herotai214]

Updated the script; Verified that #3483 worked, and able to use CLI here to apply TP, SP, Cache-DiT....

# with relevant logs like:
# Cache-DiT
�[32mINFO�[0m �[90m06-01 03:27:48�[0m �[90m[cache_dit_backend.py:1818]�[0m Cache-dit enabled successfully on GlmImagePipeline

# TP2
�[0;36m(APIServer pid=352086)�[0;0m INFO 06-01 03:30:39 [stage_utils.py:138] Stage 1 logical-to-physical device mapping: 0->4, 1->5

# HSDP2
�[32mINFO�[0m �[90m06-01 03:36:03�[0m �[90m[hsdp.py:128]�[0m HSDP Inference: replicate_size=1, shard_size=2, world_size=2, rank=0, fs_world_size=2, fs_rank=0

# SP2
�[32mINFO�[0m �[90m06-01 03:20:14�[0m �[90m[parallel_state.py:588]�[0m Building SP subgroups from explicit sp_group_ranks (sp_size=2, ulysses=2, ring=1, use_ulysses_low=True).
�[32mINFO�[0m �[90m06-01 03:20:14�[0m �[90m[parallel_state.py:630]�[0m SP group details for rank 1: sp_group=[0, 1], ulysses_group=[0, 1], ring_group=[1]
�[32mINFO�[0m �[90m06-01 03:20:14�[0m �[90m[parallel_state.py:588]�[0m Building SP subgroups from explicit sp_group_ranks (sp_size=2, ulysses=2, ring=1, use_ulysses_low=True).
�[32mINFO�[0m �[90m06-01 03:20:14�[0m �[90m[parallel_state.py:630]�[0m SP group details for rank 0: sp_group=[0, 1], ulysses_group=[0, 1], ring_group=[0]

...

================= 10 passed, 23 warnings in 1226.11s (0:20:26) =================

But after rebased and upgrading to vllm 0.22.0 this morning with the commit [Rebase] Rebase to vllm releases/v0.22.0 (#3891), I cannot use CLI to apply any feature here anymore...

Will double check and raise issue very soon....