[Bugfix] Handle numpy array outputs when generate image by lengrongfu · Pull Request #1680 · vllm-project/vllm-omni

lengrongfu · 2026-03-05T07:55:44Z

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

Purpose

Wan2.2 video pipelines return numpy arrays (output_type="np"), but image_api_utils.py assumes PIL Images and calls img.save().

We _extract_images_from_result this method to add a numpy-to-PIL conversion.

Serve:

vllm serve /home/jovyan/Wan2.2-T2V-A14B-Diffusers/ --omni --port 8091

Test:

curl  -X POST "http://localhost:8091/v1/images/generations"   -H "Content-Type: application/json"   --data '{
            "model": "/home/jovyan/Wan2.2-T2V-A14B-Diffusers/",
            "prompt": "a cat",
            "n": 1,
            "size": "256x256",
            "seed": 7
}' | jq -r '.data[0].b64_json' | base64 -d > cat.png

curl -X POST "http://localhost:8000/v1/images/generations" \
  -H "Content-Type: application/json" \
  --data '{
    "model": "/home/jovyan/Wan2.2-T2V-A14B-Diffusers/",
    "prompt": "a cat",
    "n": 2,
    "size": "256x256",
    "seed": 7
  }' \
| jq -r '.data[].b64_json' \
| nl -v 0 \
| while read i img; do
    echo "$img" | base64 -d > "cat${i}.png"
  done

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)

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 08bb4cf9b7

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".

hsliuustc0106 · 2026-03-05T11:49:53Z

why use /v1/images/generations as the endpoint?

lengrongfu · 2026-03-05T11:55:16Z

Sorry, I checked the example and it shows that this API is used. Is it not usable?

hsliuustc0106

Review

Rating: 6/10 | Verdict: ⚠️ Changes Requested

Summary

Bugfix addresses valid issue (Wan2.2 returning numpy arrays), but implementation has critical edge cases and missing tests. Code logic is unclear and needs refactoring.

Issues

Critical: IndexError on empty list - Line 1424 checks images[0] without verifying the list is non-empty. Will crash if images is empty.
Confusing logic - Line 1425 images = images[0] is unclear. Is this handling a nested array case? Needs comment explaining the expected structure.
Missing tests - No unit tests for the new _normalize_image function or the modified _extract_images_from_result.
Incomplete PR description - Checklist items not filled, no test results provided.
Edge cases not handled:
- Empty images list
- Mixed types (some PIL, some numpy)
- Unsupported dtype (e.g., bool, complex)

Highlights

✅ Addresses real issue (Wan2.2 compatibility)
✅ Handles multiple normalization cases (float [0,1] and [-1,1])
✅ Properly handles ndim > 3 case

Recommendation

Fix edge cases and add tests before merge.

Reviewed by OpenClaw with vllm-omni-skills 🦐

hsliuustc0106 · 2026-03-06T01:46:20Z

        elif hasattr(request_output, "images") and request_output.images:
            images = request_output.images
-    return images
+    if isinstance(images[0], np.ndarray):


This will raise IndexError if images is empty. Add a guard: if images and isinstance(images[0], np.ndarray):

thanks tips ~

hsliuustc0106 · 2026-03-06T01:46:20Z

            images = request_output.images
-    return images
+    if isinstance(images[0], np.ndarray):
+        images = images[0]


This logic is confusing. If images[0] is a numpy array, you replace the entire images list with images[0]. Is this handling a nested array structure like [np_array_of_frames]? Add a comment explaining the expected input format.

yes, when generate more one image this array structure is np_array_of_frames. like:

images = np.array([ [[...], [...], ...], # one png (256, 256, 3) [[...], [...], ...], # two png (256, 256, 3) [[...], [...], ...], # three png (256, 256, 3) ]) # shape: (3, 256, 256, 3)

hsliuustc0106 · 2026-03-06T01:46:20Z

        setattr(object, key, val)


+def _normalize_image(image: Any) -> Any:


The function handles floating dtypes but doesn't check for unsupported types (e.g., bool, complex). Consider adding: if not np.issubdtype(image.dtype, np.integer) and not np.issubdtype(image.dtype, np.floating): raise ValueError(f"Unsupported dtype: {image.dtype}")

thanks tips ~

hsliuustc0106 · 2026-03-06T01:46:20Z

+def _normalize_image(image: Any) -> Any:
+    """Normalize a single image output to a PIL-compatible format."""
+    if isinstance(image, np.ndarray):
+        while image.ndim > 3:


The normalization assumes float range is either [-1, 1] or [0, 1]. What about other ranges? Consider adding a debug log or warning when rescaling from [-1, 1] range to help diagnose issues.

add warning log.

hsliuustc0106 · 2026-03-06T01:46:20Z

+    if isinstance(images[0], np.ndarray):
+        images = images[0]
+    return [_normalize_image(img) for img in images]



Missing unit test for _normalize_image and _extract_images_from_result. At minimum, add tests for: (1) numpy array with float [0,1], (2) numpy array with float [-1,1], (3) numpy array with int, (4) already PIL Image, (5) empty list.

add unit test

gcanlin · 2026-03-09T02:20:30Z

@lishunyang12 PTAL. Thanks!

lishunyang12 · 2026-03-21T23:24:39Z

Will take a look.

lishunyang12

The idea is right but there are two bugs in the implementation.

Signed-off-by: rongfu.leng <lenronfu@gmail.com>

gcanlin

Could you upload the generated images?

lengrongfu · 2026-04-14T04:43:59Z

Could you upload the generated images?

Uploaded, currently, generating multiple photos via HTTP is quite inconvenient, as the response consists of JSON data. I would appreciate it if you could take a look at this PR #1673, which proposes returning the images directly as files. @gcanlin

fixed

…#1680) Signed-off-by: rongfu.leng <lenronfu@gmail.com>

lengrongfu requested a review from hsliuustc0106 as a code owner March 5, 2026 07:55

chatgpt-codex-connector Bot reviewed Mar 5, 2026

View reviewed changes

Comment thread vllm_omni/entrypoints/openai/api_server.py Outdated

lengrongfu force-pushed the fix/image-save branch from 08bb4cf to 7ea1e60 Compare March 5, 2026 08:21

hsliuustc0106 requested changes Mar 6, 2026

View reviewed changes

Gaohan123 added this to the v0.18.0 milestone Mar 14, 2026

Gaohan123 mentioned this pull request Mar 14, 2026

[Bugfix] Handle numpy array outputs in diffusion chat endpoint #1660

Closed

lishunyang12 previously requested changes Mar 21, 2026

View reviewed changes

Comment thread vllm_omni/entrypoints/openai/api_server.py

Comment thread vllm_omni/entrypoints/openai/api_server.py Outdated

hsliuustc0106 self-requested a review March 22, 2026 01:47

hsliuustc0106 added the ready label to trigger buildkite CI label Mar 22, 2026

lengrongfu added 3 commits March 25, 2026 03:15

[Bugfix] Handle numpy array outputs when generate image

d92c464

Signed-off-by: rongfu.leng <lenronfu@gmail.com>

add unit test and code comment

8f4bcd2

Signed-off-by: rongfu.leng <lenronfu@gmail.com>

opt code

8d195db

Signed-off-by: rongfu.leng <lenronfu@gmail.com>

lengrongfu force-pushed the fix/image-save branch from ae9af4a to fa13f7c Compare March 25, 2026 07:24

resolved conflicts

fa13f7c

Signed-off-by: rongfu.leng <lenronfu@gmail.com>

hsliuustc0106 requested a review from lishunyang12 April 9, 2026 07:23

gcanlin reviewed Apr 14, 2026

View reviewed changes

Gaohan123 modified the milestones: v0.18.0, v0.20.0 Apr 14, 2026

hsliuustc0106 approved these changes Apr 15, 2026

View reviewed changes

hsliuustc0106 merged commit 9e46a79 into vllm-project:main Apr 15, 2026
8 checks passed

y123456y78 pushed a commit to y123456y78/vllm-omni that referenced this pull request Apr 15, 2026

[Bugfix] Handle numpy array outputs when generate image (vllm-project…

0ec2500

…#1680) Signed-off-by: rongfu.leng <lenronfu@gmail.com>

lvliang-intel pushed a commit to lvliang-intel/vllm-omni that referenced this pull request Apr 20, 2026

[Bugfix] Handle numpy array outputs when generate image (vllm-project…

69b1014

…#1680) Signed-off-by: rongfu.leng <lenronfu@gmail.com>

lengrongfu added a commit to lengrongfu/vllm-omni that referenced this pull request May 1, 2026

[Bugfix] Handle numpy array outputs when generate image (vllm-project…

387eb32

…#1680) Signed-off-by: rongfu.leng <lenronfu@gmail.com>

clodaghwalsh17 pushed a commit to clodaghwalsh17/nm-vllm-omni-ent that referenced this pull request May 12, 2026

[Bugfix] Handle numpy array outputs when generate image (vllm-project…

41d88b0

…#1680) Signed-off-by: rongfu.leng <lenronfu@gmail.com>

daixinning pushed a commit to daixinning/vllm-omni that referenced this pull request May 28, 2026

[Bugfix] Handle numpy array outputs when generate image (vllm-project…

08cd916

…#1680) Signed-off-by: rongfu.leng <lenronfu@gmail.com>

quyifei23 pushed a commit to quyifei23/vllm-omni that referenced this pull request Jun 6, 2026

[Bugfix] Handle numpy array outputs when generate image (vllm-project…

ec819ce

…#1680) Signed-off-by: rongfu.leng <lenronfu@gmail.com>

khairulkabir1661 pushed a commit to khairulkabir1661/vllm-omni that referenced this pull request Sep 25, 2026

[Bugfix] Handle numpy array outputs when generate image (vllm-project…

0bb759f

…#1680) Signed-off-by: rongfu.leng <lenronfu@gmail.com>

		setattr(object, key, val)


		def _normalize_image(image: Any) -> Any:

Conversation

lengrongfu commented Mar 5, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Purpose

Test Plan

Test Result

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

Uh oh!

hsliuustc0106 commented Mar 5, 2026

Uh oh!

lengrongfu commented Mar 5, 2026

Uh oh!

hsliuustc0106 left a comment

Choose a reason for hiding this comment

Review

Summary

Issues

Highlights

Recommendation

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

gcanlin commented Mar 9, 2026

Uh oh!

lishunyang12 commented Mar 21, 2026

Uh oh!

lishunyang12 left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

gcanlin left a comment

Choose a reason for hiding this comment

Uh oh!

lengrongfu commented Apr 14, 2026

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

lengrongfu commented Mar 5, 2026 •

edited

Loading