[Bugfix] Fix image edit RoPE crash when explicit height/width are provided

[lishunyang12]

Purpose

Fix a server crash (AssertionError: seqlen_ro must be >= seqlen) in the Qwen-Image-Edit online serving pipeline when users provide explicit height/width in extra_body for image editing with a non-square input image.

Root cause: In pre_process_func, the input image was resized to the user-specified (height, width) (e.g., 1024x1024), but the aspect-ratio-preserving calculated_height/calculated_width (e.g., 1056x992) were stored in additional_information. Downstream, the pipeline builds img_shapes from both:

• img_shapes[0] = (1, 64, 64) from explicit 1024x1024 (noise latent)
• img_shapes[1] = (1, 66, 62) from calculated 1056x992 (image latent)

The image latent's actual seq_len (4096) exceeded the RoPE seq_len computed from calculated_height/width (4092), triggering the assertion failure.

Fix: Resize and preprocess the input image using calculated_height/calculated_width instead of the explicit height/width, consistent with the fallback/debug path behavior (line 665). This ensures img_shapes[1] always matches the actual image latent dimensions.

Also fixes two issues in the example scripts:

• The curl example crashed with Argument list too long on large images due to passing base64 data as a shell argument.
• The Python client silently ignored --height/--width CLI arguments (never sent to server).

Test Plan

1. Restart the server with Qwen-Image-Edit model.

2. Test with non-square input image + explicit height/width via curl — should succeed instead of crashing:

IMG_B64=$(base64 -w0 input.png)

cat <<EOF > request.json
{
  "messages": [{
    "role": "user",
    "content": [
      {"type": "text", "text": "Convert this image to watercolor style"},
      {"type": "image_url", "image_url": {"url": "data:image/png;base64,$IMG_B64"}}
    ]
  }],
  "extra_body": {
    "height": 1024,
    "width": 1024,
    "num_inference_steps": 50,
    "guidance_scale": 1,
    "seed": 42
  }
}
EOF

curl -s http://localhost:8092/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d @request.json | jq -r '.choices[0].message.content[0].image_url.url' \
  | cut -d',' -f2 | base64 -d > output.png

Output (explicit height=1024, width=1024):

3. Test with curl example script (auto-computed dims):

bash examples/online_serving/image_to_image/run_curl_image_edit.sh input.png "Convert to watercolor style"

Output (auto-computed dims):

4. Test with Python client (no explicit dims) — should still work as before:

python examples/online_serving/image_to_image/openai_chat_client.py \
    --input input.png --prompt "Convert to watercolor style"

Output (Python client, no explicit dims):

5. Test with Python client (explicit dims) — should work now that height/width are sent:

python examples/online_serving/image_to_image/openai_chat_client.py \
    --input input.png --prompt "Convert to watercolor style" \
    --height 1024 --width 1024

Output (Python client, explicit height=1024, width=1024):

Test Result

Before this fix, sending a non-square image (e.g., 514x556) with explicit height: 1024, width: 1024 crashes the server:

AssertionError: seqlen_ro must be >= seqlen

After this fix, the same request completes successfully and returns a valid output image.

Changes

File	Change
vllm_omni/diffusion/models/qwen_image/pipeline_qwen_image_edit.py	Resize and preprocess input image using calculated_height/width instead of explicit height/width in pre_process_func, matching the fallback path
examples/online_serving/image_to_image/run_curl_image_edit.sh	Remove hardcoded height/width from extra_body; pipe base64 via stdin and use temp file for curl payload to avoid ARG_MAX limit on large images
examples/online_serving/image_to_image/openai_chat_client.py	Wire height/width into extra_body so CLI args actually take effect

---

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, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.
• (Optional) Release notes update. If your change is user facing, please update the release notes draft.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 3757b3f0ae

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]

Allow omitting explicit height/width in chat client

After this change, edit_image forwards height/width whenever they are non-None, but main() still sets both CLI args to default 1024, so running the documented command without --height/--width now always sends square dimensions and disables the server’s auto aspect-ratio sizing path for non-square inputs. This is a behavior regression introduced by wiring these fields; if the intent is “optional explicit dims,” the parser defaults need to be None (or equivalent flag-presence detection) so omitted flags stay omitted in extra_body.

Useful? React with 👍 / 👎.

[hsliuustc0106]

@Bounty-hunter @ZJY0516 PTAL

[Bounty-hunter]

@Bounty-hunter @ZJY0516 PTAL

LGTM, the new logic is consistent with Diffusers.

[ZJY0516]

LGTM. Could you please also check if other qwen models also has same issue?

[lishunyang12]

LGTM. Could you please also check if other qwen models also has same issue?

I will keep track if this issues can apply to other models.

[lishunyang12]

Checked all Qwen pipelines:

• pipeline_qwen_image.py — text-to-image only, no image input, single img_shapes entry. Not affected.
• pipeline_qwen_image_edit_plus.py — written with separate CONDITION_IMAGE_SIZE/VAE_IMAGE_SIZE from the start, img_shapes consistently uses VAE dims. Not affected.
• pipeline_qwen_image_layered.py — height/width always set to calculated_height/calculated_width (no user override path), so no mismatch possible. Not affected.

Only pipeline_qwen_image_edit.py had this issue.