fix: surface image generation errors instead of returning empty HTTP 200

[marckarp]

Overview

A failed /v1/images/generations request returns HTTP 200 with data: [] and no error message. The client has no way to tell a failed generation from an empty one. Easy to reproduce: send any request that makes the worker raise (e.g. an out-of-range parameter or a backend failure) — the response is a successful-looking empty body.

Two layers cause it:

1. Worker (Python): the SGLang image diffusion handler catches every exception and yields {"data": [], "error": str(e)}. The OpenAI ImagesResponse schema has no error field, so the message is silently dropped during deserialization and the frontend returns HTTP 200 with empty data.
2. Frontend (Rust): even when an exception does propagate as an error event on the response stream, the images route discards it and returns a hardcoded generic 500 ("Failed to fold images stream"), so the worker's message is only visible in server logs.

What this PR changes

Handler (image_diffusion_handler.py):

• Remove the catch-and-yield: exceptions now propagate. The runtime converts a raised exception into an error event on the response stream (Annotated::from_err), which the frontend folds into a non-200 HTTP error.
• Request-validation failures (blank input_reference, malformed size) now raise the typed InvalidArgument instead of bare ValueError, matching the existing pattern in sglang_processor.py. InvalidArgument maps to HTTP 400 with the message visible to the client; other exceptions remain sanitized 500s.

Frontend (lib/llm/src/http/service/openai.rs):

• Route the images stream fold error through ErrorMessage::from_anyhow instead of the hardcoded generic 500. This follows the route's existing sanitization policy: InvalidArgument → 400 with message, internal errors → sanitized 500 (details stay server-side).

Tests: the unit test that asserted the old swallow behavior now asserts propagation; added tests for the InvalidArgument validation paths.

Verification

• Unit tests: 23/23 pass inside sglang-runtime:1.3.0 with the patched handler (the 8 behavior-dependent tests fail against the stock handler, confirming they detect the old behavior).
• Live e2e on a single-GPU FLUX.1-schnell deployment with the patched worker:
  • valid request → HTTP 200 with image (no regression)
  • worker-side failure (blank input_reference) → HTTP 500 instead of the old empty HTTP 200
• Rust half: cargo check introduces no new errors. A live 400-with-message demonstration requires a rebuilt frontend image (the released frontend predates this change), so that path relies on CI + the unit-level guarantees of from_anyhow (which existing chat/completions routes already use).

Notes for reviewers

• The _parse_size guard is currently unreachable from HTTP because the frontend's size field is a fixed enum that rejects free-form values first. It becomes load-bearing with the planned change to accept arbitrary WxH sizes (needed for diffusion models; the videos route already takes a free-form string), and is worth having defensively regardless.
• Soft overlap with fix: honor the OpenAI n parameter in the SGLang image diffusion handler #12970: both touch test_sglang_image_diffusion_handler.py in different places; whichever merges second rebases trivially. fix: honor the OpenAI n parameter in the SGLang image diffusion handler #12970's live test also demonstrated this bug: requesting n=11 returned HTTP 200 with data=[] and an empty error on the wire.

Part of a series of fixes for the SGLang image diffusion path (see #12970 for the first).

🤖 Generated with Claude Code

Part of DYN-3974

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[github-actions]

👋 Hi marckarp! Thank you for contributing to ai-dynamo/dynamo.

Just a reminder: The NVIDIA Test Github Validation CI runs an essential subset of the testing framework to quickly catch errors.Your PR reviewers may elect to test the changes comprehensively before approving your changes.

🚀

[datadog-official]

⚠️ Warnings

🚦 2 Pipeline jobs failed

Pre Merge | pre-commit     

🔧 Fix in code. This looks caused by changes in this PR.

isort failed due to code format issues in image_diffusion_handler.py requiring fixes.

Pre Merge | pre-merge-status-check     

📋 Copy prompt for your agent

CI on my pull request is failing. Help me find and fix the root cause of each failing job below — they were flagged as caused by changes in this PR, so focus on the diff. For each job, explain the failure and propose a fix.

Branch: fix/sglang-image-error-surfacing

Pre Merge | pre-commit
Commit: 4e1ea13b68d9b881ec4724c8326c0f20352b3b71
Error (code / quality):
isort failed due to code format issues in image_diffusion_handler.py requiring fixes.
CI job: https://github.com/ai-dynamo/dynamo/actions/runs/31461853106/job/93686788645

Useful? React with 👍 / 👎

_{This comment will be updated automatically if new data arrives.
🔗 Commit SHA: 4e1ea13 | Docs | Datadog PR Page | Give us feedback!}