fix(vllm): align vLLM-Omni pin with bundled vLLM 0.23.0

[tanmayv25]

Overview

release/1.3.0 bundles vLLM 0.23.0 (runtime_image_tag: v0.23.0-ubuntu2404 / cpu v0.23.0) but container/context.yaml pins vllm_omni_ref to v0.21.0rc1. vLLM-Omni is built against a specific vLLM release line and imports internal vLLM symbols that are added/renamed/removed across releases. A major/minor skew therefore crashes at import time with an opaque ImportError, before the unified-omni worker can register — so the Frontend's /v1/models stays permanently empty and no image/video/audio request is ever served.

This is the failure originally reported against 1.3.0-rc.2 (omni 0.21.x vs vLLM 0.23.0). The same alignment fix landed on main (#11345) for the 0.24 line; this backports it to release/1.3.0 for the 0.23 line.

Fix

• Bump vllm_omni_ref v0.21.0rc1 → v0.23.0rc1 (the vLLM-Omni build aligned with vLLM 0.23.0) in container/context.yaml.
• Port the vllm-project/vllm-omni#4568 cherry-pick into container/deps/vllm/install_vllm_omni.sh. In the released v0.23.0rc1, OmniRequest.__init__ declares *args after its named parameters; vLLM 0.23's positional Request(...) construction misbinds the arguments and EngineCore init fails for all vLLM workers (vLLM-Omni monkeypatches Request at import time). The commit moves *args to the front and forwards cleanly.
• Refresh the stale version reference comment in container/deps/vllm/protected_packages.txt.

Test

• A rebuild of the vllm-runtime image is required for the pin change to take effect.
• The equivalent alignment on main (omni/vLLM 0.24.0) was validated end-to-end: clean dynamo.vllm.unified_omni import, omni worker registers, /v1/models populated, and POST /v1/videos completes with a valid MP4.

[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.

[dynamo-review-agent]

The script now invokes patch, but the vLLM runtime image only installs jq before this step and does not guarantee the patch utility, so the image build can fail with command not found. Fix: install patch before running install_vllm_omni.sh.

🤖 AI Fix

Update container/templates/vllm_runtime.Dockerfile in the apt-get install RUN block immediately before the vLLM-Omni install to include patch in the apt-get install -y --no-install-recommends package list with jq.