DYN-2940: Update vLLM-Omni to v0.20.0

[ptarasiewiczNV]

Summary

• bump Dynamo container config to vLLM-Omni v0.20.0
• load vLLM-Omni v0.20 deploy configs through the model-aware loader
• fill missing default shared-memory connector edges for v0.20 deploy configs
• adapt disaggregated stage routing and stage handoff to v0.20 processor inputs
• accumulate multimodal handoff payload chunks before passing Qwen thinker output to the talker stage
• point the GLM disaggregated launch script at the v0.20 built-in deploy config path

Latest Image

• nvcr.io/nvidian/dynamo-dev/dynamo:vllm-runtime-ptarasiewicz-dyn2940-v0200-7e74ee9072d2
• digest: sha256:37edb1122b3ad9d9d6ae721ae94e9568b9d62407cea3bcade74c807e0a820015
• package check: vllm==0.20.1+cu129, vllm-omni==0.20.0, ai-dynamo==1.2.0

Validation

• python -m py_compile components/src/dynamo/vllm/omni/stage_worker.py components/src/dynamo/vllm/tests/omni/test_omni_stage_worker.py: passed
• pre-commit run --files components/src/dynamo/vllm/omni/stage_worker.py components/src/dynamo/vllm/tests/omni/test_omni_stage_worker.py: passed
• commit hooks for 7e74ee9072d2: passed
• focused host pytest for the new worker test: skipped because host vLLM-Omni deps are unavailable
• source-mounted container smoke for connector payload accumulation: passed
• built-image container smoke for connector payload accumulation: passed
• Docker build for dynamo:dyn2940-vllm-omni-v020-7e74ee9072d2: passed
• Nebius H200 Qwen3-Omni text-only validation on previous image: passed
• Nebius H200 Qwen3-Omni text+audio validation on previous image: reproduced talker handoff failure; 7e74ee9072d2 is the targeted fix and still needs rerun
• Nebius H200 GLM image validation on previous image: returned HTTP 200 with image payload; needs rerun on 7e74ee9072d2

Current Blocker

• Kubernetes rerun is blocked by expired Teleport credentials for dynamo-nebius-2.
• Updated temp manifests are ready locally with the latest image:
  • /tmp/dyn2940-qwen3-v020-job.yaml
  • /tmp/dyn2940-glm-v020-job.yaml

[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]

🌿 Fern Docs Preview: https://nvidia-preview-9da74dcc-e920-4e50-82c7-99e7bf540992.docs.buildwithfern.com/dynamo/dev