feat(trtllm): add OpenEngine sidecar

[connorcarpenter15]

Summary

• add the engine-neutral Rust dynamo-openengine-sidecar worker, configured by an OpenEngine endpoint and capability discovery rather than TRT-LLM-specific control flow
• map Dynamo preprocessed requests to OpenEngine generation and prefill/decode streams, including sampling, stopping, logprobs, priority, cache controls, ordered media, media options, and LoRA selection
• forward cancellation, drain, health, load, and KV event sources through the existing backend-common worker framework
• reuse backend-common's LoRA controller for adapter downloads, deterministic IDs, model cards, and control endpoints
• add aggregate, 1P1D, multimodal, audio, and LoRA launch examples and DGD manifests
• preserve media modality placeholders and adapter/media semantics through frontend preprocessing

Why

This adds a process-isolated TRT-LLM deployment path without replacing Dynamo's existing in-process TRT-LLM backend. The sidecar remains engine-neutral and communicates with the GPU engine exclusively through OpenEngine gRPC.

Dependencies

• feat(protocol): add revision 2 multimodal contract and packages openengine#20
• feat(openengine): add sibling server for TRT-LLM NVIDIA/TensorRT-LLM#16336
• pinned OpenEngine schema commit: cea19cb06acf03c911b84d5c147e519b60dd92a6

Validation

• Rust formatting and Clippy with warnings denied
• focused LLM preprocessor, backend-common, and fake OpenEngine sidecar tests passed
• OpenEngine descriptor, generated-code, Python/Rust fixture, and package checks passed in the dependent contract repository
• TRT-LLM fake-servicer and protocol compatibility suite: 64 passed
• computelab frontend smoke matrix passed:
  • Qwen3-0.6B aggregate text and LoRA lifecycle on 1 H100
  • Qwen3-0.6B context-first 1P1D text and LoRA on both roles on 2 NVLink H100s
  • Qwen3-VL-2B-Instruct image/video aggregate and 1P1D on 1 then 2 H100s
  • Phi-4-multimodal-instruct aggregate audio on 1 H100

Each GPU smoke verified nonempty output, terminal usage, zero post-request load, and graceful shutdown; disaggregated cases also verified the actual prefill/decode handoff.

Compatibility

The existing in-process TRT-LLM backend remains available. OpenEngine deployments are opt-in through the new sidecar binary and launch examples.

[github-actions]

🌿 Fern Docs Preview: https://nvidia-preview-019f5cba-5caa-702a-8be3-d3c68aff444d.docs.buildwithfern.com/dynamo/dev

[datadog-official]

⚠️ Warnings

🚦 8 Pipeline jobs failed

PR | backend-status-check     

PR | dynamo-runtime / rust-gpu     

PR | dynamo-status-check     

View all 8 failed jobs.

ℹ️ Info

🎯 Code Coverage (details)
• Patch Coverage: 100.00%
• Overall Coverage: 43.16% (-2.28%)

Useful? React with 👍 / 👎

_{This comment will be updated automatically if new data arrives.
🔗 Commit SHA: ac8502c | Docs | Give us feedback!}