RTVI BYOM Porting

Purpose

Port a VLM into the VSS RT-VLM service without hard-coding one checkpoint or GPU platform. Prefer configuration, then an RTVI adapter or vLLM plugin, and patch vLLM only when the supported extension points cannot express the model.

Use vss-deploy-dense-captioning for ordinary catalog-preset deployment of an already-wired model (for example cosmos-reason3). Use this skill for a custom or private checkpoint even when path 1 (configuration only) is enough, and when architecture, processor, weight mapping, runtime dependencies, or request adaptation requires repository work.

Model Contract

Before changing code, record:

• immutable model source and revision, license and access requirements;
• architecture, processor, tokenizer, quantization and context length;
• vision-token and frame-sampling assumptions;
• CUDA architecture, memory, vLLM/runtime and custom-kernel requirements.

For a private Hugging Face source, use HF_TOKEN only during authenticated download/cache population. Never print it, embed it in MODEL_PATH, or bake it into an image. Prove the cached model starts after the temporary credential is removed. If remote model code is required, review it first and pair VLM_TRUST_REMOTE_CODE=true with an exact allowlist entry. For the VSS profile, set RTVI_VLM_MODEL_PATH_ALLOWLIST; inside the service container this becomes RTVI_MODEL_PATH_ALLOWLIST. Likewise, the profile input RTVI_VLM_ALLOW_UNSAFE_MODEL_CONFIG becomes RTVI_ALLOW_UNSAFE_MODEL_CONFIG; do not enable it without reviewing the blocked config hooks.

Integration Decision

Stop at the first path that works:

1. Configuration only: use VLM_MODEL_TO_USE=vllm-compatible and set MODEL_PATH to an ngc: URI or a mounted snapshot path. A git: source is acceptable only for exploration because the develop downloader does not pin Hugging Face revisions; use a revision-pinned mounted snapshot for reproducible evidence.
2. RTVI adapter: normalize config, processor, request or response behavior inside services/rtvi/rt-vlm/ while preserving existing model behavior.
3. Plugin or shim: register architecture and deterministic weight mappings without editing vendored vLLM internals.
4. Version-gated patch: patch vLLM only after recording why the first three paths cannot work. Keep it optional and add a focused regression test.

Use VLM_MODEL_TO_USE=custom with MODEL_IMPLEMENTATION_PATH only for a custom RTVI model implementation. In the VSS Compose profile these are exposed as RTVI_VLM_MODEL_TO_USE, RTVI_VLM_MODEL_PATH, and RTVI_VLM_MODEL_IMPLEMENTATION_PATH.

For source, adapter, plugin, or custom-backend changes, build the RT-VLM image from services/rtvi/rt-vlm/, test it with standalone Compose via RTVI_IMAGE, then select the same repository and tag in the VSS profile via VSS_RT_VLM_IMAGE and VSS_RT_VLM_TAG. Record the tested image digest. A host MODEL_IMPLEMENTATION_PATH alone is insufficient: the implementation must exist at that path inside the selected image or an explicit Compose bind mount.

Read references/vllm-porting.md before changing model loading, registration, weight mapping, kernels, cache behavior or vLLM.

Hard Gates

• Do not enable eager mode by default. If it is unavoidable, capture the exact blocker, performance impact and removal condition.
• Do not bind generic model logic to H100, B200, RTX, Orin, x86 or Jetson. Capability-detect or use existing configuration boundaries.
• Keep secrets out of commands, logs, images, reports and committed files.
• Do not accept a successful load as proof of a successful port. Output must be legible and grounded for every claimed modality.

Validation

Run the smallest relevant checks in this order:

1. Static checks and focused tests for changed Python, shell and configuration.
2. Start RT-VLM through the canonical VSS Compose/profile path and verify /v1/health/ready and /v1/models. First use standalone Compose when a revision-pinned host model snapshot must be mounted; the VSS profile does not expose MODEL_ROOT_DIR on develop.
3. Smoke-test text, image and video inputs for every claimed modality. Follow the OpenAPI contract and curl examples in references/quality-gates.md: set media_type on /v1/files and /v1/generate_captions, not as a top-level field on /v1/chat/completions. Reject Git LFS pointer files before upload.
4. Retain raw redacted responses, and reject gibberish, repetition, empty generations, or ungrounded claims.
5. Run relevant caption, event-detection or summarization accuracy checks.
6. Record latency, throughput, GPU memory/utilization, eager state and custom kernel state.

Use skills/deployment/rtvi-byom-porting/scripts/byom_port_report.py to turn observed JSON facts into a compact Markdown report. Evidence is mandatory for a PASS result; the helper does not manufacture it. Run python3 skills/deployment/rtvi-byom-porting/scripts/tests/test_byom_port_report.py after changing the helper.

Completion Contract

Report the exact model revision and backend, integration path, eager-mode state, platform gating, smoke/accuracy/performance evidence, remaining blockers and the next bounded validation step.