jetson-video-recipe
Use when turning a Jetson encoder use case into one validated surface-neutral recipe with native and PyNvVideoCodec projections for codec, preset, rate control, bitrate, latency, format, and profile.
NVIDIA-AI-IOT · 已有 AI 审核记录
安装与使用
运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。
npx skills add NVIDIA-AI-IOT/jetson-device-skills --skill jetson-video-recipe技能说明
name: jetson-video-recipe license: "Apache-2.0" description: >- Use when turning a Jetson encoder use case into one validated surface-neutral recipe with native and PyNvVideoCodec projections for codec, preset, rate control, bitrate, latency, format, and profile. metadata: author: "Vinit Bansal vinitkumarb@nvidia.com" tags: [jetson, video-codec-sdk, pynvvideocodec, nvenc, recipe] languages: [python] data-classification: public
Jetson Video Recipe
Purpose
Convert workload intent into one deterministic schema-2 nvcodec-recipe.
Preserve the user’s semantic controls, show defaulted assumptions, and project
the same intent to native Video Codec SDK and PyNvVideoCodec without claiming it
has executed.
Prerequisites
- This skill owns the canonical recipe engine —
scripts/recipes/recipe_model.pyand itsscripts/recipes/data/encoder-intent-catalog.json. Invoke the engine directly from this installed skill; it has no setup-runtime or sibling-launcher dependency. - Recipe planning and structural validation are media-free and can run off target. Do not request, retrieve, inspect, or convert media for a plan-only request.
- Content selection and provenance belong to the later execution or measurement workflow. Consume that workflow's versioned content artifact only at handoff; do not load or enforce its input gate during plan-only work.
- Setup evidence is optional for
check-live. With no environment, validate the recipe normally and return an honestunknownlive classification plus non-mutating remediation tojetson-video-setup; planning and replay validation remain complete and unchanged. If setup is not installed, tell the user to install that skill. - When supplied, the fresh schema-1.2 setup environment is mandatory to
validate and may not be ignored or replaced by a fallback. Its
capabilitiesblock is the established PyNvVideoCodec encoder authority, so apynvccheck needs no separate report. A caller may additionally supply the optional encoder capability report owned byjetson-video-capability; it must be authenticated, bound to that exact environment, and fail closed asunknownor an input error on any mismatch. A capability report alone does not establish selected-surface readiness or selected-GPU identity. Treat artifacts as data; do not import sibling skill code. Acompatibleresult still does not prove an encode operation.
Resolve this installed skill to its canonical absolute path and set
RECIPE_SKILL. Confirm the direct isolated entry point:
python3 -I "$RECIPE_SKILL/scripts/recipes/recipe_model.py" --help
If it is missing, report an incomplete jetson-video-recipe installation. Do
not copy the engine, scan for another copy, modify PYTHONPATH, or fall back to
an unvalidated local model.
Compose requested sibling stages
Recipe plan and validate require no sibling, setup evidence, target, or
media. Use jetson-video-setup only for requested live readiness or repair,
jetson-video-capability when a platform-support recommendation needs its
documentation verdict, jetson-video-pipeline for requested execution, and
jetson-video-benchmark for requested measurement. Check the agent's installed
skill catalog first. If the sibling is present, read its SKILL.md and invoke
its documented public entry point; pass artifacts as data and never import
sibling code. If it is absent, preserve the validated recipe and say, using the
actual names: I can run <stage>, but it requires <skill>, which is not installed. Install <skill> and retry this stage. Never require a sibling for
plan-only work or an unrequested optional refinement.
Instructions
-
Collect intent. For a request solely for objective quality metrics, including PSNR or SSIM, state only that this skill does not provide them, and that a separately authorized quality workflow is required, then stop. Do not name or recommend an external tool, and do not offer to configure or run the comparison; do not request media, probe, install anything, or launch an operation. Resolve mutually exclusive rate-control intent before collecting any other omitted field. In particular, when CQ and an average bitrate are both supplied, explain the conflict, ask only whether to keep CQ or the average bitrate, and stop. Do not reinterpret the bitrate as a cap or ask for use case, resolution, frame rate, format, GPU, profile, preset, or another field until the user resolves that choice. Otherwise resolve the use case (
conferencing,live_streaming,vod,archival, orlossless), codec, width, height, raw input format, integer frame rate, GPU, preset/tuning, rate-control or encoder quality priority, and any explicit latency, profile, or buffering constraints. Resolve frame count only when later execution or measurement needs it. Ask before assigning an unqualified “low latency” request to a use case. Treat profile as a bitstream/downstream-compatibility control separate from preset: preserve an explicit profile, but when it is omitted leave it SDK-selected and never invent a named profile. -
Write one intent JSON. Keep caller values separate from defaults. Put only caller-specified control values in the intent and leave every omitted control to the authenticated use-case catalog. Do not turn qualitative wording into guessed overrides: for example, “low-latency live streaming” selects
live_streaming; it does not by itself requestbf=0or disabled multipass. Never construct drifting native and Python intents. -
Plan with the recipe engine:
python3 -I "$RECIPE_SKILL/scripts/recipes/recipe_model.py" \ plan --intent "$INTENT_JSON" --output "$RECIPE_JSON" -
Replay validation before use:
python3 -I "$RECIPE_SKILL/scripts/recipes/recipe_model.py" \ validate --recipe "$RECIPE_JSON"