livepeer-modules-transcode-runners

Standalone home for the Livepeer video transcode runners. This repo ships three Go HTTP runners plus shared FFmpeg/transcode logic, vendor-specific runtime images for NVIDIA, Intel, and AMD, and direct-runner smoke tooling.

For agents: start at AGENTS.md.

What this repo ships

Image	Purpose	Default endpoint
transcode-runner-nvidia / -intel / -amd	Single-rendition VOD transcode	POST /v1/video/transcode
abr-runner-nvidia / -intel / -amd	Multi-rendition ABR ladder transcode	POST /v1/video/transcode/abr
live-runner-nvidia / -intel / -amd	Live session runner for gateway-ingest RTMP and HLS push output	POST /v1/video/live/sessions
transcode-tester	Node integration smoke harness	n/a

Shared code lives in transcode-core/. Build and runtime infrastructure lives in infra/.

Build

Every gesture is Docker-first.

./build-images.sh build
./build-images.sh build transcode-runner-nvidia abr-runner-nvidia live-runner-nvidia
./build-images.sh validate
./build-images.sh clean

No host Go or host Node required.

Compose

For vendor-generic local bring-up, use:

docker compose -f infra/compose/docker-compose.runners.yml --profile nvidia up -d

For a production-oriented NVIDIA node, use:

cp infra/env/nvidia-prod.env.example .env.nvidia-prod
docker compose --env-file .env.nvidia-prod -f infra/compose/docker-compose.nvidia-prod.yml up -d

That production stack defaults to the GTX 1080 tuned preset pack in infra/presets/nvidia-gtx1080-transcode.yaml and infra/presets/nvidia-gtx1080-abr.yaml. The live runner uses infra/presets/live.yaml by default.

Live runner mode

live-runner uses a single gateway-ingest shape:

• shared RTMP ingest on one port
• direct FFmpeg ingest through a local FIFO
• HLS upload to caller-supplied S3-compatible storage

The broker must include output_credential and ingest_accept.stream_key in the session-open request.

Repo layout

.
├── AGENTS.md, CLAUDE.md
├── README.md, DESIGN.md, BUILD.md, API.md
├── OPERATIONS.md, TESTING.md, SECURITY.md
├── build-images.sh
├── go.mod, go.sum
├── abr-runner/                 # ABR ladder runner source + embedded defaults
├── live-runner/                # remote live session runtime
├── transcode-runner/           # single-rendition runner source + embedded defaults
├── transcode-core/             # shared FFmpeg / GPU / preset / HLS logic
├── transcode-tester/           # Node smoke harness
├── infra/
│   ├── compose/                # docker-compose overlays
│   ├── dockerfiles/            # all Dockerfiles; build context is repo root
│   ├── env/                    # .env.example templates
│   ├── offerings/              # runner offering manifests
│   └── presets/                # operator-editable preset YAMLs
└── docs/
    ├── exec-plans/
    └── references/

GPU support

All three vendor families are in scope:

• NVIDIA NVENC/NVDEC via CUDA 13 runtime images
• Intel QSV / VAAPI
• AMD VAAPI

Strict hardware mode is the default policy across vendors:

• jobs fail closed when the requested path would require CPU-only processing
• startup filters presets against the actual usable hardware/runtime path
• visible GPU hardware is not enough; the runner requires a working decode+encode runtime path

Current NVIDIA build note:

• CUDA 13.2.1 is supported
• FFmpeg is built without libnpp on this CUDA line due upstream API incompatibility with FFmpeg 7.1.3

What is intentionally absent

• No broker, gateway, or payment-layer code
• No checked-in secrets, keystores, or operator-local state
• No copied historical plan/doc tree from the source monorepo