# PeerTube Remote Runner — GPU Transcoding

Deploy a PeerTube remote runner with NVENC GPU transcoding. The runner pulls jobs from PeerTube over WebSocket, transcodes with the GPU, and uploads HLS streams back.

Use this when adding a new runner node, rebuilding an existing one, or re-registering after a PeerTube rebuild.

---

## Prerequisites

- PeerTube instance running and accessible (HTTP, not necessarily HTTPS)
- PeerTube admin credentials or API access to generate runner registration tokens
- Target machine with:
  - NVIDIA GPU with NVENC support (Maxwell gen 2+ / GTX 950+)
  - NVIDIA drivers installed and working (`nvidia-smi` returns output)
  - Node.js 18+ installed
  - SSH access from CC host

If the target machine needs NVIDIA drivers or Node.js, see `proxmox-create-ubuntu-vm.md` Steps 9 and 11.

---

## Inputs

Prompt the user for all of these before executing:

```
RUNNER_HOST=            # SSH alias or IP for the runner machine (e.g., cortex)
RUNNER_NAME=            # Human-readable runner name (e.g., "cortex-nvenc")
RUNNER_USER=            # User to run the service as (e.g., "zvx")
PT_URL=                 # PeerTube instance URL reachable from runner (e.g., "http://100.64.0.23:9000")
PT_HOST_HEADER=         # PeerTube's public hostname for Host header (e.g., "stream.echo6.co")
PT_ADMIN_USER=          # PeerTube admin username (e.g., "root")
PT_ADMIN_PASS=          # PeerTube admin password
INSTALL_DIR=            # Where to put runner config (default: /opt/peertube-runner)
```

---

## Step 1: Verify GPU

```bash
ssh $RUNNER_HOST 'nvidia-smi --query-gpu=name,driver_version,memory.total,encoder.stats.sessionCount --format=csv,noheader'
```

### Gate

Must return GPU name, driver version, and VRAM. If it fails:

- No output → NVIDIA drivers not installed. See `proxmox-create-ubuntu-vm.md` Step 9.
- "NVML: Driver/library version mismatch" → reboot the machine.
- "No devices found" → GPU passthrough not configured (VMs) or hardware issue.

Check NVENC specifically:

```bash
ssh $RUNNER_HOST 'nvidia-smi -q | grep -A 5 "Encoder"'
```

Must show encoder session info. If "N/A", the GPU doesn't support NVENC or drivers are too old.

---

## Step 2: Install peertube-runner

PeerTube runner is distributed via npm.

```bash
ssh $RUNNER_HOST 'which node && node --version'  # Must be 18+
```

Install the runner:

```bash
ssh $RUNNER_HOST 'sudo npm install -g @peertube/peertube-runner'
ssh $RUNNER_HOST 'which peertube-runner && peertube-runner --version'
```

### Gate

`peertube-runner --version` must return a version number. If npm install fails, check Node.js version (must be 18+).

Create config directory:

```bash
ssh $RUNNER_HOST "sudo mkdir -p $INSTALL_DIR && sudo chown $RUNNER_USER:$RUNNER_USER $INSTALL_DIR"
```

---

## Step 3: Generate Registration Token on PeerTube

Get a registration token from the PeerTube instance. This requires admin access.

### Option A: Via API

```bash
# Get OAuth client credentials
CLIENT_CREDS=$(ssh $RUNNER_HOST "curl -s $PT_URL/api/v1/oauth-clients/local -H 'Host: $PT_HOST_HEADER'")
CLIENT_ID=$(echo "$CLIENT_CREDS" | jq -r '.client_id')
CLIENT_SECRET=$(echo "$CLIENT_CREDS" | jq -r '.client_secret')

# Get admin token
TOKEN_RESP=$(ssh $RUNNER_HOST "curl -s $PT_URL/api/v1/users/token \
  -H 'Host: $PT_HOST_HEADER' \
  --data 'client_id=$CLIENT_ID&client_secret=$CLIENT_SECRET&grant_type=password&username=$PT_ADMIN_USER&password=$PT_ADMIN_PASS'")
ACCESS_TOKEN=$(echo "$TOKEN_RESP" | jq -r '.access_token')

# Generate runner registration token
REG_TOKEN=$(ssh $RUNNER_HOST "curl -s -X POST $PT_URL/api/v1/runners/registration-tokens/generate \
  -H 'Host: $PT_HOST_HEADER' \
  -H 'Authorization: Bearer $ACCESS_TOKEN' | jq -r '.registrationToken'")

echo "Registration token: $REG_TOKEN"
```

### Option B: Via PeerTube Admin UI

1. Log into PeerTube as admin
2. Administration → System → Runners
3. Click "Generate registration token"
4. Copy the token

### Gate

Must have a registration token string. If the API returns errors:

- 401 → wrong admin credentials
- 404 → PeerTube version too old (runners require v5.2+)
- Connection refused → PeerTube not reachable from runner. Check URL and network.

---

## Step 4: Register Runner

```bash
ssh $RUNNER_HOST "peertube-runner register \
  --url $PT_URL \
  --registration-token $REG_TOKEN \
  --runner-name $RUNNER_NAME"
```

### Gate

Must complete without errors. Verify registration:

```bash
ssh $RUNNER_HOST 'peertube-runner list-registered'
```

Must show the PeerTube instance URL. If registration fails:

- "Invalid registration token" → token already used or expired. Generate a new one.
- "ECONNREFUSED" → runner can't reach PeerTube. Test: `curl -s $PT_URL/api/v1/config`
- "self-signed certificate" → if PeerTube uses HTTPS with self-signed cert, use `NODE_TLS_REJECT_UNAUTHORIZED=0` (not recommended) or fix the cert.

---

## Step 5: Configure NVENC

The runner auto-detects ffmpeg capabilities, but verify NVENC is available to ffmpeg:

```bash
ssh $RUNNER_HOST 'ffmpeg -encoders 2>/dev/null | grep nvenc'
```

Must show `h264_nvenc` and `hevc_nvenc`. If missing, install ffmpeg with NVENC support:

```bash
# Ubuntu/Debian — the default ffmpeg usually includes NVENC if drivers are installed
ssh $RUNNER_HOST 'sudo apt install -y ffmpeg'

# Re-check
ssh $RUNNER_HOST 'ffmpeg -encoders 2>/dev/null | grep nvenc'
```

If still missing, the NVIDIA drivers may not include the encoding libraries. Install:

```bash
ssh $RUNNER_HOST 'sudo apt install -y libnvidia-encode-550'  # Match your driver version
```

---

## Step 6: Create systemd Service

```bash
ssh $RUNNER_HOST "sudo tee /etc/systemd/system/peertube-runner.service > /dev/null << 'EOF'
[Unit]
Description=PeerTube Remote Runner (NVENC)
After=network-online.target nvidia-persistenced.service
Wants=network-online.target
Requires=nvidia-persistenced.service

[Service]
Type=simple
User=$RUNNER_USER
Group=$RUNNER_USER
Environment=NODE_ENV=production
ExecStart=/usr/bin/peertube-runner server \
  --enable-job vod-hls-transcoding \
  --enable-job vod-audio-merge-transcoding \
  --enable-job live-rtmp-hls-transcoding \
  --enable-job video-studio-transcoding \
  --enable-job video-transcription
WorkingDirectory=/home/$RUNNER_USER
Restart=always
RestartSec=30
StandardOutput=journal
StandardError=journal
SyslogIdentifier=peertube-runner
MemoryMax=20G

[Install]
WantedBy=multi-user.target
EOF"

ssh $RUNNER_HOST 'sudo systemctl daemon-reload'
ssh $RUNNER_HOST 'sudo systemctl enable peertube-runner'
ssh $RUNNER_HOST 'sudo systemctl start peertube-runner'
```

### Gate

```bash
ssh $RUNNER_HOST 'systemctl is-active peertube-runner'
```

Must return `active`. If it fails, check logs:

```bash
ssh $RUNNER_HOST 'journalctl -u peertube-runner -n 50 --no-pager'
```

Common failures:
- "Cannot find module" → peertube-runner not installed globally, or PATH issue. Check `which peertube-runner`.
- "nvidia-persistenced.service not found" → remove the `Requires=` line if nvidia-persistenced isn't set up (it's optional but recommended).

---

## Step 7: Install Health Check

Cron script that auto-restarts the runner if it crashes or the GPU becomes inaccessible.

```bash
ssh $RUNNER_HOST "sudo tee $INSTALL_DIR/health.sh > /dev/null << 'HEALTH'
#!/bin/bash
LOG_TAG=\"peertube-runner-health\"

if ! systemctl is-active --quiet peertube-runner; then
    logger -t \$LOG_TAG \"Runner not active, restarting...\"
    systemctl restart peertube-runner
    sleep 10
fi

if ! pgrep -f \"peertube-runner server\" > /dev/null; then
    logger -t \$LOG_TAG \"Runner process not found, restarting service...\"
    systemctl restart peertube-runner
fi

if ! nvidia-smi > /dev/null 2>&1; then
    logger -t \$LOG_TAG \"GPU not accessible, restarting nvidia-persistenced and runner...\"
    systemctl restart nvidia-persistenced 2>/dev/null
    sleep 5
    systemctl restart peertube-runner
fi
HEALTH
chmod +x $INSTALL_DIR/health.sh"

# Add cron job (every 5 minutes)
ssh $RUNNER_HOST "(crontab -l 2>/dev/null | grep -v peertube-runner-health; echo '*/5 * * * * $INSTALL_DIR/health.sh') | crontab -"
```

Verify:

```bash
ssh $RUNNER_HOST 'crontab -l | grep peertube'
```

---

## Step 8: Test Transcoding

Upload a test video and verify the full pipeline works.

### Quick test via API

```bash
# Download a short test video
ssh $RUNNER_HOST 'curl -L -o /tmp/test-video.mp4 "https://test-videos.co.uk/vids/bigbuckbunny/mp4/h264/360/Big_Buck_Bunny_360_10s_1MB.mp4" 2>/dev/null'

# Upload to PeerTube (reuse ACCESS_TOKEN from Step 3)
ssh $RUNNER_HOST "curl -s -X POST $PT_URL/api/v1/videos/upload \
  -H 'Host: $PT_HOST_HEADER' \
  -H 'Authorization: Bearer $ACCESS_TOKEN' \
  -F 'videofile=@/tmp/test-video.mp4' \
  -F 'name=Runner Test Video' \
  -F 'channelId=1' \
  -F 'privacy=1' \
  -F 'waitTranscoding=true' | jq '{uuid, name, state}'"
```

### Verify GPU is processing

Within 30 seconds of upload:

```bash
ssh $RUNNER_HOST 'nvidia-smi --query-gpu=utilization.gpu,utilization.encoder,temperature.gpu,power.draw --format=csv,noheader'
```

GPU utilization and encoder utilization should be non-zero. If encoder shows 0% but GPU shows activity, NVENC isn't being used — check ffmpeg encoder detection (Step 5).

### Check runner logs

```bash
ssh $RUNNER_HOST 'journalctl -u peertube-runner -n 20 --no-pager | grep -i "transcod\|job\|error"'
```

Should show job pickup, transcoding progress, and completion.

### Clean up

```bash
# Delete test video via API (optional)
ssh $RUNNER_HOST "curl -s -X DELETE $PT_URL/api/v1/videos/<VIDEO_UUID> \
  -H 'Host: $PT_HOST_HEADER' \
  -H 'Authorization: Bearer $ACCESS_TOKEN'"

ssh $RUNNER_HOST 'rm -f /tmp/test-video.mp4'
```

---

## Verification Checklist

```bash
echo "=== PeerTube Runner Check ==="
echo ""
echo "GPU:          $(nvidia-smi --query-gpu=name --format=csv,noheader 2>/dev/null || echo 'MISSING')"
echo "NVENC:        $(ffmpeg -encoders 2>/dev/null | grep -c h264_nvenc) encoders"
echo "Runner ver:   $(peertube-runner --version 2>/dev/null || echo 'NOT INSTALLED')"
echo "Registered:   $(peertube-runner list-registered 2>/dev/null | grep -c 'http' || echo '0') instance(s)"
echo "Service:      $(systemctl is-active peertube-runner 2>/dev/null || echo 'NOT RUNNING')"
echo "Health cron:  $(crontab -l 2>/dev/null | grep -c peertube-runner || echo '0') entries"
echo "ffmpeg procs: $(pgrep -c ffmpeg 2>/dev/null || echo '0') active"
echo "GPU util:     $(nvidia-smi --query-gpu=utilization.gpu --format=csv,noheader 2>/dev/null || echo 'N/A')"
```

---

## Adding a Second PeerTube Instance

To register the same runner with another PeerTube instance:

```bash
peertube-runner register \
  --url <SECOND_PT_URL> \
  --registration-token <TOKEN> \
  --runner-name $RUNNER_NAME
```

The runner handles multiple registrations automatically — it polls all registered instances for jobs.

## Unregistering

```bash
# List registrations
peertube-runner list-registered

# Unregister from a specific instance
peertube-runner unregister --url $PT_URL
```

---

## Troubleshooting

### Runner picks up jobs but transcoding fails immediately

Check ffmpeg NVENC access:

```bash
ffmpeg -y -f lavfi -i testsrc=duration=5:size=1280x720:rate=30 -c:v h264_nvenc /tmp/nvenc-test.mp4
```

If this fails, NVENC isn't accessible to ffmpeg. Common causes: wrong driver version, missing libnvidia-encode, or GPU in use by another process that's holding all NVENC sessions.

### Runner shows 0 active jobs despite pending queue

- WebSocket connection issue. Check: `journalctl -u peertube-runner | grep -i websocket`
- Runner registered with wrong URL. Verify: `peertube-runner list-registered`
- PeerTube remote runners not enabled. Check PeerTube config: `transcoding.remote_runners.enabled` must be `true`

### GPU utilization stuck at 100% / NVENC sessions maxed

The RTX A4000 supports ~3 simultaneous NVENC sessions (consumer cards are limited to 3, pro cards vary). If all sessions are in use, new jobs queue on the runner side. This is normal — throughput is limited by NVENC session count, not GPU compute.

To increase throughput: patch the NVENC session limit (search "nvidia nvenc patch") or add a second runner node.

### Runner keeps disconnecting / restarting

- Check memory: `free -h`. The 20GB MemoryMax in the service file may be too low if processing many concurrent jobs. Increase if needed.
- Check disk space: transcoding uses temp space. Ensure `/tmp` or the runner's working directory has sufficient free space (10GB+ recommended).
- Network instability between runner and PeerTube. Use Tailscale IP instead of public URL for reliability.

### After PeerTube rebuild, runner can't connect

Registration is tied to the PeerTube instance. After a rebuild:

1. Unregister: `peertube-runner unregister --url $PT_URL`
2. Generate new registration token on the new PeerTube
3. Re-register: `peertube-runner register --url $PT_URL --registration-token <NEW_TOKEN> --runner-name $RUNNER_NAME`
4. Restart: `sudo systemctl restart peertube-runner`

---

## Quick Reference: Current Runners

| Runner | Host | GPU | PeerTube Instance | Status |
|--------|------|-----|-------------------|--------|
| cortex-nvenc | cortex (VM 150 on TOC) | RTX A4000 16GB | stream.echo6.co (CT 110) | Active |

---

## Whisper Transcription Setup

The runner also handles `video-transcription` jobs (auto-captioning via Whisper). A smart wrapper routes jobs based on audio duration:

### Smart Wrapper (`/usr/local/bin/whisper-smart` on cortex)

- **Model:** `medium` (good accuracy, fits in VRAM on float16)
- **GPU path (< 1 hour):** `--device cuda --compute_type float16` — ~3.7GB VRAM, fast
- **CPU path (>= 1 hour):** `--device cpu --compute_type int8` — ~8-11GB RAM, slow but avoids VRAM exhaustion
- **CPU serialization:** `flock --nonblock /tmp/whisper-cpu.lock` — only one CPU transcription at a time. If lock is held, wrapper exits 1 and the runner retries the job later.
- **Concurrency:** Runner config set to `concurrency = 2` — allows one GPU + one CPU job in parallel

### Symlink chain

```
/usr/local/bin/whisper-ctranslate2 → /usr/local/bin/whisper-smart (the smart wrapper)
/usr/local/bin/whisper-ctranslate2-real → /home/zvx/.local/bin/whisper-ctranslate2 (actual Python binary)
```

### Key config

- **Runner config:** `~/.config/peertube-runner-nodejs/default/config.toml` — `model = "medium"`, `concurrency = 2`
- **PeerTube config:** `/var/www/peertube/config/production.yaml` on CT 110 — `model-name: medium`
- **systemd MemoryMax:** `20G` (CPU int8 medium model peaks at ~11GB)

### Wrapper log

```bash
tail -f /tmp/whisper-wrapper.log   # Shows mode (GPU/CPU/CPU-BLOCKED), duration, args
```

---

*Last updated: 2026-02-17 — Added Whisper smart transcription setup, MemoryMax 12G→20G, concurrency 1→2*