openclaw-gpu-bridge

# @elvatis_com/openclaw-gpu-bridge

OpenClaw plugin to offload ML tasks (BERTScore + embeddings) to one or many remote GPU hosts.

## v0.2 Highlights

- Multi-GPU host pool (`hosts[]`) with:
  - round-robin or least-busy load balancing
  - automatic failover
  - periodic host health checks
- Backward compatibility with v0.1 (`serviceUrl` / `url`)
- Flexible model selection per request (`model` / `model_type`)
- GPU service model caching (on-demand loading)
- Optional transfer visibility via `/status` endpoint + batch progress logs

---

## Tools

- `gpu_health`
- `gpu_info`
- `gpu_status` (new in v0.2)
- `gpu_bertscore`
- `gpu_embed`

---

## OpenClaw Plugin Config

### v0.2 (recommended)

```json
{
  "plugins": {
    "@elvatis_com/openclaw-gpu-bridge": {
      "hosts": [
        {
          "name": "rtx-2080ti",
          "url": "http://your-gpu-host:8765",
          "apiKey": "gpu-key-1"
        },
        {
          "name": "rtx-3090",
          "url": "http://your-second-gpu-host:8765",
          "apiKey": "gpu-key-2"
        }
      ],
      "loadBalancing": "least-busy",
      "healthCheckIntervalSeconds": 30,
      "timeout": 45,
      "models": {
        "embed": "all-MiniLM-L6-v2",
        "bertscore": "microsoft/deberta-xlarge-mnli"
      }
    }
  }
}
```

### v0.1 compatibility

```json
{
  "plugins": {
    "@elvatis_com/openclaw-gpu-bridge": {
      "serviceUrl": "http://your-gpu-host:8765",
      "apiKey": "gpu-key",
      "timeout": 45
    }
  }
}
```

### Config reference

- `hosts`: array of GPU hosts (v0.2)
- `serviceUrl` / `url`: legacy single-host config
- `loadBalancing`: `round-robin` or `least-busy`
- `healthCheckIntervalSeconds`: host health polling interval
- `timeout`: request timeout for compute endpoints
- `apiKey`: fallback API key for hosts that do not define per-host key
- `models.embed`, `models.bertscore`: plugin-side default models

---

## GPU Service (Python) Setup

```bash
cd gpu-service
pip install -r requirements.txt
uvicorn g...

# @elvatis_com/openclaw-gpu-bridge

OpenClaw plugin to offload ML tasks (BERTScore + embeddings) to one or many remote GPU hosts.

## v0.2 Highlights

- Multi-GPU host pool (`hosts[]`) with:
  - round-robin or least-busy load balancing
  - automatic failover
  - periodic host health checks
- Backward compatibility with v0.1 (`serviceUrl` / `url`)
- Flexible model selection per request (`model` / `model_type`)
- GPU service model caching (on-demand loading)
- Optional transfer visibility via `/status` endpoint + batch progress logs

---

## Tools

- `gpu_health`
- `gpu_info`
- `gpu_status` (new in v0.2)
- `gpu_bertscore`
- `gpu_embed`

---

## OpenClaw Plugin Config

### v0.2 (recommended)

```json
{
  "plugins": {
    "@elvatis_com/openclaw-gpu-bridge": {
      "hosts": [
        {
          "name": "rtx-2080ti",
          "url": "http://your-gpu-host:8765",
          "apiKey": "gpu-key-1"
        },
        {
          "name": "rtx-3090",
          "url": "http://your-second-gpu-host:8765",
          "apiKey": "gpu-key-2"
        }
      ],
      "loadBalancing": "least-busy",
      "healthCheckIntervalSeconds": 30,
      "timeout": 45,
      "models": {
        "embed": "all-MiniLM-L6-v2",
        "bertscore": "microsoft/deberta-xlarge-mnli"
      }
    }
  }
}
```

### v0.1 compatibility

```json
{
  "plugins": {
    "@elvatis_com/openclaw-gpu-bridge": {
      "serviceUrl": "http://your-gpu-host:8765",
      "apiKey": "gpu-key",
      "timeout": 45
    }
  }
}
```

### Config reference

- `hosts`: array of GPU hosts (v0.2)
- `serviceUrl` / `url`: legacy single-host config
- `loadBalancing`: `round-robin` or `least-busy`
- `healthCheckIntervalSeconds`: host health polling interval
- `timeout`: request timeout for compute endpoints
- `apiKey`: fallback API key for hosts that do not define per-host key
- `models.embed`, `models.bertscore`: plugin-side default models

---

## GPU Service (Python) Setup

```bash
cd gpu-service
pip install -r requirements.txt
uvicorn g...