API Documentation

Returns per-SDK-UUID compute usage stats (successful tasks only), bucketed by hour. Rolling 3-day window. Only hours with actual activity are included.

Query Parameters

{
  "from": "2026-03-15T21",
  "to": "2026-03-16T07",
  "data": [
    { "uuid": "sdk-python-c5d145", "hour": "2026-03-15T21", "cpu_ms": 0, "gpu_ms": 72917, "cpu_tasks": 0, "gpu_tasks": 1867 }
  ]
}

Same format as above, but for failed tasks only. Tracked separately from successful task stats.

Returns basic health status and version information.

{
  "status": "healthy",
  "version": "1.1.0-abc1234",
  "timestamp": "2026-02-08T10:30:00.000Z"
}

Returns detailed health information including all service statuses and resource counts.

{
  "status": "healthy",
  "services": {
    "device_registry": { "status": "healthy", "devices": 5 },
    "resource_pool": { "status": "healthy", "available": 3 },
    "job_manager": { "status": "healthy", "active_jobs": 2 },
    "model_store": { "status": "healthy", "models": 4 },
    "websocket": { "status": "healthy", "connections": 5 }
  },
  "resources": { ... },
  "jobs": { ... }
}

Upload an ONNX or TorchScript model file. Use multipart/form-data encoding. Specify built-in preprocessing and/or postprocessing adapters to enable server-side format conversion via AWS Lambda.

Request (multipart/form-data)

{
  "id": "model_abc12345",
  "name": "my-model",
  "format": "onnx",
  "size": 4096,
  "checksum": "sha256:...",
  "uploaded_at": "2026-02-08T10:30:00.000Z",
  "preprocessing": "image_classification",
  "postprocessing": "top_k_labels",
  "input_schema": { ... },
  "labels": ["cat", "dog"]
}

Returns all models owned by the authenticated user.

{
  "models": [
    { "id": "model_abc", "name": "my-model", ... }
  ],
  "total": 1
}

Returns detailed information about a specific model.

Deletes a model. Returns 204 No Content on success.

Update input/output schema, labels, or adapter selections on an existing model without re-uploading. Only provided fields are modified.

Request Body

Returns updated model info.

Submit a single inference request. By default runs synchronously (waits for result). Use async mode for fire-and-forget.

If the model has an input_schema, the input is validated against it (returns 400 on mismatch). If the model has a postprocessing adapter (other than passthrough), the result is the business format produced by the dcmp-postprocess Lambda (e.g. { predictions: [...] }) instead of raw {data, shape}. When preprocessing is set, input is expected in the business format the adapter consumes.

Note: adapter conversion currently applies only to synchronous single inference. Asynchronous calls (async=true) and POST /inference/batch return raw {data, shape} regardless of adapter configuration.

Request Body

Async Mode: Set via query param ?async=true, header X-Async: true, or body field "async": true

{
  "job_id": "infer_abc123",
  "status": "completed",
  "result": { "data": [0.8, 0.2], "shape": [1, 2] },
  "latency_ms": 45,
  "device_id": "dev_xyz789"
}

{
  "job_id": "infer_abc123",
  "status": "queued",
  "message": "Inference job queued",
  "poll_url": "/api/v1/inference/infer_abc123"
}

Submit multiple inference requests as a batch. Always runs asynchronously.

Adapter limitation: the batch route does not currently invoke preprocessing / postprocessing adapters. Inputs must be raw {data, shape} tensors and results are returned as raw tensors, even for models that declare adapter fields.

Request Body

{
  "job_id": "batch_abc123",
  "project_id": "batch_1707384600000",
  "status": "queued",
  "total_tasks": 100
}

Get the status and results of an inference job. For batch jobs, includes progress and partial results.

Cancel a running or queued inference job.

Start a distributed training job. The system automatically distributes batches across available devices and aggregates gradients.

Request Body

data_config

training_config

resource_config

Devices are allocated proportional to job weight (total_samples * epochs). If no devices are free, the job queues until the scheduler rebalances and drains a device from an over-allocated job.

{
  "job_id": "train_abc123",
  "status": "queued",
  "estimated_devices": 3
}

Get detailed status of a training job including progress, loss, and assigned devices.

{
  "job_id": "train_abc123",
  "status": "running",
  "model_id": "model_xyz",
  "config": { "epochs": 10, "learning_rate": 0.001 },
  "progress": {
    "current_epoch": 3,
    "total_epochs": 10,
    "batches_completed": 150,
    "total_batches": 500,
    "percent_complete": 30,
    "current_loss": 0.4523,
    "best_loss": 0.4102
  },
  "assigned_devices": ["dev_a", "dev_b", "dev_c"]
}

Pause a running training job. Can be resumed later.

Resume a paused training job.

Cancel a training job. Stops all distributed training tasks.

Returns all connected devices with their status and capabilities.

Query Parameters

Returns aggregated statistics about connected devices.

Returns detailed information about a specific device including hardware, capabilities, and metrics.

Requests logs from a connected SDK device via WebSocket. Returns the device's recent log entries.

Query Parameters

Create a new project for tracking tasks.

Request Body

Returns all projects for the authenticated user.

Returns full project details including all metrics.

Returns computed analytics including latency percentiles, throughput data, and instance type comparisons.

Manually mark a realtime project as complete.

Cancel a project and all associated jobs.

Returns total compute usage for the authenticated customer including GPU/CPU hours, task counts, and estimated cost.

Returns compute usage for a specific project.

Returns hourly usage data for time-series analysis.

Query Parameters

Returns usage data for all customers. Requires X-Admin-Key header.

Returns aggregate training efficiency statistics.

Query Parameters

Returns side-by-side comparison of AWS instances vs user-contributed devices.

Chart-ready data. Types: tta, scaling, power, throughput.

Returns currently active training/inference runs with real-time metrics.

Detailed efficiency metrics for a specific job.

Returns device statistics grouped by platform (nodejs, python, windows).

Returns jobs from the last 5 days, sorted by activity.

Query Parameters

Returns up to 50 most recent projects, active first.

Server-Sent Events endpoint for real-time project updates. Pushes stats every 2 seconds.

Query Parameters

Returns individual task records across all projects.

Query Parameters

Returns time-bucketed throughput data for graphing.

Query Parameters

Training efficiency overview with recent TTA, scaling, and power data.

Training and inference comparison between AWS and user devices, formatted for dashboard charts.

Current cluster state: device allocation, CPU/GPU usage, and capacity metrics.

Start a quick inference or training benchmark.

Request Body

Start a step-load performance benchmark with configurable RPS targets.

Request Body

Get status and results of a performance benchmark.