Skip to main content

​
SDK & CLI Guide

The Agnitra CLI mirrors the Python SDK so teams can trigger optimizations, collect telemetry, and emit usage events from any environment.

​
CLI Commands

​
agnitra optimize

Optimize a TorchScript model and optionally save the optimized artifact. 
agnitra optimize \
  --model tinyllama.pt \
  --input-shape 1,16,64 \
  --output dist/tinyllama_optimized.pt \
  --target A100
Key flags:
  • --device — moves the model to a specific device (e.g. cuda:0).
  • --disable-rl — skip PPO fine-tuning passes.
  • --offline — disable control plane calls (requires enterprise license).
  • --require-license — fail if license validation is unavailable.
  • --license-seat / --license-org — override license metadata sent to the control plane.

​
agnitra-api

Start the Agentic Optimization API backed by Starlette: 
agnitra-api --host 0.0.0.0 --port 8080
Endpoints:
  • POST /optimize — synchronous or async queued optimization.
  • GET /jobs/{id} — poll async job status.
  • POST /usage — convert telemetry snapshots into marketplace usage records.
API keys are read from AGNITRA_API_KEY (and variants) and enforced for every request.

​
agnitra-dashboard

Spin up the HTML dashboard for local telemetry review: 
agnitra-dashboard --host 127.0.0.1 --port 3000

​
Python SDK Highlights

  • agnitra.optimize(model, input_tensor=...) returns a RuntimeOptimizationResult including the optimized model, usage event, and patch metadata.
  • agnitra.sdk.resolve_input_tensor synthesizes input tensors based on shape hints or example tensors on the module.
  • Usage events expose GPU hours saved, cost savings, and marketplace metadata (see agnitra/core/metering/usage_meter.py).

​
Example

from agnitra import optimize

result = optimize(
    model,
    input_tensor=sample,
    project_id="demo-project",
    metadata={"source": "notebook"}
)

usage = result.usage_event
print(f"GPU hours saved: {usage.gpu_hours_saved:.6f}")

​
Troubleshooting

  • Missing PyTorch: install torch>=2.0 or ensure CUDA libs are discoverable.
  • Control plane unavailable: pass --offline or set AGNITRA_CONTROL_PLANE_URL to the reachable endpoint.
  • Stripe/NVML optional deps: install extras (agnitra[nvml], agnitra[marketplace]) to enable GPU telemetry and marketplace dispatchers.