Skip to main content

​
Getting Started

Agnitra delivers an end-to-end optimization platform that pairs model tuning, telemetry, and usage-based billing in a single developer flow. Use the SDK or CLI to profile workloads, generate optimized TorchScript artifacts, and push structured usage records into your control plane or marketplace integrations. This quickstart walks through installing the SDK, running the CLI, and inspecting optimization telemetry.

​
1. Install the SDK

​
Python (PyPI)

Install from PyPI (recommended): 
pip install agnitra
For editable installs while developing locally: 
pip install -e .[openai,rl]
Optional extras:
  • agnitra[openai] – OpenAI Responses API client bindings.
  • agnitra[rl] – Stable Baselines3 + Gymnasium reinforcement learning add-ons.
  • agnitra[nvml] – GPU telemetry via NVIDIA NVML.
  • agnitra[marketplace] – Cloud marketplace adapters (boto3, httpx, google-auth).

​
JavaScript / TypeScript (npm)

Install the JavaScript SDK: 
npm install agnitra
# or
yarn add agnitra
Example usage: 
import { AgnitraClient } from "agnitra";

const client = new AgnitraClient({
  apiKey: process.env.AGNITRA_API_KEY!,
  baseUrl: process.env.AGNITRA_API_BASE_URL ?? "http://127.0.0.1:8080"
});

const result = await client.optimize({
  target: "A100",
  modelGraph,
  telemetry
});

console.log(result.bottleneck.expected_speedup_pct);

​
2. Optimize a Model

agnitra optimize --model tinyllama.pt --input-shape 1,16,64
The CLI loads the model, generates an optimized artifact, and prints a billing snapshot (performance uplift, GPU hours saved, billable total). Pass --output to control the destination path. From Python: 
import torch
from agnitra import optimize

model = torch.jit.load("tinyllama.pt")
sample = torch.randn(1, 16, 64)

result = optimize(model, input_tensor=sample, project_id="demo")
print(result.usage_event.total_billable)

​
3. Explore the API Surface

  • Launch the Starlette service: agnitra-api --host 127.0.0.1 --port 8080
  • POST graph + telemetry payloads to /optimize for automatic kernel suggestions.
  • Forward usage events to /usage to dispatch marketplace billing records.
Refer to the Responses API contract for the JSON schema and webhook semantics.

​
4. Next Steps