API Reference REST API

REST API

Overview

The REST API provides synchronous access to the PyTorch backend, handling static introspection and full single-pass analysis.

Endpoints

GET /health

Returns liveness and whether the model weights are currently loaded into memory.

{ "status": "ok", "model_loaded": true }

GET /architecture

Extracts the model's structure via named_parameters() without running a forward pass. Returns: Architecture metadata (layers, hidden size, vocab) and a flat list of all tensors (name, shape, dtype, param count).

POST /analyze

Runs a single forward pass on a provided sentence. Used primarily for Walkthrough mode. Body: { "sentence": "The cat sat" } Returns:

• tokens: The tokenized IDs.
• attention: The full 4D attention matrix [layer][head][from][to] (sub-0.01 values are zeroed to save bandwidth).
• hidden_states_3d: The hidden states at each layer, projected to 3D via PCA.

GET /debug/ops

Returns the static op_catalog for the loaded model, mapping operations to physical layer indices. Useful for setting programmatic breakpoints.

POST /ablate/analyze

Intervention endpoint. Zeroes out specified heads or layers using PyTorch forward hooks, runs a forward pass, and returns the modified analyze payload to allow before/after diffing of the Logit Lens.

Diagram

sequenceDiagram
    participant Client
    participant FastAPI
    participant PyTorch
    
    Client->>FastAPI: POST /analyze {sentence: "Hello"}
    FastAPI->>PyTorch: model(input_ids)
    PyTorch-->>FastAPI: logits, attentions, hidden_states
    FastAPI->>FastAPI: Apply PCA to hidden_states
    FastAPI-->>Client: JSON Payload

Loading

Related pages

• WebSocket Events
• Data Models

Further reading

• API Documentation

Navigation

Previous	Home	Next
API Reference	Home	WebSocket Events