Model attachment

ARTI can discover target layers and attach Recall branches to an existing PyTorch model.

Attach to selected layers

import arti

model = arti.ARTI.attach(
    model,
    recall={
        "layers": "model.layers.*",
        "rank": 16,
        "slots": 8,
    },
)

Attachment configuration supports explicit paths, glob patterns, per-layer dimensions, independent Recall lines, Half switches, and training settings.

Inspect resource cost

print(model.arti.summary())

Use the summary and resource preview before training or export to bound added parameters.

Save compact state

model.arti.save("assistant.recall.arti.st")

The base model and ARTI state can remain separate. Removal is reversible:

model.arti.remove()

Optional integration boundaries

• Core: PyTorch-first
• Optional: Transformers and Qwen
• Optional: PEFT
• Optional: Diffusers and Stable Diffusion

The core package does not require these frameworks.