aquin.compute is the GPU/CPU implementation. The CLI and SDK never talk to torch except through this package (and the engine daemon that wraps it).

Import policy

import aquin.compute is a no-op (__all__ = []). Import the submodule you need. That keeps python -m aquin off the GPU stack.

Device

aquin.compute.device.resolve_compute_device() picks CUDA (incl. ROCm), Metal (MPS), or CPU. vram_guard refuses oversized loads.

Model load

aquin.compute.model_loader is the single resident model. load model and causal_trace.load_model share that instance. Family aliases live in model_families. Hugging Face tokens: local env, or optional Aquin web proxy (hf_auth).

Tool map

compute/server.py is not a runnable server. It is a route index: which HTTP path the desktop would call, and which function implements it. The live HTTP listener is aquin.engine.local_server.

Area	Modules
Trace / logit lens / attribution	causal_trace
SAE train / load / stats / diff	sae, sae_train, sae_stats, sae_diff, user_sae
Features	feature_analysis, find_feature, interp_score
Evals	evals (consistency, suppression, boundary)
Steer	steer, steer_eval, steer_vector
Weight / residue	weight_diff, residual_drift, merge_analysis, weight_rank
Capture / replay	activation_capture, activation_replay, activation_store
Simulate	train_simulate, simulate_inputs
Robustness	red_team, deception_behavior, localize_collapse
Attention / layers	attention_routing, layer_analysis

Pipelines

aquin.compute.pipelines composes those tools for longer jobs (full inspection, SAE catalog metrics, publish).

Artifacts

--check writes JSON + PNG next to the cwd. --json prints the payload. Nothing here requires the Aquin cloud; R2/Supabase helpers exist only for the optional public SAE catalog.