Skip to content

Backends

Module: circuitkit.backends

The backends module provides the stability tier registry, algorithm enumeration, and the tier-query helpers. It is the single source of truth for which algorithms are production-ready.

Stability Tiers

CircuitKit ships 13 discovery algorithms across 4 backends, each with an explicit stability tier. Only 2 are validated at production scale (eap, eap-ig); acdc and ibcircuit are experimental (GPT-2 scale; ibcircuit OOMs above ~3B) and the other 9 are research (GPT-2 IOI only).

Tier Algorithms
Stable eap, eap-ig
Experimental acdc, ibcircuit
Research eap-ig-activations, eap-clean-corrupted, eap-exact, atp-gd, eap-gp, relp, peap, eap-ifr, cdt

Exports

from circuitkit.backends import (
    STABILITY,
    DISCOVERY_ALGORITHMS,
    STABLE_ALGORITHMS,
    EXPERIMENTAL_ALGORITHMS,
    RESEARCH_ALGORITHMS,
    is_stable,
    is_experimental,
    is_research,
    default_algorithm,
)

STABILITY

Dict[str, str] — maps algorithm/selector name → tier string ("stable", "experimental", "research").

Note

STABILITY includes both the 13 discovery algorithms AND the 8 compression selector keys (random, magnitude, taylor, wanda, multi_granular, gptq, awq, tacq). Use DISCOVERY_ALGORITHMS when you want only the discovery algorithms.

from circuitkit.backends import STABILITY

print(STABILITY["eap-ig"])   # "stable"
print(STABILITY["acdc"])     # "experimental"
print(STABILITY["cdt"])      # "research"

DISCOVERY_ALGORITHMS

frozenset[str] — exactly the 13 discovery algorithm names.

from circuitkit.backends import DISCOVERY_ALGORITHMS

print(sorted(DISCOVERY_ALGORITHMS))
# ['acdc', 'atp-gd', 'cdt', 'eap', 'eap-clean-corrupted', 'eap-exact',
#  'eap-gp', 'eap-ifr', 'eap-ig', 'eap-ig-activations', 'ibcircuit', 'peap', 'relp']

STABLE_ALGORITHMS, EXPERIMENTAL_ALGORITHMS, RESEARCH_ALGORITHMS

Tier subsets derived from the full ALGORITHMS/STABILITY map (not just DISCOVERY_ALGORITHMS), so STABLE_ALGORITHMS also includes the stable-tier compression selectors.

from circuitkit.backends import STABLE_ALGORITHMS, EXPERIMENTAL_ALGORITHMS

print(sorted(STABLE_ALGORITHMS))
# ['awq', 'eap', 'eap-ig', 'gptq', 'magnitude',
#  'multi_granular', 'random', 'tacq', 'taylor', 'wanda']

Functions

is_stable(algo: str) -> bool

from circuitkit.backends import is_stable

is_stable("eap-ig")   # True
is_stable("acdc")     # False

is_experimental(algo: str) -> bool

from circuitkit.backends import is_experimental
is_experimental("acdc")     # True
is_experimental("eap-ig")   # False

is_research(algo: str) -> bool

from circuitkit.backends import is_research
is_research("cdt")      # True
is_research("eap-ig")   # False

default_algorithm() -> str

Returns the default discovery algorithm ("eap-ig").

from circuitkit.backends import default_algorithm
default_algorithm()   # "eap-ig"

Tier Warnings

discover_circuit automatically emits a UserWarning when you request a non-stable algorithm:

UserWarning: Algorithm 'acdc' is experimental. May fail on larger models or non-IOI tasks. Use 'eap-ig' for production.

Research-tier algorithms get a similar warning:

UserWarning: Algorithm 'cdt' is research-quality (only validated on GPT-2 IOI). Use 'eap-ig' for production.

To suppress:

import warnings
warnings.filterwarnings("ignore", category=UserWarning, module="circuitkit")

Backend Modules

Each backend is also directly importable if you need lower-level access:

Backend Module
EAP family circuitkit.backends.eap
ACDC circuitkit.backends.acdc
IBCircuit circuitkit.backends.ibcircuit
CD-T circuitkit.backends.cdt

For normal usage, call them through discover_circuit rather than directly.

Next Steps