Run 35B MoE models in ~2.5 GB of RAM. Weights stream from SSD; a trained router predicts loads a step ahead. Mac, iPhone, Android, Windows supported
Python
"""Layer discovery and installation without tensor calculations."""
from types import SimpleNamespace
import pytest
from edge0.moe.spec import MoESpec
from edge0.streaming import install
def _spec():
return MoESpec(
num_experts=8, top_k=2, intermediate_size=64,
block_path="layers.{layer}.mlp",
)
def test_install_discovers_mixed_layers(monkeypatch):
twin = object()
resident = object()
moe = SimpleNamespace(switch_mlp=resident)
dense = SimpleNamespace()
model = SimpleNamespace(layers=[
SimpleNamespace(mlp=moe), SimpleNamespace(mlp=dense),
])
calls = []
def make_twin(shards, layer, spec, **kwargs):
calls.append((shards, layer, spec))
return twin
monkeypatch.setattr(install, "StreamingSwitchGLU", make_twin)
spec = _spec()
twins = install.install_streaming_experts(model, [], spec)
assert twins == [twin, None]
assert calls == [([], 0, spec)]
assert moe.switch_mlp is twin
assert moe._edge0_resident_switch is resident
assert not hasattr(dense, "switch_mlp")
def test_install_rejects_empty_layer_list():
model = SimpleNamespace(layers=[])
with pytest.raises(ValueError, match="resolves no layer 0"):
install.install_streaming_experts(model, [], _spec())