File size: 11,792 Bytes
a3773e2 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 | """Scientific contracts for frozen-memory SSFR replacement, on tiny CPU models."""
import copy
from pathlib import Path
import numpy as np
import pytest
import torch
from llm_memory_editability import hebbian_interface as interface
ROOT = Path(__file__).resolve().parents[1]
pytestmark = pytest.mark.skipif(
not (ROOT / interface.DEFAULT_SPEC["source_path"] / "src/hebbian").is_dir(),
reason="The pinned official hebbian-mlps source checkout is required",
)
@pytest.fixture(scope="module")
def spec():
return {
**interface.DEFAULT_SPEC,
"num_facts": 8,
"d_model": 8,
"hidden_dim": 16,
"junk_len": 2,
"junk_vocab_size": 3,
"batch_size": 8,
"mlp_epochs": 3,
"reader_steps": 4,
"eval_every": 2,
"eval_repeats": 2,
}
def components(spec, seed=53):
config = interface.make_config(spec, seed, "cpu")
from hebbian.transformer.fact_store import build_factset, build_token_embeddings
from hebbian.transformer.model import GPT
from hebbian.transformer.utils import (
copy_embeddings_to_gpt,
create_gpt_config,
insert_mlp_into_gpt,
)
factset = build_factset(config, seed=seed)
embeddings = build_token_embeddings(
factset, spec["junk_vocab_size"], embedding_init="spherical", seed=seed
)
torch.manual_seed(seed)
model = GPT(create_gpt_config(config.train_config, config.dataset_config)).eval()
copy_embeddings_to_gpt(model, embeddings)
memory = interface._new_memory(spec, "cpu")
insert_mlp_into_gpt(model, memory, embeddings)
return model, factset, embeddings
@pytest.fixture(scope="module")
def tiny_run(spec, tmp_path_factory):
output = tmp_path_factory.mktemp("hebbian-interface")
summary = interface.run_world(spec, 31, output, "cpu")
return summary, output
def test_architecture_and_only_qk_trainable(spec):
model, _, _ = components(spec)
block = model.transformer.h[0]
assert len(model.transformer.h) == 1
assert model.transformer.wpe is None
assert not model.config.use_rope
assert not block.attn_residual and not block.mlp_residual
for projection in (block.attn.c_v, block.attn.c_proj):
torch.testing.assert_close(projection.weight, torch.eye(spec["d_model"]))
assert not projection.weight.requires_grad
trainable = {name for name, value in model.named_parameters() if value.requires_grad}
assert trainable == {
"transformer.h.0.attn.c_q.weight",
"transformer.h.0.attn.c_k.weight",
}
x = torch.randn(5, spec["d_model"])
for normalization in (block.ln_2, model.transformer.ln_f):
torch.testing.assert_close(normalization(x).norm(dim=-1), torch.ones(5))
def test_query_inputs_have_no_answer_and_do_not_depend_on_mapping(spec):
interface.make_config(spec, 53, "cpu")
from hebbian.data.synthetics.factsets import BijectiveMapping
mapping_a = BijectiveMapping.from_permutation(list(range(spec["num_facts"])))
mapping_b = BijectiveMapping.from_permutation(list(reversed(mapping_a.outputs)))
torch.manual_seed(713)
rng = torch.random.get_rng_state().clone()
inputs, keys = interface.fixed_inputs(spec, mapping_a, 1234)
assert torch.equal(rng, torch.random.get_rng_state())
other, other_keys = interface.fixed_inputs(spec, mapping_b, 1234)
assert torch.equal(inputs, other) and torch.equal(keys, other_keys)
assert inputs.shape == (spec["num_facts"] * spec["eval_repeats"], 2 * spec["junk_len"] + 2)
assert torch.equal(inputs[:, spec["junk_len"]], keys)
query_id = spec["num_facts"] + spec["junk_vocab_size"]
assert inputs[:, -1].eq(query_id).all()
assert torch.equal(keys.bincount(), torch.full((spec["num_facts"],), spec["eval_repeats"]))
context = torch.cat((inputs[:, : spec["junk_len"]], inputs[:, spec["junk_len"] + 1 : -1]), 1)
assert context.ge(spec["num_facts"]).all() and context.lt(query_id).all()
batch, targets = next(iter(interface.make_batches(spec, mapping_b, num_batches=1)))
original_keys = batch[:, spec["junk_len"]]
expected = torch.tensor(mapping_b.outputs)[original_keys]
assert torch.equal(targets[:, -2], expected)
assert torch.equal(batch[:, -1], expected)
assert targets[:, :-2].eq(-100).all() and targets[:, -1].eq(-100).all()
assert batch[:, :-1][:, -1].eq(query_id).all()
def test_future_answer_cannot_change_query_logits(spec):
model, factset, _ = components(spec)
inputs, _ = interface.fixed_inputs(spec, factset.mapping, 180)
full = torch.cat((inputs, torch.zeros(len(inputs), 1, dtype=torch.long)), 1)
altered = full.clone()
altered[:, -1] = spec["num_facts"] - 1
targets = torch.zeros_like(full)
with torch.no_grad():
original_logits, _ = model(full, targets)
altered_logits, _ = model(altered, targets)
removed_logits, _ = model(inputs)
torch.testing.assert_close(original_logits[:, :-1], altered_logits[:, :-1], rtol=0, atol=0)
torch.testing.assert_close(original_logits[:, -2], removed_logits[:, 0], rtol=1e-5, atol=1e-6)
def test_standalone_distinguishes_fact_candidates_from_full_head(spec, tmp_path):
interface.make_config(spec, 1, "cpu")
from hebbian.data.synthetics.factsets import BijectiveMapping, Factset
values = torch.tensor([[1.0, 0.0], [-1.0, 0.0]])
factset = Factset(values, values, BijectiveMapping.from_permutation([0, 1]), 2, 2)
embeddings = torch.nn.Embedding.from_pretrained(torch.cat((values, torch.tensor([[0.0, 1.0]]))))
memory = torch.nn.Linear(2, 2)
with torch.no_grad():
memory.weight.copy_(torch.tensor([[0.1, 0.0], [0.0, 0.0]]))
memory.bias.copy_(torch.tensor([0.0, 2.0]))
report = interface.standalone(
memory, factset, embeddings, torch.ones(2, dtype=torch.bool), tmp_path / "scores.npz"
)
assert report["fact_accuracy"] == 1.0
assert report["full_accuracy"] == 0.0
saved = np.load(tmp_path / "scores.npz")
assert (saved["pred"] == 2).all()
def test_clean_key_readout_matches_actual_normalization_and_head(spec, tmp_path):
model, factset, embeddings = components(spec)
inputs, keys = interface.fixed_inputs(spec, factset.mapping, 612)
targets = torch.tensor(factset.mapping.outputs)[keys]
changed = torch.ones(spec["num_facts"], dtype=torch.bool)
interface.standalone(
model.transformer.h[0].mlp, factset, embeddings, changed, tmp_path / "standalone.npz"
)
stored_keys = factset.input_embeddings[keys]
def exact_keys(module, args):
replaced = args[0].clone()
replaced[:, -1] = stored_keys
return (replaced,)
hook = model.transformer.h[0].mlp.register_forward_pre_hook(exact_keys)
try:
with torch.no_grad():
logits, _ = model(inputs)
finally:
hook.remove()
expected = torch.from_numpy(np.load(tmp_path / "standalone.npz")["scores"])[keys]
torch.testing.assert_close(logits[:, 0], expected, rtol=1e-5, atol=1e-6)
assert torch.equal(logits[:, 0].argmax(-1).eq(targets), expected.argmax(-1).eq(targets))
def test_evaluation_preserves_rng_weights_mode_and_hooks(spec):
model, factset, embeddings = components(spec)
inputs, keys = interface.fixed_inputs(spec, factset.mapping, 161)
model.train()
state_hash = interface.tensor_hash(model.state_dict())
torch.manual_seed(818)
rng = torch.random.get_rng_state().clone()
interface.evaluate_reader(
model,
inputs,
keys,
torch.tensor(factset.mapping.outputs)[keys],
torch.ones(len(keys), dtype=torch.bool),
embeddings,
spec["batch_size"],
)
assert model.training
assert not model.transformer.h[0].mlp._forward_pre_hooks
assert state_hash == interface.tensor_hash(model.state_dict())
assert torch.equal(rng, torch.random.get_rng_state())
def test_complete_cross_reload_scores_and_restore(spec, tiny_run):
summary, output = tiny_run
from hebbian.transformer.utils import insert_mlp_into_gpt
world = np.load(output / "world.npz")
inputs = torch.from_numpy(world["inputs"])
keys = torch.from_numpy(world["key"])
map_a, map_b = torch.from_numpy(world["mapping_A"]), torch.from_numpy(world["mapping_B"])
changed = (map_a != map_b)[keys]
assert not summary["audits"]["B_used_for_reader_optimization_or_selection"]
assert all(
value
for key, value in summary["audits"].items()
if key != "B_used_for_reader_optimization_or_selection"
)
for objective in ("ce", "mse"):
model, _ = interface.load_checkpoint(output / f"reader-{objective}.pt")
initial_logits, _ = model(inputs)
state_before = interface.tensor_hash(model.state_dict())
original_memory = model.transformer.h[0].mlp
non_mlp_before = interface.tensor_hash(interface._state(model, non_mlp=True))
records = summary["readers"][objective]
assert set(records["endpoints"]) == {"A", "wrong_A_on_B", "B_ce", "B_mse"}
assert records["frozen_hash_before"] == records["frozen_hash_after"]
a_saved = np.load(output / f"reader-{objective}-A.npz")
wrong = np.load(output / f"reader-{objective}-wrong_A_on_B.npz")
np.testing.assert_array_equal(a_saved["scores"], wrong["scores"])
np.testing.assert_array_equal(wrong["target"], map_b[keys])
for replacement in ("ce", "mse"):
memory, _ = interface.load_checkpoint(output / f"memory-{replacement}-B.pt")
assert not any(parameter.requires_grad for parameter in memory.parameters())
insert_mlp_into_gpt(model.eval(), memory, model.transformer.wte)
with torch.no_grad():
logits, _ = model(inputs)
saved = np.load(output / f"reader-{objective}-B_{replacement}.npz")
torch.testing.assert_close(
logits[:, 0], torch.from_numpy(saved["scores"]), rtol=1e-5, atol=1e-6
)
assert records["endpoints"][f"B_{replacement}"]["n_changed"] == int(changed.sum())
assert non_mlp_before == interface.tensor_hash(interface._state(model, non_mlp=True))
insert_mlp_into_gpt(model.eval(), original_memory, model.transformer.wte)
restored_logits, _ = model(inputs)
assert state_before == interface.tensor_hash(model.state_dict())
torch.testing.assert_close(initial_logits, restored_logits, rtol=0, atol=0)
def test_B_mapping_does_not_affect_A_reader_training(spec, tiny_run, tmp_path, monkeypatch):
original, _ = tiny_run
from hebbian.data.synthetics.factsets import BijectiveMapping
from hebbian.transformer import train
original_factory = train._make_eval_factset
def alternate_B(factset, seed):
replacement = copy.copy(original_factory(factset, seed))
replacement.mapping = BijectiveMapping.from_permutation(
[(value + 1) % spec["num_facts"] for value in replacement.mapping.outputs]
)
return replacement
monkeypatch.setattr(train, "_make_eval_factset", alternate_B)
alternate = interface.run_world(spec, 31, tmp_path, "cpu")
assert original["world"]["mapping_B"] != alternate["world"]["mapping_B"]
for objective in ("ce", "mse"):
for name in ("initial_non_mlp_hash", "final_non_mlp_hash", "training_stream_hash"):
assert original["readers"][objective][name] == alternate["readers"][objective][name]
assert (
original["memories"][objective]["A"]["final_hash"]
== alternate["memories"][objective]["A"]["final_hash"]
)
assert (
original["memories"][objective]["B"]["final_hash"]
!= alternate["memories"][objective]["B"]["final_hash"]
)
|