File size: 8,734 Bytes
0b6c0a5 | 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 | """Scientific contracts for learning order, held-out knowledge and restoration."""
import copy
import json
from collections import Counter
from types import SimpleNamespace
import numpy as np
import pytest
import torch
from llm_memory_editability import sequential_transfer as st
from llm_memory_editability.realworld_composition import (
construct,
model_digest,
optimizer_for,
update,
)
def fixture_data():
def row(key, subset=None, atoms=None, role=None):
value = {
"id": key,
"answer": "answer",
"question": "question",
"encoded": {"prefix": [1, 2], "target": [3, 30], "input": [1, 2, 3]},
}
if subset:
value["subset"] = subset
if atoms:
value.update(atom_ids=atoms, sequential_role=role)
return value
atoms = [row("a0", "A"), row("a1", "A"), row("a2", "A"), row("b0", "B"), row("b1", "B")]
chains = [row("train", atoms=["a0", "a1"], role="AA")]
evaluation = [
row("test-aa", atoms=["a1", "a2"], role="AA"),
row("test-ba", atoms=["b0", "a0"], role="BA"),
row("test-ab", atoms=["a0", "b1"], role="AB"),
row("test-bb", atoms=["b0", "b1"], role="BB"),
]
return {
"atoms": atoms,
"train_compositions": chains,
"evaluation_compositions": evaluation,
"panels": {
"atomic_A": ["a0"],
"atomic_B": [],
"train_composition": ["train"],
**{r["sequential_role"]: [r["id"]] for r in evaluation},
},
}
def spec(history="sequential_composition"):
return {
"history": history,
"stage_a_steps": 7,
"stage_b_steps": 5,
"batch_size": 6,
"microbatch_size": 6,
"sampling_seed": 17,
"initialization": 23,
"unique_layers": 2,
"repeats": 1,
"learning_rate": 0.001,
"weight_decay": 0.1,
"warmup_steps": 2,
}
def test_joint_is_exact_example_and_supervision_multiset():
data = fixture_data()
records, sequential = st.training_plan(data, spec())
_, joint = st.training_plan(data, spec("joint"))
assert Counter(sequential.flatten()) == Counter(joint.flatten())
assert not np.array_equal(sequential, joint)
a = st.plan_manifest(records, sequential, spec())
b = st.plan_manifest(records, joint, spec("joint"))
for key in ("multiset_sha256", "supervised_tokens", "input_tokens", "counts", "examples"):
assert a[key] == b[key]
assert a["plan_sha256"] != b["plan_sha256"]
def test_new_facts_and_all_B_combinations_absent_from_A():
records, plan = st.training_plan(fixture_data(), spec())
assert all(records[i].get("subset") != "B" for i in plan[:7].flatten())
assert all(records[i].get("subset") in {"A", "B"} for i in plan[7:].flatten())
assert all(not row["id"].startswith("test") for row in records)
assert all("B" not in row.get("sequential_role", "") for row in records)
def test_atomic_control_has_identical_new_fact_and_replay_stream():
records, original = st.training_plan(fixture_data(), spec())
_, control = st.training_plan(fixture_data(), spec("sequential_atomic"))
np.testing.assert_array_equal(original[7:], control[7:])
assert all(records[i].get("subset") == "A" for i in control[:7].flatten())
np.testing.assert_array_equal(original[:7, :3], control[:7, :3])
def test_sham_has_identical_parent_and_only_A_continuation():
records, original = st.training_plan(fixture_data(), spec())
_, sham = st.training_plan(fixture_data(), spec("sequential_composition_sham"))
np.testing.assert_array_equal(original[:7], sham[:7])
np.testing.assert_array_equal(original[7:, :3], sham[7:, :3])
assert all(records[i].get("subset") == "A" for i in sham[7:].flatten())
def test_evaluation_panel_includes_every_necessary_atom():
groups = st.evaluation_groups(fixture_data(), full=False)
available = {row["id"] for name in ("atomic_A", "atomic_B") for row in groups[name]}
needed = {
key for role in ("AA", "BA", "AB", "BB") for row in groups[role] for key in row["atom_ids"]
}
assert needed <= available
assert {row["id"] for row in groups["BB"]} == {"test-bb"}
@pytest.mark.parametrize("leak", ("new_fact", "heldout_chain", "wrong_role"))
def test_data_contract_rejects_leakage(leak):
data = fixture_data()
if leak == "new_fact":
data["train_compositions"][0]["atom_ids"] = ["a0", "b0"]
elif leak == "heldout_chain":
data["evaluation_compositions"][0]["atom_ids"] = ["a0", "a1"]
else:
data["evaluation_compositions"][0]["sequential_role"] = "BB"
with pytest.raises(ValueError):
st.validate_data(data)
def test_learning_rate_has_same_global_history_and_explicit_stage_reset():
for step in range(1, 13):
assert len({st.learning_rate(spec(history), step) for history in st.HISTORIES}) == 1
assert st.learning_rate(spec(), 7) == pytest.approx(0.0001)
assert st.learning_rate(spec(), 8) == pytest.approx(0.0005)
assert st.learning_rate(spec(), 12) == pytest.approx(0.0001)
def test_resume_restores_optimizer_rng_and_next_update(tmp_path):
torch.set_num_threads(1)
model_config = dict(vocab_size=31, positions=32, hidden_size=8, attention_heads=2, dropout=0.1)
run_spec = spec()
records, plan = st.training_plan(fixture_data(), run_spec)
manifest = st.plan_manifest(records, plan, run_spec)
model = construct(model_config, run_spec, "cpu")
optimizer = optimizer_for(model, 0.001, 0.1)
update(model, optimizer, [records[i] for i in plan[0]], run_spec, "cpu", 30)
counts = np.zeros(len(records), dtype=np.int64)
np.add.at(counts, plan[0], 1)
st.save_checkpoint(
tmp_path / "resume.pt", model, optimizer, 1, counts, {"examples": 6}, [], manifest
)
update(model, optimizer, [records[i] for i in plan[1]], run_spec, "cpu", 30)
expected = model_digest(model)
restored = construct(model_config, run_spec, "cpu")
restored_optimizer = optimizer_for(restored, 0.001, 0.1)
state = st.restore_checkpoint(tmp_path / "resume.pt", restored, restored_optimizer, manifest)
assert state["step"] == 1 and state["counts"].sum() == 6
update(restored, restored_optimizer, [records[i] for i in plan[1]], run_spec, "cpu", 30)
assert model_digest(restored) == expected
changed = copy.deepcopy(manifest)
changed["plan_sha256"] = "wrong"
with pytest.raises(ValueError, match="sampling plan"):
st.restore_checkpoint(tmp_path / "resume.pt", restored, restored_optimizer, changed)
def test_training_worker_resumes_without_repeating_or_skipping_updates(tmp_path, monkeypatch):
torch.set_num_threads(1)
run_spec = {
**spec(),
"stage_a_steps": 2,
"stage_b_steps": 2,
"model": dict(vocab_size=31, positions=32, hidden_size=8, attention_heads=2, dropout=0.1),
"data_sha256": "unit-test",
"evaluation_nodes": [0, 2, 4],
}
data = fixture_data()
def setup(value, device):
return (
data,
SimpleNamespace(eos_token_id=30),
construct(value["model"], value, device),
torch.device(device),
)
monkeypatch.setattr(st, "_setup", setup)
monkeypatch.setattr(
st,
"evaluate_transfer",
lambda model, *args, **kwargs: (
{"fixture": {"n": 1, "digest": model_digest(model)}},
{},
),
)
uninterrupted = st.run(run_spec, tmp_path / "uninterrupted", "cpu")
original_update = st.update
calls = 0
def interrupt(*args, **kwargs):
nonlocal calls
calls += 1
if calls == 3:
raise RuntimeError("simulated interruption after saved stage A")
return original_update(*args, **kwargs)
monkeypatch.setattr(st, "update", interrupt)
with pytest.raises(RuntimeError, match="simulated interruption"):
st.run(run_spec, tmp_path / "resumed", "cpu")
monkeypatch.setattr(st, "update", original_update)
resumed = st.run(run_spec, tmp_path / "resumed", "cpu")
assert resumed["model_sha256"] == uninterrupted["model_sha256"]
assert resumed["examples"] == uninterrupted["examples"] == 24
left = torch.load(tmp_path / "uninterrupted/latest.pt", weights_only=False)
right = torch.load(tmp_path / "resumed/latest.pt", weights_only=False)
np.testing.assert_array_equal(left["counts"], right["counts"])
history = json.loads((tmp_path / "resumed/learning.json").read_text())
assert [row["step"] for row in history] == [0, 2, 4]
assert history[1]["atomic_B_epochs"] == 0
assert history[-1]["atomic_B_epochs"] > 0
|