File size: 9,319 Bytes
4126208 5e2b05f | 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 | # SPDX-License-Identifier: Apache-2.0
"""CPU tests for MiniEmbed — lookup/encode, UnifiedReceiptChain, honest LFS selfcheck."""
from __future__ import annotations
import hashlib
import json
import sys
from pathlib import Path
import numpy as np
import pytest
ROOT = Path(__file__).resolve().parents[1]
sys.path.insert(0, str(ROOT / "build" / "torch-universal"))
import szl_kernels as sk # noqa: E402
from szl_kernels.miniembed import MiniEmbed, PUBLISHED_SHA256, _is_lfs_pointer # noqa: E402
SUITE_SELFCHECK_KEYS = (
"norm_correct",
"lambda_advisory",
"energy_honest",
"cross_kernel_verify",
"spans_three_kernels",
"offline_reverify",
"tamper_detected",
"block_forward",
)
def _write_tiny_table(tmp_path: Path, *, lfs_vectors: bool = False) -> Path:
vocab = ["receipt", "chain", "lambda"]
index = {term: i for i, term in enumerate(vocab)}
rng = np.random.RandomState(0)
table = rng.randn(3, 8).astype(np.float32)
table /= np.linalg.norm(table, axis=1, keepdims=True)
if lfs_vectors:
(tmp_path / "vectors.npz").write_bytes(
b"version https://git-lfs.github.com/spec/v1\n"
b"oid sha256:" + b"ab" * 32 + b"\n"
b"size 99\n"
)
else:
np.savez_compressed(tmp_path / "vectors.npz", vectors=table)
(tmp_path / "vocab.json").write_text(
json.dumps({"vocab": vocab, "index": index}),
encoding="utf-8",
)
(tmp_path / "config.json").write_text(
json.dumps(
{
"model": "SZL-MiniEmbed",
"method": "PPMI+TruncatedSVD",
"dim": 8,
"vocab_size": 3,
"files": {
"vectors": "vectors.npz (key 'vectors', float32 [V,dim])",
"vocab": "vocab.json ({'vocab':[term...], 'index':{term:i}})",
},
}
),
encoding="utf-8",
)
return tmp_path
def test_miniembed_exported_from_package() -> None:
assert sk.MiniEmbed is MiniEmbed
assert "MiniEmbed" in sk.__all__
def test_suite_selfcheck_still_eight_of_eight() -> None:
result = sk.selfcheck()
assert result["ok"] is True, result
assert set(result["checks"]) == set(SUITE_SELFCHECK_KEYS)
assert all(result["checks"].values()), result["checks"]
assert result["kernels_touched"] == ["governed_norm", "lambda_gate", "energy_core"]
def test_torch_ext_mirrors_build_package() -> None:
build = ROOT / "build" / "torch-universal" / "szl_kernels"
ext = ROOT / "torch-ext" / "szl_kernels"
names = sorted(path.name for path in build.iterdir() if path.is_file())
assert "miniembed.py" in names
for name in names:
assert (build / name).read_bytes() == (ext / name).read_bytes(), name
def test_lookup_encode_match_npz_vocab_contract() -> None:
embed = MiniEmbed(root=ROOT)
if not embed.available:
pytest.skip(f"MiniEmbed table unavailable: {embed.unavailable_label}")
table = np.load(ROOT / "vectors.npz", allow_pickle=False)["vectors"]
vocab = json.loads((ROOT / "vocab.json").read_text(encoding="utf-8"))
term = "receipt"
row = embed.lookup(term)
assert row.dtype == np.float32
assert row.shape == (128,)
np.testing.assert_allclose(row, table[vocab["index"][term]], rtol=0, atol=0)
matrix = embed.encode(["receipt", "chain"])
assert matrix.shape == (2, 128)
np.testing.assert_allclose(matrix[0], row, rtol=0, atol=0)
assert embed.encode("lambda").shape == (128,)
def test_oov_raises_keyerror() -> None:
embed = MiniEmbed(root=ROOT)
if not embed.available:
pytest.skip(f"MiniEmbed table unavailable: {embed.unavailable_label}")
with pytest.raises(KeyError):
embed.lookup("this-term-is-not-in-the-szl-miniembed-vocab")
with pytest.raises(KeyError):
embed.encode(["receipt", "this-term-is-not-in-the-szl-miniembed-vocab"])
assert embed.neighbors("this-term-is-not-in-the-szl-miniembed-vocab") is None
def test_neighbors_match_training_receipt_when_loaded() -> None:
embed = MiniEmbed(root=ROOT)
if not embed.available:
pytest.skip(f"MiniEmbed table unavailable: {embed.unavailable_label}")
receipt = json.loads((ROOT / "TRAINING_RECEIPT.json").read_text(encoding="utf-8"))
listed = receipt["metrics_MEASURED"]["intrinsic_nearest_neighbors"]
for term, expected in listed.items():
got = embed.neighbors(term, k=6)
assert got is not None, term
assert [word for word, _ in got] == [word for word, _ in expected], term
def test_selfcheck_matches_published_hashes_when_table_is_real() -> None:
embed = MiniEmbed(root=ROOT)
result = embed.selfcheck()
if _is_lfs_pointer(ROOT / "vectors.npz"):
assert result["ok"] is False
assert result["label"] == "UNAVAILABLE_LFS"
assert result["checks"]["honest_lfs"] is True
assert result["checks"]["vectors.npz_sha256_match"] is not True
return
assert embed.available is True
assert result["ok"] is True, result
assert result["file_sha256"]["vectors.npz"] == PUBLISHED_SHA256["vectors.npz"]
assert result["file_sha256"]["vocab.json"] == PUBLISHED_SHA256["vocab.json"]
receipt = json.loads((ROOT / "TRAINING_RECEIPT.json").read_text(encoding="utf-8"))
assert result["file_sha256"]["vectors.npz"] == receipt["model"]["files"]["vectors.npz"]
assert result["file_sha256"]["vocab.json"] == receipt["model"]["files"]["vocab.json"]
def test_selfcheck_lfs_pointer_is_unavailable_not_a_pass(tmp_path: Path) -> None:
root = _write_tiny_table(tmp_path, lfs_vectors=True)
embed = MiniEmbed(root=root)
assert embed.available is False
assert embed.unavailable_label == "UNAVAILABLE_LFS"
with pytest.raises(RuntimeError, match="UNAVAILABLE_LFS"):
embed.lookup("receipt")
result = embed.selfcheck()
assert result["ok"] is False
assert result["label"] == "UNAVAILABLE_LFS"
assert result["checks"]["honest_lfs"] is True
assert result["checks"]["vectors.npz_sha256_match"] is not True
# Pointer bytes must not be reported as the published object hash.
pointer_hash = hashlib.sha256((root / "vectors.npz").read_bytes()).hexdigest()
assert pointer_hash != PUBLISHED_SHA256["vectors.npz"]
assert result["file_sha256"]["vectors.npz"] is None
def test_table_and_lookup_receipts_use_unified_chain(tmp_path: Path) -> None:
root = _write_tiny_table(tmp_path, lfs_vectors=False)
chain = sk.UnifiedReceiptChain()
embed = MiniEmbed(root=root, chain=chain)
table_rec = embed.receipt_table()
row = embed.lookup("receipt")
assert row.shape == (8,)
ok, depth, brk = chain.verify()
assert ok is True and brk == -1 and depth == 2
assert chain.kernels_touched() == ["miniembed"]
assert table_rec["kernel"] == "miniembed" and table_rec["op"] == "table"
lookup_rec = chain.tail(1)[0]
assert lookup_rec["op"] == "lookup"
assert lookup_rec["attrs"]["not_a_transformer_lm"] is True
assert "Conjecture 1" in lookup_rec["attrs"]["lambda_status"]
blob = chain.to_json()
ok2, _, _ = sk.UnifiedReceiptChain.verify_json(blob)
assert ok2 is True
mutated = json.loads(blob)
mutated[0]["attrs"]["vocab_size"] = 0
ok3, _, brk3 = sk.UnifiedReceiptChain.verify_json(json.dumps(mutated))
assert ok3 is False and brk3 == 0
def test_fixture_encode_does_not_require_gpu(tmp_path: Path) -> None:
embed = MiniEmbed(root=_write_tiny_table(tmp_path))
matrix = embed.encode(["chain", "lambda"])
assert matrix.dtype == np.float32
assert matrix.shape == (2, 8)
assert "receipt" in embed
assert len(embed) == 3
def test_source_states_not_transformer_chaski_khipu() -> None:
source = (ROOT / "build" / "torch-universal" / "szl_kernels" / "miniembed.py").read_text(
encoding="utf-8"
)
collapsed = " ".join(source.split())
assert "NOT a transformer language model" in collapsed
assert "NOT Chaski" in collapsed
assert "NOT Khipu" in collapsed
lowered = collapsed.lower()
assert "tokens/s" not in lowered
assert "tokens per second" not in lowered
assert "tokens/sec" not in lowered
def test_metadata_name_is_legal_dashed_kernel_id() -> None:
meta = json.loads(
(ROOT / "build" / "torch-universal" / "szl_kernels" / "metadata.json").read_text(
encoding="utf-8"
)
)
assert meta["name"] == "szl-kernels"
assert meta["miniembed"]["kind"] == "distributional_word_embedding_table"
def test_model_card_does_not_advertise_unsupported_feature_extraction() -> None:
readme = (ROOT / "README.md").read_text(encoding="utf-8")
lines = readme.splitlines()
assert lines[0] == "---"
front = lines[1 : lines.index("---", 1)]
tag_start = front.index("tags:") + 1
tags = set()
for line in front[tag_start:]:
if not line.startswith("- "):
break
tags.add(line[2:])
assert {"embedding-table", "not-a-transformers-config"} <= tags
assert not tags.intersection(
{"feature-extraction", "embeddings", "word-embeddings", "sklearn"}
)
config = json.loads((ROOT / "config.json").read_text(encoding="utf-8"))
assert "model_type" not in config and "auto_map" not in config
|