File size: 12,511 Bytes
9a46c96 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 | import math
import json
import os
from pathlib import Path
import sys
from types import SimpleNamespace
import pytest
import yaml
from evaluation.generative_perplexity import (
aggregate_nll,
compute_gen_ppl,
GPT2Assets,
load_offline_gpt2_large,
resolve_gpt2_assets,
TorchCausalLMScorer,
)
class TinyTokenizer:
pad_token_id = 0
def __init__(self, rows: dict[str, list[int]]) -> None:
self.rows = rows
def __call__(self, texts, **kwargs):
encoded = [self.rows[text][: kwargs["max_length"]] for text in texts]
width = max(map(len, encoded))
return {
"input_ids": [row + [0] * (width - len(row)) for row in encoded],
"attention_mask": [[1] * len(row) + [0] * (width - len(row)) for row in encoded],
}
class FakeLogitScorer:
"""Reduce fixed logits without importing or running a model framework."""
def __init__(self, logits_by_target: dict[int, list[float]]) -> None:
self.logits_by_target = logits_by_target
self.eval_called = False
def eval(self):
self.eval_called = True
return self
def score_batch(self, input_ids, attention_mask):
total = 0.0
count = 0
for row, mask in zip(input_ids, attention_mask, strict=True):
for target, valid in zip(row[1:], mask[1:], strict=True):
if valid:
logits = self.logits_by_target[target]
maximum = max(logits)
log_partition = maximum + math.log(
sum(math.exp(value - maximum) for value in logits)
)
total += log_partition - logits[target]
count += 1
return total, count
def test_ppl_uses_token_weighted_nll() -> None:
"""Catch averaging per-sample perplexities instead of all valid tokens."""
result = aggregate_nll([(math.log(2), 1), (3 * math.log(4), 3)])
assert result.perplexity == pytest.approx(
math.exp((math.log(2) + 3 * math.log(4)) / 4)
)
assert result.total_nll == pytest.approx(math.log(2) + 3 * math.log(4))
assert result.valid_token_count == 4
def test_compute_gen_ppl_is_batch_invariant_and_scores_final_short_batch() -> None:
"""Catch batch averaging or floor division silently dropping the last sample."""
texts = ["a", "b", "c"]
tokenizer = TinyTokenizer({"a": [9, 1], "b": [9, 2, 2], "c": [9, 3]})
logits = {1: [0.0] * 2, 2: [0.0] * 4, 3: [0.0] * 8}
one = compute_gen_ppl(texts, FakeLogitScorer(logits), tokenizer, batch_size=1)
two = compute_gen_ppl(texts, FakeLogitScorer(logits), tokenizer, batch_size=2)
expected_nll = math.log(2) + 2 * math.log(4) + math.log(8)
assert one.perplexity == pytest.approx(math.exp(expected_nll / 4))
assert two.perplexity == pytest.approx(one.perplexity)
assert two.valid_token_count == 4
assert two.sample_count == 3
def test_compute_gen_ppl_rejects_corpus_without_next_token_targets() -> None:
"""Catch publishing an undefined perplexity for one-token/empty targets."""
tokenizer = TinyTokenizer({"one": [9]})
with pytest.raises(ValueError, match="valid next-token"):
compute_gen_ppl(["one"], FakeLogitScorer({}), tokenizer)
def test_compute_gen_ppl_rejects_any_sample_without_a_next_token_target() -> None:
"""Catch one short row disappearing inside an otherwise valid corpus aggregate."""
tokenizer = TinyTokenizer({"short": [9], "valid": [9, 1]})
with pytest.raises(ValueError, match="sample 0.*no valid next-token"):
compute_gen_ppl(
["short", "valid"], FakeLogitScorer({1: [0.0, 0.0]}), tokenizer
)
def test_compute_gen_ppl_rejects_non_right_padded_attention_mask() -> None:
"""Catch an interior padding hole shifting a target into the score."""
class HoleTokenizer(TinyTokenizer):
def __call__(self, texts, **kwargs):
return {"input_ids": [[9, 0, 1]], "attention_mask": [[1, 0, 1]]}
with pytest.raises(ValueError, match="right-padded"):
compute_gen_ppl(
["row"], FakeLogitScorer({1: [0.0, 0.0]}), HoleTokenizer({})
)
def test_compute_gen_ppl_rejects_scorer_token_count_drift() -> None:
"""Catch padding or first-position leakage hidden behind a plausible total NLL."""
class WrongCountScorer(FakeLogitScorer):
def score_batch(self, input_ids, attention_mask):
nll, count = super().score_batch(input_ids, attention_mask)
return nll, count + 1
tokenizer = TinyTokenizer({"row": [9, 1]})
with pytest.raises(ValueError, match="attention mask.*expected 1.*reported 2"):
compute_gen_ppl(["row"], WrongCountScorer({1: [0.0, 0.0]}), tokenizer)
def test_compute_gen_ppl_rejects_context_beyond_gpt2_limit() -> None:
"""Catch silently requesting positions beyond GPT-2 Large's fixed context."""
with pytest.raises(ValueError, match="must not exceed 1024"):
compute_gen_ppl(
["row"],
FakeLogitScorer({1: [0.0, 0.0]}),
TinyTokenizer({"row": [9, 1]}),
max_length=1025,
)
@pytest.mark.parametrize(
"parts",
[[], [(float("nan"), 1)], [(1.0, 0)], [(float("inf"), 2)]],
)
def test_aggregate_nll_rejects_empty_or_nonfinite_aggregates(parts) -> None:
"""Catch non-finite or tokenless aggregate artifacts."""
with pytest.raises(ValueError):
aggregate_nll(parts)
def write_asset_manifests(root: Path, *, recorded_revision: str | None = None) -> None:
revisions = {"gpt2": "a" * 40, "gpt2-large": "b" * 40}
(root / "artifacts").mkdir()
(root / "data" / "manifests").mkdir(parents=True)
(root / "artifacts" / "data.yaml").write_text(
yaml.safe_dump({"models": revisions}), encoding="utf-8"
)
models = {}
for name, revision in revisions.items():
path = root / "data" / "raw" / name / "snapshots" / revision
path.mkdir(parents=True)
models[name] = {
"repo_id": name,
"revision": recorded_revision if name == "gpt2-large" and recorded_revision else revision,
"snapshot_path": path.relative_to(root).as_posix(),
}
(root / "data" / "manifests" / "downloads.json").write_text(
json.dumps({"schema_version": 1, "models": models}), encoding="utf-8"
)
def test_resolver_binds_model_and_tokenizer_to_pinned_local_snapshots(tmp_path: Path) -> None:
"""Catch a moving Hub alias or tokenizer/model revision mix-up."""
write_asset_manifests(tmp_path)
assets = resolve_gpt2_assets(tmp_path)
assert assets.model_id == "gpt2-large"
assert assets.model_revision == "b" * 40
assert assets.model_path.name == "b" * 40
assert assets.tokenizer_id == "gpt2"
assert assets.tokenizer_revision == "a" * 40
assert assets.tokenizer_path.name == "a" * 40
def test_resolver_rejects_download_record_at_another_revision(tmp_path: Path) -> None:
"""Catch silently evaluating from a stale or repointed local snapshot."""
write_asset_manifests(tmp_path, recorded_revision="c" * 40)
with pytest.raises(ValueError, match="gpt2-large revision"):
resolve_gpt2_assets(tmp_path)
def test_resolver_rejects_unknown_download_manifest_schema(tmp_path: Path) -> None:
"""Catch interpreting an incompatible lock format as current provenance."""
write_asset_manifests(tmp_path)
downloads_path = tmp_path / "data" / "manifests" / "downloads.json"
downloads = json.loads(downloads_path.read_text(encoding="utf-8"))
downloads["schema_version"] = 2
downloads_path.write_text(json.dumps(downloads), encoding="utf-8")
with pytest.raises(ValueError, match="schema_version"):
resolve_gpt2_assets(tmp_path)
def test_resolver_rejects_non_lowercase_hex_revision(tmp_path: Path) -> None:
"""Catch a 40-character value that is not an immutable lowercase Git revision."""
write_asset_manifests(tmp_path)
invalid = "g" * 40
configuration_path = tmp_path / "artifacts" / "data.yaml"
configuration = yaml.safe_load(configuration_path.read_text(encoding="utf-8"))
configuration["models"]["gpt2-large"] = invalid
configuration_path.write_text(yaml.safe_dump(configuration), encoding="utf-8")
downloads_path = tmp_path / "data" / "manifests" / "downloads.json"
downloads = json.loads(downloads_path.read_text(encoding="utf-8"))
source = tmp_path / downloads["models"]["gpt2-large"]["snapshot_path"]
destination = source.with_name(invalid)
source.rename(destination)
downloads["models"]["gpt2-large"].update(
revision=invalid, snapshot_path=destination.relative_to(tmp_path).as_posix()
)
downloads_path.write_text(json.dumps(downloads), encoding="utf-8")
with pytest.raises(ValueError, match="lowercase hexadecimal"):
resolve_gpt2_assets(tmp_path)
def test_resolver_rejects_symlinked_snapshot_ancestor(tmp_path: Path) -> None:
"""Catch a lock path being redirected after the download manifest was written."""
write_asset_manifests(tmp_path)
downloads_path = tmp_path / "data" / "manifests" / "downloads.json"
downloads = json.loads(downloads_path.read_text(encoding="utf-8"))
real_parent = tmp_path / "data" / "raw" / "gpt2-large"
alias = tmp_path / "data" / "raw" / "gpt2-large-alias"
alias.symlink_to(real_parent, target_is_directory=True)
revision = "b" * 40
downloads["models"]["gpt2-large"]["snapshot_path"] = (
alias / "snapshots" / revision
).relative_to(tmp_path).as_posix()
downloads_path.write_text(json.dumps(downloads), encoding="utf-8")
with pytest.raises(ValueError, match="symlink"):
resolve_gpt2_assets(tmp_path)
def test_loader_forces_local_snapshots_and_offline_mode(
tmp_path: Path, monkeypatch
) -> None:
"""Catch Transformers falling back to a network model or moving alias."""
calls = []
class Tokenizer:
pad_token_id = None
eos_token_id = 7
eos_token = "<eos>"
pad_token = None
class Model:
def to(self, device):
calls.append(("device", device))
return self
def eval(self):
calls.append(("eval",))
return self
class AutoTokenizer:
@classmethod
def from_pretrained(cls, path, **kwargs):
calls.append(("tokenizer", path, kwargs))
return Tokenizer()
class AutoModel:
@classmethod
def from_pretrained(cls, path, **kwargs):
calls.append(("model", path, kwargs))
return Model()
model_path = tmp_path / "model"
tokenizer_path = tmp_path / "tokenizer"
model_path.mkdir()
tokenizer_path.mkdir()
assets = GPT2Assets(
"gpt2-large", "b" * 40, model_path, "gpt2", "a" * 40, tokenizer_path
)
monkeypatch.setitem(
sys.modules,
"transformers",
SimpleNamespace(AutoModelForCausalLM=AutoModel, AutoTokenizer=AutoTokenizer),
)
monkeypatch.delenv("HF_HUB_OFFLINE", raising=False)
monkeypatch.delenv("TRANSFORMERS_OFFLINE", raising=False)
_, tokenizer = load_offline_gpt2_large(assets, device="cuda:1")
assert os.environ["HF_HUB_OFFLINE"] == "1"
assert os.environ["TRANSFORMERS_OFFLINE"] == "1"
assert ("tokenizer", str(tokenizer_path), {"local_files_only": True}) in calls
assert ("model", str(model_path), {"local_files_only": True}) in calls
assert ("device", "cuda:1") in calls
assert ("eval",) in calls
assert tokenizer.pad_token == "<eos>"
assert tokenizer.padding_side == "right"
assert tokenizer.truncation_side == "right"
def test_torch_scorer_shifts_masks_and_sums_causal_nll() -> None:
"""Catch framework integration using unshifted labels, padding, or mean loss."""
torch = pytest.importorskip("torch")
logits = torch.tensor(
[
[[0.0, math.log(2), 0.0], [0.0, 0.0, math.log(3)], [0.0, 0.0, 0.0]],
[[0.0, 0.0, 0.0], [100.0, -100.0, 0.0], [0.0, 0.0, 0.0]],
]
)
class TinyModel:
def eval(self):
return self
def __call__(self, *, input_ids, attention_mask):
return SimpleNamespace(logits=logits)
scorer = TorchCausalLMScorer(TinyModel(), device="cpu").eval()
total_nll, count = scorer.score_batch(
[[0, 1, 2], [0, 2, 1]], [[1, 1, 1], [1, 1, 0]]
)
assert count == 3
assert total_nll == pytest.approx(math.log(10))
|