`b85197c`

tests: cover B6 None vs 0.0 distinction; widen integration assertion

Authored by

espadonne 3 weeks ago

Status	File	+	-
M	`tests/integration/test_hf_scoring.py`	3	1
M	`tests/unit/test_scoring.py`	20	1

tests/integration/test_hf_scoring.pymodified

          # Top-k must arrive in descending probability order.
          assert np.all(np.diff(d.logprobs) <= 1e-7)
          assert d.vocab_size > 64
 -        assert math.isfinite(d.tail_logprob) or d.tail_logprob == 0.0
 +        # B6: tail_logprob is None (k covers vocab — won't happen here),
 +        # 0.0 (underflow), or a finite negative log-prob.
 +        assert d.tail_logprob is None or math.isfinite(d.tail_logprob)
      def test_dist_changes_under_adapter(self, hf_backend: HuggingFaceDifferentialBackend) -> None:
          prompt = "the adapter influences"

tests/unit/test_scoring.pymodified

              logprobs=np.array([-0.1, -1.0, -3.0], dtype=np.float32),
              vocab_size=50_257,
+         )
 -        assert dist.tail_logprob == 0.0
 +        # B6: default tail_logprob is None ("no tail recorded"), not
 +        # 0.0 (which now means "tail underflowed to zero, but exists").
 +        assert dist.tail_logprob is None
          assert dist.token_ids.shape == (3,)
 +    def test_explicit_tail_distinguishes_zero_from_none(self) -> None:
 +        """B6: 0.0 means measurable-but-tiny; None means no tail at all."""
 +        d_no_tail = TokenDist(
 +            token_ids=np.array([1], dtype=np.int64),
 +            logprobs=np.array([0.0], dtype=np.float32),
 +            vocab_size=1,
 +            tail_logprob=None,
 +        )
 +        d_underflow = TokenDist(
 +            token_ids=np.array([1], dtype=np.int64),
 +            logprobs=np.array([0.0], dtype=np.float32),
 +            vocab_size=1,
 +            tail_logprob=0.0,
 +        )
 +        assert d_no_tail.tail_logprob is None
 +        assert d_underflow.tail_logprob == 0.0
++
  class TestProtocols:
      def test_scoring_backend_runtime_checkable(self) -> None: