Skip to content

v3.11.1

Choose a tag to compare

View all tags
@anulum anulum released this 27 Mar 23:26
· 2265 commits to main since this release
v3.11.1
7d16a40

Fixed

  • NLI CUDA auto-detection: _load_nli_model() now auto-selects CUDA when torch.cuda.is_available() and device is None. Previously the model stayed on CPU unless nli_device="cuda" was passed explicitly. 6.8x latency improvement on L40S (169.5 ms → 24.9 ms).
  • director_assert() crash: passed a float to HallucinationError which expected a CoherenceScore object. Any hallucination detection via director_assert() would raise AttributeError instead of HallucinationError.

Added

  • 16 tests for integrations/dspy.py (coherence_check + director_assert)
  • 15 tests for integrations/semantic_kernel.py (DirectorAIFilter init + async call)
  • VerifiedScorer docs: atomic=True, evidence_top_k, SourceSpan dataclass, multi-span evidence
  • Privacy policy page
  • L40S GPU benchmark results (24.9 ms NLI median, 40.2 RPS)
  • Rust vs Python signal benchmark (BM25 10.2x, trend_drop 20.7x)

Measured Numbers

Metric Value
L40S NLI GPU median 24.9 ms (was 169.5 ms on CPU)
L40S NLI throughput 40.2 RPS
Heuristic median 0.088 ms
Rust BM25 (100 docs) 10.8 us (10.2x vs Python)
Rust trend_drop 0.3 us (20.7x vs Python)
AggreFact BA 75.86% (29,320 samples)

Full changelog: v3.11.0...v3.11.1

Assets 7
Loading