On torch>=2.6, `Pipeline.from_pretrained("pyannote/speaker-diarization-3.1")`
fails with "Weights only load failed ... Unsupported global: GLOBAL
torch.torch_version.TorchVersion" — PyTorch 2.6 flipped torch.load's default to
weights_only=True and its secure unpickler rejects the checkpoint's metadata
globals. This broke diarization on torch>=2.6 even when the license IS accepted
(reported on v0.3.4, RTX 4070 Ti, license accepted).
The WhisperX VAD load already solved this via
`WhisperXBackend._allow_vad_pickle_globals()` (allowlists TorchVersion,
omegaconf nodes, pyannote metadata, builtins, numpy, …). `get_diarization_pipeline`
just never called it. Reuse it before the diarization load — idempotent,
per-process, verified to register TorchVersion on torch 2.8.
Graceful fallback (silence-gap heuristic) is preserved if anything still fails.
Tests: tests/test_diarization_weights_only.py (allowlist runs before load;
no-token short-circuit). Existing diarization classification tests still pass.
Cross-platform (the torch 2.6 weights_only change affects all platforms).
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
