Troubleshooting

torch.compile warning on Python 3.13+

If you see a warning that Dynamo/torch.compile is unsupported on Python 3.13+, either:

• disable compile in config (inference.runtime.torch.compile_eikonet=false), or

• run under a supported Python version.

Shared-event PCG fallback warning

If logs show frequent fallback or “all groups fell back to diagonal”:

• inspect fallback reason counters (rows_cap, nodes_cap, tau_zero),

• increase shared_event_re.limits.max_rows and/or max_nodes,

• verify shared_event_re.model.tau_s is positive,

• enable fallback.abort_on_pcg_fallback=true during debugging.

See PCG Whitening: Convergence and Fallback Control for the full checklist.

Sampler instability at higher LR

If trajectories explode when freezing preconditioner:

• ensure adaptation phase is long enough before freeze,

• raise inference.sampler.eps,

• use gradient clipping (inference.sampler.grad_clip_norm),

• lower LR and retune gradually.