Common errors

If ns doctor reports a clean environment and you are still hitting a wall, one of the four errors below is almost certainly the culprit.

Fingerprint mismatch (E001)

Two objects that should be equal produced different fingerprints. The most common cause is an in-place mutation between fingerprinting and recording — for example callingmodel.eval() after taking the fingerprint. Reorder so the fingerprint is taken immediately before ns.record.

Diagnose with

$ python -c 'import neuronscope as ns; ns.fingerprint_diff(a, b).print()'
encoder.layer.0.buffer_running_mean: bytes differ (3 elements)
copy

Unknown adapter (E002)

You handed a framework object to a fresh install that does not have the matching extra. A pure pip install neuronscope does not carry PyTorch or JAX support — reinstall with the right extra:

$ pip install "neuronscope[torch]==0.9.2"
copy

Layer glob matched nothing (E003)

A probe was configured with a glob that resolves to an empty set of layers. This is almost always a typo. Inspect the model to see what is actually available:

$ ns introspect google-bert/bert-base-uncased --layers | head
embeddings.word_embeddings
encoder.layer.0.attention.self.query
encoder.layer.0.attention.self.key
...
copy

• Use encoder.layer.*.output instead of encoder.*.output.
• Escape square brackets — encoder.layer.[0-9].output, not [0-9].

Artifact schema too new (E004)

Someone recorded on a newer NeuronScope than you have installed. Upgrade or, if that is not possible, ask them to migrate:

$ ns migrate run.nsz --to 0.9.2 --out run-migrated.nsz
copy