FAQ

The most common cause is non-deterministic weight init inside a wrapper class the adapter does not recognize. Run ns.fingerprint(model, kind="arch") and kind="weights" separately: if arch matches and weights does not, one checkpoint has bytes the other does not. If arch also differs, the module tree is not the same — check for lazily-created buffers.

No. Fingerprints are computed on a CPU snapshot cast to a canonical dtype. Moving a model between CUDA and CPU, or between float16 and bfloat16, does not change the digest.

The dataset iterator is producing rows faster than the writer can drain the ring buffer. Either raise buffer_mb, lower recording throughput (fewer probes, wider stride), or point NS_ARTIFACT_DIR at faster storage. See common errors for a full triage tree.

Yes. Adapters exist for HuggingFace transformers, raw PyTorch modules, JAX pytrees, and any object exposing a state_dict()-like surface. See the Model concept for the full adapter matrix.

That is the intended workflow. Both artifacts share the architecture fingerprint (comparison is allowed) and differ in the weights fingerprint (which is what the report describes).

Only if the layer names align. NeuronScope does not implement a cross-framework name mapping — pass an explicit alignment= map to ns.compare() or use ns diff --alignment.

Roughly batch × tokens × hidden × layers × 2 bytes for a float16 activation recording, plus ~2% overhead for the fingerprint index. See Benchmarks for concrete numbers.

Yes — see the Cloud storage recipe. Range reads pull only the layer index and the requested tensors.

Apache 2.0 for both the library and the CLI. The Rust reader crate is dual MIT/Apache 2.0.

Not from the maintainers. NeuronScope is a local tool that writes files. Third-party hosted dashboards read the same .nsz format — see Ecosystem.