Validation

Validation compares eligible implementations with a higher-precision reference. It checks forward outputs and, when required, gradients.

records = rms_norm.validate(x, weight)
assert all(record.result.status == "pass" for record in records)

Each record identifies its implementation. Use report to print the results:

from popcorn.bench import report

for record in rms_norm.validate(x, weight):
    report(record)

Results are keyed by case, device, PyTorch version, backend version, and gradient mode. A kernel-code fingerprint invalidates records after functional changes; comments, formatting, and the running Python version do not affect it. Automatic dispatch excludes known failures. An implementation without a validation record remains available but emits UnvalidatedWarning.

validate always runs the comparison. To validate only missing cases during dispatch, pass validate=True or set POPCORN_VALIDATE=1:

output = rms_norm(x, weight, validate=True)

POPCORN_VALIDATE=1 python train.py