ai-agents-model-regression-hidden-state-indices-shift-by-1-between-transforme-fd659abd

THEN Before releasing model architecture changes, run a regression test that compares hidden state outputs against the previous release. Use a simple script to check that the shape and values of each hidden state tensor remain consistent (or are deprecated with a warning). If indices shift, fix the order or add a deprecation note.

import torch, transformers; model = AutoModelForCausalLM.from_pretrained('meta-llama/Llama-3.2-1B'); inputs = torch.arange(10).view(1,-1); out = model(inputs, output_hidden_states=True).hidden_states; # compare len and first slice of each hidden state against reference