Add TF weights
Model converted by the transformers' pt_to_tf CLI. All converted model outputs and hidden layers were validated against its PyTorch counterpart.
Maximum crossload output difference=4.530e-06; Maximum crossload hidden layer difference=6.055e-02;
Maximum conversion output difference=4.530e-06; Maximum conversion hidden layer difference=6.055e-02;
CAUTION: The maximum admissible error was manually increased to 0.1!
See GitHub PR #25558 for details, precision overridden due to hidden states being a little weird in TF; final output logits are within 1.788e-05 for all model variants/sizes.
Hey, these values look good! Don't stress too much about the hidden layer differences - an output difference of 4e-6 is very small, and suggests the port is working correctly.