fine-tuned T5 model fails to load after recent change

Directly importing a fine-tuned T5 model, like willwade/t5-small-spoken-typo fails with the following exception:

python t5.py --model willwade/t5-small-spoken-typo --prompt "grammar: A tasty apple"

  File "./mlx-examples/t5/t5.py", line 495, in <module>
    model, tokenizer = T5.from_pretrained(args.model, dtype)
                       ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "./mlx-examples/t5/t5.py", line 427, in from_pretrained
    model.load_weights(list(weights.items()))
  File "/mlx/nn/layers/base.py", line 178, in load_weights
    raise ValueError(f"Received parameters not in model: {extras}.")
ValueError: Received parameters not in model: decoder.layers.1.dense.wi.weight encoder.layers.0.dense.wi.weight decoder.layers.3.dense.wi.weight encoder.layers.1.dense.wi.weight encoder.layers.5.dense.wi.weight encoder.layers.3.dense.wi.weight encoder.layers.4.dense.wi.weight decoder.layers.0.dense.wi.weight decoder.layers.2.dense.wi.weight encoder.layers.2.dense.wi.weight decoder.layers.5.dense.wi.weight decoder.layers.4.dense.wi.weight.

After a recent commit https://github.com/ml-explore/mlx-examples/commit/d72fdeb4eeb0cf65f146ad3c3fbc48a2cd23974b, there is no longer a convert.py for T5 models.

Reverting https://github.com/ml-explore/mlx-examples/commit/d72fdeb4eeb0cf65f146ad3c3fbc48a2cd23974b and running convert.py to generate an .npz restores the expected functionality.

Start with t5/t5.py, especially T5.from_pretrained around line 427, and compare the loading behavior with commit d72fdeb4eeb0cf65f146ad3c3fbc48a2cd23974b. Reproduce the command using willwade/t5-small-spoken-typo and inspect the removed convert.py workflow; done means the fine-tuned model loads without unexpected dense.wi.weight parameters.