model._keep_in_fp32_modules
accelerate
* fix bug where weight would not be kept in fp32 * nit * address review comments * fix test
num_hidden_layers=2
test_beam_search_xla_generate_simple
T5
Tokenizer