pipeline("text-generation") + batch_size > 1 results in `For support of custom attention masks`...

#4
by michael-newsrx-com - opened

When calling pipeline with a batch_size > 1 the process fails with:

ValueError: For support of custom attention masks please set naive_attention_prefill to True in the config

Everything works fine with a batch_size=1 or batch_size not specified.

Is there a padding setting for the tokenizer/model that needs to be set?

Sign up or log in to comment