Closed kyakuno closed 1 year ago
Inference loopのshape。
input_ids.shape (12, 6)
attention_mask.shape (12, 6)
decoder_input_ids.shape (12, 1)
past_key_values[0].shape (12, 8, 0, 64)
input_ids.shape (12, 6)
attention_mask.shape (12, 6)
decoder_input_ids.shape (12, 1)
past_key_values[0].shape (12, 8, 1, 64)
decoder_input_idsには最初はpad=32000が入って、以降は前回のデコード結果が入る。
decoder_input_ids.shape (12, 1) [[32000]
[32000]
[32000]
[32000]
[32000]
[32000]
[32000]
[32000]
[32000]
[32000]
[32000]
[32000]]
decoder_input_ids.shape (12, 1) [[ 517]
[ 2]
[13785]
[23563]
[ 4793]
[ 130]
[ 2175]
[ 8270]
[ 158]
[ 1036]
[ 996]
[10080]]
Implement fugumt with ailia.tokenizer.