Transformer decoder layer, extend one step in autoregressive decoding. query_vec and aux_* may have different batch sizes, e.g., during a beam search. target_batch must be a multiple of source_batch and query_vec[i * batch_multiplier + j] corresponds to aux_vec[i], where batch_multi
(
self,
theta,
query_vec,
aux_vec,
aux_paddings,
cached_states,
time_step,
use_short_seq_opt=False,
per_step_padding=None,
*,
compute_atten_probs=False,
segment_mask=None,
aux_segment_mask=None,
)
source not stored for this graph (policy: none)