GeneZC
April 18, 2022, 12:23pm
1
# Mask is (batch_size, key_length) (non-causal) or (batch_size, key_length, key_length)
# past_key_value[0] is (batch_size, n_heads, q_len - 1, dim_per_head)
batch_size, seq_length = hidden_states.shape[:2]
real_seq_length = seq_length
if past_key_value is not None:
assert (
len(past_key_value) == 2
), f"past_key_value should have 2 past states: keys and values. Got { len(past_key_value)} past states"
real_seq_length += past_key_value[0].shape[2] if query_length is None else query_length
key_length = real_seq_length if key_value_states is None else key_value_states.shape[1]
def shape(states):
"""projection"""
return states.view(batch_size, -1, self.n_heads, self.key_value_proj_dim).transpose(1, 2)
def unshape(states):
"""reshape"""
return states.transpose(1, 2).contiguous().view(batch_size, -1, self.inner_dim)
If the past_key_value is not None, the query_length induced from SelfAttention can be the length of past_key_value + hidden_states, which is actually wrong.
Luckily, this problematic behavior would not affect the EncDecAttetion since the position_bias are always zeros.
Correct me if I am wrong.