is_causal fix for evals?
This commit is contained in:
@@ -155,6 +155,8 @@ def flashattn_forward(
|
|||||||
# during training q,k,v always have same seqlen
|
# during training q,k,v always have same seqlen
|
||||||
assert key_states.shape == query_states.shape
|
assert key_states.shape == query_states.shape
|
||||||
is_causal = True
|
is_causal = True
|
||||||
|
elif past_key_value is None:
|
||||||
|
is_causal = True
|
||||||
else:
|
else:
|
||||||
# turn off FA causal mask after first inference autoregressive iteration
|
# turn off FA causal mask after first inference autoregressive iteration
|
||||||
# only on first autoregressive step q,k,v have same seqlen
|
# only on first autoregressive step q,k,v have same seqlen
|
||||||
|
|||||||
Reference in New Issue
Block a user