Skip to content
Closed
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
8 changes: 6 additions & 2 deletions src/MaxText/layers/attention_op.py
Original file line number Diff line number Diff line change
Expand Up @@ -1367,10 +1367,14 @@ def cudnn_flash_attention(
qkv_layout = "THD_THD_THD" # Packed format: 'T3HD', 'THD_T2HD' or 'THD_THD_THD'
if decoder_segment_ids is None:
decoder_segment_ids = jnp.ones(shape=query.shape[:2], dtype=jnp.int32)
attn_mask = SequenceDescriptor.from_segment_ids_and_pos(segment_ids=decoder_segment_ids, segment_pos=None)
attn_mask = SequenceDescriptor.from_segment_ids_and_pos(
segment_ids=decoder_segment_ids, segment_pos=None, is_thd=True, is_segment_ids_reordered=False
)
# Create dummy SequenceDescriptor for lazy_init
dummy_segment_ids = jnp.ones(shape=query.shape[:2], dtype=jnp.int32)
dummy_attn_mask = SequenceDescriptor.from_segment_ids_and_pos(segment_ids=dummy_segment_ids, segment_pos=None)
dummy_attn_mask = SequenceDescriptor.from_segment_ids_and_pos(
segment_ids=dummy_segment_ids, segment_pos=None, is_thd=True, is_segment_ids_reordered=False
)
max_segments_per_seq = self.config.max_segments_per_seq
elif using_context_parallelism or self.config.dataset_type == "synthetic":
if self.attention_type == AttentionType.LOCAL_SLIDING:
Expand Down
Loading