Skip to content
Open
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
4 changes: 2 additions & 2 deletions transformer_engine/jax/attention.py
Original file line number Diff line number Diff line change
Expand Up @@ -120,11 +120,11 @@ class QKVLayout(Enum):
BSHD Format:
- BS3HD: q,k,v are interleave packed as a tensor with shape [b, s, 3, h, d].
- BSHD_BS2HD: q with shape [b, s, h, d] and kv are interleaved with shape [b, s, 2, h, d].
- BSHD_BSHD_BSHD: q,k,v are seperate tensors with shape [b, s, h, d]
- BSHD_BSHD_BSHD: q,k,v are separate tensors with shape [b, s, h, d]
THD Format: Shape is same as BSHD layout but allow multiple segments packed in a sequence.
- T3HD: q,k,v are interleave packed as a tensor with shape [b, s, 3, h, d].
- THD_T2HD: q with shape [b, s, h, d] and kv are interleaved with shape [b, s, 2, h, d].
- THD_THD_THD: q,k,v are seperate tensors with shape [b, s, h, d]
- THD_THD_THD: q,k,v are separate tensors with shape [b, s, h, d]
"""

BS3HD = NVTE_QKV_Layout.NVTE_BS3HD
Expand Down
Loading