Skip to content

Fix custom flash block sizes fallback in Ulysses attention #420

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
ninatu wants to merge 1 commit into
base: main
Choose a base branch
from
+12 −12
Open

Fix custom flash block sizes fallback in Ulysses attention #420

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
24 changes: 12 additions & 12 deletions src/maxdiffusion/models/attention_flax.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -635,19 +635,19 @@ def wrap_ulysses_attention(query, key, value):

if flash_block_sizes is not None:
if isinstance(flash_block_sizes, dict):
bq = flash_block_sizes.get("block_q", bq)
bkv = flash_block_sizes.get("block_kv", bkv)
bkv_compute = flash_block_sizes.get("block_kv_compute", bkv_compute)
bkv_compute_in = flash_block_sizes.get("block_kv_compute_in", bkv_compute_in)
heads_per_tile = flash_block_sizes.get("heads_per_tile", heads_per_tile)
vmem_limit_bytes = flash_block_sizes.get("vmem_limit_bytes", vmem_limit_bytes)
bq = flash_block_sizes.get("block_q", None) or bq
bkv = flash_block_sizes.get("block_kv", None) or bkv
bkv_compute = flash_block_sizes.get("block_kv_compute", None) or bkv_compute
bkv_compute_in = flash_block_sizes.get("block_kv_compute_in", None) or bkv_compute_in
heads_per_tile = flash_block_sizes.get("heads_per_tile", None) or heads_per_tile
vmem_limit_bytes = flash_block_sizes.get("vmem_limit_bytes", None) or vmem_limit_bytes
else:
bq = getattr(flash_block_sizes, "block_q", bq)
bkv = getattr(flash_block_sizes, "block_kv", bkv)
bkv_compute = getattr(flash_block_sizes, "block_kv_compute", bkv_compute)
bkv_compute_in = getattr(flash_block_sizes, "block_kv_compute_in", bkv_compute_in)
heads_per_tile = getattr(flash_block_sizes, "heads_per_tile", heads_per_tile)
vmem_limit_bytes = getattr(flash_block_sizes, "vmem_limit_bytes", vmem_limit_bytes)
bq = getattr(flash_block_sizes, "block_q", None) or bq
bkv = getattr(flash_block_sizes, "block_kv", None) or bkv
bkv_compute = getattr(flash_block_sizes, "block_kv_compute", None) or bkv_compute
bkv_compute_in = getattr(flash_block_sizes, "block_kv_compute_in", None) or bkv_compute_in
heads_per_tile = getattr(flash_block_sizes, "heads_per_tile", None) or heads_per_tile
vmem_limit_bytes = getattr(flash_block_sizes, "vmem_limit_bytes", None) or vmem_limit_bytes

if use_base2_exp:
query = query * LOG2E
Expand Down
Loading