Skip to content

Follow up fix custom flash block sizes fallback in Ulysses attention #437

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 2 commits into
base: main
Choose a base branch
from
+12 −17
Open

Follow up fix custom flash block sizes fallback in Ulysses attention #437

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
5e3aaa0
Follow up fix custom flash block sizes fallback in Ulysses attention
ninatu Jul 2, 2026
cd29876
Merge branch 'main' into ninatu/fix_heads_per_tile2
ninatu Jul 7, 2026
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
29 changes: 12 additions & 17 deletions src/maxdiffusion/models/attention_flax.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -363,24 +363,19 @@ def _extract_custom_block_sizes(flash_block_sizes):
if flash_block_sizes is not None:
if isinstance(flash_block_sizes, dict):
get = flash_block_sizes.get
bq = get("block_q", bq)
bkv = get("block_kv", bkv)
bkv_compute = get("block_kv_compute", bkv_compute)
bkv_compute_in = get("block_kv_compute_in", bkv_compute_in)
heads_per_tile = get("heads_per_tile", heads_per_tile)
vmem_limit_bytes = get("vmem_limit_bytes", vmem_limit_bytes)
bq = get("block_q", None) or bq
bkv = get("block_kv", None) or bkv
bkv_compute = get("block_kv_compute", None) or bkv_compute
bkv_compute_in = get("block_kv_compute_in", None) or bkv_compute_in
heads_per_tile = get("heads_per_tile", None) or heads_per_tile
vmem_limit_bytes = 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)
# A BlockSizes object carries heads_per_tile=None when the config dict omitted
# it; getattr then returns that None instead of the default, so coerce it back
# to 1 (the custom-kernel default) to keep the `heads_per_tile > 1` guards safe.
if heads_per_tile is None:
heads_per_tile = 1
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
return bq, bkv, bkv_compute, bkv_compute_in, heads_per_tile, vmem_limit_bytes


Expand Down
Loading