Skip to content

fix: honor Hygon DCU SLA topk environment setting #1197

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
starrkk wants to merge 3 commits into
base: main
Choose a base branch
from
Open

fix: honor Hygon DCU SLA topk environment setting #1197

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
be16eac
fix: honor SLA topk environment setting
Jun 22, 2026
a89f1b7
fix: validate Hygon SLA topk environment value
Jun 30, 2026
5999f50
style: format Hygon SLA topk changes
Jul 1, 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
22 changes: 20 additions & 2 deletions lightx2v_platform/ops/attn/hygon_dcu/flash_attn.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -25,6 +25,24 @@
FLASH_ATTN_AVAILABLE = False


def _get_sparse_attn_topk(default: float = 0.4) -> float:
raw_topk = os.getenv("SPARSE_ATTN_TOPK", str(default))
try:
topk_value = float(raw_topk)
except (TypeError, ValueError):
logger.warning("Invalid SPARSE_ATTN_TOPK={!r}; falling back to {:.3f}", raw_topk, default)
return default

if not 0.0 < topk_value <= 1.0:
logger.warning(
"SPARSE_ATTN_TOPK={!r} is outside (0.0, 1.0]; falling back to {:.3f}",
raw_topk,
default,
)
return default
return topk_value


@PLATFORM_ATTN_WEIGHT_REGISTER("flash_attn_hygon_dcu")
class FlashAttnHygonDcu(AttnWeightTemplate):
"""
Expand Down Expand Up @@ -114,7 +132,7 @@ def half(x):
softmax_scale = 1.0 / math.sqrt(q.shape[-1])
# Use Flash Attention 2.6.1 (ROCm version) with varlen interface
if SAPRDE_LINEAR_ATTN and int(os.getenv("USE_SLA", 0)) and q.shape[1] == k.shape[1]:
topk_value = float(os.getenv("SPARSE_ATTN_TOPK", "0.5"))
topk_value = _get_sparse_attn_topk()

q = q_flat.unsqueeze(0)
k = k_flat.unsqueeze(0)
Expand All @@ -124,7 +142,7 @@ def half(x):
q,
k,
v,
topk=0.4,
topk=topk_value,
)
else:
output = flash_attn_varlen_func(
Expand Down