Skip to content

fix(aiter): cap workspace buffer partitions by KV cache capacity to prevent OOM #20888

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

Closed
michaelzhang-ai wants to merge 1 commit into from
Closed

fix(aiter): cap workspace buffer partitions by KV cache capacity to prevent OOM #20888

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
9 changes: 8 additions & 1 deletion python/sglang/srt/layers/attention/aiter_backend.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -186,8 +186,15 @@ def __init__(
)

# aiter kernel related initialization
# Cap effective sequence length by actual KV cache capacity to avoid
# over-allocating the workspace buffer on memory-constrained GPUs.
# No single sequence can exceed max_total_num_tokens.
effective_max_seq_len = min(
self.max_context_len,
getattr(model_runner, "max_total_num_tokens", self.max_context_len),
)
self.max_num_partitions = (
self.max_context_len + _AITER_PARTITION_SIZE_ROCM - 1
effective_max_seq_len + _AITER_PARTITION_SIZE_ROCM - 1
) // _AITER_PARTITION_SIZE_ROCM

nbyes_per_qo_elem = torch.finfo(torch.float32).bits // 8
Expand Down
Loading