Skip to content

[CI] [AMD] memory_pool: Keep separate K/V allocations on ROCm to fix Mooncake PD accuracy #46875

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

[CI] [AMD] memory_pool: Keep separate K/V allocations on ROCm to fix Mooncake PD accuracy #46875

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
4 changes: 3 additions & 1 deletion python/sglang/srt/mem_cache/memory_pool.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2627,7 +2627,9 @@ def _zeros(shape):
shape, dtype=self.store_dtype, device=self.device
)

if k_shape == v_shape:
# ROCm keeps separate K/V allocations: Mooncake PD transfers
# from a V half inside a fused allocation corrupt decode KV.
if k_shape == v_shape and not _is_hip:
# Store K and V as block-aligned halves of one allocation
# so paged kernels with a single pool base can reach V
# via block offsets. Each half keeps its standalone
Expand Down
Loading