Skip to content

[PD] Optimize MHA models pp util calculation logic #17306

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

Merged
ShangmingCai merged 2 commits into from
Jan 19, 2026
Merged

[PD] Optimize MHA models pp util calculation logic #17306

Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
35e5908
[PD] Optimize MHA models pp util calculation logic
ShangmingCai Jan 18, 2026
2bdd6c6
fix
ShangmingCai Jan 18, 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
6 changes: 3 additions & 3 deletions python/sglang/srt/disaggregation/common/conn.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -168,12 +168,12 @@ def get_mha_kv_ptrs_with_pp(
num_kv_layers < dst_num_total_layers
and dst_num_total_layers % num_kv_layers != 0
):
# Case: Decode has more layers than Prefill (e.g., Decode has draft model KV while Prefill is deployed without speculative decoding)
# To prevent empty Value Cache, which leads to wrong response
# Case: Decode has draft model KV while Prefill is deployed without speculative decoding
# dst_kv_ptrs layout: [K_main..., V_main..., draft_K..., draft_V...]
multiplier_ratio = dst_num_total_layers // num_kv_layers
dst_k_ptrs = dst_kv_ptrs[start_layer:end_layer]
dst_v_ptrs = dst_kv_ptrs[
num_kv_layers + start_layer : num_kv_layers + end_layer
num_kv_layers * multiplier_ratio + start_layer : num_kv_layers * multiplier_ratio + end_layer
]
else:
# Decode pp size should be equal to prefill pp size or 1
Expand Down
Loading