Skip to content
Merged
Changes from all commits
Commits
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
11 changes: 11 additions & 0 deletions python/sglang/srt/disaggregation/common/conn.py
Original file line number Diff line number Diff line change
Expand Up @@ -164,6 +164,17 @@ def get_mha_kv_ptrs_with_pp(
if num_kv_layers == dst_num_total_layers:
dst_k_ptrs = dst_kv_ptrs[:dst_num_total_layers]
dst_v_ptrs = dst_kv_ptrs[dst_num_total_layers:]
elif (
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
# dst_kv_ptrs layout: [K_main..., V_main..., draft_K..., draft_V...]
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
]
else:
# Decode pp size should be equal to prefill pp size or 1
dst_k_ptrs = dst_kv_ptrs[start_layer:end_layer]
Expand Down
Loading