chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

Avoid kimi linear stream sync (#16186)

Browse Source 
Signed-off-by: vincentzed <207368749+vincentzed@users.noreply.github.com>
This commit is contained in:
Yi Zhong
2026-02-11 20:27:22 -05:00
committed by GitHub
parent 539bbf485c
commit dc1309fc7e
1 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
python/sglang/srt/layers/attention/hybrid_linear_attn_backend.py
View File

@@ -535,7 +535,7 @@ class MambaAttnBackendBase(AttentionBackend):
self.query_start_loc_list[bs - 1][: bs - num_padding].copy_(
self.cached_cuda_graph_decode_query_start_loc[: bs - num_padding]
)
self.query_start_loc_list[bs - 1][bs - num_padding :].copy_(
self.query_start_loc_list[bs - 1][bs - num_padding :].fill_(
bs - num_padding
)
elif forward_mode.is_target_verify():
@@ -547,7 +547,7 @@ class MambaAttnBackendBase(AttentionBackend):
self.query_start_loc_list[bs - 1][: bs - num_padding].copy_(
self.cached_cuda_graph_verify_query_start_loc[: bs - num_padding]
)
self.query_start_loc_list[bs - 1][bs - num_padding :].copy_(
self.query_start_loc_list[bs - 1][bs - num_padding :].fill_(
(bs - num_padding) * spec_info.draft_token_num
)
else: