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

[Mamba][GDN] Deduplicate spec conv-window intermediate cache via sliding window layout (#28302)

Browse Source 
Co-authored-by: luoyuan.luo <luoyuan.luo@antgroup.com>
This commit is contained in:
Yuan Luo
2026-06-17 20:42:31 -07:00
committed by GitHub
co-authored by luoyuan.luo
parent 3340f4e3da
commit 3b9db3a1f0
5 changed files with 283 additions and 20 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/model_executor/model_runner_kv_cache_mixin.py
+2
View File
@@ -350,6 +350,7 @@ class ModelRunnerKVCacheMixin:
]
),
speculative_num_draft_tokens=max_spec_draft_tokens,
speculative_eagle_topk=self.server_args.speculative_eagle_topk,
enable_mamba_extra_buffer=self.server_args.enable_mamba_extra_buffer(),
pre_alloc_size=pre_alloc_size,
enable_overlap_schedule=not self.server_args.disable_overlap_schedule,
@@ -385,6 +386,7 @@ class ModelRunnerKVCacheMixin:
enable_mamba_extra_buffer=self.server_args.enable_mamba_extra_buffer(),
enable_mamba_extra_buffer_lazy=self.server_args.enable_mamba_extra_buffer_lazy(),
speculative_num_draft_tokens=max_spec_draft_tokens,
speculative_eagle_topk=self.server_args.speculative_eagle_topk,
enable_overlap_schedule=not self.server_args.disable_overlap_schedule,
start_layer=self.start_layer,
)