EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

[Feature] Speculative decoding support lookahead (#9873)

Browse Source 
Co-authored-by: a4zhangfei <a4zhangfei@qq.com>
Co-authored-by: Qiaolin-Yu <liin1211@outlook.com>
This commit is contained in:
Zhihao Zhang
2025-09-19 07:42:41 +08:00
committed by GitHub
parent 2a2ff9a840
commit e7bc600304
30 changed files with 2058 additions and 32 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
python/sglang/srt/model_executor/model_runner.py
View File

@@ -1402,7 +1402,7 @@ class ModelRunner:
if self.is_hybrid_gdn:
max_num_reqs = min(max_num_reqs, self.server_args.max_mamba_cache_size)
if not self.spec_algorithm.is_none():
if self.spec_algorithm.is_eagle() or self.spec_algorithm.is_standalone():
if self.is_draft_worker:
self.max_total_num_tokens = self.server_args.draft_runner_cache_size
max_num_reqs = self.server_args.max_num_reqs