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

model: Support Hybrid Mamba2 NemotronHForCausalLM (nvidia/NVIDIA-Nemotron-Nano-9B-v2) (#10909)

Browse Source 
Signed-off-by: Netanel Haber <nhaber@nvidia.com>
This commit is contained in:
Netanel Haber
2025-10-08 19:37:38 +03:00
committed by GitHub
parent c882b5ae75
commit d6837aea4d
35 changed files with 3280 additions and 854 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
python/sglang/srt/managers/scheduler.py
View File

@@ -1770,7 +1770,7 @@ class Scheduler(
chunked_req_to_exclude.add(self.chunked_req)
self.tree_cache.cache_unfinished_req(self.chunked_req, chunked=True)
# chunked request keeps its rid but will get a new req_pool_idx
if self.tp_worker.worker.model_runner.is_hybrid_gdn:
if self.tp_worker.worker.model_runner.mambaish_config is not None:
self.req_to_token_pool.free(
self.chunked_req.req_pool_idx, free_mamba_cache=False
)