[Qwen3-next] support mamba radix cache for overlap scheduler (#14792)

This commit is contained in:
Hanming Lu
2025-12-14 18:54:16 -08:00
committed by GitHub
parent 36e7c8c59f
commit e61dabf5e4
30 changed files with 1414 additions and 204 deletions
+3 -2
View File
@@ -405,7 +405,7 @@ class Scheduler(
# Hybrid memory pool
self.is_hybrid_swa = self.tp_worker.is_hybrid_swa
self.is_ssm_model = (
self.is_hybrid_ssm = (
self.tp_worker.model_runner.hybrid_gdn_config is not None
or self.tp_worker.model_runner.mamba2_config is not None
)
@@ -772,6 +772,7 @@ class Scheduler(
eviction_policy=server_args.radix_eviction_policy,
enable_metrics=self.enable_metrics,
enable_kv_cache_events=self.enable_kv_cache_events,
enable_mamba_extra_buffer=server_args.enable_mamba_extra_buffer(),
)
if (
@@ -808,7 +809,7 @@ class Scheduler(
self.tree_cache = SWARadixCache(
params=params, sliding_window_size=self.sliding_window_size
)
elif self.is_ssm_model:
elif self.is_hybrid_ssm:
from sglang.srt.mem_cache.mamba_radix_cache import MambaRadixCache
self.tree_cache = MambaRadixCache(params)