[Bugfix][KV Offload] Propagate EAGLE mode to SimpleCPU coordinator (#49071)
Signed-off-by: Yifan Qiao <yifanqiao@inferact.ai> Co-authored-by: OpenAI Codex <codex@openai.com>
Yifan Qiao committed
Jul 19, 2026 at 23:39 UTC
9bc266d923560f1c8bc1cd6f50e841f0fabacc71
1 file changed
+3
-3
vllm/v1/simple_kv_offload/manager.py
+3
-3
@@ -117,12 +117,13 @@ class SimpleCPUOffloadScheduler:
117
"lazy" if lazy_offload else "eager",
118
)
119
120
- # TODO (yifan): maybe need to enable kv_cache_events and metrics_collector here.
120
+ spec_config = vllm_config.speculative_config
121
+ use_eagle = spec_config is not None and spec_config.use_eagle()
122
self.cpu_coordinator: KVCacheCoordinator = get_kv_cache_coordinator(
123
kv_cache_config=self.cpu_kv_cache_config,
124
max_model_len=vllm_config.model_config.max_model_len,
125
max_in_flight_tokens=vllm_config.max_in_flight_tokens,
125
- use_eagle=False,
126
+ use_eagle=use_eagle,
127
enable_caching=True,
128
enable_kv_cache_events=self.enable_kv_cache_events,
129
dcp_world_size=dcp_world_size,
@@ -131,7 +132,6 @@ class SimpleCPUOffloadScheduler:
132
hash_block_size=self.hash_block_size,
133
)
134
self.cpu_block_pool: BlockPool = self.cpu_coordinator.block_pool
134
-
135
# GPU block pool reference - bound after scheduler builds kv_cache_manager
136
self._gpu_block_pool: BlockPool | None = None
137