A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 7, 00:15 UTC
[Bugfix] Clear stale CG keys after memory profiling (#36416)
Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>
Matthew Bonanni committed
Mar 9, 2026 at 11:56 UTC
00c4cb5606ae4f7ba80485f4a2756df33a2d4065
1 file changed
+3
vllm/v1/worker/gpu_model_runner.py
+3
@@ -5644,6 +5644,9 @@ class GPUModelRunner(
5644
for instance in list(CUDAGraphWrapper._all_instances):
5645
if id(instance) in original_pools:
5646
instance.graph_pool = original_pools[id(instance)]
5647
+ for key_set in self.cudagraph_dispatcher.cudagraph_keys.values():
5648
+ key_set.clear()
5649
+ self.cudagraph_dispatcher.keys_initialized = False
5650
self.maybe_remove_all_loras(self.lora_config)
5651
self._cleanup_profiling_kv_cache()
5652
compilation_counter.num_cudagraph_captured = saved_num_cudagraph_captured