MCPcopy Create free account
hub / github.com/PaddlePaddle/FastDeploy / clear_cache

Method clear_cache

fastdeploy/worker/gpu_model_runner.py:2813–2830  ·  view source on GitHub ↗

Clear cached data from shared inputs and forward metadata

(self, profile=False)

Source from the content-addressed store, hash-verified

2811 return get_stop(self.share_inputs["not_need_stop"]).item()
2812
2813 def clear_cache(self, profile=False):
2814 """Clear cached data from shared inputs and forward metadata"""
2815 create_cache_tensor = profile or not (
2816 self.fd_config.cache_config.num_cpu_blocks > 0
2817 or self.fd_config.cache_config.kvcache_storage_backend
2818 or self.fd_config.scheduler_config.splitwise_role != "mixed"
2819 )
2820 local_rank = self.local_rank % self.parallel_config.tensor_parallel_size
2821
2822 if not create_cache_tensor:
2823 for name, tensor in self.cache_kvs_map.items():
2824 unset_data_ipc(tensor, name, True, False)
2825 self.cache_ready_signal.value[local_rank] = 0
2826 self.cache_kvs_map.clear()
2827 self.share_inputs.pop("caches", None)
2828 if self.forward_meta is not None:
2829 self.forward_meta.clear_caches()
2830 paddle.device.cuda.empty_cache()
2831
2832 def clear_parameters(self, pid):
2833 """Dynamic model loader use to clear parameters use for RL"""

Callers 3

profile_runMethod · 0.95
clear_parametersMethod · 0.95
sleepMethod · 0.95

Calls 4

itemsMethod · 0.80
popMethod · 0.80
clearMethod · 0.45
clear_cachesMethod · 0.45

Tested by

no test coverage detected