[core][torch.compile] discard the compile for profiling (#7796)
This commit is contained in:
@@ -1097,6 +1097,10 @@ class GPUModelRunnerBase(ModelRunnerBase[TModelInputForGPU]):
|
||||
device=self.device)
|
||||
self.execute_model(model_input, kv_caches, intermediate_tensors)
|
||||
torch.cuda.synchronize()
|
||||
|
||||
# reset and discard the guard and compiled bytecode for profiling runs
|
||||
torch._dynamo.reset()
|
||||
|
||||
return
|
||||
|
||||
def remove_all_loras(self):
|
||||
|
||||
Reference in New Issue
Block a user