When evaluating the performance of Torch's Dynamo, I encountered the following issues:
Initially, I observed negative optimization results. However, I later discovered that adding warm-up in the evaluator resulted in positive optimization.
I was unable to run the CPU version, even after moving the tensor to the CPU.
When dealing with slightly larger matrices, I encountered bugs that prevented successful compilation. This issue has also been reported by others in the PyTorch repository. As of last month, the community had not provided a solution.
When evaluating the performance of Torch's Dynamo, I encountered the following issues:
Initially, I observed negative optimization results. However, I later discovered that adding warm-up in the evaluator resulted in positive optimization.
I was unable to run the CPU version, even after moving the tensor to the CPU.
When dealing with slightly larger matrices, I encountered bugs that prevented successful compilation. This issue has also been reported by others in the PyTorch repository. As of last month, the community had not provided a solution.