Hi, i don’t understand why the performance is worse than that of the original model. I compile and optimize the model for CUDA, and the model is a RNN define with pytorch.
Here are the performance: TVM unoptimized stats: {‘mean’: 4.033793111110652, ‘median’: 3.8734833333364804, ‘std’: 0.8640352986937846, ‘min’: 3.23199999999512, ‘max’: 8.233496666677334} TVM optimized stats: {‘mean’: 9.483390888888403, ‘median’: 9.281806666664732, ‘std’: 1.1006184318633645, ‘min’: 8.021686666658448, ‘max’: 11.621050000000347}
The measurements are in ms (milli seconds), and we can notice as the model optimized by TVM have a worse performance. How is possible?