torch.export of ResNet with dynamic height fails due to constraint violation about pytorch

[zhxchen17]

One more (unrelated) issue is that performance of exported model, when called from python, is a bit slower than both eager evaluation, and torch.compile (which is the fastest, ignoring compilation time), although it's a small model. I was hoping it would be similar in performance to torch.compile -- just curious if this is expected?

Note that if you run exported model directly, it won't be performant by default because we are just running it op by op. To get it optimized, it's better to use a backend such as torch._inductor.aot_compile.

[lopuhin]

One more (unrelated) issue is that performance of exported model, when called from python, is a bit slower than both eager evaluation, and torch.compile (which is the fastest, ignoring compilation time), although it's a small model. I was hoping it would be similar in performance to torch.compile -- just curious if this is expected?