Antoni Viros

Ok! I'll make the PR then, just figured I'd ask before.

The same happens to me, but when using the simple Commit command it doesn't fail.

Can you suggest an alternative for accuracy debugging with symbolic shapes? Given that we're compiling for token generation, the compilation goes from minutes with symbolic shapes to hours without them,...

The outputs on static shapes are the same at least to 1e-4.

I did a study with generating 10 tokens with static shapes and dynamic shapes, and this is the result:  What I have noticed is that with dynamic...

Is there any value in sharing the profiler's traces to help debug, or how can I help otherwise?

I'll keep updating this when I get more results: For now, "aot_eager" backend returns exactly the same results as pure eager mode. I'm working on writing a repro, but I...

When trying `assume_static_by_default` and using `mark_dynamic` on the sequence length of the decoder input tensor, the inductor portion breaks with the following error: ``` [2023-03-22 16:02:32,094] torch._inductor.compile_fx: [INFO] Step 3:...

The spam is fixed! The error is not though...

I've been trying to break the graph with print statements to figure out where the accuracy error happens and make a repro, but as soon as I've put a few...