-
Notifications
You must be signed in to change notification settings - Fork 21.7k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Dynamo failure due to non constants input of aten.lift_fresh_copy
#113331
Comments
aten.lift_fresh_copy
aten.lift_fresh_copy
Further investigation, here is the debug_trace of this fake tensor:
Looks like in pytorch/torch/_subclasses/fake_tensor.py Lines 1876 to 1885 in addb8e2
make_constant does not turn on.
cc @jgong5 |
Hi @eellison, could you kindly help to take a look of this issue? |
This is is because dynamo currently doesn't initialize its fake tensors with constant. i.e. we have |
I think it is fine to However, we need to do a lot of work to ensure that we know when the tensor is going to be constant. Further: In the case of compiling the exported graph, the tensor source is a It's even harder to prove it is constant in this case, compared with if it originates in a call to |
Thanks @jon-chuang for the comment. @ezyang @wconstab do we have any idea for how to fix this issue? Appreciate if anyone can help looking into this issue. |
In this case, the input could be fake! Just treat it normally in that case. Fixes #113331 Signed-off-by: Edward Z. Yang <ezyang@meta.com> [ghstack-poisoned]
In this case, the input could be fake! Just treat it normally in that case. Fixes #113331 Signed-off-by: Edward Z. Yang <ezyangmeta.com> cc avikchaudhuri gmagogsfm zhxchen17 tugsbayasgalan angelayi suo ydwu4 [ghstack-poisoned]
In this case, the input could be fake! Just treat it normally in that case. Fixes #113331 Signed-off-by: Edward Z. Yang <ezyangmeta.com> cc avikchaudhuri gmagogsfm zhxchen17 tugsbayasgalan angelayi suo ydwu4 [ghstack-poisoned]
In this case, the input could be fake! Just treat it normally in that case. Fixes #113331 Signed-off-by: Edward Z. Yang <ezyangmeta.com> cc avikchaudhuri gmagogsfm zhxchen17 tugsbayasgalan angelayi suo ydwu4 [ghstack-poisoned]
🐛 Describe the bug
For DistilBert defined in Transformer, there is runtime definition of mask fill value.
It cause the error when invoke
torch.compile
, the detail failure msg is:Example code to reproduce this failure:
Versions
cc @ezyang @msaroufim @wconstab @bdhirsh @anijain2305 @zou3519 @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @chenyang78 @aakhundov @kadeng
The text was updated successfully, but these errors were encountered: