Requires for torch.tensor before casting #31755
Conversation
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
| @@ -762,7 +762,7 @@ def torch_int(x): | |||
|
|
|||
| import torch | |||
|
|
|||
| return x.to(torch.int64) if torch.jit.is_tracing() else int(x) | |||
| return x.to(torch.int64) if torch.jit.is_tracing() and isinstance(x, torch.Tensor) else int(x) | |||
There was a problem hiding this comment.
Here you can use the torch_int utility instead
There was a problem hiding this comment.
do you mean ensuring shift_size is a torch.Tensor when jit tracing in the modeling directly ?
| @@ -774,7 +774,7 @@ def torch_float(x): | |||
|
|
|||
| import torch | |||
|
|
|||
| return x.to(torch.float32) if torch.jit.is_tracing() else int(x) | |||
| return x.to(torch.float32) if torch.jit.is_tracing() and isinstance(x, torch.Tensor) else int(x) | |||
There was a problem hiding this comment.
I think since we're explicitly calling torch_float where float was used to be it should be fine, no?
|
@amyeroberts do you think this fix could be included in a patch release ? |
|
@echarlaix I'm happy to include it in a patch release towards the end of the week |
thanks a lot! |
Fixes ONNX export for swin, swin-donut and clap models
transformers/src/transformers/models/swin/modeling_swin.py
Line 643 in 82486e5
coming from :
introduced in #31311
as
torch_intis expecting atorch.Tensor:transformers/src/transformers/utils/generic.py
Line 765 in 82486e5
also I think we should be able to have here
cc @merveenoyan @xenova