Fix parsing error caused by dot-separated variable debugNames #1109
Conversation
Signed-off-by: Cheng Hang <calvinhance@gmail.com>
| LOG_DEBUG("input name:" << input->debugName()); | ||
| if (input->debugName().find(".") != std::string::npos) { | ||
| auto pos = input->debugName().find("."); | ||
| auto newName = input->debugName().replace(pos, 1, "_"); |
There was a problem hiding this comment.
This seems hacky, @bowang007 thoughts? We might just be doing something wrong upstream causing these issues.
There was a problem hiding this comment.
what happens is that when there is any fallback we will copy the Metadata of the inputs of the lowered graph (include debugname()) to the inputs of the new graph through this line As a result, the inputs' names' are same with the Lowered Graph inputs, which has some values as x.1, these names cannot be deserialized by torch.jit.load(). This explains why we are not getting this error when there is no fallback, since all the inputs are substituted by input_0, input_1 when there is no fallback.
Another solution to this error would be simply deleting the mapping here. I did some tests locally which could help resolve this issue. However, not sure if this deletion would induce other issues, will run the test cases to ensure it works.
|
raised another PR to fix it in upstream code. #1148 |
|
Closing in favor of #1148 |
Description
Sometimes when fallback is triggered, the debugName of parameter names could contain
., e.g.,input.1andinput_ids.1. Such torchscript files, after conversion and saving to disk, cannot be parsed successfully. We propose to replace such.with_, which can be a solution.Fixes #1112
Type of change
Please delete options that are not relevant and/or add your own.
Checklist: