-
Notifications
You must be signed in to change notification settings - Fork 81
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
[RFC][Frontend] Add a OneFlow Frontend #24
base: main
Are you sure you want to change the base?
Conversation
Update 0024-add-oneflow-frontend.md
Update 0024-add-oneflow-frontend.md
res50_module = resnet50() | ||
pretrain_models = flow.load(model_path) | ||
res50_module.load_state_dict(pretrain_models) |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Does OneFlow support any input model from a file? If so, I'd like also to suggest that you also cover tvmc
on your plans/implementation at some point.
Example implementation - https://github.com/apache/tvm/blob/e691c7f83892d7242d0992c78cec1e2f8953a9e3/python/tvm/driver/tvmc/frontends.py#L174-L198
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Thanks for your suggestions! We have added this task to Future possibilities.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Thank you for your suggestions! We can't support this feature yet, but it is in our future development plan. After OneFlow supports this feature, we will add it to tvmc
.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Perhaps this isn't an issue that is covered in the scope of this RFC, but I'm wondering how independent the importers are from one another, and if it's worthwhile to look at a more general framework for third-party importers.
Are we aiming to support all operators? What does the mapping from OneFlow operators to TVM look like? Will there be gaps that a user will have to be concerned with? What is the plain for maintenance of the new format to import?
- Handle large models easily | ||
- Almost zero runtime overhead & linear speedup | ||
- Support automatic mixed precision | ||
- ... |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Is this just a placeholder? We should fill out the rest of the motivations or remove the ellipses.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Yes,I will delete it。
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Perhaps this isn't an issue that is covered in the scope of this RFC, but I'm wondering how independent the importers are from one another, and if it's worthwhile to look at a more general framework for third-party importers.
Are we aiming to support all operators? What does the mapping from OneFlow operators to TVM look like? Will there be gaps that a user will have to be concerned with? What is the plain for maintenance of the new format to import?
Yes, we will support all commonly used operators. OneFlow will release version 1.0 soon, and our operator interface is aligned with the Pytorch API.
Add a OneFlow frontend for TVM, also I have created a RFC topic in forum where we can discuss at same time.
The related pull request is apache/tvm#8790 .
@tqchen @comaniac @junrushao1994