【Hackathon 9th No.113】torch._C._fft.fft_irfft API转换 torch.fft.irfft #323
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
Thanks for your contribution! |
lixinqi
reviewed
Nov 4, 2025
| # TODO | ||
| def fft_irfft_to_irfft(self, gm): | ||
| def replace_in_graph(graph_mod): | ||
| # 在 GraphModule 上注册稳定实现，codegen 可以使用 self.irfft（有时） |
lixinqi
reviewed
Nov 4, 2025
|
|
||
| return gm | ||
|
|
||
| # def check_unstable_api(self, gm): |
Collaborator
There was a problem hiding this comment.
check_unstable_api 这里的逻辑复原，放到下一次pr来更改。
check_unstable_api是裁判员，而unstable_to_stable是运动员。我们不在同一个pr里同时改裁判员和运动员
Contributor
Author
ES_resul.png
|
lixinqi
reviewed
Nov 5, 2025
Collaborator
lixinqi
left a comment
lixinqi
left a comment
There was a problem hiding this comment.
本Pr得改一下这里
https://github.com/PaddlePaddle/GraphNet/blob/develop/graph_net/torch/fx_graph_serialize_util.py#L22 这样新抽取的计算图就不会有torch._C._fft.fft_irfft了
| self.unstable_api = unstable_api | ||
|
|
||
| def my_backend(gm, sample_inputs): | ||
| gm = self.unstable_to_stable(gm) |
|
|
||
| def my_backend(gm, sample_inputs): | ||
| gm = self.unstable_to_stable(gm) | ||
| gm = self.fft_irfft_to_irfft(gm) |
Collaborator
There was a problem hiding this comment.
这里的逻辑写到unstable_to_stable里。
Collaborator
|
@G2uge 可以扫码进群，研发大哥在群里等着你 |
Contributor
Author
|
lixinqi
approved these changes
Nov 6, 2025
Collaborator
|
hi, @G2uge
|
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Description
我对接口的检测逻辑进行了修改，原因是 PyTorch 2.5.1 及以上版本的 FX 代码生成机制会在生成代码字符串时保留原始 C API 路径（如 torch._C._fft.fft_irfft），即使节点 target 已经被替换为稳定 API，基于字符串的检测也会出现误报，无法准确反映模型实际运行时是否还会调用不稳定 API。为此，我将检测逻辑改为遍历 FX Graph 的所有节点 target，只要节点 target 包含不允许的 API 就报错。这样修改后，检测逻辑直接检查 FX Graph 中每个节点实际要调用的 target，而不是依赖代码字符串。即使 gm.code或其他生成的代码文本中还存在 torch._C._fft.fft_irfft这样的不稳定 API 名称，只要节点的 target 已经被替换为稳定的 torch.fft.irfft，模型实际运行时就会调用稳定接口，而不会再调用不稳定的实现。因此，这种检测方式能够准确反映模型实际运行时的安全性，避免了因代码生成机制导致的误判。换句话说，即使代码字符串中出现了不稳定 API 的名字，只要节点 target 已经替换，实际运行时调用的就是稳定接口，模型的安全性和正确性都能得到保障。
在 fft_irfft_to_irfft的实现中，我递归遍历主 GraphModule 及其所有子模块，对每个节点进行检查，如果发现调用了不稳定的 fft_irfft 接口，就将其 target 替换为稳定的 torch.fft.irfft。替换后重新编译图，确保后续模型运行时只会调用稳定 API。