refactor fp8 linear #210
refactor fp8 linear #210
Conversation
Summary of ChangesHello @akaitsuki-ii, I'm Gemini Code Assist1! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed! This pull request significantly refactors the implementation of FP8 linear operations within the Highlights
Using Gemini Code AssistThe full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips. Invoking Gemini You can request assistance from Gemini at any point by creating a comment using either
Customization To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a Limitations & Feedback Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for Github and other Google products, sign up here. You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension. Footnotes
|
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Code Review
This pull request refactors the FP8 linear layer implementation by replacing the previous monkey-patching approach with a more explicit and modular design. This is a significant improvement for code clarity and maintainability. The new implementation uses a dedicated fp8_linear function and LoRAFP8Linear module, which are consistently applied across various models. I've identified one critical issue that will cause a runtime error and a few medium-severity suggestions to improve the robustness of the module selection logic.
| ) | ||
| if config.use_fp8_linear: | ||
| enable_fp8_linear(dit) | ||
| dit.enable_fp8_linear() |
There was a problem hiding this comment.
This call to dit.enable_fp8_linear() will raise an AttributeError at runtime. The dit object, an instance of WanS2VDiT, does not have the enable_fp8_linear method defined.
To fix this, you should implement the enable_fp8_linear method in the WanS2VDiT class in diffsynth_engine/models/wan/wan_s2v_dit.py. It seems this was overlooked during the refactoring.
You can add a method similar to the one in WanDiT:
from diffsynth_engine.models.basic.lora import LoRAFP8Linear
...
class WanS2VDiT(WanDiT):
...
def enable_fp8_linear(self):
target_names = ["blocks"]
for name, module in self.named_modules():
if any(name.startswith(f"{t}.") for t in target_names) and isinstance(module, nn.Linear):
self.set_submodule(name, LoRAFP8Linear.from_linear(module))Note that you'll also need to import LoRAFP8Linear in diffsynth_engine/models/wan/wan_s2v_dit.py.
| def enable_fp8_linear(self): | ||
| target_names = ["blocks", "single_blocks"] | ||
| for name, module in self.named_modules(): | ||
| if any([t in name for t in target_names]) and isinstance(module, nn.Linear): |
There was a problem hiding this comment.
The condition any([t in name for t in target_names]) to identify target modules is a bit loose and could lead to incorrect module replacement if a module name incidentally contains one of the target names (e.g., a new module named archived_blocks). Using startswith would be more robust and ensure that only modules within the specified parent modules are matched.
| if any([t in name for t in target_names]) and isinstance(module, nn.Linear): | |
| if any(name.startswith(f"{t}.") for t in target_names) and isinstance(module, nn.Linear): |
| def enable_fp8_linear(self): | ||
| target_names = ["transformer_blocks"] | ||
| for name, module in self.named_modules(): | ||
| if any([t in name for t in target_names]) and isinstance(module, nn.Linear): |
There was a problem hiding this comment.
The condition any([t in name for t in target_names]) to identify target modules is a bit loose and could lead to incorrect module replacement if a module name incidentally contains one of the target names. Using startswith would be more robust and ensure that only modules within the specified parent modules are matched.
| if any([t in name for t in target_names]) and isinstance(module, nn.Linear): | |
| if any(name.startswith(f"{t}.") for t in target_names) and isinstance(module, nn.Linear): |
| def enable_fp8_linear(self): | ||
| target_names = ["blocks"] | ||
| for name, module in self.named_modules(): | ||
| if any([t in name for t in target_names]) and isinstance(module, nn.Linear): |
There was a problem hiding this comment.
The condition any([t in name for t in target_names]) to identify target modules is a bit loose and could lead to incorrect module replacement if a module name incidentally contains one of the target names. Using startswith would be more robust and ensure that only modules within the specified parent modules are matched.
| if any([t in name for t in target_names]) and isinstance(module, nn.Linear): | |
| if any(name.startswith(f"{t}.") for t in target_names) and isinstance(module, nn.Linear): |
No description provided.