[BugFix] Fix Llama4 Calibration #2101
Conversation
Summary of ChangesHello @dsikka, 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 addresses a bug in the Llama4 calibration process by refactoring the model's forward pass. The changes aim to improve the accuracy and consistency of expert routing and output accumulation, while also enhancing compatibility with recent versions of 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 forward pass of SequentialLlama4TextMoe to fix an issue with MoE calibration. The new implementation correctly handles different transformers library versions and uses a dense pass for calibration. However, it introduces a couple of issues. A debug print statement has been left in the code. More importantly, the calibrate_all_experts flag is now ignored, causing the module to always perform a dense pass and breaking the sparse routing functionality expected when the flag is False. My review includes suggestions to address these points.
| def forward(self, hidden_states: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]: | ||
| hidden_states = hidden_states.reshape(-1, self.hidden_dim) | ||
| router_scores, router_logits = self.router(hidden_states) # transformers>=4.54 | ||
| router_logits = self.router(hidden_states) | ||
| # support transformers 4.53 and greater | ||
| if isinstance(router_logits, tuple): | ||
| router_logits = router_logits[-1] | ||
|
|
||
| out = self.shared_expert(hidden_states) | ||
|
|
||
| for expert_index in range(self.num_experts): | ||
| # find expert scores | ||
| expert_score = router_scores[:, expert_index].unsqueeze(-1) | ||
| top_token_mask = expert_score[:, 0] > 0 | ||
|
|
||
| # llama4 applies scores before expert relu | ||
| expert_in = hidden_states * expert_score | ||
| router_top_value, router_indices = torch.topk(router_logits, self.top_k, dim=1) | ||
|
|
||
| # calibrate experts | ||
| if self.calibrate_all_experts: | ||
| expert_out = self.experts[expert_index](expert_in)[top_token_mask] | ||
| else: | ||
| expert_out = self.experts[expert_index](expert_in[top_token_mask]) | ||
| router_scores = ( | ||
| torch.full_like(router_logits, float("-inf")) | ||
| .scatter_(1, router_indices, router_top_value) | ||
| .transpose(0, 1) | ||
| ) | ||
| router_scores = torch.sigmoid(router_scores.float()).to(hidden_states.dtype) | ||
|
|
||
| # accumulate output | ||
| out[top_token_mask] += expert_out | ||
| out = self.shared_expert(hidden_states) | ||
| for i in range(self.num_experts): | ||
| out += self.experts[i](hidden_states) * router_scores[i].reshape(-1, 1) | ||
|
|
||
| return out, router_logits |
There was a problem hiding this comment.
The refactored forward method now implements a dense pass where all tokens are processed by all experts, which is a valid approach for calibration. However, the self.calibrate_all_experts flag is no longer used. This is a regression, as the module will now behave incorrectly when calibrate_all_experts is False by not performing the expected sparse routing.
Please reintroduce the conditional logic based on self.calibrate_all_experts. The new dense implementation should be within the if self.calibrate_all_experts: block. An else block should be added to implement the sparse expert routing, which is the expected behavior when not calibrating all experts. You can look at other MoE implementations in the codebase for reference on implementing an efficient sparse forward pass.
| self.router = original.router | ||
| self.shared_expert = original.shared_expert | ||
| self.calibrate_all_experts = calibrate_all_experts | ||
| print(self.calibrate_all_experts, "CALIBRATE ALL EXPERTS") |
There was a problem hiding this comment.
This print statement appears to be for debugging and should be removed before merging.
SUMMARY:
"please provide a brief summary"
TEST PLAN:
"please outline how the changes were tested"