Bugfix: Fix accuracy degradation caused by EPLB

[Mercykid-bash]

No description provided.

[gemini-code-assist]

Code Review

This pull request introduces several bug fixes related to the Expert Placement and Load Balancing (EPLB) mechanism for Mixture-of-Experts (MoE) models, aimed at resolving accuracy degradation. The key changes include: using in-place copy_ to correctly update expert weights, padding expert maps before updating to handle dynamic shape changes, and ensuring that up-to-date weight scales are used during computation. These changes appear to correctly address the root causes of the accuracy issues. However, I've identified a potential issue in how the CPU-side expert map is updated, which could lead to runtime errors. My detailed feedback is in the comment below.

[gemini-code-assist]

The update to self.expert_map_per_layer_cpu using copy_ might lead to a runtime error if updated_expert_map has a different shape than the existing tensor in self.expert_map_per_layer_cpu[layer_id]. The logic to calculate pad_len just before this suggests that the shape of updated_expert_map can indeed be smaller. While the device tensor self.expert_map_per_layer is correctly updated with a padded version, this CPU copy is not, which is inconsistent and unsafe.

To prevent potential crashes, I recommend reverting this line to use clone(), which was the original implementation and is safer as it replaces the tensor entirely, handling any shape changes gracefully.

Suggested change

	self.expert_map_per_layer_cpu[layer_id].copy_(updated_expert_map)
	self.expert_map_per_layer_cpu[layer_id] = updated_expert_map.clone()

[github-actions]

👋 Hi! Thank you for contributing to the vLLM Ascend project. The following points will speed up your PR merge:‌‌

• A PR should do only one thing, smaller PRs enable faster reviews.
• Every PR should include unit tests and end-to-end tests ‌to ensure it works and is not broken by other future PRs.
• Write the commit message by fulfilling the PR description to help reviewer and future developers understand.

If CI fails, you can run linting and testing checks locally according Contributing and Testing.

[github-actions]

This pull request has conflicts, please resolve those before we can evaluate the pull request.