[NPU] [Features] [Bugfix] Support mindiesd adaln

[jiangmengyu18]

PLEASE FILL IN THE PR DESCRIPTION HERE ENSURING ALL CHECKLIST ITEMS (AT THE BOTTOM) HAVE BEEN CONSIDERED.

Purpose

1. accelerate adaln by mindiesd.
2. fix bug about adaln implemented by torch_npu.

The following is a detailed explanation of the bug about adaln implemented by torch_npu.

First of all, torch_npu.npu_layer_norm_eval is an optimized implementation of torch.nn.functional.layer_norm on NPU. The reason for using torch_npu.npu_layer_norm_eval here is that it is exactly equivalent to AdaLN when batch_size = 1, but it may cause precision issues when batch_size > 1.

Since AdaLN normalizes the last dimension of x, which is d, normalized_shape must be [d], and the shapes of weight and bias must also be [d]. However, the shapes of scale_result and shift_result are [b, 1, d]. Although torch_npu.npu_layer_norm_eval does not raise an error when the shapes of weight and bias are [b, 1, d], it will take weight[0][0] and bias[0][0] for broadcasting, instead of using the full weight and bias. It will lead to precision issues when batch_size > 1.

For example:

error result:

reproduce the error result a second time:

correct reuslt:

Test Plan

• Qwen-Image-Edit-2509

Test Result

The table below shows the time consumption of AdaLN when using native, torch_npu, and mindiesd, respectively.

native	torch_npu	mindiesd
36.8s	36.7s	35.8s

Performance:
native:

torch_npu:

mindiesd:

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan. Please providing the test scripts & test commands. Please state the reasons if your codes don't require additional test scripts. For test file guidelines, please check the test style doc
• The test results. Please pasting the results comparison before and after, or e2e results.
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model. Please run mkdocs serve to sync the documentation editions to ./docs.
• (Optional) Release notes update. If your change is user facing, please update the release notes draft.

BEFORE SUBMITTING, PLEASE READ https://github.com/vllm-project/vllm-omni/blob/main/CONTRIBUTING.md (anything written below this line will be removed by GitHub Actions)

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 8b3b5a0972

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[gcanlin]

Look good now. Thanks!
cc @hsliuustc0106

[lishunyang12]

Bugfix looks right -- the old code passing scale/shift as weight/bias to npu_layer_norm_eval was definitely wrong for batch > 1. Left one comment on the exception handling.

[lishunyang12]

Duplicated line. Remove one.

[jiangmengyu18]

you mean the rebundant logger.warning_once? I have pull request to solve it.

[jiangmengyu18]

cc @lishunyang12

[lishunyang12]

Bugfix looks correct — applying scale/shift outside npu_layer_norm_eval avoids the silent broadcasting issue. Left a nit about a duplicated line.