[EPLB][Nightly][Bugfix] Get expert from moe layer only

[shenchuxiaofugui]

What this PR does / why we need it?

1. If the model has dense layers, the current code will attempt to obtain the routing experts of the dense layers, which will cause an error. This should be fixed by modifying the code to skip the dense layers when obtaining the routing experts.
2. The global_expert_map that the function directly outputs a affects the performance of dsv3.2.

Does this PR introduce any user-facing change?

How was this patch tested?

DeepSeek V3.1 conversation is normal.
del":"dsr1","choices":[{"index":0,"message":{"role":"assistant","content":"Of course. Here is a clear and comprehensive explanation of deep learning.\n\n### What is Deep Learning? (In a Nutshell)\n\nDeep learning is a subset of machine learning that uses artificial neural networks with many layers ("deep" networks) to learn and make intelligent decisions from vast amounts of data.\n\nThink of it as teaching a computer to recognize patterns by example, much like how a human child learns to identify a cat by seeing many pictures of cats.\n\n---\n\n### The Core Idea: Mimicking the Brain (Very Loosely)\n\nAt the heart of deep learning are Artificial Neural Networks (ANNs), which are inspired by the structure of the human brain.\n\n* Neurons: The building blocks are artificial neurons (or nodes), which are simple computational units.\n* Layers: These neurons are arranged in layers:\n * Input Layer: Receives the raw data (e.g., pixels of an image, words of a sentence).\n * Hidden Layers: These are the "deep" part. A network can have dozens or even hundreds of these layers. Each layer processes the input from the previous layer, extracts increasingly complex features, and passes it on.\n * Output Layer: Produces the final result (e.g., the label "cat," the translated sentence, a probability score).\n\n\nA simple visualization of a deep neural network with multiple hidden layers.\n\n### How Does it Learn?\n\nThe "learning" happens through a process","refus

aime precision test (dsv3.1)

baseline without eplb

dataset	version	metric	mode	vllm-api-general-chat
aime2024	604a78	accuracy	gen	66.67

eplb

dataset	version	metric	mode	vllm-api-general-chat
aime2024	604a78	accuracy	gen	70.00

dsv3.2 performance

• vLLM version: v0.13.0
• vLLM main: vllm-project/vllm@11b6af5

[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.

[gemini-code-assist]

Code Review

This pull request correctly fixes a bug in get_global_expert_map where an incorrect layer index was used to access MoE layers. The change ensures that the correct offset (num_dense_layers) is applied, making the layer access consistent with other parts of the code. I have added a suggestion to further improve the code's readability and maintainability by refactoring the loop to avoid code duplication, which will help prevent similar bugs in the future.