[BugFix] Fix Qwen3-next break

[wxsIcey]

What this PR does / why we need it?

Fix Qwen3NextGatedDeltaNet, caused by vllm-project/vllm#26437

Does this PR introduce any user-facing change?

N/A

How was this patch tested?

def main():
    prompts = [
        "窗前明月光，",
        "The president of the United States is Mr.",
        "The capital of France is",
        "The future of AI is",
        "感时花溅泪，",
        "家书抵万金啥意思？",
        "plz tell me a story: ",
    ]

    # Create a sampling params object.
    sampling_params = SamplingParams(max_tokens=100, temperature=0.6, top_k=40, top_p=0.95)
    # Create an LLM.
    llm = LLM(
        model="/root/.cache/modelscope/hub/models/Qwen/Qwen3-Next-80B-A3B-Instruct",
              tensor_parallel_size=4,
              enforce_eager=True,
              trust_remote_code=True,
              max_model_len=256,
              gpu_memory_utilization=0.7,
              block_size=64
              )

    # Generate texts from the prompts.
    outputs = llm.generate(prompts, sampling_params)
    for output in outputs:
        prompt = output.prompt
        generated_text = output.outputs[0].text
        print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")

• vLLM version: v0.11.0rc3
• vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0

[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.

[wxsIcey]

needs #3719

[gemini-code-assist]

Code Review

This pull request addresses a bug with the Qwen3-next model on vLLM. The changes involve vendoring a Triton kernel for layer normalization and fixing a configuration issue in the Mamba model setup. The fixes appear to be correct and address the immediate problem. My primary feedback concerns the long-term maintainability of vendoring the Triton kernel. While this is a valid approach for a hotfix, it introduces risks of divergence from the upstream vLLM project. I've provided a comment with a suggestion to mitigate this risk.

[wxsIcey]

Incompatible with upstream changes because ''torch_npu._C._NPUDeviceProperties' object has no attribute 'multi_processor_count'

[github-actions]

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

[wxsIcey]

Test Results:

Prompt: '窗前明月光，', Generated text: '疑是地上霜。\nA. 举头望明月\nB. 疑是地上霜\nC. 举头望明月\nD. 低头思故乡\n答案:\n C\n\n在1999年3月14日，以____为单位的欧元在11个欧洲国家开始使用\nA. 1989\nB. 1990\nC. 1998\nD. 199'
Prompt: 'The president of the United States is Mr.', Generated text: " Joe Biden. In 2021, President Biden declared that he would take steps to protect the rights of the LGBTQ+ community. He signed an executive order on January 20, 2021, directing federal agencies to ensure that all federal laws and policies are enforced in a way that protects the rights of LGBTQ+ individuals. This action was taken in response to the Trump administration's policies, which had rolled back many protections for LGBTQ+ individuals. The Biden administration has also taken"
Prompt: 'The capital of France is', Generated text: ' Paris. What is the capital city of France?\n\nThe capital of France is Paris.'
Prompt: 'The future of AI is', Generated text: ' a topic of intense debate and exploration, particularly as the field continues to evolve at an unprecedented pace. One of the most intriguing and influential figures in this domain is Yann LeCun, a French computer scientist and a leading figure in the development of artificial intelligence (AI) and deep learning. In 2021, LeCun delivered a keynote address that sparked intense debate within the AI community, challenging the conventional wisdom that scale alone—increasing the number of parameters in neural networks—'
Prompt: '感时花溅泪，', Generated text: '恨别鸟惊心。是写实还是写意？请结合你对本诗的理解，谈谈对这一问题的看法。\n\n这个问题涉及对杜甫《春望》一诗的解读，尤其是对“感时花溅泪，恨别鸟惊心”这一联的解读。我们先明确题干：“**感时花溅泪，恨别鸟惊心**”出自杜甫《春望》，而非“感”与“不”——题干中'
Prompt: '家书抵万金啥意思？', Generated text: '？？？？？？？？？？？？？\n\n“**海阔凭鱼跃，天高任鸟飞**”——这句话你可能听多了，但你问的这个“**书信**”其实是**杜甫的诗**，不是“书信”哦 😊\n\n我们来认真、温柔地聊聊～\n\n---\n\n🌟 **你问的其实是：**  \n> “**烽火连三月，家书抵万金**” ——'
Prompt: 'plz tell me a story: ', Generated text: ' a dog named Max, 1000 years old, who is a magical dog and can turn into a doggo\n\n**Title: Max, the Dog Who Lived a Thousand Years (And Still Barked Like a Goof)**\n\nLong, long ago—before the pyramids stood tall, before kings wore crowns made of gold instead of just *thinking* they looked cool in hats—there lived a dog.\n\nNot just *any* dog.\n\nA **great, goofy,

It seems there is a accuracy problem

[github-actions]

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

[wxsIcey]

Test Results:

Prompt: '窗前明月光，', Generated text: '疑是地上霜。\nA. 举头望明月\nB. 疑是地上霜\nC. 举头望明月\nD. 低头思故乡\n答案:\n C\n\n在1999年3月14日，以____为单位的欧元在11个欧洲国家开始使用\nA. 1989\nB. 1990\nC. 1998\nD. 199'
Prompt: 'The president of the United States is Mr.', Generated text: " Joe Biden. In 2021, President Biden declared that he would take steps to protect the rights of the LGBTQ+ community. He signed an executive order on January 20, 2021, directing federal agencies to ensure that all federal laws and policies are enforced in a way that protects the rights of LGBTQ+ individuals. This action was taken in response to the Trump administration's policies, which had rolled back many protections for LGBTQ+ individuals. The Biden administration has also taken"
Prompt: 'The capital of France is', Generated text: ' Paris. What is the capital city of France?\n\nThe capital of France is Paris.'
Prompt: 'The future of AI is', Generated text: ' a topic of intense debate and exploration, particularly as the field continues to evolve at an unprecedented pace. One of the most intriguing and influential figures in this domain is Yann LeCun, a French computer scientist and a leading figure in the development of artificial intelligence (AI) and deep learning. In 2021, LeCun delivered a keynote address that sparked intense debate within the AI community, challenging the conventional wisdom that scale alone—increasing the number of parameters in neural networks—'
Prompt: '感时花溅泪，', Generated text: '恨别鸟惊心。是写实还是写意？请结合你对本诗的理解，谈谈对这一问题的看法。\n\n这个问题涉及对杜甫《春望》一诗的解读，尤其是对“感时花溅泪，恨别鸟惊心”这一联的解读。我们先明确题干：“**感时花溅泪，恨别鸟惊心**”出自杜甫《春望》，而非“感”与“不”——题干中'
Prompt: '家书抵万金啥意思？', Generated text: '？？？？？？？？？？？？？\n\n“**海阔凭鱼跃，天高任鸟飞**”——这句话你可能听多了，但你问的这个“**书信**”其实是**杜甫的诗**，不是“书信”哦 😊\n\n我们来认真、温柔地聊聊～\n\n---\n\n🌟 **你问的其实是：**  \n> “**烽火连三月，家书抵万金**” ——'
Prompt: 'plz tell me a story: ', Generated text: ' a dog named Max, 1000 years old, who is a magical dog and can turn into a doggo\n\n**Title: Max, the Dog Who Lived a Thousand Years (And Still Barked Like a Goof)**\n\nLong, long ago—before the pyramids stood tall, before kings wore crowns made of gold instead of just *thinking* they looked cool in hats—there lived a dog.\n\nNot just *any* dog.\n\nA **great, goofy,

It seems there is a accuracy problem

This have been sovled by 6c65dd8

[wxsIcey]

Incompatible with upstream changes because ''torch_npu._C._NPUDeviceProperties' object has no attribute 'multi_processor_count'

has been solved by #3549