fix dllm long-context

[grimoire]

Thanks for your contribution and we appreciate it a lot. The following instructions would make your pull request more healthy and more easily receiving feedbacks. If you do not understand some items, don't worry, just make the pull request and seek help from maintainers.

Motivation

Please describe the motivation of this PR and the goal you want to achieve through this PR.

Modification

Please briefly describe what modification is made in this PR.

BC-breaking (Optional)

Does the modification introduce changes that break the backward-compatibility of the downstream repositories?
If so, please describe how it breaks the compatibility and how the downstream projects should modify their code to keep compatibility with this PR.

Use cases (Optional)

If this PR introduces a new feature, it is better to list some use cases here, and update the documentation.

Checklist

1. Pre-commit or other linting tools are used to fix the potential lint issues.
2. The modification is covered by complete unit tests. If not, please add more unit tests to ensure the correctness.
3. If the modification has a dependency on downstream projects of a newer version, this PR should be tested with all supported versions of downstream projects.
4. The documentation has been modified accordingly, like docstring or example tutorials.

[lvhan028]

I conducted aime2025 evalution with opencompass. The generation of many prompts cannot be finished by 'stop' but by 'length'. There are lots of repetition patterns in the generated text.

Here is the minimum reproducible path:

from lmdeploy import pipeline, PytorchEngineConfig, GenerationConfig


if __name__ == '__main__':
    prompt = 'The 9 members of a baseball team went to an ice cream parlor after their game. Each player had a singlescoop cone of chocolate, vanilla, or strawberry ice cream. At least one player chose each flavor, and the number of players who chose chocolate was greater than the number of players who chose vanilla, which was greater than the number of players who chose strawberry. Let $N$ be the number of different assignments of flavors to players that meet these conditions. Find the remainder when $N$ is divided by 1000.\nRemember to put your final answer within \\boxed{}.'

    prompt = [
        dict(role='user', content=prompt)
    ]

    with pipeline('JetLM/SDAR-30B-A3B-Chat', backend_config=PytorchEngineConfig(tp=2)) as pipe:
        response = pipe(prompt, gen_config=GenerationConfig(do_sample=True, temperature=0.6, max_new_tokens=16384))
        print(response)

B.T.W, if do_sample is turned off, the prediction works well.

[grimoire]

@lvhan028 Fixed, torch topk is not stable, unmasking in different rank would leads to different branch.

[RunningLeon]

LGTM