[Kunlunxin] Fix duplicate generation

[0songHan]

PR Category

Vendor

PR Type

Bug Fixes

Description

Fix duplicate / garbled generation during multi-card inference on Kunlunxin.

The issue was caused by TP-inconsistent random sampling, which could lead to different tensor-parallel ranks selecting different next tokens and eventually producing duplicated or garbled outputs. This PR ensures TP-consistent random sampling on Kunlunxin.

Related Issues

N/A

Changes

• Fix TP-consistent random sampling on Kunlunxin.
• Resolve duplicate / garbled generation during multi-card inference.

Testing

Validated on the following models with multi-card inference:

• Qwen3.6-27B
• Qwen3.5-35B-A3B

Checklist

• I have run the existing tests and they pass
• I have added tests for my changes (if applicable)
• I have updated the documentation (if applicable)

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.

---

0songHan seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}