[Kunlunxin] Fix duplicate generation

[0songHan]

PR Category

Vendor

PR Type

Bug Fixes

Description

This PR fixes the duplicate generation issue on Kunlunxin by patching the sampler RNG implementation to ensure consistent sampling behavior during inference.

Related Issues

N/A

Changes

• Patch the sampler RNG implementation for Kunlunxin.
• Fix the duplicate generation issue during model inference.
• Keep the implementation transparent to existing inference workflows.

Testing

Validated on Kunlunxin hardware with:

• Qwen3.5-35B-A3B
• Qwen3.6-27B

Verified that the duplicate generation issue is resolved on both models.

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]

All committers have signed the CLA.

[0songHan]

The Lint Check is failing due to an existing lint issue intests/e2e_tests/serving/test_bge_m3.py, which is not modified by this PR. Could you please confirm whether this PR can be merged despite the unrelated lint failure?