[Kunlunxin] Fix garbled output with CUDA Graph

[0songHan]

PR Category

PR Type

Description

Fix a garbled output issue when CUDA Graph is enabled in the Kunlunxin attention backend. The issue was caused by uninitialized values in the padded region of the attention output buffer.

Related Issues

None.

Changes

• Initialize the padding region of the attention output to zero when num_actual_tokens is smaller than the output buffer size.

Testing

• Tested using the Zhiyuan test dataset.
• Ran python run_eval_lms.py and verified that the generated results.json contains correct results.

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.