llama.cpp always produces malformed output on Loongarch starting from b6353

[darkgeek]

Hi Loong64 community,

I've opened an issue in the llama.cpp repo reporting that the application produces garbage output regardless of the model used.

We've found that disabling Flash Attention (-fa off) resolves the issue. However, this problem appears specific to the Loong64 architecture now, as we haven't observed a similar problem on other architectures like ARM64.

Could anyone familiar with Loong64 and the llama.cpp build/implementation please help investigate this potential Flash Attention bug? Thanks!

[MQ-mengqing]

ggml-org/llama.cpp#16958