FlashAttention requires compute capability 8.0 or higher.

[CaspervanAarle]

I am trying to offload my compute to a gRPC server on my newly setup server. Now it seems that my RTX 2060 card should be supported, however i cannot overcome the issue that arises in the dockerized version.

FATAL: FlashAttention requires building with sm version sm80-sm90, but was built for < 8.0

From what i have learned, the 2060 has Compute Capability of 7.5, but the error indicates it tries to apply flash attention which requires 8.0 or higher. The readme indicates it supports 2060 but i have not found a way to make this work.

[liuliu]

You need to add the flag --no-flash-attention when launching it with 20xx card to disable flash attention. Afterwards, it will behave as expected.

[CaspervanAarle]

Thank you very much, I see it was my mistake.

[TokenChingy]

Hi, this doesn't seem to work for me when launching on WSL Ubuntu with a 2060 Super. Adding --no-flash-attention to the docker image as per the README file still threw this error.

Specifically loading SkyReels v1 Hunyuan I2V 544p (5-bit, SVDQuant).