masahi opened a new pull request, #15423: URL: https://github.com/apache/tvm/pull/15423
[Flash attention v2](https://github.com/Dao-AILab/flash-attention) was released last week and it is claimed to give good speed up compared to existing implementations. I integrated flash v2 into our Relax CUTLASS BYOC and got 2-3 img / sec speed up on SD v1.5 compared to our existing flow that uses the xformer kernel. The original code depends on libtorch so it is hard to integrate. A version without torch dep was created in https://github.com/tlc-pack/libflash_attn and I'm adding it as a submodule. @tqchen @vinx13 @yzh119 @sunggg -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
