vectorch-ai
is there any way or chance to use flash_attn 1.x ? to support more gpus
i think it depend on flash_attn 2.x,will ouput runtime error with unsupport gpus.
Yes, you are correct. Currently, we only support Ampere or newer architecture. The flash_attn1 implementation does indeed work for Turing GPUs. At the moment, we don't have plans to integrate flash_attn1, but we do intend to extend support to more GPUs. One of our ongoing efforts is to integrate flashinfer as an attention implementation, which supports Turing gpu and shows promising perf results. thanks, stay tuned.
Yes, you are correct. Currently, we only support Ampere or newer architecture. The flash_attn1 implementation does indeed work for Turing GPUs. At the moment, we don't have plans to integrate flash_attn1, but we do intend to extend support to more GPUs. One of our ongoing efforts is to integrate flashinfer as an attention implementation, which supports Turing gpu and shows promising perf results. thanks, stay tuned.
that would be great!thanks for your work😀