Hi authors, thanks for the great work!
I noticed in the README that MagiAttention is required for grouped causal attention, and according to its documentation, it only supports Hopper and Blackwell GPUs. This hardware requirement seems quite restrictive.
I’m wondering whether MagiAttention can be replaced by other libraries (e.g., FlashAttention or similar implementations). Even if this leads to some efficiency degradation, it would greatly improve usability for users without access to newer GPUs.
Moreover, I noticed that in the supplementary material, the experiments were conducted on A800 GPUs. I’m curious whether the authors have any workaround or internal solution that allows MagiAttention to run on A100/A800 GPUs.
Any clarification would be greatly appreciated. Thanks again for your excellent work!
Hi authors, thanks for the great work!
I noticed in the README that MagiAttention is required for grouped causal attention, and according to its documentation, it only supports Hopper and Blackwell GPUs. This hardware requirement seems quite restrictive.
I’m wondering whether MagiAttention can be replaced by other libraries (e.g., FlashAttention or similar implementations). Even if this leads to some efficiency degradation, it would greatly improve usability for users without access to newer GPUs.
Moreover, I noticed that in the supplementary material, the experiments were conducted on A800 GPUs. I’m curious whether the authors have any workaround or internal solution that allows MagiAttention to run on A100/A800 GPUs.
Any clarification would be greatly appreciated. Thanks again for your excellent work!