Hi author, thank you for sharing your excellent work!
I have a question about the FLOPs counting.
The code writes that you count the FLOPs with the repo mamba-minimal , but in mamba's issue#110 the author of mamba claimed that the actual implementation needs much more FLOPs than simply scan in a for loop.
Can you tell me what approach you have adopted in calculating the FLOPs?
Also, what about the throughput of Simba?
Hi author, thank you for sharing your excellent work!
I have a question about the FLOPs counting.
The code writes that you count the FLOPs with the repo mamba-minimal , but in mamba's issue#110 the author of mamba claimed that the actual implementation needs much more FLOPs than simply scan in a for loop.
Can you tell me what approach you have adopted in calculating the FLOPs? Also, what about the throughput of Simba?
Thank you.