Post
2547
More context for your Pascal GPU or older!
These are great news for all the users with GTX 10XX, P40...
Flash Attention implementation for older NVIDIA GPUs without requiring Tensor Cores has come to llama.cpp in the last few days, and should be merged in the next version of KoboldCpp, you can already try it with another fork or by building it.
[Mentioned KCPP fork] https://github.com/Nexesenex/kobold.cpp/releases/latest
[PR] https://github.com/ggerganov/llama.cpp/pull/7188
You should expect less VRAM usage for the same context, allowing you to experience higher contexts with your current GPU.
There have also been reported final tokens/second speed improvements for inference, so that's also grand!
If you have tried it, I'd like to hear your experiences with --flashattention so far, especially for this implementation and for the large number of Pascal (GTX 10XX, P40...) cards.
Discussion linked bellow, with more links to relevant information:
LWDCLS/LLM-Discussions#11
Cheers!
These are great news for all the users with GTX 10XX, P40...
Flash Attention implementation for older NVIDIA GPUs without requiring Tensor Cores has come to llama.cpp in the last few days, and should be merged in the next version of KoboldCpp, you can already try it with another fork or by building it.
[Mentioned KCPP fork] https://github.com/Nexesenex/kobold.cpp/releases/latest
[PR] https://github.com/ggerganov/llama.cpp/pull/7188
You should expect less VRAM usage for the same context, allowing you to experience higher contexts with your current GPU.
There have also been reported final tokens/second speed improvements for inference, so that's also grand!
If you have tried it, I'd like to hear your experiences with --flashattention so far, especially for this implementation and for the large number of Pascal (GTX 10XX, P40...) cards.
Discussion linked bellow, with more links to relevant information:
LWDCLS/LLM-Discussions#11
Cheers!