4 ms·
Does it support flash attention? Use tensor cores? Can I write custom kernels? UPD. found no evidence that it supports tensor cores, so it's going to be many t
by lostmsu 11mo ago
Does it support flash attention? Use tensor cores? Can I write custom kernels?
UPD. found no evidence that it supports tensor cores, so it's going to be many times slower than implementations that do.
- mikepapadim 11mo agoYes, when you use the PTX backend it supports Tensor Cores.It has also implementation for flash attention. You can also write your own kernels, have a look here: https://github.com/beehive-lab/GPULlama3.java/blob/main/src/main/java/org/beehive/gpullama3/tornadovm/kernels/TransformerComputeKernels.java https://github.com/beehive-lab/GPULlama3.java/blob/main/src/... https://github.com/beehive-lab/GPULlama3.java/blob/main/src/main/java/org/beehive/gpullama3/tornadovm/kernels/TransformerComputeKernelsLayered.java https://github.com/beehive-lab/GPULlama3.java/blob/main/src/...
- lostmsu 11mo agoTornadoVM GitHub has no mentions of tensor cores or WMMA instructions. The only mention of tensor cores is in 2024 and states they are not used: https://github.com/beehive-lab/TornadoVM/discussions/393 https://github.com/beehive-lab/TornadoVM/discussions/393
- mikepapadim 11mo agohttps://github.com/beehive-lab/TornadoVM/pull/732 https://github.com/beehive-lab/TornadoVM/pull/732 https://github.com/beehive-lab/TornadoVM/pull/313 https://github.com/beehive-lab/TornadoVM/pull/313
- lostmsu 11mo agoI believe these are SIMD. Tensor cores require MMA family of instructions. Ask me how I know. :) https://github.com/m4rs-mt/ILGPU/compare/master...lostmsu:ILGPU:wgmma/make-descriptor https://github.com/m4rs-mt/ILGPU/compare/master...lostmsu:IL... Good article: https://alexarmbr.github.io/2024/08/10/How-To-Write-A-Fast-Matrix-Multiplication-From-Scratch-With-Tensor-Cores.html#tensor-core-vs-ffma https://alexarmbr.github.io/2024/08/10/How-To-Write-A-Fast-M...