Hi Christian, On 2023-10-17 23:10, Christian Kastner wrote:
About a year ago, you shared this graph:
https://lists.debian.org/debian-ai/2022/09/msg00029.html
It would be very helpful if you could update this graph for 5.7, so that
we get a current picture of what's missing breadth-wise.
I've attached an updated chart. I marked some optional dependencies with a dotted line, but I also omitted some optional dependencies to simplify the diagram. Libraries that have been packaged are marked with a check, and I've numbered the next four libraries that I suggest we package. I don't think we're going to see GPU-accelerated PyTorch in Ubuntu 24.04, but I've made my peace with that.
Other libraries and tools in ROCm 5.7 that have not been packaged for Debian and that are also not on the attached chart include:
- rocWMMA: helper library for using matrix multiply-accumulate instructions on CDNA 1/2/3 and RDNA 3 GPUs
- hipTensor: tensor primitives library - MIVisionX: computer vision toolkit - rpp: computer vision performance primitives - MIGraphX: graph optimization library - rocm_bandwidth_test: system validation tool - amdsmi: extension of rocm-smi to EPYC CPUs - rocr_debug_agent: provider of additional debugging information - ROCmValidationSuite: system validation tool - hipfort: Fortran bindings for the ROCm librariesHowever, not all of those tools and libraries are mature enough to be packaged for Debian.
Sincerely, Cory BloorP.S. There is now a hipblas backend to llama.cpp [1]. That's the second application I know of that the Debian ROCm packages are useful for (the first being Blender). We should probably package the Tensile benchmarking tool so that it's easier to tune GEMMs on Debian. The performance of rocBLAS can vary significantly depending on whether the problem size has been tuned or not [2].
[1]: https://github.com/ggerganov/llama.cpp/pull/1087 [2]: https://old.reddit.com/r/ROCm/comments/17m2qo0/how_to_tune_an_mi100/
Attachment:
pytorch-2023.png
Description: PNG image
