RAPIDS support discontinued for systems running OS versions lower than Ubuntu 20.04 or Rocky Linux 8

EliHei2 / segger_dev

a cutting-edge cell segmentation model specifically designed for single-molecule resolved spatial omics datasets. It addresses the challenge of accurately segmenting individual cells in complex imaging datasets, leveraging a unique approach based on graph neural networks (GNNs).

https://elihei2.github.io/segger_dev/

MIT License

28 stars 1 forks source link

RAPIDS support discontinued for systems running OS versions lower than Ubuntu 20.04 or Rocky Linux 8 #14

Open quail768 opened 6 days ago

quail768 commented 6 days ago

Hi, Cool tool! I'd like to try this with CUDA12 and RAPIDS support. However it appears that that the cudf-cu12 version which is required by segger which is 24.8 can't be installed on OSs with Ubuntu 18.04 or below with glibc 2.17 as stated here:https://github.com/rapidsai/cudf/issues/16804#issuecomment-2351092184

I tried installing a local glibc v2.31 but it doesn't seem to fix this issue

I work on a computing cluster with REHL Fedora where the default version OS version is glibc 2.17. So I can't upgrade the OS unfortunately.

The main installation works fine. However this seems to be a little complicated. Any thoughts on potential work-arounds or would you suggest to go ahead without CUDA and RAPIDS support?

EliHei2 commented 6 days ago

hey @quail768 thanks for posting the issue, we're aware of this issue as also mentioned in #9 and are working on it to get a stable installation of rapids, for the current release, rapids is not necessary to get through the pipeline. For the next releases that would highly depend on RAPIDS, we make sure to provide proper installation instructions and/or a docker container. pining @andrewmoorman to be aware.

Anyways, you might have success by pip install cu{df/vs/graph/ml}-cu12==24.4.* --no-cache-dir

quail768 commented 2 days ago

Gotcha thanks @EliHei2. I was able to use a container and the installation for RAPIDS works fine now!

I also had a few questions about the cell type abundance embedding. Is that required or is it optional? In your experience, how much does adding the cell type abundance embedding affect the final output?

For the step where you annotate the cells using a scRNAseq reference, in the benchmarking analysis. Is it okay to use an integrated scRNA-seq dataset with cells from different samples as a reference or does that step require raw counts from a single scRNA-seq run from the same sample?