Inquiry Regarding Running RFdiffusion Colab Notebook (diffusion.ipynb) on ARM64 Architecture (Jupyter Notebook Attempt Failed)

[schoyeon]

Dear RFdiffusion / DGL Team,

I am writing to you with great interest in the RFdiffusion project. I've been attempting to run the code from your Colab Notebook (https://colab.research.google.com/github/sokrypton/ColabDesign/blob/v1.1.1/rf/examples/diffusion.ipynb#scrollTo=pZQnHLuDCsZm) on an ARM64 architecture environment.

Unfortunately, I have encountered difficulties and failed to execute the code using a Jupyter Notebook on my ARM64-based system. I was hoping you could provide some guidance. Is there a recommended way to get this code running on an ARM64 environment? I would be most grateful if you could share any specific installation instructions, dependency configurations, or recommended execution procedures for ARM64.

Furthermore, any insights into potential challenges or considerations I should be aware of when targeting this architecture would be immensely helpful.

Thank you for your time and assistance.

[rclune]

What errors are you running into?

If you have a google account you can run the notebook from your web browser via Google CoLab, which should avoid any issues with your computer architecture. The generated files will be compressed into a zip file that you can download once the kernel is done running.

[schoyeon]

What errors are you running into?

If you have a google account you can run the notebook from your web browser via Google CoLab, which should avoid any issues with your computer architecture. The generated files will be compressed into a zip file that you can download once the kernel is done running.

Hello @rclune,

Thank you for your quick response and for the suggestion about Google Colab.

However, my primary need is to execute this code on my local server, which runs on a linux-aarch64 (ARM64) architecture. This is crucial for my research as the server provides dedicated and powerful computational resources, and I aim to utilize its GPU capabilities for accelerating complex deep learning tasks like RFdiffusion.

Given this, I'm trying to find a way to adapt the code from the Colab notebook (https://colab.research.google.com/github/sokrypton/ColabDesign/blob/v1.1.1/rf/examples/diffusion.ipynb#scrollTo=tSgCPxIZ1T_A) to run effectively on my aarch64 server.

Could you please confirm if there's a recommended way to get this code running on an ARM64 environment, specifically utilizing GPU acceleration if available on ARM? I would be most grateful for any guidance on specific installation instructions, dependency configurations, or recommended execution procedures for ARM64 platforms.

Thank you for your time and assistance.

[rclune]

I would recommend using the Docker image for RFdiffusion or installing it on your system (see this installation guide) rather than using the notebook. The linked installation guide has instructions for ARM systems.

[schoyeon]

I would recommend using the Docker image for RFdiffusion or installing it on your system (see this installation guide) rather than using the notebook. The linked installation guide has instructions for ARM systems.

Hello @rclune,

Thank you for your further recommendations and for pointing to the Docker image and the installation guide. I truly appreciate your continued support.

I have thoroughly reviewed both suggestions:

1. RFdiffusion Docker Image (https://hub.docker.com/r/rosettacommons/rfdiffusion): I checked this Docker Hub page. It explicitly states: "Please note that this image is not compatible with systems using the ARM architecture." This confirms that the official pre-built image is unfortunately not suitable for my linux-aarch64 server.

2. Installation Guide (https://sites.google.com/omsf.io/rfdiffusion/getting-started/installation): I also reviewed the installation guide, specifically the "ARM-Based (Apple Silicon) Architectures" section. This section indeed addresses the ModuleNotFoundError: No module named 'torchdata.datapipes' problem. It suggests using pip install torchdata==0.9.0 to resolve it. However, upon attempting this, I found that torchdata==0.9.0 is not available for linux-aarch64 on PyPI (ERROR: No matching distribution found for torchdata==0.9.0). This blocks the installation following this guide. (We also consistently encountered PackagesNotFoundError for various dgl and torchaudio versions when trying conda install for aarch64.)

Summary: It appears that the primary challenge remains the fundamental unavailability or incompatibility of dgl and torchdata package builds for the linux-aarch64 architecture through standard package managers (Conda/Pip), which is essential for RFdiffusion.

Could you please confirm if there's any other verified working method or a specific aarch64 native Docker image / package combination that is known to reliably run RFdiffusion on linux-aarch64? My apologies for the persistence, but I'm trying to ensure RFdiffusion can run on this specific hardware.

Thank you again for your patience and dedicated assistance.

[rclune]

There is currently not a Rosetta Commons Docker image that supports this architecture nor are we aware of installation instructions that will work with linux-aarch64 systems. We are looking into how to update the dependencies to avoid this issue, but currently do not have a timeline for this change.

[schoyeon]

Thank you for your prompt and informative reply.

I understand the current technical limitations regarding support for the linux-aarch64 architecture. While it's unfortunate to hear there isn't a timeline for a fix, I appreciate you clarifying the situation.

Thank you again for your time and effort.