Are you interested in doing a PyTorch version?

[BaiLiping]

Hi Sionna team. Are you guys interested in making a PyTorch version? I happen to have some time on my hand, so if that is something you are interested in, I can make arrangement for that project.

[SebastianCa]

Hi @BaiLiping,

I’ve moved this thread to Discussions as it is not related to an issue with the Sionna code.

[perceptiveshawty]

this project being built on tensorflow is genuinely devastating

[BaiLiping]

I think we can probably organize some sort of working group to just convert their core code into PyTorch. With the help of Sionna team, things would be easier.

[Nicolas-Schlegel]

I feel the same way. Maybe I don't have enough experience with tensorflow, but trying to do anything that is bit more off the beaten path seems really difficult with it.

[jhoydis]

Hi,

We appreciate your interest in potentially porting Sionna to PyTorch. After careful consideration and extensive research, we've decided to maintain our focus on TensorFlow for the time being. Here’s a detailed explanation of our reasoning:

1. Performance Considerations
   In the current release, torch.compile faces challenges with complex-valued tensors, as evidenced by ongoing issues in the PyTorch repository (e.g., this issue). As a result, the PyTorch version would need to rely exclusively on eager mode, which has the drawback that large simulations run much slower compared to TensorFlow with XLA and consume a prohibitive amount of memory.
   Our benchmarking tests reveal significant performance gaps: For some multiuser MIMO simulations over fading channels, TensorFlow with XLA outperforms PyTorch in eager mode by a factor of 1.5. In a simple end-to-end constellation shaping training over an AWGN channel, TensorFlow with XLA demonstrates a 4x speed advantage.

2. Feature Parity
   We believe that TensorFlow currently provides all the necessary tools and capabilities for Sionna's development. While we remain open to new possibilities, we haven't identified any crucial features exclusive to PyTorch that would justify a transition at this time.

3. Resource Allocation
   Porting Sionna to PyTorch would require a substantial investment of time and resources. Currently, our development roadmap includes several high-priority features and improvements that we believe will bring more immediate value.
   Transitioning to PyTorch would also render Sionna incompatible with TensorFlow. This shift would place a significant burden on our current users, many of whom have developed extensive codebases around Sionna.

Looking ahead, we continue to monitor developments in the PyTorch ecosystem and remain open to reevaluating this decision in the future. Our primary goal is to ensure that Sionna remains a high-performance, feature-rich tool for our community.

We appreciate your interest and welcome any further questions or insights you may have regarding this discussion.