Description
🚀 The feature, motivation and pitch
Hi PyTorch Team,
I’m currently working on projects on newer GPUs, such as the RTX 5090 that require CUDA 12.8. However, many of my dependencies and external packages are not compatible with the latest PyTorch releases (e.g., 2.7), and upgrading breaks a number of pipelines.
Unfortunately, earlier PyTorch versions like 2.3 do not provide wheels compatible with newer CUDA versions (e.g., CUDA 12.3 or 12.4). This causes difficulty for many users who:
• Depend on stable older PyTorch versions for reproducibility or compatibility reasons;
• Need to utilize newer GPUs (e.g., 5090) which require updated CUDA runtimes;
Are there any plans to provide updated builds of previous PyTorch versions (like 2.3 or 2.4) that support newer CUDA toolkits such as 12.8?
In my experiments, the same project shows large discrepancies in training performance between PyTorch 2.3 and PyTorch 2.7 — in some cases, accuracy drops dramatically.
Could you clarify what changes in the backend (e.g., kernels, precision behavior, compiler flags, etc.) might cause such differences?
Is there a recommended way to debug these regressions?
Alternatives
No response
Additional context
No response