The outdated PyTorch version that was automatically installed for the first time on the Mac version of Desktop resulted in some functions of the k-sampler not supporting MPS.

[Nuo-Mituan]

When I was running the workflow to the k-sampler through Desktop, I found an error in aten:: nlinalg.solve-ex.result. After checking, I found out that my Mac version of Desktop automatically installed pytorch version 2.5.1 for the first time, but now the latest stable version of pytorch is 2.9.1. After trying to update it, the error in aten:: nlinalg.solve-ex.result no longer exists
Why can't the Desktop version automatically install the latest stable version of Pytorch when it first starts

┆Issue is synchronized with this Notion page by Unito

[dosubot]

Hi @Nuo-Mituan! I'm Dosu and I’m helping the desktop team.

The Desktop app for ComfyUI on Mac installs PyTorch version 2.5.1 by default because it uses a pinned requirements file (macos.compiled) that explicitly specifies torch==2.5.1 and related packages. This pinning is done to ensure stability and compatibility across platforms, especially given the complexity of PyTorch and MPS support on Mac. The Desktop app only updates dependencies like PyTorch when a new desktop release is published, not automatically on first launch. If you need the latest PyTorch version, you can manually update it in the .venv environment after installation, or use the portable/manual installation method for more control over dependencies. This approach is common for desktop apps that bundle complex dependencies, and the process may change as the app matures.
Source: macos.compiled requirements file
Source: PR #1488
Source: VirtualEnvironment installer logic

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}

   

[Nuo-Mituan]

I did experience a situation where Desktop couldn't open properly after updating PyTorch through (.venv), but after a while, Desktop automatically fixed it and installed the corresponding version of dependencies based on the updated PyTorch
The PyTorch team has been working on adapting MPS
Keeping the pytorch version updated on Desktop may be a good way to reduce some users giving up using Desktop due to version incompatibility