Offer official Pytorch Vulkan backend on pytorch.org

### 🚀 The Feature Pytorch can be compiled to use a Vulkan backend, I believe the use case is for Android. This thread requests/discusses official support for a Vulkan backend in Pytorch. The target audience for this is essentially all hardware that supports Vulkan. <img width="1001" height="403" alt="Image" src="https://github.com/user-attachments/assets/40161f69-aef7-4fdc-ac4f-a9903ff7636a" /> ### 🚀 Use Cases The following hardware is either unsupported or poorly supported by their hardware vendors in regards to ML APIs - AMD's "AI Max" APUs - Can have 100gb+ video RAM - ~Currently no support for ROCm and cannot practically be used with common AI tools~ - ROCm 7.0/7.1 underperforms compared to Vulkan [[1]](https://github.com/ggml-org/llama.cpp/discussions/10879) [[2]](https://www.phoronix.com/review/amd-rocm-7-strix-halo/4) - AMD appear to be dragging their feet on tangible improvements to ROCm - RDNA4 got MVP support for ROCm 6 months after launch and consumers are unsure when it will be practically usable (OOM issues, segfaults, crashes, etc) - Extending support to AMD cards with poor/no ROCm - ~Currently RDNA4 cards have no (practical) support for ROCm and cannot be used with common AI tools~ - ROCm 7.0 has started to bring support for RDNA4 however it's plagued with crashes and underperforms compared to Vulkan - ~Running AI workloads on Windows for AMD hardware~ - Better/less error prone support for AMD hardware on Windows - Intel GPUs - Qu

Posted by alshdavid on Aug 8, 2025, 11:19 PM

Tags

• support (Software)

Related issues

• Add 65536 size support for maps
• TypeScript 7.1 Iteration Plan
• nvim-treeconfig: upstream nvim-treesitter
• Add the "action" and "method" attributes to <button>
• [FEATURE] Support multiple Connector accounts (same connector, different accounts) in Claude and Claude Code on the web (claude.ai/code)
• > We were imposed GitHub for so many exercises!