Add GPU support for stem separation

[tucktuckg00se]

Description

Motivation

Stem separation is currently compute-heavy, especially on older PCs and on larger files. CPU export works, but performance can be slow enough to
impact usability. We already expose a device choice in the UI and have ONNX Runtime in place, so the next step is making GPU execution a first-
class supported path.

Goals

• Support GPU-backed stem separation where ONNX Runtime execution providers are available
• Keep CPU export as the fallback on every platform
• Make failures explicit and predictable instead of silently dropping to an unexpected path
• Preserve existing hosted-plugin stability improvements, especially around runtime loading on Windows

Proposed Scope

• Add ONNX Runtime GPU execution provider support to the build/package flow
• Detect available providers at runtime and expose supported choices in the STEM panel
• Initialize the ONNX session with the selected provider when available
• Show clear warnings/errors when the requested GPU path is unavailable or misconfigured
• Document platform-specific requirements for GPU inference

Platform Notes

Windows

• Likely first target
• Needs careful packaging/runtime loading so the plugin uses the intended ONNX Runtime and provider DLLs instead of conflicting DAW-loaded copies
• CUDA/DirectML support should be evaluated explicitly rather than assumed

macOS

• Investigate CoreML support and practical compatibility/performance tradeoffs
• Confirm behavior in AU/VST3 hosts, not just standalone

Linux

• Evaluate CUDA provider support and deployment expectations
• Keep CPU fallback straightforward for systems without compatible drivers/runtime

Acceptance Criteria

• A supported GPU-capable machine can complete stem separation using GPU instead of CPU
• The selected device is reflected accurately in runtime behavior
• Unsupported or broken GPU setups fail with a clear message and allow switching back to CPU
• Hosted plugin builds remain stable in major DAWs
• README/release notes document GPU requirements and fallback behavior

Non-Goals

• Removing CPU support
• Requiring GPU hardware for stem separation
• Shipping a GPU path that silently degrades or is hard to diagnose

Use Case

Stem separation currently only uses CPU. GPU acceleration would be much faster.

Alternatives Considered

No response

[simbio-sounds]

I can't believe the rate of development you are going with this, its incredible!! Keep up the good work

[tucktuckg00se]

An experimental pre-release with GPU stem separation is available for testing: https://github.com/tucktuckg00se/INTERSECT/releases/tag/v0.14.0-rc.1

If you try it out please reply with:

• OS and version
• GPU vendor / model
• Which ONNX Runtime bundle you installed (from SET > Stem Separation > ONNX Runtime)
• Whether stem export ran successfully (and which model you used)
• If it failed, the message from the header status bar (click to copy)

Thanks!