[enhancement]: Multi-threaded upscaling in the unified canvas / batch starts

[ghost]

Is there an existing issue for this?

• I have searched the existing issues

Contact Details

What should this feature add?

Hi there,

I run invokeai on an ubuntu-linux with an AMD 5600x 6 core cpu, 32 gig internal memory and a Radeon 7900XT with 20 gigs of VRAM. I love the unified canvas, but using it in SDXL with the auto upscaling enabled when the bounding box is lower than the optimized model size was killing my cpu/gpu.

When I look in top i see mostly a single cpu core being maxed out when it's upscaling, this slows down m y system and makes the feature basically unusable. I

I would like to have, if possible, better multi-threaded upscaling.

Alternatives

thought there also was a GPU problem, so I uninstalled pytorch 5.6 and force installed the version for rocm6.0 (from AMD). The last action greatly improved my rendenering and invokeai seems to run ok on it (I do get version errors, but haven not encountered any issues).

I do think switching to ROCm 6.0 greatly improved the situation. I shall install two versions of invokeai and test it. But the CPU-usage is crazy high.

Additional Content

No response