Out of device memory when running on BMG 24GB

[piotrgrubicki]

1. Set a folder with images as source.
2. Create one prompt (shown below)
3. Click on other images to see the annotations.
4. Change project.
5. After a while return to this project.
   The inference pane stops responding. UR_RESULT_ERROR_OUT_OF_DEVICE_MEMORY in the logs (attached).
   Restarting the app = the same effect.

oom.log

As expected, I am getting the same when running the Win app on ARL with 16 GB.
If we cannot overcome the memory issue, we need to communicate to the user minimum HW requirements.

[mpryahin]

A likely root cause: SAM's decoder allocates mask tensors at the original input resolution (here), with h, w sourced from the input image.

At 4K (3840×2160) with 3 categories × 40 masks, this single float32 tensor is ~4 GB can easily result in OOM.

We need to cap input resolution before inference and upscale masks/boxes back to original coordinates afterward.

@rajeshgangireddy Could you please take a look?