[bug]: Clearing queues in mid generation does not free VRAM, even when configured to do so #6759

Is there an existing issue for this problem?

• I have searched the existing issues

Operating system

Linux

GPU vendor

Nvidia (CUDA)

GPU model

RTX 4000 ADA SFF

GPU VRAM

20GB

Version number

4.2.8rc2

Browser

Brave v1.68.141

Python dependencies

No response

What happened

Might be related to #6613

Here, Invoke is configured with "vram: 0.1" in invokeai.yaml in order to always free up VRAM after generation.

Clicking "Cancel and Clear All Items" during image generation does however not free VRAM: It just sticks at what it used to need during image generation (e.g. 7 GB)

However, waiting until the queue finished clears VRAM as expected: Not entirely though - might be a different bug - but at least down to 1.8 GB on my system

What you expected to happen

No matter if a queue finished or has been aborted via the "Cancel and Clear All Items" button, VRAM should be freed.

How to reproduce the problem

Normal working operation (yet a bit too much VRAM usage)

1. Configure InvokeAI with 'vram: 0.1' which is supposed to free VRAM > 100 MB after any generation
2. Generate any image and observe VRAM via nvidia-smi --loop-ms=1000. It will rise 7 GB on my system (SDXL)
   3.Once everything in the queue has been generated, VRAM drops down significantly. On my system down to 1.8 GB. Still way more than expected but "quiet low" though

Trigger the bug (excessive VRAM usage)

1. Configure InvokeAI with 'vram: 0.1' which is supposed to free VRAM > 100 MB after any generation
2. Generate any image and observe VRAM via nvidia-smi --loop-ms=1000. It will rise 7 GB on my system (SDXL)
   3.Abort the generation by clicking on the 'Cancel and Clear All Items'. VRAM usage stays at 7 GB

Additional context

No response

Discord username

systemofapwne