Performance regression after recent updates: DynamicVRAM / AIMDO causes extreme slowdowns when models are on HDD + feature request for “prefer GPU / disable dynamic memory” option

### Feature Idea

Hello,
About two months ago ComfyUI worked perfectly on my setup. After the latest updates, generation became extremely slow: a simple test image now takes ~15 minutes instead of ~10 seconds. The GPU (RTX 4090 24 GB) shows very low utilization, even though it is correctly detected as the primary device.
Hardware:

GPU: NVIDIA RTX 4090 24 GB
CPU: Intel Xeon E5-2690 v2 (AVX1)
RAM: 64 GB DDR3
Models storage: 4 TB HDD (models occupy ~1 TB)
System drive: 1 TB SATA SSD
OS: Windows 11 25H2

Previously the same models and workflows (including animation) ran instantly. Now the system spends a lot of time reading from the slow HDD and using system RAM instead of keeping models in VRAM.
There are multiple similar reports in the community about models being reloaded from disk on every generation, low GPU usage, and large performance drops after the introduction/changes related to DynamicVRAM / comfy-aimdo. I have 24 GB of VRAM, which is more than enough for my needs, so aggressive dynamic offloading to RAM/disk only hurts performance because of the slow HDD.
Request:

Please add a clear option (UI checkbox or easily discoverable launch flag) such as “Prefer GPU / Keep models in VRAM / Disable DynamicVRAM” or “Use GPU by default”.
Allow users with sufficient VRAM to easily disable the new dynamic memory management and restore the previous behavior (models stay resident in VRAM).

I have tried various launch arguments, but there is currently no simple and reliable toggle in the interface. I would greatly appreciate it if such an option could be added.
Thank you!

### Existing Solutions

_No response_

### Other

_No response_