Compute Units outage mistaken for memories outage

### Expected Behavior

Surpasses into UniVRAM’s, CPU memories, and even swaps when Compute Units are entirely in use

### Actual Behavior

Since i’ve to merge different SDiff 1.5 checkpoints in 0.31x doing so entirely backfires itself unable to surpass memories into the actual UniVRAM’s, CPU memories, and swaps when Compute Units are entirely in use
Due to software limits, i can only use the first 7.45GiB Compute Units as the only Comfy memories, no matter i also have 14.9GiB UniVRAM’s, 46.9GiB CPU memories, and 93GiB swaps actually it’s entirely unusable just due to software bottlenecks somewhy

### Steps to Reproduce

(1) Make a workflow generating wallpapers by different SDiff 1.5 checkpoints
(2) Merge those checkpoints into a new training result checkpoint
(3) See if your Compute Units are entirely in use but unable to surpass into the actual memories

### Debug Logs

```
HIP error: out of memory
HIP kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect.
For debugging consider passing AMD_SERIALIZE_KERNEL=3
Compile with `TORCH_USE_HIP_DSA` to enable device-side assertions.
```

### Other
```
HW: ASRock AMD X300 DeskMini × AMD R4750G APU × 16GB UniVRAM’s
Total memories: 7.45GiB Compute Units (Only this slice is usable on Comfy)
> 14.9GiB UniVRAM’s (This is the actual VRAM’s in DDR4 HW)
> 46.9GiB CPU memories
> 93GiB swaps
OS: Artix AMD64 / Liquorix kernel AMD64 / BtrFS root / ZFS working dirs
Borwser app: Brave nightly
Image results are ok, everything is wrong after image results
```