Performance Regression & OOM in v0.6.0 compared to v0.5.1 (RTX 4080)

### Custom Node Testing

- [ ] I have tried disabling custom nodes and the issue persists (see [how to disable custom nodes](https://docs.comfy.org/troubleshooting/custom-node-issues#step-1%3A-test-with-all-custom-nodes-disabled) if you need help)

### Expected Behavior

Describe the bug
I am experiencing significant performance degradation and Out of Memory (OOM) issues after updating from v0.5.1 to v0.6.0. With the exact same workflow and settings, the behavior has changed drastically.
Comparison:
v0.5.1: Stable, no OOM. Performance was acceptable (approx. 700s for 10 images, ~70s/img).
v0.6.0: Frequent OOM. Even when it runs, it is much slower (approx. 400s for only 2 images, ~200s/img).
Environment:
GPU: NVIDIA RTX 4080 (16GB VRAM)
RAM: 128GB
Model: QwenEdit 2511 BF16
Weight Type: Loaded using fp8_e4m3 (f8_em4)
Workflow: Identical across both versions.
Additional Context:
I also noticed the following warning in the console which might be related to node logic or memory management:
WARNING: ProcessString.IS_CHANGED() got an unexpected keyword argument 'input_string'

### Actual Behavior

After upgrading from v0.5.1 to v0.6.0, I encountered a significant performance regression and frequent Out of Memory (OOM) issues using the same workflow and hardware (RTX 4080 16GB, 128GB RAM).
Comparison:
v0.5.1: Stable and efficient. 10 images were completed in about 700 seconds (~70s/img).
v0.6.0: Frequent OOM errors. When it does run, it is extremely slow, taking 400 seconds for just 2 images (~200s/img).
Technical Details:
Model: QwenEdit 2511 BF16
Weight type: fp8_e4m3
Console Warning: WARNING: ProcessString.IS_CHANGED() got an unexpected keyword argument 'input_string'
It seems like memory management or VRAM offloading logic has changed significantly in the new version, leading to much higher resource consumption.

### Steps to Reproduce

Update the environment to version 0.6.0.
Load a workflow using the QwenEdit 2511 BF16 model.
Set the model weight type to fp8_e4m3.
Run a batch generation (e.g., 2 or more images).
Monitor VRAM usage and generation time. Compare the results with version 0.5.1.
Check the console for the ProcessString.IS_CHANGED warning and observe the OOM crash during the sampling process.

### Debug Logs

```powershell
WARNING: ProcessString.IS_CHANGED() got an unexpected keyword argument 'input_string'
HTTP Request: POST https:///v1/chat/completions "HTTP/1.1 200 OK"
Requested to load WanVAE
Unloaded partially: 621.46 MB freed, 10947.53 MB remains loaded, 252.01 MB buffer reserved, lowvram patches: 596
loaded completely; 291.73 MB usable, 242.03 MB loaded, full load: True
Requested to load QwenImageTEModel_
Unloaded partially: 6626.12 MB freed, 4321.41 MB remains loaded, 378.09 MB buffer reserved, lowvram patches: 760
loaded completely; 8875.12 MB usable, 7910.29 MB loaded, full load: True
Unloaded partially: 1674.54 MB freed, 2646.86 MB remains loaded, 378.09 MB buffer reserved, lowvram patches: 791
loaded partially; 11821.09 MB usable, 11568.99 MB loaded, 7914.96 MB offloaded, 252.01 MB buffer reserved, lowvram patches: 0
100%|████████████████████████████████████████████████████████████████████████████████████| 8/8 [03:39<00:00, 27.47s/it]
loaded partially; 11821.09 MB usable, 11568.99 MB loaded, 7914.96 MB offloaded, 252.01 MB buffer reserved, lowvram patches: 0
12%|██████████▍ | 1/8 [02:19<16:19, 139.95s/it]
```

### Other

_No response_

Start by reproducing the v0.5.1 versus v0.6.0 comparison with the QwenEdit 2511 BF16 workflow, fp8_e4m3 weights, and an RTX 4080. Check the ProcessString.IS_CHANGED warning alongside the VRAM and timing logs; done means identifying the regression and confirming that the same batch no longer OOMs or suffers the reported slowdown.

Written by the indexing model from the issue text.