Possible VRAM regression when running a MiniMax H3 workflow on the latest development version

### 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

The MiniMax H3 workflow should complete successfully with similar VRAM usage to the previous ComfyUI version when using the same settings.

### Actual Behavior

After updating to the latest development version, system RAM usage becomes much lower, but VRAM usage increases significantly. The MiniMax H3 workflow now reaches the VRAM limit much more easily and may fail with a CUDA out-of-memory error.

### Steps to Reproduce

1. Update ComfyUI to the latest development version.
2. Start ComfyUI on Windows with the usual launch arguments.
3. Load the MiniMax H3 workflow.
4. Use the same model, resolution, batch size, precision, and other settings as before the update.
5. Queue the workflow and monitor system RAM and VRAM usage.
6. Compared with the previous ComfyUI version, system RAM usage is significantly lower, but VRAM usage increases much more easily.
7. The workflow may reach the VRAM limit and fail with a CUDA out-of-memory error.

### Debug Logs

```powershell
[2026-09-06 22:32:55.092] got prompt
[2026-09-06 22:32:55.521] Model MiniMaxH3VideoVAE prepared for dynamic VRAM loading. 4965MB Staged.
[2026-09-06 22:32:59.438] Model MiniMaxH3TEModel_ prepared for dynamic VRAM loading. 25882MB Staged.
[2026-09-06 22:33:17.187] Model MiniMaxH3 prepared for dynamic VRAM loading. 63208MB Staged.
[2026-09-06 22:39:50.373] Interrupting prompt
[2026-09-06 22:39:50.920] Comfy model compiler graph breaks: 2
[2026-09-06 22:39:51.673] Processing interrupted
[2026-09-06 22:39:51.794] Prompt executed in 416.67 seconds

[2026-09-06 22:49:59.089] got prompt
[2026-09-06 22:50:20.955] Model MiniMaxH3 prepared for dynamic VRAM loading. 63208MB Staged.
[2026-09-06 22:57:06.352] Interrupting prompt
[2026-09-06 22:57:06.352] Comfy model compiler graph breaks: 2
[2026-09-06 22:57:07.956] Processing interrupted
[2026-09-06 22:57:08.046] Prompt executed in 428.92 seconds

No explicit CUDA out-of-memory traceback was recorded in this log, but the workflow repeatedly reached a state where VRAM usage became too high and the generation had to be interrupted.
```

### Other

Environment:

- OS: Windows
- GPU: NVIDIA GeForce RTX 5090 D
- VRAM: 32 GB
- System RAM: 128 GB
- Python: 3.13.14

The issue occurs with the same MiniMax H3 workflow and settings that were used successfully before updating ComfyUI.

No source file or test is identified in the report. Start by reproducing the MiniMax H3 workflow on the current and previous ComfyUI versions with identical settings while monitoring system RAM and VRAM; done means the regression is isolated and the workflow completes without the unexplained VRAM increase.

Written by the indexing model from the issue text.