Minimax-h3 REF2VA. Low VRAM usage and much pinning on 5090. VRAM screenshots included. Is this correct?

### Custom Node Testing

- [x] 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)

### Your question

Running Minimax-h3 ref2va int8convrot , I seem huge amount of RAM pinning and V low VRAM usage on my 5090. It is much slower than FLF2VA. It seems spend it's time shoveling stuff into my sizeable but slow DDR4 RAM. Is this right?

Are these screenshots expected?

Beginning:

Middle

VAE

By comparison, FLF2VA, the entire process:

middle:

vae:

### Logs

```powershell
Relevant bits of logs for this, happy to provide more if needed.

[INFO] got prompt
[INFO] Model MiniMaxH3VideoVAE prepared for dynamic VRAM loading. 4965MB Staged. 0 patches attached. Force pre-loaded 128 weights: 348 KB.
[INFO] Model MiniMaxH3VideoVAE prepared for dynamic VRAM loading. 4965MB Staged. 0 patches attached. Force pre-loaded 128 weights: 348 KB.
[INFO] Model MiniMaxH3TEModel_ prepared for dynamic VRAM loading. 14956MB Staged. 0 patches attached. Force pre-loaded 410 weights: 4572 KB.
[INFO] 0 models unloaded.
[INFO] Model MiniMaxH3 prepared for dynamic VRAM loading. 19995MB Staged. 208 patches attached. Force pre-loaded 210 weights: 1175 KB.
100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 8/8 [01:57<00:00, 14.72s/it]
[INFO] Model MiniMaxH3AudioVAE prepared for dynamic VRAM loading. 576MB Staged. 0 patches attached. Force pre-loaded 401 weights: 539 KB.
[INFO] Model MiniMaxH3VideoVAE prepared for dynamic VRAM loading. 4965MB Staged. 0 patches attached. Force pre-loaded 128 weights: 348 KB.
[INFO] Prompt executed in 141.14 seconds

[INFO] Total VRAM 32109 MB, total RAM 110037 MB
[INFO] pytorch version: 2.9.1+cu130
[INFO] Set vram state to: NORMAL_VRAM
[INFO] Device: cuda:0 NVIDIA GeForce RTX 5090 : cudaMallocAsync
[INFO] Using async weight offloading with 2 streams
[INFO] Enabled pinned memory 99033.0
[INFO] working around nvidia conv3d memory bug.
[INFO] Using pytorch attention
[INFO] aimdo: /project/src/control.c:276:INFO:comfy-aimdo inited for GPU: NVIDIA GeForce RTX 5090 (VRAM: 32109 MB)
[INFO] DynamicVRAM support detected and enabled
[INFO] Python version: 3.12.3 (main, Mar 23 2026, 19:04:32) [GCC 13.3.0]
[INFO] ComfyUI version: 0.33.1
[INFO] comfy-aimdo version: 0.4.13
[INFO] comfy-kitchen version: 0.2.31
[INFO] comfyui-frontend-package version: 1.48.7
[INFO] comfyui-workflow-templates version: 0.11.41
[INFO] comfyui-embedded-docs version: 0.5.9
[INFO] comfy-kitchen version: 0.2.31
[INFO] comfy-aimdo version: 0.4.13
```

### Other

Wondering if it relates to PR 15027 and all the pinning, which I asked about https://github.com/Comfy-Org/ComfyUI/issues/15556

Reproduce the Minimax-h3 REF2VA run on the reported RTX 5090 setup and compare its VRAM and pinned-RAM behavior with FLF2VA. Start with the dynamic VRAM, async offloading, pinned-memory, and conv3d-related log entries, then review PR 15027 and issue 15556; done means establishing whether the observed behavior is expected or identifying a reproducible regression.

Written by the indexing model from the issue text.