MiniMax Music 3 CUDA Graphs emit repeated CUDAMallocAsyncAllocator “uncaptured free” warnings on Windows/RTX 5090

### 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

MiniMax Music 3 should finish AR sampling and continue to the diffusion/audio decoding stages without CUDA Graph allocation-lifetime warnings.

### Actual Behavior

Immediately after MiniMax Music 3 finishes AR sampling, PyTorch emits a large burst of warnings:

```
AR sampling: 100%|█████████████████████████████████████████████████████████████| 1501/1501 [00:33<00:00, 44.89it/s]

[W813 17:37:39.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264EA67400
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
```

This repeats 76 times with 76 different allocation addresses, all at the same timestamp immediately after AR sampling: 1501/1501.
The workflow doesn't crash. It continues through the remaining stages and produces output:

```
[INFO] Requested to load MiniMaxMusic3
[INFO] Model MiniMaxMusic3 prepared for dynamic VRAM loading. 4685MB Staged. 0 patches attached. Force pre-loaded 73 weights: 580 KB.
100%|██████████████████████████████████████████████████████████████████████████| 30/30 [00:20<00:00, 1.43it/s]
[INFO] Requested to load MiniMaxMusic3DAV
[INFO] loaded completely; 206.60 MB loaded, full load: True
[INFO] Prompt executed in 68.44 seconds
```

I couldn't notice an obvious failure in the generated result (the default workflow prompt).

### Steps to Reproduce

1. Launch ComfyUI normally, leaving CUDA Graphs enabled.
2. Load and run a MiniMax Music 3 workflow.
3. Allow the autoregressive stage to complete.
4. Observe the console immediately after AR sampling reaches 100%.
5. A large burst of Attempting uncaptured free of a captured allocation warnings appears before the diffusion stage runs.

### Debug Logs

```powershell
Environment:
- OS: Windows 11, build 10.0.26200.9168
- GPU: NVIDIA GeForce RTX 5090, 32607 MiB VRAM
- NVIDIA driver: 610.88
- ComfyUI: v0.33.0
- Commit: 2f35f4a08176d993cded35dac3332be4f7287f41
- Python: 3.12.9
- PyTorch: 2.12.1+cu130
- CUDA runtime: 13.0
- comfy-kitchen: 0.2.31
- Launch command: python main.py --listen --port 8191 --enable-triton-backend

Debug log:

[INFO] got prompt
[INFO] Requested to load MiniMaxMusic3TEModel
[INFO] Model MiniMaxMusic3TEModel prepared for dynamic VRAM loading. 17605MB Staged. 0 patches attached. Force pre-loaded 154 weights: 674 KB.
AR sampling: 77%|█████████████████████████████████████████████████▉ | 2304/3001 [00:55<00:16, 41.59it/s]
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ED3C000
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ED3BE00
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ECE1C00
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ECE1A00
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ED3D800
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ED3D600
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ED3D400
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ED3D200
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ECE0400
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ECE0200
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ECDD400
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ECDD200
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ECE2400
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ECE2200
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ECDEC00
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ECDEA00
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ED3E000
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ED3DE00
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 0000002644966C00
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ED5EC00
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 0000002644967000
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 0000002644966E00
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ED3CC00
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ED3CA00
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ECDE000
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ECDDE00
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ED3E400
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ED3E200
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ED3AE00
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ED3AC00
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ECDE800
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ECDE600
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ECDF000
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ECDEE00
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ECDF400
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ECDF200
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ED3DC00
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ED3DA00
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ECE0800
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ECE0600
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ECDD000
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ED3E600
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ED3B800
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ED3B000
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ECDDC00
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ECDDA00
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ECDE400
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ECDE200
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ECE1400
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ECE1200
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ECE2000
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ECE1E00
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ED3C400
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ED3C200
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ECE1000
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ECE0E00
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ED3AA00
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ED3A800
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ECDFC00
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ECDFA00
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ED3C800
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ED3C600
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ED3D000
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ED3CE00
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ECE0000
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ECDFE00
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ECE0C00
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ECE0A00
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ECDF800
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ECDF600
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ECE1800
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ECE1600
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ED3BC00
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ED3BA00
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ECDD800
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
[W813 19:06:00.000000000 CUDAMallocAsyncAllocator.cpp:307] Warning: Attempting uncaptured free of a captured allocation with address 000000264ECDD600
This is technically allowed, but may indicate you are losing the last user-visible tensor through which the allocation can be accessed, so you'll have no way to view the data after future replays of the owning graph. (function freeAsync)
```

### Other

This appears to involve the CUDA Graph support introduced alongside MiniMax Music 3 in PR #15570:
[https://github.com/Comfy-Org/ComfyUI/pull/15570](https://github.com/Comfy-Org/ComfyUI/pull/15570)

Start by reproducing the MiniMax Music 3 workflow on Windows with CUDA Graphs enabled and observe the transition after AR sampling completes. Trace the allocation cleanup around that transition; done means the repeated CUDAMallocAsyncAllocator uncaptured-free warnings no longer appear while diffusion and audio decoding still produce output.

Written by the indexing model from the issue text.