Full-training the controlnet in VACE-14B, it is OOM even in 8*H200.

Hi，when I full-training the controlnet in VACE-14B, it is OOM even in 8*H200 (144G).

[rank2]: x_out = torch.view_as_real(x_out * freqs).flatten(2)
[rank2]: torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 986.00 MiB. GPU 2 has a total capacity of 139.81 GiB of which 149.69 MiB is free. Including non-PyTorch memory, this process has 139.66 GiB memory in use. Of the allocated memory 133.10 GiB is allocated by PyTorch, and 3.84 GiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables)

0%| | 1/1271 [06:15<132:36:14, 375.89s/it]
[rank6]: Traceback (most recent call last):

Start by reviewing the full-training setup for the VACE-14B controlnet on 8 H200 GPUs and reproduce the CUDA out-of-memory failure. Inspect the traceback around torch.view_as_real(x_out * freqs) and compare the reported allocation and reserved memory. Done means full training proceeds without running out of GPU memory, with the relevant configuration and reproduction results documented.