LTX 2.3 doesn't work when using two RTX 5090s.

I've even tried with 120 frames and get a CUDA error, but it works fine when using just an RTX 5090 in a separate instance on Vast or RunPod.

[wtf.json](https://github.com/user-attachments/files/29481294/wtf.json)

INFO] Prompt executed in 48.26 seconds
[INFO] got prompt
[INFO] Requested to load LTXAV
[INFO] loaded partially; 1260.18 MB usable, 140.04 MB loaded, 28649.31 MB offloaded, 1120.13 MB buffer reserved, lowvram patches: 1653
[INFO] Patching torch settings: torch.backends.cuda.matmul.allow_fp16_accumulation = True
100%|██████████| 9/9 [00:39<00:00, 4.38s/it]
[INFO] Patching torch settings: torch.backends.cuda.matmul.allow_fp16_accumulation = False
[INFO] Requested to load AudioVAE
[INFO] loaded completely; 693.46 MB loaded, full load: True
[INFO] Requested to load VideoVAE
[INFO] loaded partially; 1120.44 MB usable, 1079.68 MB loaded, 305.26 MB offloaded, 40.50 MB buffer reserved, lowvram patches: 0
[LTX Reference Conditioning] strength=0, bypassing and clearing state.
[INFO] Requested to load LTXAV
[INFO] loaded partially; 269.54 MB usable, 0.00 MB loaded, 28789.36 MB offloaded, 1120.13 MB buffer reserved, lowvram patches: 1660
[INFO] Patching torch settings: torch.backends.cuda.matmul.allow_fp16_accumulation = True
0%| | 0/3 [00:00

Start by reproducing the attached wtf.json workflow with two RTX 5090s, then trace the failure from custom_nodes/10S-Comfy-nodes/latent_tiled_sampler.py into comfy/samplers.py and comfy/ldm/lightricks/av_model.py. Compare the two-GPU run with the reported single-GPU setup; done means the LTX 2.3 workflow completes without the device-0 out-of-memory errors.

Written by the indexing model from the issue text.