Multiple GPUs loading sub-models

I seem to have set it up this way without success. I'm running demo_3db_mesh_alignment. Are there any other solutions? Thank you. I have already set the parameters in ppipeline.yaml to downsample_ss_dist: 8 and compile_model: true.
output = inference._pipeline.run(
inference.merge_mask_to_rgba(image, mask),
None,
seed=10,decode_formats=["gaussian"] # skip mesh decoding to save memory
)
still：
Error: At "Decoding sparse latent..."
Omitted content...
ret["gaussian"] = self.models["slat_decoder_gs"](slat)
Omitting...
torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 3.15 GiB. GPU 0 has a total capacity of 23.64 GiB of which 3.10 GiB is free.