[RTX4090] Bus error(core dumped) when running embedder 4

[daxinY]

1. Specific problem encountered: When executing python3 render.py --config configs/waymo_val_173.yaml mode diffusion command, an error "Bus error(core dumped)" is generated. After import pdb single-step debugging, it is determined that the problem occurs in "def forward(self, batch: Dict, force_zero_embeddings: Optional[List] = None) -> Dict in class GeneralConditioner(nn.Module)", specifically when executing "### for i in range(batch[embedder.input_key].shape[0]): emb_out_1 =embedder(batch[embedder.input_key][i].unsqueeze(0)) emb_out_1s.append(emb_out_1)". When training in embedder 1, 2, and 3, it runs normally. When embedder is 4, the code "for i in range(batch[embedder.input_key].shape[0])" cannot be executed. The specific graphics card configuration is to execute the "render.py" code on a single RTX4090 with 24GB video memory;
2. When I try to execute this code alone, the same error message is displayed. When executing the 4th embedder, the operation ends directly. What's going on?

Environment information

• Docker environment: nvidia/cuda:12.1.0-devel-ubuntu20.04
• Graphics card: NVIDIA RTX 4090 (24GB)
• Python version: 3.9.5
• PyTorch version: 2.4.0+cu121