What are the graphics memory requirements for this model?

[NexF]

I use the NVIDIA L20 (have 48GB memory), and it has an "CUDA out of memory" error. so how big the graphics memory requirements for this model is?

2025/08/27 02:21:51 - Round 0 in_view_ids: [0], tar_view_ids: [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15]
100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 20/20 [00:21<00:00,  1.06s/it]
image 0 generated: (1024, 1024)
2025/08/27 02:22:14 - Do depth map prediction for input views
Traceback (most recent call last):
  File "/workspace/spatialgen/SpatialGen/src/inference_sd.py", line 1924, in <module>
    main()
  File "/workspace/spatialgen/SpatialGen/src/inference_sd.py", line 1908, in main
    rooms_infer_res_dict = log_val_autoregressive_from_dataloader(
...... ......
  File "/root/miniconda3/envs/spatialgen/lib/python3.10/site-packages/torch/nn/modules/linear.py", line 116, in forward
    return F.linear(input, self.weight, self.bias)
torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 3.12 GiB. GPU

[NexF]

I use scripts/infer_spatialgen_t2s.sh script

[zixue365]

支持多个GPU卡部署吗

[NexF]

not support now.
I have NVIDIA L20 * 8, but this scripts only use one of it

[fangchuan]

We test our example scripts on NVIDIA-H20 (96G mem) .
scripts/infer_spatialgen_t2s.sh uses 30G mem for single-image-to-scene generation, while scripts/infer_spatialgen_t2s.sh takes about 40G mem to do text-to-scene generation, due to the huge mem consumption of FLUX.

I use scripts/infer_spatialgen_t2s.sh script

[fangchuan]

We are working hard to make the inference more affordable, we will update the code upon finishing the test

[whk6688]

Is there only command？ Will you support upload picture via UI ， then to show render result？

[ZhouBay-TF]

Can the experts successfully run according to the requirements? I have been stuck because diffusers and ImageReward depend on different versions of transformers. Can you share your running requirements file.thanks

[fangchuan]

Is there only command？ Will you support upload picture via UI ， then to show render result？

We will put the gradio on huggingface soon

[fangchuan]

Can the experts successfully run according to the requirements? I have been stuck because diffusers and ImageReward depend on different versions of transformers. Can you share your running requirements file.thanks

Can you share the error reported on your computer, that will give us more clues to help you fix it

[ZhouBay-TF]

not support now. I have NVIDIA L20 * 8, but this scripts only use one of it

#5 (comment)
Hello, have you encountered this problem or how did you solve it

[atcelen]

Hello,
I’m working on building a training pipeline for the UNetMVMM2DConditionModel class. This discussion made me curious, did you use model or activation sharding across multiple GPUs during training (e.g., ZeRO-3)? I noticed that the paper mentioned training with RTX 4090s and wanted to understand how you managed memory efficiency. Thanks!