Lightweight test with 8GB GPU

Read the reported demo.py loading path and compare it with the command and environment details in this issue. There is no requested file change or defined completion condition; a contribution would need an agreed documentation location and scope for recording the 8GB GPU workflow.

I have successfully run lingbot-map in my environment.
Here is informations, hope it helps others.

Environment

Modifications

1. Use only 21 frames of example/courthouse.
2. In load_model(demo.py), load the model to CPU first, then transfer to GPU to avoid memory peak.
3. Use Chrome instead of Firefox. Firefox may be incompatible with Viser, causing the point cloud to fail to render.

CMD and Log

python demo.py --model_path ../models/lingbot-map-long.pt --image_folder example/courthouse_0_20 --use_sdpa --offload_to_cpu --mask_sky

Loading 21 images...
Loading images: 100%|████████████████████████████████████████████████████████████████████████████████████████████| 21/21 [00:00<00:00, 509.37it/s]
Preprocessed images to 518x294 using canonical crop mode
/home/jcy/miniconda3/envs/lingbot-map/lib/python3.10/site-packages/tvm_ffi/_optional_torch_c_dlpack.py:181: UserWarning: Failed to JIT torch c dlpack extension, EnvTensorAllocator will not be enabled.
We recommend installing via pip install torch-c-dlpack-ext
warnings.warn(
Building model...
pretrained_path:
Failed to load pretrained weights: [Errno 2] No such file or directory: ''
Loading checkpoint: ../models/lingbot-map-long.pt
Checkpoint loaded.
Total load time: 9.4s
Casting aggregator to torch.bfloat16 (heads kept in fp32)
Input: 21 frames, shape (21, 3, 294, 518)
Mode: streaming
GPU mem after load: alloc=2.85 GB, reserved=2.88 GB
Running streaming inference (dtype=torch.bfloat16)...
Streaming inference: 100%|████████████████████████████████████████████████████████████████████████████████████████| 21/21 [00:02<00:00, 4.63it/s]
Inference done in 5.4s
GPU peak during inference: 6.02 GB (reserved peak 6.42 GB)
Moving results to CPU...
╭────── viser (listening *:8080) ───────╮
│ ╷ │
│ HTTP │ http://localhost:8080 │
│ Websocket │ ws://localhost:8080 │
│ ╵ │
╰───────────────────────────────────────╯
Generating sky masks from image array...
100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████| 21/21 [00:00<00:00, 784.82it/s]
Sky segmentation applied successfully

Screenshot