Dasong Gao

_**PS**_: For my hypothetical use case above, I was able to workaround by: ```python def func2(x_, p_): return SE3.Identity(x_.shape[:-1]).retr(x_).act(p_).sum() print(torch.autograd.functional.jacobian(func2, (x, p))) ``` Output: ``` (tensor([[1000., 1000., 1000., 0., 0.,...

The problem seems to be that when `CopyRAI` is looking for suitable computers, `IIToIIParallel` instead of `IIToRAIParallel` is selected since the second argument can be cast into `II`, which outputs...

Thanks for your interest. Theoretically, you can as long as the model fits into your memory. However, I would recommend downsizing the image if possible.

您好，`--env`后面需要列出training environment的顺序，例如Nordland数据集：`--env spring summer fall winter`。

您好，`--dataset`应为`tartanair`，`airloop/AirLoop-main/data`应该通过`config/config.yaml`或`--dataset-root`设置。不好意思，之前确实有一些问题，请尝试最新的README和代码。

Try to put the file in a directory and set `DATA_PATH` to the directory.

As a related issue, access to `tile_bins` in the kernel below sometimes also throws OOM at the line: https://github.com/nerfstudio-project/gsplat/blob/bae873819fe2d50b4bf097117842d15f17d3a572/gsplat/cuda/csrc/forward.cu#L286 The reason is that the size of `tile_bins` returned by `get_tile_bin_edges_tensor`...

Was able to build and (kind of) run `benchmarks/basic.sh` on Jetson Orin Nano 8GB (Jetpack 6.1, CUDA 12.6, [torch==2.3.0+cu12.4](https://forums.developer.nvidia.com/t/pytorch-for-jetson/72048)). `gsplat` built successfully out of box in 20-30min with 4 threads....