torch.compile -> torch_compile from misc

[liruilong940607]

Fix the issue of torch.compile() introduced in #3200 is not supported on windows

[gradeeterna]

Hey, just updated on Windows and getting these errors with splatfacto, Nerfstudio v1.1.2 and gsplat v1.0. Any ideas?

`3 errors detected in the compilation of "C:/Users/user/anaconda3/envs/nerfstudio/lib/site-packages/gsplat/cuda/csrc/rasterization.cu".
rasterization.cu
ninja: build stopped: subcommand failed.

Printing profiling stats, from longest to shortest duration in seconds
Trainer.train_iteration: 16.7488
VanillaPipeline.get_train_loss_dict: 16.7478
Traceback (most recent call last):
File "C:\Users\user\anaconda3\envs\nerfstudio\lib\runpy.py", line 196, in _run_module_as_main
return _run_code(code, main_globals, None,
File "C:\Users\user\anaconda3\envs\nerfstudio\lib\runpy.py", line 86, in run_code
exec(code, run_globals)
File "C:\Users\user\anaconda3\envs\nerfstudio\Scripts\ns-train.exe_main.py", line 7, in
File "D:\nerfstudiogs\nerfstudio\nerfstudio\scripts\train.py", line 262, in entrypoint
main(
File "D:\nerfstudiogs\nerfstudio\nerfstudio\scripts\train.py", line 247, in main
launch(
File "D:\nerfstudiogs\nerfstudio\nerfstudio\scripts\train.py", line 189, in launch
main_func(local_rank=0, world_size=world_size, config=config)
File "D:\nerfstudiogs\nerfstudio\nerfstudio\scripts\train.py", line 100, in train_loop
trainer.train()
File "D:\nerfstudiogs\nerfstudio\nerfstudio\engine\trainer.py", line 261, in train
loss, loss_dict, metrics_dict = self.train_iteration(step)
File "D:\nerfstudiogs\nerfstudio\nerfstudio\utils\profiler.py", line 112, in inner
out = func(*args, **kwargs)
File "D:\nerfstudiogs\nerfstudio\nerfstudio\engine\trainer.py", line 496, in train_iteration
_, loss_dict, metrics_dict = self.pipeline.get_train_loss_dict(step=step)
File "D:\nerfstudiogs\nerfstudio\nerfstudio\utils\profiler.py", line 112, in inner
out = func(*args, **kwargs)
File "D:\nerfstudiogs\nerfstudio\nerfstudio\pipelines\base_pipeline.py", line 301, in get_train_loss_dict
model_outputs = self._model(ray_bundle) # train distributed data parallel model if world_size > 1
File "C:\Users\user\anaconda3\envs\nerfstudio\lib\site-packages\torch\nn\modules\module.py", line 1518, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
File "C:\Users\user\anaconda3\envs\nerfstudio\lib\site-packages\torch\nn\modules\module.py", line 1527, in _call_impl
return forward_call(*args, **kwargs)
File "D:\nerfstudiogs\nerfstudio\nerfstudio\models\base_model.py", line 143, in forward
return self.get_outputs(ray_bundle)
File "D:\nerfstudiogs\nerfstudio\nerfstudio\models\splatfacto.py", line 733, in get_outputs
render, alpha, info = rasterization(
File "C:\Users\user\anaconda3\envs\nerfstudio\lib\site-packages\gsplat\rendering.py", line 212, in rasterization
proj_results = fully_fused_projection(
File "C:\Users\user\anaconda3\envs\nerfstudio\lib\site-packages\gsplat\cuda_wrapper.py", line 260, in fully_fused_projection
return _FullyFusedProjection.apply(
File "C:\Users\user\anaconda3\envs\nerfstudio\lib\site-packages\torch\autograd\function.py", line 539, in apply
return super().apply(*args, **kwargs) # type: ignore[misc]
File "C:\Users\user\anaconda3\envs\nerfstudio\lib\site-packages\gsplat\cuda_wrapper.py", line 692, in forward
radii, means2d, depths, conics, compensations = _make_lazy_cuda_func(
File "C:\Users\user\anaconda3\envs\nerfstudio\lib\site-packages\gsplat\cuda_wrapper.py", line 12, in call_cuda
return getattr(_C, name)(*args, **kwargs)
AttributeError: module 'gsplat_cuda' has no attribute 'fully_fused_projection_fwd' `

[liruilong940607]

@gradeeterna It shows the compiling is failed. Any error message before this to show the compiling errors?

[gradeeterna]

@liruilong940607 Yeah there was a huge wall of errors when running ns-train splatfacto, starting with:

Traceback (most recent call last):
File "C:\Users\user\anaconda3\envs\nerfstudio\lib\site-packages\gsplat\cuda_backend.py", line 41, in
from gsplat import csrc as C
ImportError: cannot import name 'csrc' from 'gsplat' (C:\Users\user\anaconda3\envs\nerfstudio\lib\site-packages\gsplat_init.py)

Attached the full errors as a txt below.

errors.txt

Also tried setting up a fresh conda env and installing from scratch, tinycudann etc seemed to build fine, but get these errors when I try ns-train. Downgraded to nerfstudio v1.1.0 and gsplat v0.1.11 for now.

Thanks!

[liruilong940607]

Thanks for sharing the error log! this is very helpful! I'm gonna look into it

[Ben-Mack]

Hi @gradeeterna , I'm having the same error as you did, but I tried to downgrade nerfstudio & gsplat but still same error

pip install nerfstudio==1.1.0
pip install gplat==0.1.11

Can you share how did you do the downgrade?

[gradeeterna]

@Ben-Mack I'm pretty sure that's all I did to downgrade a few days ago, but I was also struggling with this again yesterday.

I eventually got the latest versions of nerfstudio 1.1.2 and gsplat 1.0 to work following the updated Windows installation guide here - https://github.com/nerfstudio-project/nerfstudio/blob/main/docs/quickstart/installation.md

After the vcvarsall.bat part, I also had to run "set DISTUTILS_USE_SDK=1" to get gsplat to build successfully.

Also if installing nerfstudio from source rather than pip install, skip "pip install --upgrade pip setuptools" as that updates setuptools from v69 to v70 which wasn't working with splatfacto.