RichardScottOZ

Thanks @Huite

I was just wondering about this today - if you had some reasonable number of terabytes of generic rasters - how can you start approaching this sort of icechunk-esque speedup...

> VirtualiZarr is the way, don't necessarily need Icechunk. Check out Tom's in-dev open_virtual_mfdataset which is the xarray analog for open_mfdataset. I'm not clear yet on how specific format is...

Don't think I would want to make millions multiple times anyway.

> VirtualiZarr is the way, don't necessarily need Icechunk. Check out Tom's in-dev open_virtual_mfdataset which is the xarray analog for open_mfdataset. I'm not clear yet on how specific format is...

I just tried this example pretty much https://projectpythia.org/kerchunk-cookbook/notebooks/generating_references/GeoTIFF.html With own private data e.g. anon is false give key and secret. The s3fs stuff is fine as you would expect. Not...

> ```python > pid = grid.find_closest_point(pyvista.PolyData(np.array([-0.4, 0.1, 0]))) > ``` ```python --------------------------------------------------------------------------- TypeError Traceback (most recent call last) Cell In[48], line 1 ----> 1 pid = mesh4.find_closest_point(p) File ~\AppData\Local\miniconda3\envs\pangeo\lib\site-packages\pyvista\core\dataset.py:2193, in...

@rekram1-node - if you want anything tested, let me know. I did try several recent versions too to check - same behaviour. Tried a lot of things - symlinks, aliases...

Thanks - so theoretically if I upgrade CUDA toolkit to 12.8 and make a torch 2.7 environment this should work?

> > Did you compile successfully with torch 2.6.0 cu126? > > I was eager to try cuda 12.6, but torch2.6.0+cu126 and flash_attn-2.7.4-cp310 didn't work. "CUDA error: device-side assert triggered"...