Sergii Kharagorgiev

it's rather 2.1, not even in 2.0: https://pytorch.org/docs/2.0/generated/torch.jit.ScriptModule.html?highlight=load_state_dict#torch.jit.ScriptModule.load_state_dict ok, will test it and make a PR. One side effect of this feature -- it might might simplify support for loading...

> what do you mean by quantization settings? Eg 8 bit? Yes, with 4-bit quantization, bitsandbytes just dequanizes weights and multiplies with torch: ``` output = torch.nn.functional.linear(A, F.dequantize_4bit(B, quant_state).to(A.dtype).t(), bias)...

found one issue with cones -- in ConeGeometry there are some degenerate triangles (because of the way it's build out of CylinderGeometry). A check for input triangles solves it: https://github.com/coolvision/three-bvh-csg/commit/8f7d1d300454f9e61ffb19f953aca17182416b16#diff-fafbaae52166e47a0cb45ec83abbb8b4008ee65b84fc346b7cc1aafa2d022f10R199-R202

> Regarding the degenerate triangles in the cone geometry - can you elaborate a bit more on why cones get a degenerate triangle and where it shows up? Is this...

> There's a task involved here to remove most of the manual device setting throughout the codebase (e.g. `to.(device=)` and `tensor([], device=`), as torch handles most of this by default...