Guillaume Lagrange
Guillaume Lagrange
Yeah I've been stumbling into issues with scalars when looking at different ops to add support for recently... Even for node inputs/outputs this might cause trouble since ONNX spec is...
Was running the checks locally and `test record::serde::ser::tests::test_param_serde` just failed.. will investigate & fix. /edit: Previously the `fmt::Debug` captured by the test had the vector as `Vec([F32(1.0), F32(1.0), F32(1.0), ...]...
To close this PR I'll handle the other tensor element types, but I've opened a new issue regarding the other improvements suggested in previous discussions.
> If we do #1773 alone, then we can deprecate serialization. So you don't need to do other accumulating item types. Sure, we can limit the current PR to `Vec`...
This belongs to cubecl now 😅 so I'll close due to the massive changes in this area.
Hmm not quite clear at first glance what could be the cause, looks like it could not find the operation for the CUDA backend (so perhaps it couldn't find the...
From the error in the OP it just seems that `tch` is not accessing the right dll. ``` ... 00007FFC0F40310400007FFC0F4030B0 torch_cpu.dll!c10::impl::OperatorEntry::lookup [ @ ]\n00007FFC0FF8858C00007FFC0FF11E90 torch_cpu.dll!at::_ops::xlogy__Tensor::redispatch [ @ ]\n00007FFC100562EE00007FFC10056210 torch_cpu.dll!at::_ops::empty_memory_format::redispatch [...
Yeah our wgpu backend is not optimized in all areas. We know that our vision ops are not as great but this is on our list :)
As a reference, the plan is not to support everything such that it can be a 1-to-1 map with PyTorch. The `PyTorchFileRecorder` is there to help make the transition easier...
The community is always one step ahead 😄 We've actually discussed adding SAM to our models and this was in the plans following the release. We still haven't decided whether...