torch.cat is missing since update to PyTorch 1.13

[sbrunk]

Before the update to PyTorch 1.13 in b2b8eb9, the PyTorch bindings included the following functions

@Namespace("at") public static native @ByVal Tensor cat(@ByVal TensorArrayRef tensors, @Cast("int64_t") long dim/*=0*/);
@Namespace("at") public static native @ByVal Tensor cat(@ByVal TensorArrayRef tensors);

as well as a few variants like cat_out, all coming from ATen/ops/cat.h, which has the following diff:

-TORCH_API inline at::Tensor cat(at::TensorList tensors, int64_t dim=0) {
+inline at::Tensor cat(const at::ITensorListRef & tensors, int64_t dim=0) {
    return at::_ops::cat::call(tensors, dim);
}

So the type changed from TensorList to ITensorListRef which seems to be skipped in the mappings:

javacpp-presets/pytorch/src/main/java/org/bytedeco/pytorch/presets/torch.java

Line 2369 in c0d15db

"at::ITensorListRef", "std::array<c10::FunctionalityOffsetAndMask,c10::num_functionality_keys>").skip())

Could this be the reason why cat is not mapped anymore?

[saudet]

Good catch! Done in commit 10ddc85. Thanks

[sbrunk]

Thanks for adding! Could you perhaps re-trigger the CI? The windows build timed out and I think that prevented publishing the snapshots.

[sbrunk]

Works like a charm with the new snapshot. :)

val t = torch.Tensor(Seq(1,2,3))                                                                                                                                                                                                 
// t: Tensor[Int32] = dtype=int32, shape=[3], device=CPU 
// [1, 2, 3]
torch.cat(Seq(t,t))
// res0: Tensor[Int32] = dtype=int32, shape=[6], device=CPU 
// [1, 2, 3, 1, 2, 3]

So I think we can close this.