chore: import upstream snapshot with attribution

csrc/torch_bindings.cpp

@@ -0,0 +1,39 @@
+// Provides torch::Tensor for ops.h (previously included transitively via
+// cache.h, which is no longer included here after cache ops moved to
+// _C_stable_libtorch).
+#include <torch/all.h>
+#include "ops.h"
+#include "core/registration.h"
+#include <torch/library.h>
+#include <torch/version.h>
+
+// Note on op signatures:
+// The X_meta signatures are for the meta functions corresponding to op X.
+// They must be kept in sync with the signature for X. Generally, only
+// functions that return Tensors require a meta function.
+//
+// See the following links for detailed docs on op registration and function
+// schemas.
+// https://docs.google.com/document/d/1_W62p8WJOQQUzPsJYa7s701JXt0qf2OfLub2sbkHOaU/edit#heading=h.ptttacy8y1u9
+// https://github.com/pytorch/pytorch/blob/main/aten/src/ATen/native/README.md#annotations
+
+#ifdef USE_ROCM
+TORCH_LIBRARY_FRAGMENT(CONCAT(TORCH_EXTENSION_NAME, _custom_ar), custom_ar) {
+  // Quick Reduce all-reduce kernels (ROCm-only; stays on legacy _C).
+  custom_ar.def(
+      "qr_all_reduce(int fa, Tensor inp, Tensor out, int quant_level, bool "
+      "cast_bf2half) -> ()");
+  custom_ar.impl("qr_all_reduce", torch::kCUDA, &qr_all_reduce);
+
+  custom_ar.def("init_custom_qr", &init_custom_qr);
+  custom_ar.def("qr_destroy", &qr_destroy);
+  custom_ar.def("qr_get_handle", &qr_get_handle);
+
+  custom_ar.def("qr_open_handles(int _fa, Tensor[](b!) handles) -> ()");
+  custom_ar.impl("qr_open_handles", torch::kCPU, &qr_open_handles);
+
+  custom_ar.def("qr_max_size", &qr_max_size);
+}
+#endif
+
+REGISTER_EXTENSION(TORCH_EXTENSION_NAME)