Add aten::multi_margin_loss and its variants (#895)

- [x] multi_margin_loss
- [x] multi_margin_loss.out
- [x] multi_margin_loss_backward
- [x] multi_margin_loss_backward.grad_input

---------

Co-authored-by: Yutao Xu <[email protected]>

src/ATen/native/xpu/LossMultiMargin.cpp

@@ -0,0 +1,64 @@
+#include <ATen/core/Tensor.h>
+#include <ATen/native/xpu/sycl/MultiMarginLossKernels.h>
+
+#include <ATen/ops/empty.h>
+#include <xpu/ATen/ops/multi_margin_loss_backward_native.h>
+#include <xpu/ATen/ops/multi_margin_loss_native.h>
+
+namespace at::native {
+
+Tensor& multi_margin_loss_xpu_out(
+    const Tensor& self,
+    const Tensor& target,
+    const Scalar& p,
+    const Scalar& margin,
+    const std::optional<Tensor>& weight,
+    int64_t reduction,
+    Tensor& out) {
+  xpu::multi_margin_loss_kernel(
+      self, target, p, margin, weight, reduction, out);
+  return out;
+}
+
+Tensor multi_margin_loss_xpu(
+    const Tensor& self,
+    const Tensor& target,
+    const Scalar& p,
+    const Scalar& margin,
+    const std::optional<Tensor>& weight,
+    int64_t reduction) {
+  auto out = at::empty({0}, self.options());
+  xpu::multi_margin_loss_kernel(
+      self, target, p, margin, weight, reduction, out);
+  return out;
+}
+
+Tensor& multi_margin_loss_xpu_backward_out(
+    const Tensor& grad_output,
+    const Tensor& self,
+    const Tensor& target,
+    const Scalar& p,
+    const Scalar& margin,
+    const std::optional<Tensor>& weight,
+    int64_t reduction,
+    Tensor& grad_input) {
+  xpu::multi_margin_loss_backward_kernel(
+      grad_output, self, target, p, margin, weight, reduction, grad_input);
+  return grad_input;
+}
+
+Tensor multi_margin_loss_xpu_backward(
+    const Tensor& grad_output,
+    const Tensor& self,
+    const Tensor& target,
+    const Scalar& p,
+    const Scalar& margin,
+    const std::optional<Tensor>& weight,
+    int64_t reduction) {
+  auto grad_input = at::empty({0}, self.options());
+  xpu::multi_margin_loss_backward_kernel(
+      grad_output, self, target, p, margin, weight, reduction, grad_input);
+  return grad_input;
+}
+
+} // namespace at::native

src/ATen/native/xpu/XPUFallback.template

@@ -199,8 +199,6 @@ TORCH_LIBRARY_IMPL(aten, XPU, m) {
     "lu_unpack.out",
     "multilabel_margin_loss_backward",
     "multilabel_margin_loss_forward",
-    "multi_margin_loss",
-    "multi_margin_loss_backward",
     "ormqr",
     "rrelu_with_noise",
     "_scaled_dot_product_efficient_attention",