chore: import upstream snapshot with attribution

2026-07-13 13:24:13 +08:00
commit 1037506f2e
6050 changed files with 1731598 additions and 0 deletions
@@ -0,0 +1,56 @@
+/**
+ * Copyright (c) Facebook, Inc. and its affiliates.
+ *
+ * This source code is licensed under the MIT license found in the
+ * LICENSE file in the root directory of this source tree.
+ */
+
+#include <torch/extension.h>
+#include <vector>
+
+std::vector<at::Tensor> dynamicconv_cuda_forward(
+    at::Tensor input,
+    at::Tensor filters,
+    int padding_l);
+
+std::vector<at::Tensor> dynamicconv_cuda_backward(
+    at::Tensor gradOutput,
+    int padding_l,
+    at::Tensor input,
+    at::Tensor filters);
+
+
+#define CHECK_CUDA(x) AT_ASSERTM(x.type().is_cuda(), #x " must be a CUDA tensor")
+#define CHECK_CONTIGUOUS(x) AT_ASSERTM(x.is_contiguous(), #x " must be contiguous")
+#define CHECK_INPUT(x) CHECK_CUDA(x); CHECK_CONTIGUOUS(x)
+
+std::vector<at::Tensor> dynamicconv_forward(
+    at::Tensor input,
+    at::Tensor filters,
+    int padding_l) {
+
+    CHECK_INPUT(input);
+    CHECK_INPUT(filters);
+
+    return dynamicconv_cuda_forward(input, filters,
+            padding_l);
+}
+
+std::vector<at::Tensor> dynamicconv_backward(
+    at::Tensor gradOutput,
+    int padding_l,
+    at::Tensor input,
+    at::Tensor filters) {
+
+    CHECK_INPUT(gradOutput);
+    CHECK_INPUT(input);
+    CHECK_INPUT(filters);
+
+    return dynamicconv_cuda_backward(gradOutput, padding_l,
+            input, filters);
+}
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+    m.def("forward", &dynamicconv_forward, "dynamicconv forward (CUDA)");
+    m.def("backward", &dynamicconv_backward, "dynamicconv backward (CUDA)");
+}