fix bert

aten/src/ATen/native/Checkpoint.cpp

@@ -137,6 +137,30 @@ Tensor checkpoint_to(at::Tensor const& a, c10::TensorOptions const& b, bool c, b
   return CheckpointTensorImpl::make("to", rt, {a})[0];
 }
 
+Tensor checkpoint_to(at::Tensor const& a, at::Tensor const& b, bool c, bool d, c10::optional<c10::MemoryFormat> e) {
+  rematerialize_function_t rt =
+    [=](const Tensors& vec) -> Tensors {
+      return {vec.at(0).to(b, c, d, e)};
+    };
+  return CheckpointTensorImpl::make("to", rt, {a})[0];
+}
+
+Tensor checkpoint_to(at::Tensor const& a, c10::ScalarType b, bool c, bool d, c10::optional<c10::MemoryFormat> e) {
+  rematerialize_function_t rt =
+    [=](const Tensors& vec) -> Tensors {
+      return {vec.at(0).to(b, c, d, e)};
+    };
+  return CheckpointTensorImpl::make("to", rt, {a})[0];
+}
+
+Tensor checkpoint_to(at::Tensor const& a, c10::Device b, c10::ScalarType c, bool d, bool e, c10::optional<c10::MemoryFormat> f) {
+  rematerialize_function_t rt =
+    [=](const Tensors& vec) -> Tensors {
+      return {vec.at(0).to(b, c, d, e, f)};
+    };
+  return CheckpointTensorImpl::make("to", rt, {a})[0];
+}
+
 Tensor checkpoint_div(const Tensor& a, const Tensor& b) {
   rematerialize_function_t rt =
     [=](const Tensors& vec) -> Tensors {
@@ -703,6 +727,40 @@ Tensor checkpoint_sum_dim_IntList(const Tensor& a, c10::ArrayRef<long> b, bool c
   return CheckpointTensorImpl::make("sum_dim_IntList", rt, {a})[0];
 }
 
+Tensor& checkpoint_transpose_(at::Tensor& a, long b, long c) {
+  mutate_function_t mt =
+    [=](const Tensors& vec) {
+      Tensor a_ = vec.at(0);
+      a_.transpose_(b, c);
+    };
+  CheckpointTensorImpl::mutate("transpose_", mt, {a}, {0});
+  return a;
+}
+
+Tensor checkpoint_transpose(at::Tensor const& a, long b, long c) {
+  rematerialize_function_t rt =
+    [=](const Tensors& vec) -> Tensors {
+      return {at::transpose(vec.at(0), b, c)};
+    };
+  return CheckpointTensorImpl::make("transpose", rt, {a})[0];
+}
+
+Tensor checkpoint_gelu(at::Tensor const& a) {
+  rematerialize_function_t rt =
+    [=](const Tensors& vec) -> Tensors {
+      return {at::gelu(vec.at(0))};
+    };
+  return CheckpointTensorImpl::make("gelu", rt, {a})[0];
+}
+
+Tensor checkpoint_matmul(at::Tensor const& a, at::Tensor const& b) {
+  rematerialize_function_t rt =
+    [=](const Tensors& vec) -> Tensors {
+      return {at::matmul(vec.at(0), vec.at(1))};
+    };
+    return CheckpointTensorImpl::make("matmul", rt, {a, b})[0];
+}
+
 Tensor checkpoint_threshold(const Tensor& a, c10::Scalar b, c10::Scalar c) {
   rematerialize_function_t rt =
     [=](const Tensors& vec) -> Tensors {

aten/src/ATen/native/native_functions.yaml

@@ -2157,8 +2157,17 @@
 - func: matmul(Tensor self, Tensor other) -> Tensor
   use_c10_dispatcher: full
   variants: function, method
+  dispatch:
+    CPU, CUDA: matmul
+    Checkpoint: checkpoint_matmul
 
 - func: matmul.out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)
+  #use_c10_dispatcher: full
+  #variants: function, method
+  #dispatch:
+  #  CPU, CUDA: matmul_out
+  #  Checkpoint: checkpoint_matmul_out
+  # there is some pt bug which disallow the above code.
 
 - func: matrix_rank.tol(Tensor self, float tol, bool symmetric=False) -> Tensor
   use_c10_dispatcher: full
@@ -2906,6 +2915,7 @@
   dispatch:
     CPU: gelu_cpu
     CUDA: gelu_cuda
+    Checkpoint: checkpoint_gelu
 
 - func: gelu_backward(Tensor grad, Tensor self) -> Tensor
   use_c10_dispatcher: full
@@ -3426,6 +3436,9 @@
   use_c10_dispatcher: full
   variants: function, method
   device_guard: False
+  dispatch:
+    CPU, CUDA: transpose
+    Checkpoint: checkpoint_transpose
 
 - func: transpose.Dimname(Tensor(a) self, Dimname dim0, Dimname dim1) -> Tensor(a)
   variants: function, method
@@ -3441,6 +3454,9 @@
   use_c10_dispatcher: full
   variants: method
   device_guard: False
+  dispatch:
+    CPU, CUDA: transpose_
+    Checkpoint: checkpoint_transpose_
 
 - func: _mkldnn_transpose_(Tensor(a!) self, int dim0, int dim1) -> Tensor(a!)
   use_c10_dispatcher: full
@@ -4461,16 +4477,25 @@
   use_c10_dispatcher: full
   variants: method
   device_guard: False
+  dispatch:
+    CPU, CUDA: to
+    Checkpoint: checkpoint_to
 
 - func: to.dtype(Tensor self, ScalarType dtype, bool non_blocking=False, bool copy=False, MemoryFormat? memory_format=None) -> Tensor
   use_c10_dispatcher: full
   variants: method
   device_guard: False
+  dispatch:
+    CPU, CUDA: to
+    Checkpoint: checkpoint_to
 
 - func: to.other(Tensor self, Tensor other, bool non_blocking=False, bool copy=False, MemoryFormat? memory_format=None) -> Tensor
   use_c10_dispatcher: full
   variants: method
   device_guard: False
+  dispatch:
+    CPU, CUDA: to
+    Checkpoint: checkpoint_to
 
 - func: meshgrid(Tensor[] tensors) -> Tensor[]
   use_c10_dispatcher: full