[Cherry-pick] Fix amp error cp (#46272)

* [Eager] Fix ocr (#46124)

* fix linspace error in amp

* fix log

* fix amp error

* fix ocr error which caused by amp

* add more check

* rename dtype ns

* [Eager Bug fix]Fix Detection (#46147)

* fix linspace error in amp

* fix log

* fix amp error

* Revert "Simplify size op impl (#45808)"

This reverts commit c252b1de.

* fix_seg

* fix detection
Co-authored-by: NChen Weihang <sunny_cwh@163.com>
Co-authored-by: NChen Weihang <sunny_cwh@163.com>

paddle/fluid/eager/auto_code_generator/eager_generator.cc

@@ -1797,6 +1797,15 @@ static std::pair<std::string, std::string> GenerateForwardFunctionContents(
     generated_function_body += amp_context;
     generated_function_body += "\n";
   }
+
+  if (!forward_inplace_map.empty()) {
+    generated_function_body +=
+        "  auto current_level = egr::Controller::Instance().GetAMPLevel();\n";
+    generated_function_body +=
+        "  "
+        "egr::Controller::Instance().SetAMPLevel(paddle::imperative::AmpLevel::"
+        "O0);\n";
+  }
   // forward ins insert
   const char* FWD_INS_MAP_TEMPLATE =
       "  std::map<std::string, "
@@ -1999,6 +2008,10 @@ static std::pair<std::string, std::string> GenerateForwardFunctionContents(
     }
     trace_op_body_str += out_tensor_str;
   }
+  if (!forward_inplace_map.empty()) {
+    trace_op_body_str +=
+        "  egr::Controller::Instance().SetAMPLevel(current_level);\n";
+  }
   trace_op_body_str += "\n";
   VLOG(6) << "Converted Output VarBase to EagerVariable(s)";
   /* ------ END Generate TraceOp ----- */

paddle/fluid/pybind/eager_method.cc

@@ -47,7 +47,9 @@ typedef SSIZE_T ssize_t;
 #include "pybind11/numpy.h"
 #include "pybind11/pybind11.h"
 #pragma GCC diagnostic ignored "-Wmissing-field-initializers"
+#include "paddle/fluid/eager/amp_utils.h"
 #include "paddle/fluid/eager/api/generated/eager_generated/forwards/dygraph_functions.h"
+#include "paddle/fluid/eager/eager_amp_auto_cast.h"
 #include "paddle/fluid/framework/python_headers.h"
 #include "paddle/fluid/memory/allocation/mmap_allocator.h"
 #include "paddle/fluid/pybind/tensor_py.h"
@@ -1171,6 +1173,17 @@ static PyObject* tensor_method__setitem_eager_tensor(TensorObject* self,
       // Release gil and do tracing
       py::gil_scoped_release release;
       // use inplace set_value_ operator
+      if (value_tensor.initialized() &&
+          (self->tensor.dtype() != value_tensor.dtype())) {
+        paddle::small_vector<std::vector<paddle::experimental::Tensor>,
+                             egr::kSlotSmallVectorSize>
+            tmps = {{self->tensor}, {value_tensor}};
+        auto amp_dtype = egr::GetAmpDestDtype("set_value", tmps);
+        self->tensor = egr::EagerAmpAutoCast(
+            self->tensor.name(), self->tensor, amp_dtype, "set_value");
+        value_tensor = egr::EagerAmpAutoCast(
+            value_tensor.name(), value_tensor, amp_dtype, "set_value");
+      }
       self->tensor = set_value__dygraph_function(
           self->tensor, value_tensor, {}, {}, {}, attrs);
     }

paddle/phi/kernels/cpu/scale_kernel.cc

@@ -58,6 +58,7 @@ PD_REGISTER_KERNEL(scale,
                    float,
                    double,
                    phi::dtype::bfloat16,
+                   phi::dtype::float16,
                    uint8_t,
                    int8_t,
                    int16_t,

paddle/phi/kernels/cpu/set_value_grad_kernel.cc

@@ -26,4 +26,5 @@ PD_REGISTER_KERNEL(set_value_grad,
                    double,
                    int,
                    int64_t,
-                   bool) {}
+                   bool,
+                   phi::dtype::float16) {}

paddle/phi/kernels/cpu/set_value_kernel.cc

@@ -26,7 +26,8 @@ PD_REGISTER_KERNEL(set_value,
                    double,
                    int,
                    int64_t,
-                   bool) {}
+                   bool,
+                   phi::dtype::float16) {}
 PD_REGISTER_KERNEL(set_value_with_tensor,
                    CPU,
                    ALL_LAYOUT,
@@ -35,4 +36,5 @@ PD_REGISTER_KERNEL(set_value_with_tensor,
                    double,
                    int,
                    int64_t,
-                   bool) {}
+                   bool,
+                   phi::dtype::float16) {}

paddle/phi/kernels/funcs/eigen/scale.cc

@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 #include "paddle/phi/common/bfloat16.h"
 #include "paddle/phi/common/complex.h"
+#include "paddle/phi/common/float16.h"
 #include "paddle/phi/kernels/funcs/eigen/eigen_function.h"
 
 namespace phi {
@@ -41,6 +42,7 @@ struct EigenScale<Eigen::DefaultDevice, T> {
 template struct EigenScale<Eigen::DefaultDevice, float>;
 template struct EigenScale<Eigen::DefaultDevice, double>;
 template struct EigenScale<Eigen::DefaultDevice, dtype::bfloat16>;
+template struct EigenScale<Eigen::DefaultDevice, dtype::float16>;
 template struct EigenScale<Eigen::DefaultDevice, uint8_t>;
 template struct EigenScale<Eigen::DefaultDevice, int8_t>;
 template struct EigenScale<Eigen::DefaultDevice, int16_t>;

paddle/phi/kernels/gpu/set_value_grad_kernel.cu

@@ -26,4 +26,5 @@ PD_REGISTER_KERNEL(set_value_grad,
                    double,
                    int,
                    int64_t,
-                   bool) {}
+                   bool,
+                   phi::dtype::float16) {}

paddle/phi/kernels/gpu/set_value_kernel.cu

@@ -26,7 +26,8 @@ PD_REGISTER_KERNEL(set_value,
                    double,
                    int,
                    int64_t,
-                   bool) {}
+                   bool,
+                   paddle::platform::float16) {}
 PD_REGISTER_KERNEL(set_value_with_tensor,
                    GPU,
                    ALL_LAYOUT,
@@ -35,4 +36,5 @@ PD_REGISTER_KERNEL(set_value_with_tensor,
                    double,
                    int,
                    int64_t,
-                   bool) {}
+                   bool,
+                   paddle::platform::float16) {}