[cherry pick] Refine param conversion logic in layer.to (#38068)

优化layer.to实现逻辑，相关pr：
Remove additional warnning in layer.to ( #36700)
Refine param conversion logic in layer.to ( #36862)
Fix Layer.to() of device bug ( #37156)

paddle/fluid/pybind/pybind.cc

@@ -527,8 +527,13 @@ PYBIND11_MODULE(core_noavx, m) {
 
 #if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
   m.def("cudnn_version", &platform::CudnnVersion);
+  m.def("gpu_memory_available", []() {
+    size_t available = 0;
+    size_t total = 0;
+    paddle::platform::GpuMemoryUsage(&available, &total);
+    return available;
+  });
 #endif
-
 #ifdef PADDLE_WITH_NCCL
   m.def("nccl_version", &GetNCCLVersion);
 #endif

python/paddle/fluid/dygraph/amp/auto_cast.py

@@ -119,10 +119,7 @@ def _in_amp_guard():
 
 
 @dygraph_only
-def pure_fp16_initialize(enable_pure_fp16, models, optimizers):
-    if not enable_pure_fp16:
-        return models, optimizers
-
+def pure_fp16_initialize(models):
     for idx in range(len(models)):
         for layer in models[idx].sublayers(include_self=True):
             layer._casted_by_pure_fp16 = True
@@ -132,43 +129,7 @@ def pure_fp16_initialize(enable_pure_fp16, models, optimizers):
                         paddle.nn.BatchNorm, paddle.nn.LayerNorm)):
                     continue
                 layer.to(dtype='float16')
-
-    for idx_opt in range(len(optimizers)):
-        # update _param_groups
-        if getattr(optimizers[idx_opt], '_param_groups', None) and isinstance(
-                optimizers[idx_opt]._param_groups[0], dict):
-            for param_group in optimizers[idx_opt]._param_groups:
-                for i, param in enumerate(param_group['params']):
-                    for idx_model in range(len(models)):
-                        for layer in models[idx_model].sublayers(
-                                include_self=True):
-                            if id(param) in layer._parameters_transform_map:
-                                param_group['params'][
-                                    i] = layer._parameters_transform_map[id(
-                                        param)][0]
-            for param_group in optimizers[idx_opt]._parameter_list:
-                params = param_group['params']
-                for i, param in enumerate(params):
-                    for idx_model in range(len(models)):
-                        for layer in models[idx_model].sublayers(
-                                include_self=True):
-                            if id(param) in layer._parameters_transform_map:
-                                params[i] = layer._parameters_transform_map[id(
-                                    param)][0]
-        # update _parameter_list
-        else:
-            for i, param in enumerate(optimizers[idx_opt]._parameter_list):
-                for idx_model in range(len(models)):
-                    for layer in models[idx_model].sublayers(include_self=True):
-                        if id(param) in layer._parameters_transform_map:
-                            optimizers[idx_opt]._parameter_list[
-                                i] = layer._parameters_transform_map[id(param)][
-                                    0]
-                            if hasattr(optimizers[idx_opt], '_param_groups'):
-                                optimizers[idx_opt]._param_groups[
-                                    i] = layer._parameters_transform_map[id(
-                                        param)][0]
-    return models, optimizers
+    return models
 
 
 def check_models(models):
@@ -397,8 +358,7 @@ def amp_decorate(models,
             "optimizers must be either a single optimizer or a list of optimizers."
         )
 
-    models, optimizers = pure_fp16_initialize(
-        enable_pure_fp16=True, models=models, optimizers=optimizers)
+    models = pure_fp16_initialize(models=models)
 
     # supprot master_weight    
     for idx_opt in range(len(optimizers)):

python/paddle/fluid/dygraph/layers.py

@@ -121,9 +121,6 @@ class Layer(core.Layer):
         self._forward_pre_hooks = collections.OrderedDict()
         self._forward_post_hooks = collections.OrderedDict()
 
-        self._parameters_transform_map = {}
-        self._buffers_transform_map = {}
-
         self._casted_by_pure_fp16 = False
 
         self._state_dict_hooks = collections.OrderedDict()
@@ -1473,24 +1470,14 @@ class Layer(core.Layer):
             if param is not None:
                 with no_grad():
                     param_applied = func(param, device, dtype, blocking)
-                    assert param.is_leaf
-                    param_applied.stop_gradient = param.stop_gradient
-                    self._parameters[key] = param_applied
 
                 if param.grad is not None:
                     with no_grad():
                         grad_applied = func(param._grad_ivar(), device, dtype,
                                             blocking)
 
-                        grad_applied.stop_gradient = param._grad_ivar(
-                        ).stop_gradient
-                        self._parameters[key]._set_grad_ivar(grad_applied)
-
-            self._parameters_transform_map[id(param)] = [param_applied, key]
-
         for key, buf in self._buffers.items():
             self._buffers[key] = func(buf, device, dtype, blocking)
-            self._buffers_transform_map[id(buf)] = [self._buffers[key], key]
 
     def to(self, device=None, dtype=None, blocking=None):
         '''
@@ -1568,24 +1555,54 @@ class Layer(core.Layer):
             if dtype is None:
                 dtype = t.dtype
 
-            new_t = t._copy_to(device, blocking)
-            if isinstance(t, framework.ParamBase):
-                if dtype is not None and dtype != t.dtype:
-                    framework._dygraph_tracer().trace_op(
-                        type='cast',
-                        inputs={'X': new_t},
-                        outputs={'Out': new_t},
-                        attrs={
-                            'in_dtype': t.dtype,
-                            'out_dtype': convert_np_dtype_to_dtype_(dtype)
-                        })
+            if type(dtype) is str:
+                dtype = convert_np_dtype_to_dtype_(dtype)
+
+            # 1. gpu place need to determine whether the memory is sufficient for allocation:
+            if t.place.is_gpu_place():
+                # for gpu, minimum memory allocation unit is 256 bytes.
+                size_dtype = core.size_of_dtype(dtype)
+                # Note(zhangbo): Paddle GPU minimum memory allocation unit is 256 bytes, waiting_alloc_memory will comput ‘t’ occupied memory space.
+                # Coefficient 1.2 is used to avoid OOM that may occur in this critical state when the memory is just enough.
+                waiting_alloc_memory = (
+                    (np.prod(t.shape) * size_dtype) / 256 + 1) * 256 * 1.2
+                gpu_memory_available = core.gpu_memory_available()
+                if gpu_memory_available < waiting_alloc_memory:
+                    # Copy param / Tensor to cpu
+                    t_used = t._copy_to(paddle.CPUPlace(),
+                                        blocking)  # k-v type will error
+                    # Release mem of t
+                    t.value().get_tensor()._clear()
+                else:
+                    t_used = t
+            else:
+                t_used = t
+
+            # 2. cast param / Tensor to dtype
+            if dtype is not None and dtype != t_used.dtype:
+                with paddle.fluid.framework._dygraph_place_guard(
+                        place=t_used.place):
+                    t_casted = t_used.cast(dtype=dtype)
+            else:
+                t_casted = t_used
+
+            # 3. Copy casted cpu param / Tensor to device
+            if device is not None and not t_casted.place._equals(device):
+                new_t = t_casted._copy_to(device, blocking)
             else:
-                if dtype is not None and dtype != t.dtype:
-                    new_t = new_t.cast(dtype=dtype)
+                new_t = t_casted
 
-            return new_t
+            # 4. share Tensor to origin param / Tensor
+            dst_tensor = t.value().get_tensor()
+            src_tensor = new_t.value().get_tensor()
+            dst_tensor._share_data_with(src_tensor)
 
+            return t
+
+        with warnings.catch_warnings():
+            warnings.filterwarnings("ignore", category=UserWarning)
             self._apply(transform, device, dtype, blocking)
+
         self._dtype = dtype
 
     # [aliases] Compatible with old method names