[AMP] Support AMP-O2 for bfloat16 (#45541)

* support bfloat16 for amp_decorate

* add check_finite for bf16

* fix bug

* add ut

* add ut

* refine code

paddle/phi/kernels/gpu/amp_kernel.cu

@@ -357,7 +357,8 @@ PD_REGISTER_KERNEL(check_finite_and_unscale,
                    phi::CheckFiniteAndUnscaleKernel,
                    float,
                    double,
-                   phi::dtype::float16) {}
+                   phi::dtype::float16,
+                   phi::dtype::bfloat16) {}
 
 PD_REGISTER_KERNEL(update_loss_scaling,
                    GPU,
@@ -365,6 +366,7 @@ PD_REGISTER_KERNEL(update_loss_scaling,
                    phi::UpdateLossScalingKernel,
                    float,
                    double,
-                   phi::dtype::float16) {
+                   phi::dtype::float16,
+                   phi::dtype::bfloat16) {
   kernel->InputAt(1).SetBackend(phi::Backend::ALL_BACKEND);
 }

python/paddle/amp/auto_cast.py

@@ -81,21 +81,23 @@ def auto_cast(enable=True,
 def decorate(models,
              optimizers=None,
              level='O1',
+             dtype='float16',
              master_weight=None,
              save_dtype=None):
     """
     Decorate models and optimizers for auto-mixed-precision. When level is O1(amp), the decorate will do nothing. 
-    When level is O2(pure fp16), the decorate will cast all parameters of models to FP16, except BatchNorm and LayerNorm.
+    When level is O2(pure float16/bfloat16), the decorate will cast all parameters of models to float16/bfloat16, except BatchNorm and LayerNorm.
     
-    Commonly, it is used together with `auto_cast` to achieve Pure fp16 in imperative mode.
+    Commonly, it is used together with `auto_cast` to achieve Pure float16/bfloat16 in imperative mode.
 
     Args:
         models(Layer|list of Layer, optional): The defined models by user, models must be either a single model or a list of models. Default is None.
         optimizers(Optimizer|list of Optimizer, optional): The defined optimizers by user, optimizers must be either a single optimizer or a list of optimizers. Default is None.
         level(str, optional): Auto mixed precision level. Accepted values are "O1" and "O2": O1 represent mixed precision, the decorator will do nothing; 
-             O2 represent Pure fp16, the decorator will cast all parameters of models to FP16, except BatchNorm and LayerNorm. Default is O1(amp)
+             O2 represent Pure float16/bfloat16, the decorator will cast all parameters of models to float16/bfloat16, except BatchNorm and LayerNorm. Default is O1(amp)
+        dtype(str, optional): Whether to use 'float16' or 'bfloat16'. Default is 'float16'.
         master_weight(bool, optinal): For level='O2', whether to use multi-precision during weight updating. If master_weight is None, in O2 level optimizer will use multi-precision. Default is None.
-        save_dtype(float, optional): The save model parameter dtype when use `paddle.save` or `paddle.jit.save`,it should be float16, float32, float64 or None.
+        save_dtype(float, optional): The save model parameter dtype when use `paddle.save` or `paddle.jit.save`,it should be float16, bfloat16, float32, float64 or None.
              The save_dtype will not change model parameters dtype, it just change the state_dict dtype. When save_dtype is None, the save dtype is same as model dtype. Default is None.
 
     Examples:
@@ -145,4 +147,5 @@ def decorate(models,
             output = model(data)
             print(output.dtype) # FP16
     """
-    return amp_decorate(models, optimizers, level, master_weight, save_dtype)
+    return amp_decorate(models, optimizers, level, dtype, master_weight,
+                        save_dtype)

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

@@ -76,7 +76,7 @@ AMP_RELATED_FLAGS_SETTING = {
     'FLAGS_cudnn_batchnorm_spatial_persistent': 1,
 }
 
-PURE_FP16_WHITE_LIST = {' '}
+PURE_FP16_WHITE_LIST = {''}
 PURE_FP16_BLACK_LIST = {
     'lookup_table',
     'lookup_table_v2',
@@ -91,7 +91,10 @@ PURE_FP16_BLACK_LIST = {
 }
 
 BF16_WHITE_LIST = {'conv2d', 'matmul_v2'}
-BF16_BLACK_LIST = {' '}
+BF16_BLACK_LIST = {''}
+
+PURE_BF16_WHITE_LIST = {''}
+PURE_BF16_BLACK_LIST = {''}
 
 _g_amp_state_ = None
 
@@ -118,8 +121,12 @@ def _update_list(custom_white_list,
             _white_list = copy.copy(PURE_FP16_WHITE_LIST)
             _black_list = copy.copy(PURE_FP16_BLACK_LIST)
     else:
+        if level == 'O1':
             _white_list = copy.copy(BF16_WHITE_LIST)
             _black_list = copy.copy(BF16_BLACK_LIST)
+        else:
+            _white_list = copy.copy(PURE_BF16_WHITE_LIST)
+            _black_list = copy.copy(PURE_BF16_BLACK_LIST)
     if custom_white_list and custom_black_list:
         for op_name in custom_white_list:
             if op_name in custom_black_list:
@@ -198,6 +205,16 @@ def pure_fp16_initialize(models):
     return models
 
 
+@dygraph_only
+def pure_bf16_initialize(models):
+    for idx in range(len(models)):
+        for layer in models[idx].sublayers(include_self=True):
+            layer._to_impl(dtype='bfloat16',
+                           include_sublayers=False,
+                           floating_only=True)
+    return models
+
+
 def check_models(models):
     for model in models:
         if not isinstance(model, paddle.nn.Layer):
@@ -424,6 +441,7 @@ class StateDictHook(object):
 def amp_decorate(models,
                  optimizers=None,
                  level='O1',
+                 dtype='float16',
                  master_weight=None,
                  save_dtype=None):
     """
@@ -436,9 +454,10 @@ def amp_decorate(models,
         models(Layer|list of Layer, optional): The defined models by user, models must be either a single model or a list of models. Default is None.
         optimizers(Optimizer|list of Optimizer, optional): The defined optimizers by user, optimizers must be either a single optimizer or a list of optimizers. Default is None.
         level(str, optional): Auto mixed precision level. Accepted values are "O1" and "O2": O1 represent mixed precision, the decorator will do nothing; 
-             O2 represent Pure fp16, the decorator will cast all parameters of models to FP16, except BatchNorm and LayerNorm. Default is O1(amp)
+             O2 represent Pure fp16/bf16, the decorator will cast all parameters of models to FP16/BF16, except BatchNorm and LayerNorm. Default is O1(amp)
+        dtype(str, optional): Whether to use 'float16' or 'bfloat16'. Default is 'float16'.
         master_weight(bool, optinal): For level='O2', whether to use multi-precision during weight updating. If master_weight is None, in O2 level optimizer will use multi-precision. Default is None.
-        save_dtype(float, optional): The save model parameter dtype when use `paddle.save` or `paddle.jit.save`,it should be float16, float32, float64 or None.
+        save_dtype(float, optional): The save model parameter dtype when use `paddle.save` or `paddle.jit.save`,it should be float16, bfloat16, float32, float64 or None.
              The save_dtype will not change model parameters dtype, it just change the state_dict dtype. When save_dtype is None, the save dtype is same as model dtype. Default is None.
 
     Examples:
@@ -510,8 +529,12 @@ def amp_decorate(models,
     else:
         raise TypeError(
             "models must be either a single model or a list of models.")
-
+    if dtype == 'float16':
         models = pure_fp16_initialize(models=models)
+    elif dtype == 'bfloat16':
+        models = pure_bf16_initialize(models=models)
+    else:
+        raise TypeError("dtype only support float16 or bfloat16.")
 
     if optimizers is not None:
         # check optimizers
@@ -538,7 +561,7 @@ def amp_decorate(models,
                     optimizers[idx_opt]._multi_precision = True
 
     if save_dtype is not None:
-        if not (save_dtype in ['float16', 'float32', 'float64']):
+        if not (save_dtype in ['float16', 'bfloat16', 'float32', 'float64']):
             raise ValueError(
                 "save_dtype can only be float16 float32 or float64, but your input save_dtype is %s."
                 % save_dtype)

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

@@ -132,6 +132,8 @@ class AmpScaler(object):
             self._found_inf = to_variable(np.array([0]).astype(np.bool_))
             self._temp_found_inf_fp16 = to_variable(
                 np.array([0]).astype(np.bool_))
+            self._temp_found_inf_bf16 = to_variable(
+                np.array([0]).astype(np.bool_))
             self._temp_found_inf_fp32 = to_variable(
                 np.array([0]).astype(np.bool_))
             self._scale = to_variable(
@@ -262,6 +264,7 @@ class AmpScaler(object):
                 optimizer._param_groups[0], dict):
             param_grads = []
             param_grads_fp16 = []
+            param_grads_bf16 = []
             param_grads_fp32 = []
             for group in optimizer._param_groups:
                 for param in group['params']:
@@ -270,6 +273,9 @@ class AmpScaler(object):
                         if param._grad_ivar(
                         ).dtype == core.VarDesc.VarType.FP16:
                             param_grads_fp16.append(param._grad_ivar())
+                        elif param._grad_ivar(
+                        ).dtype == core.VarDesc.VarType.BF16:
+                            param_grads_bf16.append(param._grad_ivar())
                         else:
                             param_grads_fp32.append(param._grad_ivar())
         else:
@@ -281,6 +287,10 @@ class AmpScaler(object):
                 param for param in param_grads
                 if param.dtype == core.VarDesc.VarType.FP16
             ]
+            param_grads_bf16 = [
+                param for param in param_grads
+                if param.dtype == core.VarDesc.VarType.BF16
+            ]
             param_grads_fp32 = [
                 param for param in param_grads
                 if param.dtype == core.VarDesc.VarType.FP32
@@ -293,6 +303,10 @@ class AmpScaler(object):
                 _legacy_C_ops.check_finite_and_unscale(
                     param_grads_fp16, self._scale, float_status,
                     param_grads_fp16, self._temp_found_inf_fp16)
+            if len(param_grads_bf16):
+                _legacy_C_ops.check_finite_and_unscale(
+                    param_grads_bf16, self._scale, float_status,
+                    param_grads_bf16, self._temp_found_inf_bf16)
             if len(param_grads_fp32):
                 _legacy_C_ops.check_finite_and_unscale(
                     param_grads_fp32, self._scale, float_status,
@@ -302,12 +316,16 @@ class AmpScaler(object):
                 _legacy_C_ops.check_finite_and_unscale(
                     param_grads_fp16, self._scale, param_grads_fp16,
                     self._temp_found_inf_fp16)
+            if len(param_grads_bf16):
+                _legacy_C_ops.check_finite_and_unscale(
+                    param_grads_bf16, self._scale, param_grads_bf16,
+                    self._temp_found_inf_bf16)
             if len(param_grads_fp32):
                 _legacy_C_ops.check_finite_and_unscale(
                     param_grads_fp32, self._scale, param_grads_fp32,
                     self._temp_found_inf_fp32)
 
-        self._found_inf = self._temp_found_inf_fp16 or self._temp_found_inf_fp32
+        self._found_inf = self._temp_found_inf_fp16 or self._temp_found_inf_bf16 or self._temp_found_inf_fp32
 
         optimizer_state["state"] = OptimizerState.UNSCALED
 

python/paddle/fluid/framework.py

@@ -1169,6 +1169,9 @@ def convert_np_dtype_to_dtype_(np_dtype):
         core.VarDesc.VarType: the data type in Paddle.
 
     """
+    if np_dtype == "bfloat16":
+        dtype = np.uint16
+    else:
         dtype = np.dtype(np_dtype)
     if dtype == np.float32:
         return core.VarDesc.VarType.FP32

python/paddle/fluid/tests/unittests/collective/fleet/test_imperative_auto_mixed_precision.py

@@ -1310,6 +1310,10 @@ class TestBf16(unittest.TestCase):
         paddle.seed(100)
         input = paddle.uniform((2, 4, 8, 8), dtype='float32', min=-1., max=1.)
         conv = paddle.nn.Conv2D(4, 6, (3, 3))
+        if amp_level == 'O2':
+            conv = paddle.amp.decorate(models=conv,
+                                       level=amp_level,
+                                       dtype='bfloat16')
         with paddle.amp.auto_cast(enable=enable_amp,
                                   level=amp_level,
                                   dtype='bfloat16'):

python/paddle/fluid/tests/unittests/collective/fleet/test_imperative_auto_mixed_precision_for_eager.py

@@ -1297,6 +1297,10 @@ class TestBf16(unittest.TestCase):
         paddle.seed(100)
         input = paddle.uniform((2, 4, 8, 8), dtype='float32', min=-1., max=1.)
         conv = paddle.nn.Conv2D(4, 6, (3, 3))
+        if amp_level == 'O2':
+            conv = paddle.amp.decorate(models=conv,
+                                       level=amp_level,
+                                       dtype='bfloat16')
         with paddle.amp.auto_cast(enable=enable_amp,
                                   level=amp_level,
                                   dtype='bfloat16'):