Merge branch 'optimize_refactor_backward_compat' into 'optimize_refactor'

added backward compatibility to be able to load checkpoints

See merge request ADLR/megatron-lm!200

megatron/checkpointing.py

@@ -205,12 +205,16 @@ def load_checkpoint(model, optimizer, lr_scheduler, load_arg='load'):
     try:
         state_dict = torch.load(checkpoint_name, map_location='cpu')
     except ModuleNotFoundError:
+        from megatron.fp16_deprecated import loss_scaler
         # For backward compatibility.
         print_rank_0(' > deserializing using the old code structure ...')
         sys.modules['fp16.loss_scaler'] = sys.modules[
-            'megatron.fp16.loss_scaler']
+            'megatron.fp16_deprecated.loss_scaler']
+        sys.modules['megatron.fp16.loss_scaler'] = sys.modules[
+            'megatron.fp16_deprecated.loss_scaler']
         state_dict = torch.load(checkpoint_name, map_location='cpu')
         sys.modules.pop('fp16.loss_scaler', None)
+        sys.modules.pop('megatron.fp16.loss_scaler', None)
     except BaseException:
         print_rank_0('could not load the checkpoint')
         sys.exit()

megatron/fp16_deprecated/loss_scaler.py

+# coding=utf-8
+# Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""For backward compatibility, we need the class definitions to deserialize."""
+
+class LossScaler:
+    def __init__(self, scale=1):
+        self.cur_scale = scale
+
+class DynamicLossScaler:
+    def __init__(self,
+                 init_scale=2**32,
+                 scale_factor=2.,
+                 scale_window=1000,
+                 min_scale=1,
+                 delayed_shift=1,
+                 consecutive_hysteresis=False):
+        self.cur_scale = init_scale
+        self.cur_iter = 0
+        self.last_overflow_iter = -1
+        self.scale_factor = scale_factor
+        self.scale_window = scale_window
+        self.min_scale = min_scale
+        self.delayed_shift = delayed_shift
+        self.cur_hysteresis = delayed_shift
+        self.consecutive_hysteresis = consecutive_hysteresis
+