Improve overflow handling (#2944)

Co-authored-by: Jeff Rasley <jerasley@microsoft.com>

deepspeed/runtime/fp16/loss_scaler.py

@@ -16,6 +16,8 @@
 #    https://github.com/NVIDIA/Megatron-LM/blob/master/fp16/loss_scaler.py
 #Commit: 93ab4bea59dc5cbf97c079d313741866af4deac9
 
+import torch
+
 INITIAL_LOSS_SCALE = 'init_scale'
 SCALE_WINDOW = 'scale_window'
 DELAYED_SHIFT = 'delayed_shift'
@@ -35,6 +37,7 @@ class LossScalerBase:
     """
     def __init__(self, cur_scale):
         self.cur_scale = cur_scale
+        self.dynamic = False
 
     @property
     def loss_scale(self):
@@ -117,6 +120,7 @@ def __init__(self,
         self.cur_hysteresis = delayed_shift
         self.consecutive_hysteresis = consecutive_hysteresis
         self.raise_error_at_min_scale = raise_error_at_min_scale
+        self.dynamic = True
 
     # `params` is a list / generator of torch.Variable
     def has_overflow_serial(self, params):
@@ -170,6 +174,18 @@ def update_scale(self, overflow):
         self.cur_iter += 1
 
 
+# Although loss scaling is only defined for fp16, yet for backwards compatibility
+# we still create a scaler for other dtypes (fp32, bf16) which does not perform any scaling.
+def CreateLossScaler(dtype, static_loss_scale, dynamic_scaling, dynamic_loss_args):
+    if dtype == torch.half and dynamic_scaling:
+        if dynamic_loss_args is None:
+            return DynamicLossScaler()
+        return DynamicLossScaler(**dynamic_loss_args)
+
+    loss_scale_value = static_loss_scale if dtype == torch.half else 1.0
+    return LossScaler(scale=loss_scale_value)
+
+
 ##############################################################
 # Example usage below here -- assuming it's in a separate file
 ##############################################################

deepspeed/runtime/zero/stage3.py

@@ -10,7 +10,7 @@
 
 from deepspeed.runtime import ZeROOptimizer
 from deepspeed.utils import logger
-from deepspeed.runtime.fp16.loss_scaler import LossScaler, DynamicLossScaler
+from deepspeed.runtime.fp16.loss_scaler import CreateLossScaler
 from deepspeed.runtime.comm.coalesced_collectives import reduce_scatter_coalesced
 from deepspeed.runtime.utils import inf, get_global_norm, is_model_parallel_parameter
 from deepspeed.runtime.zero.partition_parameters import *
@@ -332,18 +332,11 @@ def __init__(self,
         #exit(0)
 
         # we may have a way of fusing dynamic scale. Do not support for now
-        if self.dtype == torch.float or not dynamic_loss_scale:
-            loss_scale_value = 1.0 if self.dtype == torch.float else static_loss_scale
-
-            self.dynamic_loss_scale = False
-            self.loss_scaler = LossScaler(scale=loss_scale_value)
-        else:
-            if dynamic_loss_args is None:
-                self.loss_scaler = DynamicLossScaler()
-            else:
-                self.loss_scaler = DynamicLossScaler(**dynamic_loss_args)
-
-            self.dynamic_loss_scale = True
+        self.loss_scaler = CreateLossScaler(dtype=self.dtype,
+                                            static_loss_scale=static_loss_scale,
+                                            dynamic_scaling=dynamic_loss_scale,
+                                            dynamic_loss_args=dynamic_loss_args)
+        self.dynamic_loss_scale = self.loss_scaler.dynamic
 
         self.debug_fp16_grads = [{} for _ in self.fp16_groups]
 
@@ -1844,11 +1837,10 @@ def _overflow_clean_up(self, prev_scale):
         see_memory_usage('After overflow after clearing gradients', force=False)
 
         if dist.get_rank() == 0:
-            logger.info(
-                "[deepspeed] OVERFLOW! Rank {} Skipping step. Attempted loss scale: {}, "
-                "reducing to {}".format(dist.get_rank(),
-                                        prev_scale,
-                                        self.loss_scale))
+            overflow_msg = f"[deepspeed] OVERFLOW! Rank {dist.get_rank()} Skipping step."
+            if self.dtype == torch.half:
+                overflow_msg += f" Attempted loss scale: {prev_scale}, reducing to {self.loss_scale}"
+            logger.info(overflow_msg)
 
     @instrument_w_nvtx
     def _overflow_check_and_loss_scale_update(self):

deepspeed/runtime/zero/stage_1_and_2.py

@@ -9,7 +9,7 @@
 from collections import OrderedDict
 
 from deepspeed.runtime import ZeROOptimizer
-from deepspeed.runtime.fp16.loss_scaler import LossScaler, DynamicLossScaler
+from deepspeed.runtime.fp16.loss_scaler import CreateLossScaler
 from deepspeed.runtime.utils import (bwc_tensor_model_parallel_rank,
                                      get_global_norm,
                                      empty_cache,
@@ -506,21 +506,11 @@ def __init__(self,
         self.external_loss_scale = None
 
         # we may have a way of fusing dynamic scale. Do not support for now
-        if self.dtype == torch.float or self.dtype == torch.bfloat16 or not dynamic_loss_scale:
-            loss_scale_value = 1.0 if (
-                (self.dtype == torch.float) or
-                (self.dtype == torch.bfloat16)) else static_loss_scale
-
-            self.dynamic_loss_scale = False
-            self.loss_scaler = LossScaler(scale=loss_scale_value)
-            cur_iter = 0
-        else:
-            if dynamic_loss_args is None:
-                self.loss_scaler = DynamicLossScaler()
-            else:
-                self.loss_scaler = DynamicLossScaler(**dynamic_loss_args)
-
-            self.dynamic_loss_scale = True
+        self.loss_scaler = CreateLossScaler(dtype=self.dtype,
+                                            static_loss_scale=static_loss_scale,
+                                            dynamic_scaling=dynamic_loss_scale,
+                                            dynamic_loss_args=dynamic_loss_args)
+        self.dynamic_loss_scale = self.loss_scaler.dynamic
 
         see_memory_usage("Before initializing optimizer states", force=True)
         self.initialize_optimizer_states()
@@ -1788,11 +1778,10 @@ def step(self, closure=None):
         self._update_scale(self.overflow)
         if self.overflow:
             if dist.get_rank() == 0:
-                logger.info(
-                    "[deepspeed] OVERFLOW! Rank {} Skipping step. Attempted loss scale: {}, "
-                    "reducing to {}".format(dist.get_rank(),
-                                            prev_scale,
-                                            self.loss_scale))
+                overflow_msg = f"[deepspeed] OVERFLOW! Rank {dist.get_rank()} Skipping step."
+                if self.dtype == torch.half:
+                    overflow_msg += f" Attempted loss scale: {prev_scale}, reducing to {self.loss_scale}"
+                logger.info(overflow_msg)
 
             see_memory_usage('After overflow before clearing gradients')
             self.zero_grad(set_to_none=True)