Change zero_grad() argument to match pytorch (#2741)

deepspeed/runtime/fp16/fused_optimizer.py

@@ -130,14 +130,14 @@ class FP16_Optimizer(DeepSpeedOptimizer):
 
         return
 
-    def zero_grad(self, set_grads_to_None=True):
+    def zero_grad(self, set_to_none=False):
         """
         Zero FP16 parameter grads.
         """
         # For speed, set model fp16 grad to None by default
         for group in self.fp16_groups:
             for p in group:
-                if set_grads_to_None:
+                if set_to_none:
                     p.grad = None
                 else:
                     if p.grad is not None:

deepspeed/runtime/fp16/unfused_optimizer.py

@@ -110,7 +110,7 @@ class FP16_UnfusedOptimizer(DeepSpeedOptimizer):
 
         self.initialize_optimizer_states()
 
-    def zero_grad(self, set_grads_to_None=True):
+    def zero_grad(self, set_to_none=False):
         """
         Zero FP16 parameter grads.
         """
@@ -118,7 +118,7 @@ class FP16_UnfusedOptimizer(DeepSpeedOptimizer):
         # For speed, set model fp16 grad to None by default
         for group in self.fp16_groups:
             for p in group:
-                if set_grads_to_None:
+                if set_to_none:
                     p.grad = None
                 else:
                     if p.grad is not None:

deepspeed/runtime/zero/stage3.py

@@ -1509,7 +1509,7 @@ class DeepSpeedZeroOptimizer_Stage3(ZeROOptimizer):
         return params_in_partition, params_not_in_partition, first_offset
 
     @instrument_w_nvtx
-    def zero_grad(self, set_grads_to_None=True):
+    def zero_grad(self, set_to_none=False):
         """
         Zero FP16 parameter grads.
         """
@@ -1519,7 +1519,7 @@ class DeepSpeedZeroOptimizer_Stage3(ZeROOptimizer):
         # For speed, set model fp16 grad to None by default
         for group in self.fp16_groups:
             for p in group:
-                if set_grads_to_None:
+                if set_to_none:
                     if p.grad is not None and p.grad.is_cuda:
                         p.grad.record_stream(torch.cuda.current_stream())
                     p.grad = None
@@ -1708,7 +1708,7 @@ class DeepSpeedZeroOptimizer_Stage3(ZeROOptimizer):
         self.fp32_partitioned_groups_flat[sub_group_id].grad = single_grad_partition
 
         # release all the gradient since we have already created a necessary copy in dp_grad_partition
-        self.zero_grad()
+        self.zero_grad(set_to_none=True)
 
         for grad in filter(lambda g: g.is_cuda, self.averaged_gradients[sub_group_id]):
             grad.record_stream(torch.cuda.current_stream())
@@ -1816,7 +1816,7 @@ class DeepSpeedZeroOptimizer_Stage3(ZeROOptimizer):
 
     def _overflow_clean_up(self, prev_scale):
         see_memory_usage('After overflow before clearing gradients', force=False)
-        self.zero_grad()
+        self.zero_grad(set_to_none=True)
 
         if self.offload_optimizer:
             self.reset_cpu_buffers()

deepspeed/runtime/zero/stage_1_and_2.py

@@ -753,7 +753,7 @@ class DeepSpeedZeroOptimizer(ZeROOptimizer):
         # No need to keep the gradients anymore.
         # All gradients required by the step
         # are in self.averaged_gradients
-        self.zero_grad()
+        self.zero_grad(set_to_none=True)
         see_memory_usage(f"End ipg_epilogue")
 
     # resets all partition to no reduced
@@ -1526,7 +1526,7 @@ class DeepSpeedZeroOptimizer(ZeROOptimizer):
 
         return params_in_partition, params_not_in_partition, first_offset
 
-    def zero_grad(self, set_grads_to_None=True):
+    def zero_grad(self, set_to_none=False):
         """
         Zero FP16 parameter grads.
         """
@@ -1534,7 +1534,7 @@ class DeepSpeedZeroOptimizer(ZeROOptimizer):
         # For speed, set model fp16 grad to None by default
         for group in self.bit16_groups:
             for p in group:
-                if set_grads_to_None:
+                if set_to_none:
                     p.grad = None  # epilogue and in step
                 else:
                     if p.grad is not None:
@@ -1766,7 +1766,7 @@ class DeepSpeedZeroOptimizer(ZeROOptimizer):
                                             self.loss_scale))
 
             see_memory_usage('After overflow before clearing gradients')
-            self.zero_grad()
+            self.zero_grad(set_to_none=True)
             if self.cpu_offload:
                 self.reset_cpu_buffers()
             else: