after experiments, before new arg 'distrib-opt-comm-buf-1m'.

megatron/optimizer/distrib_optimizer.py

@@ -362,11 +362,11 @@ class DistributedOptimizer(MixedPrecisionOptimizer):
                 sub_view_items.append((model_index, dtype, sub_views))
 
         # >>>
-        from lutil import pax
-        pax(0, {
-            "gbuf_view_items" : [(a,b,c.shape) for a,b,c in gbuf_view_items],
-            "sub_view_items" : [(a,b,c.shape) for a,b,c in sub_view_items],
-        })
+        # from lutil import pax
+        # pax(0, {
+        #     "gbuf_view_items" : [(a,b,"%d / %s" % (len(c), [ d.nelement() for d in c ])) for a,b,c in gbuf_view_items],
+        #     "sub_view_items" : [(a,b,"%d / %s" % (len(c), [ d.nelement() for d in c ])) for a,b,c in sub_view_items],
+        # })
         # <<<
 
         return sub_view_items
@@ -432,10 +432,10 @@ class DistributedOptimizer(MixedPrecisionOptimizer):
         data_parallel_group = mpu.get_data_parallel_group()
 
         sub_numel = 1 * 1048576
-        sub_view_items = self.get_model_grad_buffer_dp_views_SUB(sub_numel)
+        gbuf_view_items = self.get_model_grad_buffer_dp_views_SUB(sub_numel)
         for model_index, dtype, gbuf_views in gbuf_view_items:
-            gbuf = self.models[model_index]._grad_buffers[dtype].data
-            gbuf /= data_parallel_world_size
+            # gbuf = self.models[model_index]._grad_buffers[dtype].data
+            # gbuf /= data_parallel_world_size
             torch.distributed.reduce_scatter(
                 gbuf_views[data_parallel_rank],
                 gbuf_views,
@@ -444,7 +444,7 @@ class DistributedOptimizer(MixedPrecisionOptimizer):
         timers('backward-params-all-reduce').stop()
     def reduce_model_grads(self, *args):
         # >>>
-        # return
+        return
         # <<<
         # self.reduce_model_grads_0(*args)
         self.reduce_model_grads_1(*args)
@@ -475,6 +475,49 @@ class DistributedOptimizer(MixedPrecisionOptimizer):
                     param.detach().copy_(param.main_grad)
 
         timers('backward-params-all-gather').stop()
+    # def gather_model_params_1(self, args, timers):
+
+    #     timers('backward-params-all-gather').start()
+
+    #     data_parallel_rank = mpu.get_data_parallel_rank()
+    #     data_parallel_group = mpu.get_data_parallel_group()
+
+    #     # All-gather updated main params.
+    #     # - All grad buffer views are guaranteed to have the same num elements
+    #     #   across all data parallel ranks, with grad buffer padding that is done
+    #     #   in distributed.py. Thus, all sub-views will have consistent start/end
+    #     #   indexes across data parallel ranks.
+    #     gbuf_view_items = self.get_model_grad_buffer_dp_views()
+
+    #     # sub_view_numel = 1 * 1024
+    #     # sub_view_numel = 1 * 131072
+    #     sub_view_numel = 256 * 1048576
+    #     for model_index, dtype, gbuf_views in gbuf_view_items:
+
+    #         # ** Sanity check. ** (should be unnecessary; see comment above)
+    #         view_numel = gbuf_views[0].nelement()
+    #         for view in gbuf_views:
+    #             assert view.nelement() == view_numel
+
+    #         for start_index in range(0, view_numel, sub_view_numel):
+
+    #             end_index = min(view_numel, start_index + sub_view_numel)
+    #             sub_views = [ t[start_index:end_index] for t in gbuf_views ]
+
+    #             torch.distributed.all_gather(
+    #                 sub_views,
+    #                 sub_views[data_parallel_rank],
+    #                 group = data_parallel_group,
+    #             )
+
+    #     # Each model param now contains its updated values in its
+    #     # '.main_grad' field.
+    #     for model in self.models:
+    #         for dtype, param_map in model._grad_buffer_param_index_map.items():
+    #             for param in param_map:
+    #                 param.detach().copy_(param.main_grad)
+
+    #     timers('backward-params-all-gather').stop()
     def gather_model_params_1(self, args, timers):
 
         timers('backward-params-all-gather').start()
@@ -487,26 +530,15 @@ class DistributedOptimizer(MixedPrecisionOptimizer):
         #   across all data parallel ranks, with grad buffer padding that is done
         #   in distributed.py. Thus, all sub-views will have consistent start/end
         #   indexes across data parallel ranks.
-        gbuf_view_items = self.get_model_grad_buffer_dp_views()
 
-        # sub_view_numel = 1 * 1024
-        # sub_view_numel = 1 * 131072
-        sub_view_numel = 1 * 1048576
+        # sub_numel = 1 * 1024
+        # sub_numel = 1 * 131072
+        sub_numel = 1024 * 1048576
+        gbuf_view_items = self.get_model_grad_buffer_dp_views_SUB(sub_numel)
         for model_index, dtype, gbuf_views in gbuf_view_items:
-
-            # ** Sanity check. ** (should be unnecessary; see comment above)
-            view_numel = gbuf_views[0].nelement()
-            for view in gbuf_views:
-                assert view.nelement() == view_numel
-
-            for start_index in range(0, view_numel, sub_view_numel):
-
-                end_index = min(view_numel, start_index + sub_view_numel)
-                sub_views = [ t[start_index:end_index] for t in gbuf_views ]
-
             torch.distributed.all_gather(
-                    sub_views,
-                    sub_views[data_parallel_rank],
+                gbuf_views,
+                gbuf_views[data_parallel_rank],
                 group = data_parallel_group,
             )
 
@@ -641,7 +673,7 @@ class DistributedOptimizer(MixedPrecisionOptimizer):
     #     timers('backward-params-all-gather').stop()
     def gather_model_params(self, *args):
         # >>>
-        return
+        # return
         # <<<
         # self.gather_model_params_0(*args)
         self.gather_model_params_1(*args)