split optimization ops on pserver to independenty blocks

python/paddle/fluid/distribute_transpiler.py

@@ -25,6 +25,8 @@ LOOKUP_TABLE_TYPE = "lookup_table"
 LOOKUP_TABLE_GRAD_TYPE = "lookup_table_grad"
 RPC_CLIENT_VAR_NAME = "RPC_CLIENT_VAR"
 
+GLOBAL_BLOCK_IDX = 0
+
 
 class VarBlock:
     def __init__(self, varname, offset, size):
@@ -368,8 +370,8 @@ class DistributeTranspiler:
             else:
                 recv_inputs.append(single_trainer_var)
 
-        # step3
+        optimize_block = None
-        optimize_block = pserver_program.create_block(0)
+
         # step 4
         # Create a union-find data structure from optimize ops,
         # If two ops are connected, we could add these two ops
@@ -415,29 +417,34 @@ class DistributeTranspiler:
             else:
                 self._append_pserver_non_opt_ops(block, op)
 
-        append_block = optimize_block
         # append lr decay ops to the child block if exists
+        lr_decay_block = None
         lr_ops = self._get_lr_ops()
         if len(lr_ops) > 0:
+            lr_decay_block = pserver_program.create_block(GLOBAL_BLOCK_IDX)
             for _, op in enumerate(lr_ops):
-                self._append_pserver_non_opt_ops(append_block, op)
+                self._append_pserver_non_opt_ops(lr_decay_block, op)
-
-            append_block = pserver_program.create_block(append_block.idx)
 
         # append op to the current block
-        per_opt_block = append_block
+        per_opt_block = None
+        pre_block_idx = GLOBAL_BLOCK_IDX
+        if lr_decay_block is not None:
+            pre_block_idx = lr_decay_block.idx
         for idx, opt_op in enumerate(opt_op_on_pserver):
+            per_opt_block = pserver_program.create_block(pre_block_idx)
+            if optimize_block is None:
+                optimize_block = per_opt_block
             for _, op in enumerate(self.optimize_ops):
                 # optimizer is connected to itself
-                if ufind.is_connected(op, opt_op) and \
+                if ufind.is_connected(op, opt_op) and op not in global_ops:
-                    op not in global_ops:
                     __append_optimize_op__(op, per_opt_block)
-            if idx == len(opt_op_on_pserver) - 1 and global_ops:
-                per_opt_block = pserver_program.create_block(append_block.idx)
 
         # append global ops
+        opt_state_block = None
+        if global_ops:
+            opt_state_block = pserver_program.create_block(per_opt_block.idx)
         for glb_op in global_ops:
-            __append_optimize_op__(glb_op, per_opt_block)
+            __append_optimize_op__(glb_op, opt_state_block)
 
         # NOT USED: single block version:
         #
@@ -451,10 +458,11 @@ class DistributeTranspiler:
         prefetch_block = None
         if self.has_distributed_lookup_table:
             pserver_index = self.pserver_endpoints.index(endpoint)
-            self._create_table_optimize_block(pserver_index, pserver_program,
+            table_opt_block = self._create_table_optimize_block(
-                                              append_block)
+                pserver_index, pserver_program, opt_state_block or
+                pserver_program.global_block())
             prefetch_block = self._create_prefetch_block(
-                pserver_index, pserver_program, optimize_block)
+                pserver_index, pserver_program, table_opt_block)
 
         # NOTE: if has_distributed_lookup_table is False, then prefetch_block will
         # not be executed, so it's safe to use optimize_block to hold the place
@@ -724,6 +732,8 @@ class DistributeTranspiler:
             outputs=outputs,
             attrs=table_opt_op.attrs)
 
+        return table_opt_block
+
     # ====================== private transpiler functions =====================
     def _create_vars_from_blocklist(self,
                                     program,