[Inference] optimize device memory occupation (#64493)

Author: yuanlehome

paddle/fluid/pir/transforms/general/constant_folding_pass.cc

@@ -227,6 +227,13 @@ class ConstantFoldingPattern : public pir::RewritePattern {
       }
     }
     rewriter.EraseOp(op);
+
+    // NOTE(liuyuanle): Here, we release one useless variable after another to
+    // effectively reduce peak memory usage.
+    if (deleted_vars_.size() > 0) {
+      scope_->EraseVars(deleted_vars_);
+      deleted_vars_.clear();
+    }
     VLOG(4) << "constant_folding_pass applied rewrite on [" << op->name()
             << "] op";
   }
@@ -306,6 +313,8 @@ class ConstantFoldingPattern : public pir::RewritePattern {
     if (op->operand_source(index).use_count() > 1) {
       from_op->set_attribute(kAttrIsPersistable,
                              rewriter.array_attr({rewriter.bool_attr(true)}));
+    } else {
+      deleted_vars_.push_back(var_name);
     }
     return from_op;
   }
@@ -397,6 +406,7 @@ class ConstantFoldingPattern : public pir::RewritePattern {
   phi::Place place_;
   paddle::framework::Scope* scope_;
   paddle::framework::interpreter::ExecutionConfig* exe_config_;
+  mutable std::vector<std::string> deleted_vars_;
 };
 
 class ConstantFoldingPatternForTrain : public ConstantFoldingPattern {

paddle/fluid/pir/transforms/general/dead_code_elimination_pass.cc

@@ -75,6 +75,7 @@ class DeadCodeEliminationPass : public pir::Pass {
         deleted_vars->push_back(constant_tensor_op.tensor_name());
       }
       op->Erase();
+      VLOG(4) << "erase op: " << op->name();
       (*num_erasers)++;
     }