Merge branch 'develop' into up32_2

pyproject.toml

@@ -123,29 +123,6 @@ ignore = [
     "PLC0414",
 ]
 
-# UP018
-"python/paddle/distributed/fleet/meta_optimizers/localsgd_optimizer.py" = ["UP018"]
-"python/paddle/distributed/fleet/meta_optimizers/sharding/utils.py" = ["UP018"]
-"python/paddle/distributed/fleet/meta_optimizers/sharding_optimizer.py" = ["UP018"]
-"python/paddle/distributed/fleet/utils/pp_parallel_adaptor.py" = ["UP018"]
-"python/paddle/incubate/autograd/utils.py" = ["UP018"]
-"python/paddle/incubate/optimizer/gradient_merge.py" = ["UP018"]
-"python/paddle/nn/functional/loss.py" = ["UP018"]
-"python/paddle/distributed/passes/auto_parallel_gradient_merge.py" = ["UP018"]
-"python/paddle/hapi/dynamic_flops.py" = ["UP018"]
-"python/paddle/incubate/optimizer/pipeline.py" = ["UP018"]
-"python/paddle/static/nn/metric.py" = ["UP018"]
-"test/cinn/ops/test_fill_constant_op.py" = ["UP018"]
-"test/collective/fleet/parallel_dygraph_transformer.py" = ["UP018"]
-"test/collective/fleet/test_dgc_op.py" = ["UP018"]
-"test/dygraph_to_static/transformer_dygraph_model.py" = ["UP018"]
-"test/legacy_test/fleet_heter_ps_training.py" = ["UP018"]
-"test/legacy_test/test_fake_quantize_op.py" = ["UP018"]
-"test/legacy_test/test_generate_mask_labels_op.py" = ["UP018"]
-"test/legacy_test/test_imperative_transformer_sorted_gradient.py" = ["UP018"]
-"test/legacy_test/test_progressbar.py" = ["UP018"]
-"test/legacy_test/test_tensor_fill_.py" = ["UP018"]
-
 # B017
 "test/auto_parallel/spmd_rules/test_reshape_rule.py" = ["B017"]
 "test/dygraph_to_static/test_assert.py" = ["B017"]

python/paddle/distributed/passes/pipeline_scheduler_pass.py

@@ -406,6 +406,15 @@ def apply_pass(main_program, startup_program, pass_name, pass_attr={}):
         "FThenB",
         "1F1B",
     ], f"pipeline scheduler only support FThenB and 1F1B, but recieve {pass_name}"
+
+    if pass_name == "1F1B":
+        # TODO(Ruibiao): Move FLAGS_1f1b_backward_forward_overlap and
+        # FLAGS_mp_async_allreduce_in_backward to auto parallel Strategy
+        # after these two optimizations are available.
+        pass_attr["enable_backward_forward_overlap"] = int(
+            os.environ.get("FLAGS_1f1b_backward_forward_overlap", 0)
+        )
+
     pipeline_pass = new_pass("pipeline_scheduler_" + pass_name, pass_attr)
     pass_context = PassContext()
     pipeline_pass.apply([main_program], [startup_program], pass_context)