[AutoParallel] update pipeline pass for while control_flow (#54224)

* [AutoParallel] update while control_flow with pipeline

* update process group instantiate

* fix micro_bsz for reshard

* update api for micro batch size

* add strategy for dp optimization

paddle/fluid/distributed/fleet_executor/message_bus.cc

@@ -216,8 +216,8 @@ bool MessageBus::SendInterRank(int64_t dst_rank,
   brpc::Channel channel;
   brpc::ChannelOptions options;
   options.protocol = "baidu_std";
-  options.connect_timeout_ms = 1000;
-  options.timeout_ms = 1000;
+  options.connect_timeout_ms = 100000;
+  options.timeout_ms = 100000;
   options.max_retry = 5;
   PADDLE_ENFORCE_EQ(
       channel.Init(dst_addr_for_brpc, &options),

python/paddle/distributed/auto_parallel/constants.py

@@ -124,9 +124,9 @@ set_field_default_config(QAT, "not_quant_pattern", ['skip_quant'])
 set_field_default_config(QAT, "algo", None)
 set_field_default_config(QAT, "onnx_format", True)
 
-# #########################################
+#########################################
 # auto tuning configuration
-# #########################################
+#########################################
 TUNING = "tuning"
 set_field_default_config(TUNING, "enable", False)
 set_field_default_config(TUNING, "profile_start_step", 1)
@@ -147,3 +147,12 @@ set_field_default_config(DATASET, "num_shards", 1)
 FUSED_PASSES = "fused_passes"
 set_field_default_config(FUSED_PASSES, "enable", False)
 set_field_default_config(FUSED_PASSES, "fused_passes_list", [])
+
+#########################################
+# data parallel configuration
+#########################################
+DP_OPTIMIZATION = "dp_optimization"
+set_field_default_config(DP_OPTIMIZATION, "enable", False)
+set_field_default_config(DP_OPTIMIZATION, "fuse_all_reduce_ops", True)
+set_field_default_config(DP_OPTIMIZATION, "fuse_grad_size_in_MB", 32)
+set_field_default_config(DP_OPTIMIZATION, "overlap_comm_cacl", True)

python/paddle/distributed/auto_parallel/static/dist_loader.py

@@ -58,6 +58,7 @@ class DistributedDataLoaderFromGenerator(DistributedDataLoaderBase):
         split_data=True,
         data_parallel_world_size=[],
         data_parallel_rank=[],
+        acc_steps=1,
     ):
         self.dataset = dataset
         self.feed_list = feed_list
@@ -77,6 +78,7 @@ class DistributedDataLoaderFromGenerator(DistributedDataLoaderBase):
         assert len(data_parallel_rank) == len(feed_list)
         self.dp_world_sizes = data_parallel_world_size
         self.dp_ranks = data_parallel_rank
+        self.acc_steps = acc_steps
 
         if isinstance(dataset, IterableDataset):
             self.dataset_kind = _DatasetKind.ITER
@@ -141,9 +143,11 @@ class DistributedDataLoaderFromGenerator(DistributedDataLoaderBase):
             if isinstance(self.dataset, IterableDataset):
                 steps_per_epoch = None
             elif self.batch_size is None:
-                steps_per_epoch = len(self.dataset)
+                steps_per_epoch = len(self.dataset) // self.acc_steps
             else:
-                steps_per_epoch = len(self.dataset) // self.batch_size
+                steps_per_epoch = (
+                    len(self.dataset) // self.batch_size // self.acc_steps
+                )
         except:
             raise ValueError(
                 "Please set `steps_per_epoch` or implement `__len__` method in dataset class."

python/paddle/distributed/auto_parallel/static/engine.py

@@ -235,6 +235,11 @@ class Engine:
         self._planned_mode = None
         self._dygraph_mode = False
         self._tuning = self._strategy.tuning
+        self._acc_steps = 1
+        if self._strategy.gradient_merge.enable:
+            self._acc_steps = self._strategy.gradient_merge.k_steps
+        elif self._strategy.pipeline.enable:
+            self._acc_steps = self._strategy.pipeline.accumulate_steps
 
         self.history = None
 
@@ -400,7 +405,12 @@ class Engine:
         if self.main_program._pipeline_opt:
             assert "tasks" in self.main_program._pipeline_opt["fleet_opt"]
             fleet_opt = self.main_program._pipeline_opt["fleet_opt"]
+            fwd_task = None
+            if self._strategy.pipeline.schedule_mode == "1F1B":
+                fwd_task = fleet_opt["tasks"][1]
+            elif self._strategy.pipeline.schedule_mode == "stream":
                 fwd_task = fleet_opt["tasks"][0]
+            assert fwd_task is not None
             fwd_prog = fwd_task.get_program()
             fwd_block = fwd_prog.global_block()
 
@@ -450,8 +460,6 @@ class Engine:
             ), "user_fetches must be a list, but receive {}".format(
                 type(user_fetches).__name__
             )
-        else:
-            user_fetches = []
         fetch_names = []
         fetch_indices = []
 
@@ -478,7 +486,7 @@ class Engine:
                 _process_fetch_group("metrics_" + str(i), var_list)
         if mode == "predict":
             _process_fetch_group("outputs", fetch_vars["outputs"])
-        for usr_fetch in user_fetches:
+        for usr_fetch in user_fetches or []:
             var_name = _to_name_str(usr_fetch)
             fetch(var_name)
         user_fetches_collection = [
@@ -931,6 +939,7 @@ class Engine:
         self._inputs_spec, self._labels_spec = self._prepare_data_spec(
             train_data, train_sample_split, batch_size
         )
+        micro_batch_size = self._validate_batch_size(batch_size)
         if not self._has_prepared[self._mode]:
             self._prepare_program(self._mode)
         else:
@@ -940,7 +949,7 @@ class Engine:
             dataset=train_data,
             capacity=70,
             iterable=False,
-            batch_size=batch_size,
+            batch_size=micro_batch_size,
             epochs=epochs,
             steps_per_epoch=steps_per_epoch,
             collate_fn=collate_fn,
@@ -951,7 +960,7 @@ class Engine:
         cbks = config_callbacks(
             callbacks,
             engine=self,
-            batch_size=batch_size,
+            batch_size=micro_batch_size,
             epochs=epochs,
             steps=train_dataloader._steps,
             log_freq=log_freq,
@@ -959,7 +968,7 @@ class Engine:
             save_dir=save_dir,
             verbose=verbose,
             metrics=self._metrics_name(),
-            acc_step=self._k_steps,
+            acc_step=self._acc_steps,
         )
 
         cbks.on_begin('train')
@@ -977,7 +986,7 @@ class Engine:
                     )
                 except core.EOFException:
                     break
-                lr = auto_utils.get_lr(self._optimizer)
+                lr = auto_utils.get_lr(self.optimizer)
                 logs = self._prepare_logger(
                     outs,
                     epoch,
@@ -1074,6 +1083,7 @@ class Engine:
         self._inputs_spec, self._labels_spec = self._prepare_data_spec(
             valid_data, valid_sample_split, batch_size
         )
+        micro_batch_size = self._validate_batch_size(batch_size)
         if not self._has_prepared[self._mode]:
             self._prepare_program(self._mode)
         else:
@@ -1083,7 +1093,7 @@ class Engine:
             dataset=valid_data,
             capacity=70,
             iterable=False,
-            batch_size=batch_size,
+            batch_size=micro_batch_size,
             steps_per_epoch=steps,
             collate_fn=collate_fn,
         )
@@ -1093,7 +1103,7 @@ class Engine:
         cbks = config_callbacks(
             callbacks,
             engine=self,
-            batch_size=batch_size,
+            batch_size=micro_batch_size,
             log_freq=log_freq,
             verbose=verbose,
             metrics=self._metrics_name(),
@@ -1180,6 +1190,7 @@ class Engine:
         self._inputs_spec, self._labels_spec = self._prepare_data_spec(
             test_data, test_sample_split, batch_size
         )
+        micro_batch_size = self._validate_batch_size(batch_size)
         if not self._has_prepared[self._mode]:
             self._prepare_program(self._mode)
         else:
@@ -1189,7 +1200,7 @@ class Engine:
             dataset=test_data,
             capacity=70,
             iterable=False,
-            batch_size=batch_size,
+            batch_size=micro_batch_size,
             steps_per_epoch=steps,
             collate_fn=collate_fn,
         )
@@ -1242,6 +1253,7 @@ class Engine:
         self._inputs_spec, self._labels_spec = self._prepare_data_spec(
             dataset, sample_split, batch_size
         )
+        micro_batch_size = self._validate_batch_size(batch_size)
         if not self._has_prepared[self._mode]:
             self._prepare_program(self._mode)
         else:
@@ -1250,7 +1262,7 @@ class Engine:
         dataloader = self._prepare_dataloader(
             dataset,
             return_list=False,
-            batch_size=batch_size,
+            batch_size=micro_batch_size,
             shuffle=shuffle,
             drop_last=drop_last,
             collate_fn=collate_fn,
@@ -1284,6 +1296,7 @@ class Engine:
         self._inputs_spec, self._labels_spec = self._prepare_data_spec(
             dataset, sample_split, batch_size
         )
+        micro_batch_size = self._validate_batch_size(batch_size)
         if not self._has_prepared[self._mode]:
             self._prepare_program(self._mode)
         else:
@@ -1297,7 +1310,7 @@ class Engine:
             return_list=False,
             use_multiprocess=use_multiprocess,
             drop_last=drop_last,
-            batch_size=batch_size,
+            batch_size=micro_batch_size,
             epochs=epochs,
             steps_per_epoch=steps_per_epoch,
             collate_fn=collate_fn,
@@ -1399,14 +1412,6 @@ class Engine:
         steps_per_epoch=None,
     ):
 
-        if self._strategy.gradient_merge and batch_size is not None:
-            assert (
-                batch_size % self._k_steps == 0
-            ), "Requires batch_size:[{}] to be divisible by k_steps:[{}].".format(
-                batch_size, self._k_steps
-            )
-            batch_size //= self._k_steps
-
         dist_context = self._dist_contexts[self._mode]
         dist_main_prog = dist_context.dist_main_programs[self._cur_rank]
         dist_startup_prog = dist_context.dist_startup_programs[self._cur_rank]
@@ -1468,14 +1473,6 @@ class Engine:
         collate_fn=None,
     ):
 
-        if self._strategy.gradient_merge and batch_size is not None:
-            assert (
-                batch_size % self._k_steps == 0
-            ), "Requires batch_size:[{}] to be divisible by k_steps:[{}].".format(
-                batch_size, self._k_steps
-            )
-            batch_size //= self._k_steps
-
         dist_context = self._dist_contexts[self._mode]
         dist_main_prog = dist_context.dist_main_programs[self._cur_rank]
         dist_startup_prog = dist_context.dist_startup_programs[self._cur_rank]
@@ -1515,6 +1512,9 @@ class Engine:
                 split_data=self._strategy.split_data,
                 data_parallel_world_size=self._dp_world_sizes,
                 data_parallel_rank=self._dp_ranks,
+                acc_steps=1
+                if not self._strategy.pipeline.enable
+                else self._acc_steps,
             )
         self._prepare_reader(feed_list)
         return dataloader
@@ -1526,9 +1526,18 @@ class Engine:
         )
         self._optimization_tuning(self._mode, tune_data, batch_size)
 
+    def _validate_batch_size(self, batch_size):
+        if batch_size is None:
+            return None
+        assert (
+            batch_size % self._acc_steps == 0
+        ), "Requires batch_size:[{}] to be divisible by acc_steps:[{}].".format(
+            batch_size, self._acc_steps
+        )
+        return batch_size // self._acc_steps
+
     def _validate_spec(self, specs):
         specs = auto_utils.to_list(specs)
-        self._k_steps = self._strategy.gradient_merge.k_steps
         if specs is not None:
             for i, spec in enumerate(specs):
                 if not isinstance(spec, InputSpec):
@@ -1541,14 +1550,14 @@ class Engine:
                             i, spec
                         )
                     )
-                if self._k_steps > 1:
+                if self._acc_steps > 1:
                     shape = list(spec.shape)
                     assert (
-                        shape[0] % self._k_steps == 0
+                        shape[0] % self._acc_steps == 0
                     ), "Requires batch_size[{}] to be divisible by k_steps[{}].".format(
-                        spec.shape[0], self._k_steps
+                        spec.shape[0], self._acc_steps
                     )
-                    shape[0] //= self._k_steps
+                    shape[0] //= self._acc_steps
                     spec.shape = shape
         return specs or []
 
@@ -1579,7 +1588,6 @@ class Engine:
             mode in self._dist_contexts
         ), f"{mode} model is not ready, please call `prepare()` first."
         self.to_mode(mode)
-        self._optimizer = self._dist_contexts[mode]._serial_optimizer
 
     def to_mode(self, mode):
         assert mode in [

python/paddle/distributed/auto_parallel/static/parallelizer.py

@@ -499,21 +499,12 @@ class AutoParallelizer:
                         break
                 if is_pipeline:
                     with paddle.static.program_guard(dist_main_prog):
-                        paddle.distributed.barrier(get_process_group(0))
+                        paddle.distributed.barrier()
 
             # Traverse different rank programs and traverse each op of them,
             # instantiate communication by process_mapping.
             all_process_groups = get_all_process_groups()
             for process_group in all_process_groups:
-                if len(_g_process_group_map) > 0:
-                    tmp = paddle.to_tensor([1], dtype="int32")
-                    paddle.distributed.all_reduce(
-                        tmp, sync_op=True, group=_g_process_group_map[0]
-                    )
-                    paddle.device.cuda.synchronize()
-
-                if rank not in process_group.ranks:
-                    continue
                 process_group.instantiate()
 
             # Copy distributed info to the default context

python/paddle/distributed/auto_parallel/static/parallelizer_v2.py

@@ -105,13 +105,6 @@ class Parallelizer:
                     time.time() - time0, self._mode
                 )
             )
-            # Do reshard process
-            time0 = time.time()
-            micro_bsz = (
-                1
-                if not self._strategy.pipeline.enable
-                else self._strategy.pipeline.micro_batch_size
-            )
             set_grad_var_shape(dist_main_prog, self._dist_context)
             resharder = Resharder(
                 dist_main_prog,
@@ -119,7 +112,6 @@ class Parallelizer:
                 rank,
                 self._dist_context,
                 dist_params_grads,
-                micro_bsz,
             )
             resharder.reshard()
             self._logger.debug(
@@ -169,13 +161,19 @@ class Parallelizer:
                 )
             )
             time0 = time.time()
+            # Do reshard process
+            micro_bsz = (
+                1
+                if not self._strategy.pipeline.enable
+                else self._strategy.pipeline.micro_batch_size
+            )
             resharder = Resharder(
                 dist_main_prog,
                 dist_startup_prog,
                 rank,
                 self._dist_context,
                 [],
-                1,
+                micro_bsz,
             )
             resharder.reshard()
             self._logger.debug(
@@ -305,11 +303,14 @@ class Parallelizer:
             return
 
         # data parallel optimization
-        config = {}
+        if self._strategy.dp_optimization.enable:
+            config = copy.deepcopy(self._strategy.dp_optimization.to_dict())
             config["dist_context"] = self._dist_context
             config["global_rank"] = rank
             config["use_sharding"] = self._strategy.sharding.enable
-        dp_pass = new_pass("auto_parallel_data_parallel_optimization", config)
+            dp_pass = new_pass(
+                "auto_parallel_data_parallel_optimization", config
+            )
             dp_pass.apply([main_program], [startup_program], self._pass_context)
 
         if self._strategy.sharding.enable:

python/paddle/distributed/auto_parallel/static/process_group.py

@@ -49,6 +49,12 @@ def clear_all_process_groups():
     _g_process_group_map[0] = ProcessGroup(0, [])
 
 
+def remove_process_group(ring_id):
+    global _g_process_group_map
+    if ring_id in _g_process_group_map:
+        _g_process_group_map.pop(ring_id)
+
+
 def new_process_group(ranks, group_id=None, force_new_group=False):
 
     global _g_process_group_map

python/paddle/distributed/auto_parallel/strategy.py

@@ -132,6 +132,12 @@ class FusedPassesConfig(BaseConfig):
         super().__init__(category, config_dict)
 
 
+class DPOptimizationConfig(BaseConfig):
+    def __init__(self, config_dict=None):
+        category = constants.DP_OPTIMIZATION
+        super().__init__(category, config_dict)
+
+
 class Strategy(BaseConfig):
     """
     The `Strategy` object is used to configure the parallelization and optimization behaviors.
@@ -206,3 +212,6 @@ class Strategy(BaseConfig):
 
         config_dict = self._config_dict.get(constants.FUSED_PASSES, None)
         self.fused_passes = FusedPassesConfig(config_dict)
+
+        config_dict = self._config_dict.get(constants.DP_OPTIMIZATION, None)
+        self.dp_optimization = DPOptimizationConfig(config_dict)

python/paddle/distributed/communication/stream/all_reduce.py

@@ -122,9 +122,9 @@ def all_reduce(
             tensor, op, group, sync_op, use_calc_stream
         )
     else:
-        # assert (
-        #     group is None
-        # ), "Group can not be used in static graph mode for now."
+        assert (
+            group is None
+        ), "Group can not be used in static graph mode for now."
         return _all_reduce_in_static_mode(
             tensor, op, group, sync_op, use_calc_stream
         )

python/paddle/distributed/passes/auto_parallel_pipeline.py

@@ -14,6 +14,9 @@
 
 import os
 
+from paddle.distributed.auto_parallel.static.process_group import (
+    remove_process_group,
+)
 from paddle.distributed.fleet.fleet_executor_utils import TaskNode
 from paddle.fluid import core
 from paddle.fluid.framework import Parameter, Program
@@ -50,6 +53,14 @@ class PipelinePass(PassBase):
         self._gen_bsz = self.get_attr("generation_batch_size")
         self._program = main_program
 
+        self._cur_rank = int(os.getenv("PADDLE_TRAINER_ID", 0))
+        trainer_endpoints = os.getenv("PADDLE_TRAINER_ENDPOINTS", "").split(',')
+        self._nrank = len(trainer_endpoints)
+
+        # compute current pp stage
+        self._pp_stages = len(self._dist_context.process_meshes)
+        self._cur_pp_stage = self._get_pp_stage(self._cur_rank)
+
         if self._mode == "1F1B":
             raise NotImplementedError("1F1B has not been implemented")
         elif self._mode == "F-Then-B":
@@ -70,6 +81,10 @@ class PipelinePass(PassBase):
             send_vars = []
             # insert sync ops
             for index, op in enumerate(list(block.ops)):
+                # NOTE: pipeline might hang when dynamic_shape is True
+                if op.type in ['send_v2', 'recv_v2']:
+                    op._set_attr("dynamic_shape", False)
+                # set send op on comm stream
                 if op.type == 'send_v2':
                     # step1: set 'use_calc_stream' False
                     op._set_attr("use_calc_stream", False)
@@ -176,21 +191,13 @@ class PipelinePass(PassBase):
     def _get_pp_stage(self, rank):
         pp_idx = None
         for idx, process_mesh in enumerate(self._dist_context.process_meshes):
-            if rank in process_mesh.processes:
+            if rank in process_mesh.process_ids:
                 pp_idx = idx
                 break
         return pp_idx
 
     def _task_stream(self):
-        cur_rank = int(os.getenv("PADDLE_TRAINER_ID", 0))
-        trainer_endpoints = os.getenv("PADDLE_TRAINER_ENDPOINTS", "").split(',')
-        nrank = len(trainer_endpoints)
         num_of_functionality = 5
-
-        # compute current pp stage
-        pp_stages = len(self._dist_context.process_meshes)
-        cur_pp_stage = self._get_pp_stage(cur_rank)
-
         start_prog = Program()
         cond_prog = Program()
         end_prog = Program()
@@ -198,6 +205,7 @@ class PipelinePass(PassBase):
         recv_prog = Program()
 
         cond_var_name = None
+        # record the varnames related to the while cond vars and communicate by nccl
         send_vars_name = set()
         recv_vars_name = {}
         for ib, src_block in enumerate(self._program.blocks):
@@ -222,38 +230,25 @@ class PipelinePass(PassBase):
                             src_block, end_block, op, force_create=True
                         )
             elif ib == 1:
+                # NOTE: for ernie generation
+                # The while block will be split to two separate blocks:
+                #     while{transformer_layer(send_block), generation_and_broadcast(recv_block)}
+                # The send_block:
+                #     include all ops about tansformer layers computation
+                #     execlude the nccl op about the while cond var(the last pp stage).
+                # The recv_block:
+                #     include all computation ops about generation and while cond var
+                #     execlude the nccl op about the while cond var(the pp stages exclude the last one)
+                # the nccl op about the while cond var:
+                #     put these varnames in the recv task node and do communication with brpc instead of nccl.
                 send_block = send_prog.block(0)
                 recv_block = recv_prog.block(0)
 
                 is_after_send_op = False
                 is_after_recv_op = False
-                for op in src_block.ops:
+                for i, op in enumerate(src_block.ops):
                     if op.type == "send_v2" and not is_after_send_op:
                         is_after_send_op = True
-                        if cur_pp_stage == pp_stages - 1:
-                            if op.type in ["c_sync_calc_stream", "nop"]:
-                                continue
-                            if (
-                                op.type not in ["recv_2", "assign"]
-                                and op.has_attr('op_namescope')
-                                and "/auto_parallel/reshard"
-                                in op.attr('op_namescope')
-                            ):
-                                if (
-                                    len(op.desc.input_arg_names()) > 0
-                                    and "@RESHARD"
-                                    not in op.desc.input_arg_names()[0]
-                                ):
-                                    send_vars_name.add(
-                                        op.desc.input_arg_names()[0]
-                                    )
-                                    continue
-                                if op.type == "send_v2":
-                                    continue
-                        self._create_program(
-                            src_block, send_block, op, force_create=True
-                        )
-                        continue
 
                     if (
                         is_after_send_op
@@ -261,45 +256,21 @@ class PipelinePass(PassBase):
                         and op.type == "recv_v2"
                     ):
                         is_after_recv_op = True
-                        if op.has_attr(
-                            'op_namescope'
-                        ) and "/auto_parallel/reshard" in op.attr(
-                            'op_namescope'
-                        ):
-                            var_name = op.desc.output_arg_names()[0]
-                            index = var_name.find("@")
-                            if index > 0:
-                                old_var_name = var_name[:index]
-                            else:
-                                old_var_name = var_name
-                            recv_vars_name[var_name] = old_var_name
-                            if not src_block._find_var_recursive(old_var_name):
-                                src_var = src_block._var_recursive(var_name)
-                                recv_block.create_var(
-                                    type=src_var.type,
-                                    name=old_var_name,
-                                    shape=src_var.shape,
-                                    dtype=src_var.dtype,
-                                    lod_level=src_var.lod_level,
-                                    persistable=src_var.persistable,
-                                    error_clip=src_var.error_clip,
-                                    stop_gradient=src_var.stop_gradient,
-                                    is_data=src_var.is_data,
-                                    belong_to_optimizer=src_var.belong_to_optimizer,
-                                )
-                            continue
-
-                        self._create_program(
-                            src_block, recv_block, op, force_create=True
-                        )
-                        continue
 
                     if not is_after_send_op or not is_after_recv_op:
-                        if cur_pp_stage == pp_stages - 1:
-                            if op.type in ["c_sync_calc_stream", "nop"]:
+                        if self._cur_pp_stage == self._pp_stages - 1:
+                            # NOTE: the c_sync_calc_stream about c_allgather cannot be removed
+                            if (
+                                op.type == "c_sync_calc_stream"
+                                and src_block.ops[i + 1].type == "send_v2"
+                            ):
                                 continue
+                            if op.type == "nop":
+                                continue
+                            # HACKCODE: the varname of send_v2 op, cast op should be recorded for brpc comm
                             if (
-                                op.type not in ["recv_2", "assign"]
+                                op.type
+                                not in ["recv_2", "assign", "c_allgather"]
                                 and op.has_attr('op_namescope')
                                 and "/auto_parallel/reshard"
                                 in op.attr('op_namescope')
@@ -312,19 +283,27 @@ class PipelinePass(PassBase):
                                     send_vars_name.add(
                                         op.desc.input_arg_names()[0]
                                     )
+                                    if op.type == "send_v2":
+                                        remove_process_group(op.attr("ring_id"))
                                     continue
                                 if op.type == "send_v2":
+                                    remove_process_group(op.attr("ring_id"))
                                     continue
                         self._create_program(
                             src_block, send_block, op, force_create=True
                         )
+                        continue
 
                     if is_after_send_op and is_after_recv_op:
+                        # HACKCODE: the varname of recv_v2 op, assign op should be recorded for brpc comm
                         if op.has_attr(
                             'op_namescope'
                         ) and "/auto_parallel/reshard" in op.attr(
                             'op_namescope'
                         ):
+                            if op.type in ["send_v2", "recv_v2"]:
+                                remove_process_group(op.attr("ring_id"))
+                            # remove the suffix of "@RESHARD"
                             var_name = op.desc.output_arg_names()[0]
                             index = var_name.find("@")
                             if index > 0:
@@ -356,6 +335,7 @@ class PipelinePass(PassBase):
                         self._create_program(
                             src_block, recv_block, op, force_create=True
                         )
+                        continue
             else:
                 raise Exception("Only support generation condition.")
 
@@ -397,52 +377,52 @@ class PipelinePass(PassBase):
             vars_to_shape = recv_task_node_var_shape
 
         start_task_node = TaskNode(
-            rank=cur_rank,
+            rank=self._cur_rank,
             max_run_times=self._acc_steps,
             node_type="Start",
-            task_id=int(cur_rank * num_of_functionality + 0),
+            task_id=int(self._cur_rank * num_of_functionality + 0),
             program=start_prog,
             lazy_initialize=True,
         )
         cond_task_node = TaskNode(
-            rank=cur_rank,
+            rank=self._cur_rank,
             max_run_times=self._acc_steps,
             node_type="Cond",
-            task_id=int(cur_rank * num_of_functionality + 1),
+            task_id=int(self._cur_rank * num_of_functionality + 1),
             program=cond_prog,
             cond_var_name=cond_var_name,
             lazy_initialize=True,
         )
         send_task_node = TaskNode(
-            rank=cur_rank,
+            rank=self._cur_rank,
             max_run_times=self._acc_steps,
             node_type="Compute",
-            task_id=int(cur_rank * num_of_functionality + 2),
+            task_id=int(self._cur_rank * num_of_functionality + 2),
             program=send_prog,
             lazy_initialize=True,
         )
         recv_task_node = TaskNode(
-            rank=cur_rank,
+            rank=self._cur_rank,
             max_run_times=self._acc_steps,
             node_type="Compute",
-            task_id=int(cur_rank * num_of_functionality + 3),
+            task_id=int(self._cur_rank * num_of_functionality + 3),
             program=recv_prog,
             lazy_initialize=True,
             vars_to_dtype=vars_to_dtype,
             vars_to_shape=vars_to_shape,
         )
         end_task_node = TaskNode(
-            rank=cur_rank,
+            rank=self._cur_rank,
             max_run_times=self._acc_steps,
             node_type="Compute",
-            task_id=int(cur_rank * num_of_functionality + 4),
+            task_id=int(self._cur_rank * num_of_functionality + 4),
             program=end_prog,
             lazy_initialize=True,
         )
 
         # add dependencies for task nodes intra stage
         inf = -1
-        pp_buff_size = int(pp_stages - cur_pp_stage)
+        pp_buff_size = int(self._pp_stages - self._cur_pp_stage)
         start_task_node.add_downstream_task(
             cond_task_node.task_id(), self._gen_bsz
         )
@@ -551,12 +531,12 @@ class PipelinePass(PassBase):
         # add dependencies for task nodes inter stage
         # get upstream ranks and downstream ranks of cur_rank
         up_down_streams = self._dist_context.up_down_streams
-        pp_upstream_ranks = up_down_streams.ups(cur_rank)
-        pp_downstream_ranks = up_down_streams.downs(cur_rank)
+        pp_upstream_ranks = up_down_streams.ups(self._cur_rank)
+        pp_downstream_ranks = up_down_streams.downs(self._cur_rank)
 
         for upstream_rank in pp_upstream_ranks:
             upstream_pp_stage = self._get_pp_stage(upstream_rank)
-            if upstream_pp_stage < pp_stages - 1:
+            if upstream_pp_stage < self._pp_stages - 1:
                 upstream_task_id = int(upstream_rank * num_of_functionality + 2)
                 send_task_node.add_upstream_task(upstream_task_id)
                 print(
@@ -579,7 +559,7 @@ class PipelinePass(PassBase):
                     2,
                 )
         for downstream_rank in pp_downstream_ranks:
-            if cur_pp_stage < pp_stages - 1:
+            if self._cur_pp_stage < self._pp_stages - 1:
                 downstream_task_id = int(
                     downstream_rank * num_of_functionality + 2
                 )
@@ -607,7 +587,7 @@ class PipelinePass(PassBase):
                 )
 
         task_id_to_rank = {}
-        for i in range(nrank):
+        for i in range(self._nrank):
             for j in range(num_of_functionality):
                 task_id_to_rank[int(i * num_of_functionality + j)] = i
         self._program._pipeline_opt = {

python/paddle/fluid/executor.py

@@ -1462,12 +1462,6 @@ class Executor:
             if "fleet_opt" in program._pipeline_opt:
                 # Move prepare here for port conflict with nccl in startup program
                 if self._fleet_executor is None:
-                    # Temporary manual enable standalone executor for fleet executor,
-                    # delete this code after the FLAGS is removed.
-                    if 'tasks' in program._pipeline_opt["fleet_opt"]:
-                        set_flags(
-                            {"FLAGS_fleet_executor_with_standalone": True}
-                        )
                     self._fleet_executor = _prepare_fleet_executor()
                 return self._run_using_fleet_executor(
                     program=program,

test/auto_parallel/CMakeLists.txt

@@ -58,6 +58,10 @@ if(WITH_DISTRIBUTE AND WITH_GPU)
   py_test_modules(test_engine_callbacks MODULES test_engine_callbacks)
   set_tests_properties(test_engine_callbacks
                        PROPERTIES LABELS "RUN_TYPE=EXCLUSIVE" TIMEOUT 50)
+  py_test_modules(test_pass_generation_pipeline MODULES
+                  test_pass_generation_pipeline)
+  set_tests_properties(test_pass_generation_pipeline
+                       PROPERTIES LABELS "RUN_TYPE=EXCLUSIVE" TIMEOUT 50)
   # End of unittests WITH multi cards and timeout
 
   # NOTE(zyl): unittests WITH multi cards and WITHOUT timeout

test/auto_parallel/generation_pipeline_pass_unittest.py

@@ -98,14 +98,11 @@ class GEN(nn.Layer):
         while cur_step < total_step:
 
             out = self.mlp(model_kwargs['input'])
-            model_kwargs['res'] = out
             paddle.increment(cur_step)
+            out_assign = auto.shard_op(paddle.assign, _g_mesh)(out)
 
-            auto.shard_op(paddle.assign, _g_mesh)(model_kwargs['input'], out)
-
-        output = F.gelu(model_kwargs['input'], approximate=True)
-
-        return output, cur_step
+        model_kwargs['output'] = paddle.assign(out_assign)
+        return model_kwargs['output'], cur_step
 
 
 def get_model():
@@ -125,35 +122,33 @@ class TestGenerationPipeline(unittest.TestCase):
         pipeline = strategy.pipeline
         pipeline.enable = True
         pipeline.schedule_mode = "stream"
-        pipeline.generation_batch_size = 4
-        pipeline.accumulate_steps = 4
+        pipeline.generation_batch_size = 2  # equal to the number of pp stages
+        pipeline.accumulate_steps = 20  # the number of all sample
         engine = auto.Engine(model, strategy=strategy)
 
         engine.prepare(
             inputs_spec=paddle.static.InputSpec(
-                shape=[2, 1024], name='input', dtype='float32'
+                shape=[20, 1024], name='input', dtype='float32'
             ),
             labels_spec=paddle.static.InputSpec(
-                shape=[2, 1024], name='label', dtype='float32'
+                shape=[20, 1024], name='label', dtype='float32'
             ),
             mode="eval",
         )
 
-        train_data = MyDataset(50 * 2)
+        train_data = MyDataset(20)
         train_dataloader = engine._prepare_dataloader_from_generator(
             dataset=train_data,
-            capacity=70,
+            capacity=20,
             iterable=False,
-            batch_size=2,
+            batch_size=1,  # micro_batch_size
             epochs=1,
-            steps_per_epoch=100,
         )
-        engine._prepare_reader()
 
         fleet_opt = engine.main_program._pipeline_opt['fleet_opt']
         assert len(fleet_opt['tasks']) == 5
         assert fleet_opt['inference_generation']
-        assert fleet_opt['num_micro_batches'] == 4
+        assert fleet_opt['num_micro_batches'] == 20
         num_task_in_rank = 5
         for idx, (task_id, rank_id) in enumerate(
             fleet_opt['task_id_to_rank'].items()
@@ -170,7 +165,6 @@ class TestGenerationPipeline(unittest.TestCase):
         except paddle.fluid.core.EOFException:
             print("test done")
             train_dataloader._inner_dataloader.reset()
-            train_dataloader._inner_dataloader.start()
 
 
 if __name__ == "__main__":