adapter flex_checkpoint

Author: xingmingyyj

paddlenlp/trainer/trainer.py

@@ -31,6 +31,7 @@
 import warnings
 from collections import OrderedDict
 from collections.abc import Mapping
+
 from pathlib import Path
 from typing import Any, Callable, Dict, List, Optional, Tuple, Union
 
@@ -163,9 +164,12 @@
     should_skip_data,
     speed_metrics,
     split_parallel_config,
+    init_optimizer,
 )
 from .training_args import TrainingArguments
 from .unified_checkpoint import UnifiedCheckpointHandler
+from .unified_checkpoint.utils import generate_base_static_name
+from paddle.distributed.checkpoint.sharded_tensor import ShardedTensor, build_sharded_state_dict
 from .utils import reshard as reshard_util
 from .utils.async_save import AsyncSaver
 
@@ -197,7 +201,6 @@
 if is_datasets_available():
     import datasets
 
-
 try:
     from paddle.distributed.fleet.utils import mix_precision_utils
 except:
@@ -914,7 +917,7 @@ def train(
         self._memory_tracker.start()
 
         if not self.args.enable_auto_parallel:
-            if not self.args.should_load_sharding_stage1_model:
+            if not self.args.should_load_sharding_stage1_model and not self.args.using_flex_checkpoint:
                 self._load_from_checkpoint(resume_from_checkpoint)
 
             if self.args.should_load_sharding_stage1_model:
@@ -934,14 +937,32 @@ def train(
                 if delay_optimizer_creation:
                     self.create_optimizer_and_scheduler(num_training_steps=max_steps)
                 self._load_optimizer_and_scheduler(resume_from_checkpoint)
-            else:
+            elif not self.args.using_flex_checkpoint:
                 model = self._wrap_model(self.model_wrapped)
                 # for the rest of this function `model` is the outside model, whether it was wrapped or not
                 if model is not self.model:
                     self.model_wrapped = model
                 if delay_optimizer_creation:
                     self.create_optimizer_and_scheduler(num_training_steps=max_steps)
                 self._load_optimizer_and_scheduler(resume_from_checkpoint)
+            else:
+                assert self.args.using_flex_checkpoint, "default using flex_checkpoint!"
+
+                model = self._wrap_model(self.model_wrapped)
+                if model is not self.model:
+                    self.model_wrapped = model
+
+                if delay_optimizer_creation:
+                    self.create_optimizer_and_scheduler(num_training_steps=max_steps)
+                    
+                if resume_from_checkpoint is not None:
+                    model_sharded_state_dict = self.model.sharded_state_dict()
+                    self.optimizer.sharded_state_dict(model_sharded_state_dict)
+                    init_optimizer(self.optimizer)
+                    optimizer_sharded_state_dict = self.optimizer.sharded_state_dict(model_sharded_state_dict)
+                    sharded_state_dict = {**model_sharded_state_dict, **optimizer_sharded_state_dict}
+                    dist.load_state_dict(sharded_state_dict, resume_from_checkpoint)
+                    self._load_scheduler(resume_from_checkpoint)
         else:
             model = self.model_wrapped
             if delay_optimizer_creation:
@@ -1342,6 +1363,7 @@ def fused_allreduce_gradients_no_sync(paramlist, hcg):
                             logger.warning(
                                 f"optimizer not run, scale_before: {scale_before_value[0]}, scale_after: {scale_after_value[0]}"
                             )
+                  
                     elif isinstance(self.optimizer, HybridParallelOptimizer):
                         self.optimizer._step(parameters_list)
                     else:
@@ -1593,7 +1615,7 @@ def _maybe_log_save_evaluate(self, tr_loss, model, epoch, ignore_keys_for_eval,
             if num_steps == 0:
                 logs["loss"] = 0.0
             else:
-                logs["loss"] = round(tr_loss_scalar / num_steps, 8)
+                logs["loss"] = round(tr_loss_scalar / num_steps, 8)                
 
             logs["learning_rate"] = float("{0:.3e}".format(self._get_learning_rate()))
             logs["global_step"] = int(self.state.global_step)
@@ -1968,7 +1990,6 @@ def apply_decay_param_fun(x):
                 grad_clip=nn.ClipGradByGlobalNorm(self.args.max_grad_norm) if self.args.max_grad_norm > 0 else None,
                 **optimizer_kwargs,
             )
-
         return self.optimizer
 
     def _apply_to_optimizer(self, action):
@@ -2212,7 +2233,7 @@ def _wrap_model(self, model, training=True):
         in_tensor_parallel_mode = self.args.tensor_parallel_degree > 1
         in_sep_parallel_mode = self.args.sep_parallel_degree > 1
         in_cp_parallel_mode = self.args.context_parallel_degree > 1
-
+         
         # Multi-gpu training
         if self.args.world_size > 1 and (not self.args.use_hybrid_parallel):
             # MOE use DDP to broadcaset parameters.
@@ -2234,7 +2255,6 @@ def _wrap_model(self, model, training=True):
                 mix_precision_utils.MixPrecisionLayer(model, dtype=self.amp_dtype)
                 assert self.optimizer is not None, "optimizer is empty!"
                 self.optimizer = mix_precision_utils.MixPrecisionOptimizer(self.optimizer)
-
         # Pipeline mode
         if in_pipeline_parallel_mode:
             if self.args.amp_master_grad:
@@ -2279,20 +2299,18 @@ def get_expected_keys(inputs, keys):
                     "Using default prepare pipeline inputs func, only support input_ids and labels as inputs."
                 )
                 model._prepare_pipeline_inputs_func = _prepare_pipeline_inputs_func
-
+            
             assert self.optimizer is not None, "Pipeline mode need decorate optimizer, pelease init optimizer."
             if self.args.amp_master_grad:
                 self.optimizer = mix_precision_utils.MixPrecisionOptimizer(self.optimizer)
             self.optimizer = fleet.distributed_optimizer(self.optimizer)
-
             if (
                 hasattr(self.args, "enable_sharding_comm_overlap")
                 and self.args.enable_sharding_comm_overlap
                 and self.args.unified_checkpoint
                 and "split_param" in split_parallel_config(self.args.sharding_parallel_config)
             ):
                 model.register_sharding_comm_overlap_hook(self.optimizer)
-
         # No pipeline mode, sharding only
         if not in_pipeline_parallel_mode and in_sharding_parallel_mode:
             # Sharded DDP!
@@ -2306,7 +2324,6 @@ def get_expected_keys(inputs, keys):
                     model = paddle.distributed.fleet.meta_parallel.TensorParallel(
                         model, hcg, strategy=fleet.fleet._user_defined_strategy
                     )
-
             if ShardingOption.SHARD_OP in self.args.sharding:
                 if self.args.amp_master_grad:
                     mix_precision_utils.MixPrecisionLayer(model, dtype=self.amp_dtype)  # return value has no use
@@ -2325,7 +2342,6 @@ def get_expected_keys(inputs, keys):
                     level = "p_g_os"
 
                 from paddle.distributed.sharding import group_sharded_parallel
-
                 # add dp_group and exclude_layer params
                 # https://www.paddlepaddle.org.cn/documentation/docs/zh/develop/api/paddle/distributed/sharding/group_sharded_parallel_cn.html#group-sharded-parallel
                 extra_kwargs = {}
@@ -2348,6 +2364,7 @@ def get_expected_keys(inputs, keys):
                     offload=cpu_offload,
                     **extra_kwargs,
                 )
+
                 if ShardingOption.SHARD_GRAD_OP in self.args.sharding and self.args.amp_master_grad:
                     assert hasattr(optimizer, "use_main_grad"), (
                         "Current installed paddle doesn't support sharding stage 2 with main grad, "
@@ -2373,7 +2390,6 @@ def get_expected_keys(inputs, keys):
             if self.args.amp_master_grad:
                 self.optimizer = mix_precision_utils.MixPrecisionOptimizer(self.optimizer)
             self.optimizer = fleet.distributed_optimizer(self.optimizer)
-
         # stage1 has v1 and v2 version
         if in_sharding_parallel_mode and ShardingOption.SHARD_OP in self.args.sharding:
             if "split_param" in self.args.sharding_parallel_config:
@@ -2388,7 +2404,6 @@ def get_expected_keys(inputs, keys):
                     and "enable_stage1_broadcast_overlap" in self.args.sharding_parallel_config
                 ):
                     self.optimizer._set_broadcast_overlap(True, model)
-
         return model
 
     def _prepare_input(self, data: Union[paddle.Tensor, Any]) -> Union[paddle.Tensor, Any]:
@@ -2700,6 +2715,10 @@ def _save_checkpoint(self, model, metrics=None):
         else:
             self.save_model(output_dir)
 
+        model_sharded_state_dict  = self.model.sharded_state_dict()
+        if self.args.using_flex_checkpoint:
+            os.makedirs(output_dir, exist_ok=True)
+
         # Determine the new best metric / best model checkpoint
         if metrics is not None and self.args.metric_for_best_model is not None:
             metric_to_check = self.args.metric_for_best_model
@@ -2763,23 +2782,30 @@ def _save_checkpoint(self, model, metrics=None):
                             signal_dir,
                         )
                     else:
-                        if self.dp_group.rank > 0:  # this should only work for MoE saving
-                            self._save_ckpt_func(
-                                self._filter_moe_no_sync_optimizer_params(),
-                                os.path.join(output_dir, optimizer_name),
-                                saved_signal_path,
-                            )
-
-                        else:
-                            state_dict = self.optimizer.state_dict()
-                            save_path = os.path.join(output_dir, optimizer_name)
-                            if self.args.use_async_save:
-                                assert not strtobool(os.getenv("FLAG_LLM_PDC", "False")), "Dont support FLAG_LLM_PDC"
-                                self._async_optimizer_saver.run(
-                                    state_dict, save_path, saved_signal_path=saved_signal_path
+                        if not self.args.using_flex_checkpoint:
+                            if self.dp_group.rank > 0:  # this should only work for MoE saving
+                                self._save_ckpt_func(
+                                    self._filter_moe_no_sync_optimizer_params(),
+                                    os.path.join(output_dir, optimizer_name),
+                                    saved_signal_path,
                                 )
+
                             else:
-                                self._save_ckpt_func(state_dict, save_path, saved_signal_path)
+                                state_dict = self.optimizer.state_dict()
+                                save_path = os.path.join(output_dir, optimizer_name)
+                                if self.args.use_async_save:
+                                    assert not strtobool(os.getenv("FLAG_LLM_PDC", "False")), "Dont support FLAG_LLM_PDC"
+                                    self._async_optimizer_saver.run(
+                                        state_dict, save_path, saved_signal_path=saved_signal_path
+                                    )
+                                else:
+                                    self._save_ckpt_func(state_dict, save_path, saved_signal_path)
+                        else:
+                            optimizer_sharded_state_dict = self.optimizer.sharded_state_dict(model_sharded_state_dict)
+                            dist.save_state_dict(
+                                {**model_sharded_state_dict, **optimizer_sharded_state_dict},
+                                output_dir,
+                            )
                 else:
                     if self.args.unified_checkpoint and "async_save" in self.args.unified_checkpoint_config:
                         global_rank = paddle.distributed.get_rank() if paddle.distributed.get_world_size() > 1 else -1
@@ -2800,7 +2826,7 @@ def _save_checkpoint(self, model, metrics=None):
                             output_dir,
                             signal_dir,
                         )
-                    else:
+                    elif not self.args.using_flex_checkpoint:
                         if self.args.data_parallel_rank > 0 and self.args.use_expert_parallel:
                             self._save_ckpt_func(
                                 self._filter_moe_no_sync_optimizer_params(),
@@ -2814,6 +2840,13 @@ def _save_checkpoint(self, model, metrics=None):
                                 saved_signal_path,
                             )
 
+                    else:
+                        optimizer_sharded_state_dict = self.optimizer.sharded_state_dict(model_sharded_state_dict)
+                        dist.save_state_dict(
+                            {**model_sharded_state_dict, **optimizer_sharded_state_dict},
+                            output_dir,
+                        )
+                
                 # FIXME: maybe only save one copy
                 paddle.save(self.lr_scheduler.state_dict(), os.path.join(output_dir, SCHEDULER_NAME))
 
@@ -3077,13 +3110,32 @@ def _save(
                 with open(path, "w") as f:
                     json.dump(model_meta, f)
 
+    def _load_scheduler(self, checkpoint):
+        if checkpoint is None:
+            self.runtime_timer.stop()
+            return
+
+        if not self.args.ignore_load_lr_and_optim:
+            if distributed_isfile(os.path.join(checkpoint, SCHEDULER_NAME)):
+                self.lr_scheduler.set_state_dict(
+                    paddle.load(distributed_file(os.path.join(checkpoint, SCHEDULER_NAME)))
+                )
+            else:
+                raise ValueError(f"scheduler-file not found, scheduler:{os.path.join(checkpoint, SCHEDULER_NAME)}")
+
+            if self.do_grad_scaling and distributed_isfile(os.path.join(checkpoint, SCALER_NAME)):
+                self.scaler.load_state_dict(
+                    paddle.load(distributed_file(os.path.join(checkpoint, SCALER_NAME)), return_numpy=True)
+                )
+
     def _load_optimizer_and_scheduler(self, checkpoint):
         """If optimizer and scheduler states exist, load them."""
         self.runtime_timer.start("checkpoint loading time")
         if checkpoint is None:
             self.runtime_timer.stop()
             return
 
+
         logger.info("Loading optimizer and scheduler...")
         if (not self.args.should_load_sharding_stage1_model) and self.args.ignore_load_lr_and_optim:
             self.runtime_timer.stop()
@@ -3118,6 +3170,7 @@ def _load_optimizer_and_scheduler(self, checkpoint):
                     and "split_param" in split_parallel_config(self.args.sharding_parallel_config)
                 ):
                     model = self.model_wrapped
+
                 opt_state_dict = self.unified_checkpoint_handler.load_unified_optimizer(
                     model=model,
                     optimizer=self.optimizer,
@@ -3149,18 +3202,7 @@ def _load_optimizer_and_scheduler(self, checkpoint):
             optimizer_name = _add_variant(PADDLE_OPTIMIZER_NAME, self.args.optimizer_name_suffix)
             raise ValueError(f"optimizer-state-dict not found, opt: {os.path.join(checkpoint, optimizer_name)}.")
 
-        if not self.args.ignore_load_lr_and_optim:
-            if distributed_isfile(os.path.join(checkpoint, SCHEDULER_NAME)):
-                self.lr_scheduler.set_state_dict(
-                    paddle.load(distributed_file(os.path.join(checkpoint, SCHEDULER_NAME)))
-                )
-            else:
-                raise ValueError(f"scheduler-file not found, scheduler:{os.path.join(checkpoint, SCHEDULER_NAME)}")
-
-            if self.do_grad_scaling and distributed_isfile(os.path.join(checkpoint, SCALER_NAME)):
-                self.scaler.load_state_dict(
-                    paddle.load(distributed_file(os.path.join(checkpoint, SCALER_NAME)), return_numpy=True)
-                )
+        self._load_scheduler(checkpoint)
 
         if self.args.offload_optim:
             logger.info("Offloading optimizer state...")