parallelize writing of layer checkpoint files across data parallel in…

…stances (#1419)

* parallelize layer checkpoints across data parallel groups

* use partition_uniform to determine start/end index values

* formatting fix

* config: add option for parallel write of layer checkpoints in pipeline stage

* yapf fixes

* enable parallel layer write according to config param

* avoid extraneous makedir when rank 0 writes all layers

Co-authored-by: Olatunji Ruwase <[email protected]>

deepspeed/runtime/config.py

@@ -669,6 +669,19 @@ def get_checkpoint_tag_validation_mode(checkpoint_params):
         )
 
 
+def get_checkpoint_parallel_write_pipeline(checkpoint_params):
+    par_write_params = checkpoint_params.get(CHECKPOINT_PARALLEL_WRITE, {})
+    par_write_pipeline = par_write_params.get(
+        CHECKPOINT_PARALLEL_WRITE_PIPELINE_STAGE,
+        CHECKPOINT_PARALLEL_WRITE_PIPELINE_STAGE_DEFAULT)
+    if par_write_pipeline in [True, False]:
+        return par_write_pipeline
+    else:
+        raise DeepSpeedConfigError(
+            "checkpoint::parallel_write::pipeline_stage "
+            f"value of '{par_write_pipeline}' is invalid, expecting: true or false")
+
+
 def get_dataloader_drop_last(param_dict):
     return get_scalar_param(param_dict,
                             DATALOADER_DROP_LAST,
@@ -887,6 +900,8 @@ def _initialize_params(self, param_dict):
         self.load_universal_checkpoint = checkpoint_params.get(
             LOAD_UNIVERSAL_CHECKPOINT,
             LOAD_UNIVERSAL_CHECKPOINT_DEFAULT)
+        par_write_pipe = get_checkpoint_parallel_write_pipeline(checkpoint_params)
+        self.checkpoint_parallel_write_pipeline = par_write_pipe
 
         self.aio_config = get_aio_config(param_dict)
 

deepspeed/runtime/constants.py

@@ -367,6 +367,9 @@ class ValidationMode:
 # "checkpoint": {
 #   tag_validation=["Ignore"|"Warn"|"Fail"]
 #   load_universal=false
+#   parallel_write: {
+#     pipeline_stage: [True|False]
+#   }
 # }
 CHECKPOINT = "checkpoint"
 CHECKPOINT_TAG_VALIDATION = "tag_validation"
@@ -380,6 +383,10 @@ class ValidationMode:
 LOAD_UNIVERSAL_CHECKPOINT = "load_universal"
 LOAD_UNIVERSAL_CHECKPOINT_DEFAULT = False
 
+CHECKPOINT_PARALLEL_WRITE = "parallel_write"
+CHECKPOINT_PARALLEL_WRITE_PIPELINE_STAGE = "pipeline_stage"
+CHECKPOINT_PARALLEL_WRITE_PIPELINE_STAGE_DEFAULT = False
+
 #########################################
 # Drop the last incomplete Batch
 # #########################################

deepspeed/runtime/engine.py

@@ -2924,7 +2924,11 @@ def save_checkpoint(self, save_dir, tag=None, client_state={}, save_latest=True)
             self._create_checkpoint_file(save_dir, tag, False)
             self._save_moe_checkpoint(save_dir, tag, client_state=client_state)
 
-        if self.save_non_zero_checkpoint:
+        # We distribute the task of saving layer checkpoint files among
+        # data parallel instances, so all procs should call _save_checkpoint.
+        # All procs then call module_state_dict(), but only procs of data
+        # parallel rank 0 save the general model params.
+        if not self.has_moe_layers:
             self._create_checkpoint_file(save_dir, tag, False)
             self._save_checkpoint(save_dir, tag, client_state=client_state)
 
@@ -3091,12 +3095,18 @@ def _create_zero_checkpoint_files(self, save_dir, tag):
     def _save_checkpoint(self, save_dir, tag, client_state={}):
 
         save_path = self._get_ckpt_name(save_dir, tag)
+
+        zero_optimizer_state = self.zero_optimization() or self.bfloat16_enabled()
+
         # A hack to save the checkpointing directory. Pipeline parallelism overrides
         # module_state_dict() and uses this path to save the model. module_state_dict()
-        # then instead just returns None.
+        # then instead just returns None.  The module_state_dict() implementation in
+        # PipelineEngine expects the save path to be set in self._curr_ckpt_path.
         self._curr_ckpt_path = os.path.join(save_dir, tag)
-        zero_optimizer_state = self.zero_optimization() or self.bfloat16_enabled()
-        state = dict(module=self.module_state_dict(),
+        module = self.module_state_dict()
+        self._curr_ckpt_path = None
+
+        state = dict(module=module,
                      buffer_names=self._get_buffer_names(),
                      optimizer=self.optimizer.state_dict()
                      if self.optimizer and not zero_optimizer_state else None,
@@ -3114,9 +3124,9 @@ def _save_checkpoint(self, save_dir, tag, client_state={}):
                      ds_version=version)
         state.update(client_state)
 
-        log_dist(message=f'Saving model checkpoint: {save_path}', ranks=[0, 1])
-        self.checkpoint_engine.save(state, save_path)
-        self._curr_save_path = None
+        if self.save_non_zero_checkpoint:
+            log_dist(message=f'Saving model checkpoint: {save_path}', ranks=[0, 1])
+            self.checkpoint_engine.save(state, save_path)
 
     def _get_buffer_names(self):
         buffer_names = []

deepspeed/runtime/pipe/engine.py

@@ -182,6 +182,8 @@ def __init__(self, has_bool_tensors=False, *super_args, **super_kwargs):
             self.module.activation_checkpoint_interval = self._config.pipeline[
                 'activation_checkpoint_interval']
 
+        self.module.checkpoint_parallel_write_pipeline = self._config.checkpoint_parallel_write_pipeline
+
         if self.is_last_stage():
             self.loss_model = self.module.loss_fn
 

deepspeed/runtime/pipe/module.py

@@ -562,13 +562,28 @@ def ckpt_layer_path_list(self, ckpt_dir, local_layer_idx):
         return ckpt_files
 
     def save_state_dict(self, save_dir, checkpoint_engine):
-        if self._grid.data_parallel_id != 0:
-            return
+        # Processes having the same model parallel rank on different data parallel instances
+        # have identical layer weights.  We can distribute the task of saving the layer weights
+        # among the data parallel ranks.  For example, if a pipeline stage has 9 layers and
+        # if there are 2 data parallel instances, rank 0 will save the first 5 layers and
+        # rank 1 will save the last 4.
+        dp_rank = self._grid.data_parallel_id
+        dp_size = self._grid.data_parallel_size
+        num_layers = len(self.forward_funcs)
+        if self.checkpoint_parallel_write_pipeline:
+            # spread layers evenly across data parallel ranks
+            offsets = ds_utils.partition_uniform(num_layers, dp_size)
+            start, end = offsets[dp_rank], offsets[dp_rank + 1]
+        else:
+            # data parallel rank 0 writes all layers
+            if dp_rank != 0:
+                return
+            start, end = 0, num_layers
+        layer_list = self.forward_funcs[start:end]
 
         os.makedirs(save_dir, exist_ok=True)
-        layer_offset = self._local_start
-        for idx, layer in enumerate(self.forward_funcs):
-            model_ckpt_path = self.ckpt_layer_path(save_dir, idx)
+        for idx, layer in enumerate(layer_list):
+            model_ckpt_path = self.ckpt_layer_path(save_dir, start + idx)
             if not hasattr(layer, 'state_dict'):
                 continue
             # We pass cloned tensors to torch.save() to avoid checkpoint bloat which occurs because torch.save()