[Performance] Optimize unified checkpoint save/load speed.

paddlenlp/trainer/plugins/unified_checkpoint.py

@@ -30,6 +30,7 @@
 from paddlenlp.transformers.model_utils import (
     PretrainedModel,
     _load_state_dict_into_model,
+    faster_set_state_dict,
     get_parameter_dtype,
     load_state_dict,
     unwrap_model,
@@ -64,9 +65,10 @@
 from paddlenlp.utils.log import logger
 
 if is_safetensors_available():
-    from safetensors import safe_open
+    # from safetensors import safe_open
     from safetensors.numpy import save_file as safe_save_file
 
+    from paddlenlp.utils.safetensors import fast_safe_open as safe_open
 
 FP32_MASTER = "fp32_master_0"
 optimizer_scalar_name = [
@@ -195,7 +197,6 @@
     Returns:
         None
     """
-
     if paddle.distributed.get_world_size() <= 1:
         load_single_card_checkpoint(args, model, resume_from_checkpoint)
         return
@@ -221,7 +222,6 @@
         pretrained_model_name_or_path=resume_from_checkpoint,
         index_filename=os.path.join(resume_from_checkpoint, index_filename),
     )
-
     loaded_keys = sharded_metadata["all_checkpoint_keys"]
 
     model_state_dict = get_expected_state_dict(model)
@@ -265,7 +265,9 @@
             else:
                 tp_actions = model.get_tensor_parallel_convert_actions(model.config, loaded_keys, ignore_error=True)
         # Here we use expected_keys to optimize weights loading for pipeline model. Only works for safetensors
-        state_dict = load_state_dict(shard_file, tp_actions if pre_tensor_parallel_split else None, expected_keys)
+        state_dict = load_state_dict(
+            shard_file, tp_actions if pre_tensor_parallel_split else None, expected_keys, device="expected"
+        )
 
         if not pre_tensor_parallel_split:
             # Since we load all keys but we only need one of pipeline stages
@@ -278,11 +280,12 @@
                 None, model.config, state_dict=state_dict, ignore_error=len(resolved_archive_file) > 1
             )
 
-        error_msgs += _load_state_dict_into_model(model, state_dict, "")
+        # error_msgs += _load_state_dict_into_model(model, state_dict, "")
+        error_msgs += faster_set_state_dict(model, state_dict, strict_dtype=False)
 
         # force memory release
         del state_dict
-        gc.collect()
+        # gc.collect()
 
     if len(error_msgs) > 0:
         error_msg = "\n\t".join(error_msgs)
@@ -336,6 +339,7 @@
             tp_actions = model_to_save.get_tensor_parallel_convert_actions(
                 model_to_save.config, state_dict.keys(), is_split=False, ignore_error=True
             )
+        logger.info("Unified model tensor parallel weights in shards")
         state_dict = merge_tensor_parallel_with_shard(state_dict, tp_actions, all_filter_keys)
 
     # build index json file
@@ -489,6 +493,7 @@
     # This should always be a list but, just to be sure.
     if not isinstance(resolved_archive_file, list):
         resolved_archive_file = [resolved_archive_file]
+
     if len(resolved_archive_file) > 1:
         resolved_archive_file = tqdm(resolved_archive_file, desc="Loading optimizer shards")
 
@@ -536,10 +541,10 @@
                         tp_actions = mapping_optimizer_tp_actions(tp_actions, expected_keys)
 
                     # Here we use expected_keys to optimize weights loading for pipeline model. Only works for safetensors
-                    state_dict = load_state_dict(shard_file, tp_actions, expected_keys)
+                    state_dict = load_state_dict(shard_file, tp_actions, expected_keys, device="expected")
                 else:
                     # for pipeline model, we don't need to use tp_actions
-                    state_dict = load_state_dict(shard_file, None, expected_keys)
+                    state_dict = load_state_dict(shard_file, None, expected_keys, device="expected")
 
             returned_state_dict.update(state_dict)
             # force memory release
@@ -552,7 +557,6 @@
         state_dict_master_weight = load_resolved_archive_file(
             resolved_archive_file_mw, sharded_metadata_mw, expected_keys_mw, is_master_weights=True
         )
-
     # rename optimizer param
     for key in list(state_dict_optim.keys()):
         key_name = key.split("/")
@@ -561,13 +565,13 @@
             key_name = "_".join([static_name, FP32_MASTER, key_name[1]])
         else:
             key_name = "_".join([static_name, key_name[1]])
-        returned_optim_state_dict[key_name] = state_dict_optim[key]
+        returned_optim_state_dict[key_name] = state_dict_optim.pop(key)
         returned_optim_state_dict[key_name].name = key_name
 
     if has_master_weights:
         for key in list(state_dict_master_weight.keys()):
             static_name = struct2static_name_mappings[key]
-            returned_optim_state_dict["master_weights"][static_name] = state_dict_master_weight[key]
+            returned_optim_state_dict["master_weights"][static_name] = state_dict_master_weight.pop(key)
             returned_optim_state_dict["master_weights"][static_name].name = "_".join([static_name, FP32_MASTER])
 
     returned_optim_state_dict = nested_copy_place(
@@ -639,6 +643,7 @@
             tp_actions = model.get_tensor_parallel_convert_actions(
                 model.config, model_keys, is_split=False, ignore_error=True
             )
+        logger.info("Unified optimizer tensor parallel in shards")
         optim_state_dict = merge_tensor_parallel_for_optimizer(
             optim_state_dict,
             tp_actions,
@@ -647,6 +652,7 @@
         paddle.device.cuda.empty_cache()
 
         if master_weights is not None:
+            logger.info("Unified master weight tensor parallel in shards")
             master_weights = merge_tensor_parallel_for_optimizer(
                 master_weights,
                 tp_actions,
@@ -702,7 +708,6 @@
 def check_unified_checkpoint(args, model, resume_from_checkpoint, safe_serialization=False):
     index_filename = select_model_weight_index(args, model, resume_from_checkpoint, safe_serialization, local=False)
     index_filename = os.path.join(resume_from_checkpoint, index_filename)
-
     # Find index json file and distribute this file in global group.
     if distributed_isfile(index_filename):
         distributed_file(index_filename)
@@ -1604,7 +1609,9 @@
     tp_group = hcg.get_model_parallel_group()
     pp_group = hcg.get_pipe_parallel_group()
 
-    logger.info("Unified checkpoint generating sharded_index json files.")
+    logger.info(
+        f"Unified checkpoint: generating sharded_index json files for {'optimizer or master weight' if is_optimizer else 'model weight'}."
+    )
 
     if tp_group.nranks > 1:
         dist.all_gather_object(index_file_list, index_file, tp_group)
@@ -1713,8 +1720,6 @@
 
 
 def merge_tensor_parallel_with_shard(state_dict, tp_actions, all_filter_keys):
-    logger.info("Unified checkpoint merge tensor parallel in shards")
-
     hcg = fleet.get_hybrid_communicate_group()
     tp_group = hcg.get_model_parallel_group()
     tp_rank = tp_group.rank
@@ -1740,7 +1745,7 @@
                 action = tp_actions.pop(key)
                 tensor = action(ret) if is_dst else None
             else:
-                tensor = tensor._copy_to(paddle.CPUPlace(), False) if is_dst else None
+                tensor = tensor._copy_to(paddle.CUDAPinnedPlace(), False) if is_dst else None
 
             if is_dst:
                 state_dict_to_save[key] = tensor
@@ -1753,8 +1758,7 @@
 
 
 def merge_tensor_parallel_for_optimizer(state_dict, tp_actions, all_filter_keys):
-    logger.info("Unified optimizer tensor parallel in shards")
-
+    # Core function for UC
     hcg = fleet.get_hybrid_communicate_group()
     tp_group = hcg.get_model_parallel_group()
     tp_rank = tp_group.rank
@@ -1773,14 +1777,14 @@
                 # for example: beta1, beta2
                 if tensor.numel().item() == 1:
                     tensor = (
-                        tensor._copy_to(paddle.CPUPlace(), False) if is_dst else None
+                        tensor._copy_to(paddle.CUDAPinnedPlace(), False) if is_dst else None
                     )  # Need broadcast when loaded
                 else:
                     ret = distributed_gather(tensor, dst=j, group=tp_group, offload=False)
                     action = tp_actions[model_key]
                     tensor = action(ret) if is_dst else None
             else:
-                tensor = tensor._copy_to(paddle.CPUPlace(), False) if is_dst else None
+                tensor = tensor._copy_to(paddle.CUDAPinnedPlace(), False) if is_dst else None
 
             if is_dst:
                 state_dict_to_save[filter_keys[i]] = tensor
@@ -1892,7 +1896,10 @@
             outputs[key] = nested_copy_place(inputs[key], place, blocking)
         return outputs
     if isinstance(inputs, paddle.Tensor):
-        inputs = inputs if inputs.place == place else inputs._copy_to(place, blocking)
+        if inputs.place._equals(place):
+            return inputs
+        else:
+            return inputs._copy_to(place, blocking)
     return inputs
 
 

paddlenlp/trainer/trainer.py

@@ -2406,6 +2406,7 @@
             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()
             return

paddlenlp/transformers/conversion_utils.py

@@ -285,8 +285,12 @@
 
     if isinstance(weight_list[0], np.ndarray):
         return np.concatenate([reorder[i] for i in index], axis=axis)
+    else:
+        tensor = paddle.concat([reorder[i] for i in index], axis=axis)
 
-    return paddle.concat([reorder[i] for i in index], axis=axis)._copy_to(paddle.CPUPlace(), False)
+        if tensor.place.is_gpu_place():
+            tensor = tensor._copy_to(paddle.CUDAPinnedPlace(), False)
+        return tensor
 
 
 def naive_fuse_split_tp(
@@ -361,12 +365,18 @@
         if isinstance(weight_list[0], np.ndarray):
             return np.concatenate(weight_list, axis=-1)
         else:
-            return paddle.concat(weight_list, axis=-1)._copy_to(paddle.CPUPlace(), False)
+            tensor = paddle.concat(weight_list, axis=-1)
+            if tensor.place.is_gpu_place():
+                tensor = tensor._copy_to(paddle.CUDAPinnedPlace(), False)
+            return tensor
     else:
         if isinstance(weight_list[0], np.ndarray):
             return np.concatenate(weight_list, axis=0)
         else:
-            return paddle.concat(weight_list, axis=0)._copy_to(paddle.CPUPlace(), False)
+            tensor = paddle.concat(weight_list, axis=0)
+            if tensor.place.is_gpu_place():
+                tensor = tensor._copy_to(paddle.CUDAPinnedPlace(), False)
+            return tensor
 
 
 def normal_fuse_split_tp(weight, tensor_parallel_degree, tensor_parallel_rank=None, is_column=True):

paddlenlp/transformers/model_utils.py

@@ -108,10 +108,13 @@
 
 
 if is_safetensors_available():
-    from safetensors import safe_open
+
+    # from safetensors import safe_open
     from safetensors.numpy import load_file as safe_load_file
     from safetensors.numpy import save_file as safe_save_file
 
+    from paddlenlp.utils.safetensors import fast_safe_open as safe_open
+
 
 def prune_linear_layer(layer: nn.Linear, index: paddle.Tensor, dim: int = 0) -> nn.Linear:
     """
@@ -312,7 +315,7 @@
 
 
 def load_state_dict(
-    checkpoint_file: Union[str, os.PathLike], tensor_parallel_split_mapping=None, fliter_dict_keys=None
+    checkpoint_file: Union[str, os.PathLike], tensor_parallel_split_mapping=None, fliter_dict_keys=None, device="cpu"
 ):
     """
     Reads a PaddlePaddle checkpoint file, returning properly formatted errors if they arise.
@@ -345,11 +348,16 @@
                         weight = tensor_parallel_split_mapping[key](py_safe_slice_)
                     else:
                         weight = py_safe_slice_[:]
+                    if device == "expected":
+                        with device_guard():
+                            weight = paddle.Tensor(weight, zero_copy=True)
+                        weight = weight._copy_to(paddle.framework._current_expected_place(), False)
                     state_dict[key] = weight
 
-            for k in list(state_dict.keys()):
-                with device_guard():
-                    state_dict[k] = paddle.Tensor(state_dict.pop(k), zero_copy=True)
+            if device == "cpu":
+                for k in list(state_dict.keys()):
+                    with device_guard():
+                        state_dict[k] = paddle.Tensor(state_dict.pop(k), zero_copy=True)
 
             return state_dict
 
@@ -671,8 +679,10 @@
     return missing_keys, unexpected_keys
 
 
-def faster_set_state_dict(model, state_dict):
+def faster_set_state_dict(model, state_dict, strict_dtype=True):
     # the state_dict will be destroied.
+    unused_keys = set(state_dict.keys())
+    unset_keys = set(model.state_dict().keys())
     with paddle.no_grad():
         for k, v in model.state_dict().items():
             if k in state_dict:
@@ -682,8 +692,10 @@
                         f"faster_set_state_dict need state dict with paddle.Tensor, but got {type(v_new)}"
                     )
                 # 2. cast param / Tensor to dtype
+                #
                 if v.dtype != v_new.dtype:
-                    raise ValueError(f"for key: {k}, expect dtype {v.dtype}, but got {v_new.dtype}")
+                    if strict_dtype or (not v.is_floating_point() or not v_new.is_floating_point()):
+                        raise ValueError(f"for key: {k}, expect dtype {v.dtype}, but got {v_new.dtype}")
                 # check shape
                 if list(v.shape) != list(v_new.shape):
                     raise ValueError(f"for key: {k}, expect shape {v.shape}, but got {v_new.shape}")
@@ -699,9 +711,22 @@
                 else:
                     new_t = v_new
 
+                if not strict_dtype and v.dtype != new_t.dtype:
+                    new_t = new_t.astype(v.dtype)
+
                 # 4. share Tensor to origin param / Tensor
                 src_tensor = new_t.value().get_tensor()
                 dst_tensor._share_data_with(src_tensor)
+                unset_keys.remove(k)
+                unused_keys.remove(k)
+
+    error_msgs = []
+    # if len(unset_keys) > 0:
+    #    error_msgs.append(f"Those weight of model is not initialized: {list(unset_keys)}")
+    if len(unused_keys) > 0:
+        error_msgs.append(f"Those state dict keys are not using in model: {list(unused_keys)}")
+
+    return error_msgs
 
 
 def _load_state_dict_into_model(model_to_load, state_dict, start_prefix):
@@ -733,22 +758,16 @@
     def is_0d_or_1d(tensor):
         return len(tensor.shape) == 0 or list(tensor.shape) == [1]
 
-    expected_place = paddle.framework._current_expected_place()
     for key, value in model_to_load.state_dict().items():
-        if key in state_dict:
+        if key in list(state_dict.keys()):
             if isinstance(state_dict[key], np.ndarray):
                 raise ValueError(
                     "convert_state_dict_dtype expected paddle.Tensor not numpy.ndarray, plase convert numpy.ndarray to paddle.Tensor"
                 )
             # confirm parameter cast is executed on the same device as model
             # TODO: cast(FP32 -> FP16) has diff on different devices, need to fix it
             if state_dict[key].is_floating_point() and state_dict[key].dtype != value.dtype:
-                value_pop = state_dict.pop(key)
-                value_new_place = (
-                    value_pop if value_pop.place == expected_place else value_pop._copy_to(expected_place, False)
-                )
-                state_dict[key] = paddle.cast(value_new_place, value.dtype)._copy_to(value_pop.place, False)
-                del value_new_place
+                state_dict[key] = paddle.cast(state_dict.pop(key), value.dtype)
             # unified 0d and 1d tensor
             if is_0d_or_1d(value) and is_0d_or_1d(state_dict[key]):
                 if list(value.shape) != list(state_dict[key].shape):