Merge branch 'PaddlePaddle:develop' into dev_20250108_fix_requirements

Author: DrownFish19

paddlenlp/trainer/auto_trainer.py

@@ -105,7 +105,15 @@ def loss_func(loss, outputs):
         model = kwargs["model"]
         for param in model.parameters():
             if not param._is_initialized():
-                param.initialize()
+                try:
+                    param.initialize()
+                except Exception as e:
+                    # NOTE(zhangwl):maybe param is not initialized and param init_func is set in later.user need set_init_func before auto_trainer
+                    logger.warning(
+                        f"AutoTrainer requires all parameters to be initialized when auto_trainer init, but failed to initialize parameter {param.name} {param}.\n"
+                        + "Please check param init func.\n"
+                        + f"The original exception message is:\n{str(e)}"
+                    )
         kwargs["model"] = model
 
         super().__init__(*args, **kwargs)

paddlenlp/trainer/utils/ckpt_converter.py

@@ -41,7 +41,13 @@
 
 class CheckpointConverter:
     def __init__(
-        self, hybrid_parallel_ckpt_path, state_dict, parameter_to_structured_name, trainging_args=None, patch_dict=None
+        self,
+        hybrid_parallel_ckpt_path,
+        state_dict,
+        parameter_to_structured_name,
+        trainging_args=None,
+        patch_dict=None,
+        local_view_pattern: list | bool = None,
     ):
         self.use_dist = True if paddle.distributed.get_world_size() > 1 else False
         self.path = hybrid_parallel_ckpt_path
@@ -85,6 +91,17 @@ def __init__(
                 self.auto_parallel_state_dict[self.patch_dict[k]] = self.auto_parallel_state_dict[k]
             for k in del_keys:
                 self.auto_parallel_state_dict.pop(k)
+        # solve the problem of inconsistent parameter names in moe automatic parallel mode.
+        if hasattr(trainging_args, "moe_group") and trainging_args.moe_group:
+            if local_view_pattern is False:
+                self.local_view_pattern_list = None
+            else:
+                if isinstance(local_view_pattern, list):
+                    self.local_view_pattern_list = local_view_pattern
+                else:
+                    self.local_view_pattern_list = ["experts"]
+        else:
+            self.local_view_pattern_list = None
 
         flags = [
             ["tp degree", self.tp_degree],
@@ -497,6 +514,46 @@ def gen_metadata_and_prepare_source_state_dict(self):
             else:
                 return self.gen_metadata_for_tp_sharded_tensor()
 
+    def rename_local_view_state_dict(self, state_dict, file_name):
+        """
+        Rename the key for local views to the key for global views, and return the renamed `state_dict`.
+        """
+        if self.local_view_pattern_list is None:
+            return state_dict
+        # case 1: moe_group is mp_group
+        if self.tp_degree > 1 and self.sharding_degree <= 1:
+            (tp_rank, pp_rank, sharding_rank) = self.get_distribution_rank_from_file_name(file_name)
+            expert_name_old2new = {}
+            for pattern in self.local_view_pattern_list:
+                expert_pattern = rf"({pattern}\.)(\d+)"
+                # extract all experts IDs
+                expert_ids = set()
+                for state_name in state_dict.keys():
+                    res = re.search(expert_pattern, state_name)
+                    if res:
+                        expert_ids.add(int(res.group(2)))
+                expert_num = len(expert_ids)
+                # construct old name to new name mapping
+                for state_name in state_dict.keys():
+                    res = re.search(expert_pattern, state_name)
+                    if res:
+                        new_expert_id = int(res.group(2)) % expert_num + tp_rank * expert_num
+                        expert_name_old2new[state_name] = re.sub(
+                            expert_pattern, f"{res.group(1)}{new_expert_id}", state_name
+                        )
+            # rename state_dict
+            renamed_state_dict = {
+                expert_name_old2new[state_name]
+                if state_name in expert_name_old2new
+                else state_name: state_dict[state_name]
+                for state_name in state_dict.keys()
+            }
+
+            return renamed_state_dict
+        # TODO: add support for sharding
+        else:
+            return state_dict
+
     def load_state_dict_and_rename(self):
         """
         Parse the distributed information from the names of the checkpoint files and evenly parse out the distributed information for each weight/optimizer state
@@ -741,11 +798,10 @@ def load_state_dict_and_rename(self):
                         model_state_file_name = self.get_model_state_file_from(file_name)
                         assert model_state_file_name is not None
                         model_state_keys = global_file_to_state_dict_keys_mapping[model_state_file_name]
-                        renamed_state_dict = self.rename_using_optimizer_state_order(model_state_keys, state_dict)
-                        self.get_sharded_tensor_infos(file, renamed_state_dict, cur_rank_sharded_tensor_infos)
-                        self.cur_rank_loaded_state_dict[file_name] = renamed_state_dict
-                    else:
-                        self.get_sharded_tensor_infos(file_name, state_dict, cur_rank_sharded_tensor_infos)
+                        state_dict = self.rename_using_optimizer_state_order(model_state_keys, state_dict)
+                    renamed_state_dict = self.rename_local_view_state_dict(state_dict, file_name)
+                    self.get_sharded_tensor_infos(file_name, renamed_state_dict, cur_rank_sharded_tensor_infos)
+                    self.cur_rank_loaded_state_dict[file_name] = renamed_state_dict
             else:
                 for file, state_dict in self.cur_rank_loaded_state_dict.items():
                     # The rule for renaming is to change the master_weights name in the optimizer state to the model weight name,
@@ -897,6 +953,9 @@ def rename(old_name, parameter_to_structured_name):
             return None
 
         for key, value in state_dict.items():
+            # NOTE: Skip the parameters that are not initialized，which are not in the current rank.
+            if value is None or (isinstance(value, paddle.Tensor) and not value._is_initialized()):
+                continue
             if key in parameter_to_structured_name.values():
                 new_name = key
             else:
@@ -909,7 +968,9 @@ def rename(old_name, parameter_to_structured_name):
     def rename_using_optimizer_state_order(self, model_state_keys, optimizer_state_dict):
         name_mapping = {}
         suffix_bucket = {}
-        assert len(optimizer_state_dict) % len(model_state_keys) == 0
+        # TODO: After adapting to sharding, remove the code below.
+        if self.is_sharding_stage3 or (self.sharding_degree > 1 and self.sharding_stage1_v == 2):
+            assert len(optimizer_state_dict) % len(model_state_keys) == 0
         for suffix in OPTIMIZER_STATE_NAME_SUFFIX:
             suffix_bucket[suffix] = []
         for opt_name, opt_value in optimizer_state_dict.items():
@@ -927,10 +988,27 @@ def rename_using_optimizer_state_order(self, model_state_keys, optimizer_state_d
         for suffix, old_names in suffix_bucket.items():
             if len(old_names) == 0:
                 continue
-            assert len(old_names) == len(model_state_keys)
-            for i in range(len(old_names)):
-                name_mapping[old_names[i]] = model_state_keys[i] + suffix
-
+            # TODO: After adapting to sharding, remove the code below.
+            if self.is_sharding_stage3 or (self.sharding_degree > 1 and self.sharding_stage1_v == 2):
+                assert len(old_names) == len(model_state_keys)
+
+            # NOTE: Handle the case where the number of master_weight elements is not equal to the number of model_state_keys.
+            if suffix != ".master_weight":
+                for i in range(len(old_names)):
+                    name_mapping[old_names[i]] = model_state_keys[i] + suffix
+            else:
+                for i in range(len(old_names)):
+                    param = old_names[i][:-14]
+                    index = -1
+                    for idx, opt_name in enumerate(suffix_bucket[".moment1"]):
+                        if param == opt_name[:-24]:
+                            index = idx
+                            break
+                    if index >= 0:
+                        name_mapping[old_names[i]] = model_state_keys[index] + suffix
+                    else:
+                        raise RuntimeError(f"Can't find {param} in optimizer state dict.")
+        # rename state dict
         renamed_state_dict = {}
         for k, v in optimizer_state_dict.items():
             renamed_state_dict[name_mapping[k]] = v

paddlenlp/transformers/__init__.py

@@ -209,6 +209,9 @@
 from .xlm.modeling import *
 from .xlm.tokenizer import *
 from .xlm.configuration import *
+from .xlm_roberta.modeling import *
+from .xlm_roberta.tokenizer import *
+from .xlm_roberta.configuration import *
 from .gau_alpha.modeling import *
 from .gau_alpha.tokenizer import *
 from .gau_alpha.configuration import *

paddlenlp/transformers/auto/configuration.py

@@ -113,6 +113,7 @@
         ("unimo", "UNIMOConfig"),
         ("visualglm", "VisualGLMConfig"),
         ("xlm", "XLMConfig"),
+        ("xlm-roberta", "XLMRobertaConfig"),
         ("xlnet", "XLNetConfig"),
         ("yuan", "YuanConfig"),
     ]
@@ -202,6 +203,7 @@
         ("unimo", "UNIMO"),
         ("visualglm", "VisualGLM"),
         ("xlm", "XLM"),
+        ("xlm-roberta", "XLMRoberta"),
         ("xlnet", "XLNet"),
         ("yuan", "Yuan"),
     ]

paddlenlp/transformers/auto/modeling.py

@@ -94,6 +94,7 @@
         ("UNIMO", "unimo"),
         ("XLNet", "xlnet"),
         ("XLM", "xlm"),
+        ("XLMRoberta", "xlm_roberta"),
         ("GPT", "gpt"),
         ("GLM", "glm"),
         ("MT5", "mt5"),

paddlenlp/transformers/auto/tokenizer.py

@@ -99,6 +99,7 @@
             ("squeezebert", "SqueezeBertTokenizer"),
             ("t5", "T5Tokenizer"),
             ("xlm", "XLMTokenizer"),
+            ("xlm_roberta", "XLMRobertaTokenizer"),
             ("xlnet", "XLNetTokenizer"),
             ("bert_japanese", "BertJapaneseTokenizer"),
             ("bigbird", "BigBirdTokenizer"),

paddlenlp/transformers/xlm_roberta/__init__.py

@@ -0,0 +1,17 @@
+# Copyright (c) 2025 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from .configuration import *
+from .modeling import *
+from .tokenizer import *

paddlenlp/transformers/xlm_roberta/configuration.py

@@ -0,0 +1,160 @@
+# coding=utf-8
+# Copyright (c) 2025 PaddlePaddle Authors. All Rights Reserved.
+# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" XLM-RoBERTa configuration"""
+
+from ..model_utils import PretrainedConfig
+
+__all__ = ["PRETRAINED_INIT_CONFIGURATION", "XLMRobertaConfig"]
+
+PRETRAINED_INIT_CONFIGURATION = {
+    "hf-internal-testing/tiny-random-onnx-xlm-roberta": {
+        "attention_probs_dropout_prob": 0.1,
+        "bos_token_id": 0,
+        "classifier_dropout": None,
+        "eos_token_id": 2,
+        "hidden_act": "gelu",
+        "hidden_dropout_prob": 0.1,
+        "hidden_size": 4,
+        "initializer_range": 0.02,
+        "intermediate_size": 37,
+        "layer_norm_eps": 1e-05,
+        "max_position_embeddings": 514,
+        "model_type": "xlm-roberta",
+        "num_attention_heads": 4,
+        "num_hidden_layers": 5,
+        "output_past": True,
+        "pad_token_id": 1,
+        "position_embedding_type": "absolute",
+        "dtype": "float32",
+        "type_vocab_size": 1,
+        "use_cache": True,
+        "vocab_size": 250002,
+    },
+}
+
+
+class XLMRobertaConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`XLMRobertaModel`] or a [`TFXLMRobertaModel`]. It
+    is used to instantiate a XLM-RoBERTa model according to the specified arguments, defining the model architecture.
+    Instantiating a configuration with the defaults will yield a similar configuration to that of the XLMRoBERTa
+    [xlm-roberta-base](https://huggingface.co/xlm-roberta-base) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 30522):
+            Vocabulary size of the XLM-RoBERTa model. Defines the number of different tokens that can be represented by
+            the `inputs_ids` passed when calling [`XLMRobertaModel`] or [`TFXLMRobertaModel`].
+        hidden_size (`int`, *optional*, defaults to 768):
+            Dimensionality of the encoder layers and the pooler layer.
+        num_hidden_layers (`int`, *optional*, defaults to 12):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 12):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        intermediate_size (`int`, *optional*, defaults to 3072):
+            Dimensionality of the "intermediate" (often named feed-forward) layer in the Transformer encoder.
+        hidden_act (`str` or `Callable`, *optional*, defaults to `"gelu"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
+            `"relu"`, `"silu"` and `"gelu_new"` are supported.
+        hidden_dropout_prob (`float`, *optional*, defaults to 0.1):
+            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
+        attention_probs_dropout_prob (`float`, *optional*, defaults to 0.1):
+            The dropout ratio for the attention probabilities.
+        max_position_embeddings (`int`, *optional*, defaults to 512):
+            The maximum sequence length that this model might ever be used with. Typically set this to something large
+            just in case (e.g., 512 or 1024 or 2048).
+        type_vocab_size (`int`, *optional*, defaults to 2):
+            The vocabulary size of the `token_type_ids` passed when calling [`XLMRobertaModel`] or
+            [`TFXLMRobertaModel`].
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        layer_norm_eps (`float`, *optional*, defaults to 1e-12):
+            The epsilon used by the layer normalization layers.
+        position_embedding_type (`str`, *optional*, defaults to `"absolute"`):
+            Type of position embedding. Choose one of `"absolute"`, `"relative_key"`, `"relative_key_query"`. For
+            positional embeddings use `"absolute"`. For more information on `"relative_key"`, please refer to
+            [Self-Attention with Relative Position Representations (Shaw et al.)](https://arxiv.org/abs/1803.02155).
+            For more information on `"relative_key_query"`, please refer to *Method 4* in [Improve Transformer Models
+            with Better Relative Position Embeddings (Huang et al.)](https://arxiv.org/abs/2009.13658).
+        is_decoder (`bool`, *optional*, defaults to `False`):
+            Whether the model is used as a decoder or not. If `False`, the model is used as an encoder.
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the model should return the last key/values attentions (not used by all models). Only
+            relevant if `config.is_decoder=True`.
+        classifier_dropout (`float`, *optional*):
+            The dropout ratio for the classification head.
+
+    Examples:
+
+    ```python
+    >>> from paddlenlp.transformers import XLMRobertaConfig, XLMRobertaModel
+
+    >>> # Initializing a XLM-RoBERTa xlm-roberta-base style configuration
+    >>> configuration = XLMRobertaConfig()
+
+    >>> # Initializing a model (with random weights) from the xlm-roberta-base style configuration
+    >>> model = XLMRobertaModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "xlm-roberta"
+
+    def __init__(
+        self,
+        vocab_size=30522,
+        hidden_size=768,
+        num_hidden_layers=12,
+        num_attention_heads=12,
+        intermediate_size=3072,
+        hidden_act="gelu",
+        hidden_dropout_prob=0.1,
+        attention_probs_dropout_prob=0.1,
+        max_position_embeddings=512,
+        type_vocab_size=2,
+        initializer_range=0.02,
+        layer_norm_eps=1e-12,
+        pad_token_id=1,
+        bos_token_id=0,
+        eos_token_id=2,
+        position_embedding_type="absolute",
+        use_cache=True,
+        classifier_dropout=None,
+        **kwargs,
+    ):
+        kwargs["return_dict"] = kwargs.pop("return_dict", False)
+        super().__init__(pad_token_id=pad_token_id, bos_token_id=bos_token_id, eos_token_id=eos_token_id, **kwargs)
+
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.hidden_act = hidden_act
+        self.intermediate_size = intermediate_size
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.max_position_embeddings = max_position_embeddings
+        self.type_vocab_size = type_vocab_size
+        self.initializer_range = initializer_range
+        self.layer_norm_eps = layer_norm_eps
+        self.position_embedding_type = position_embedding_type
+        self.use_cache = use_cache
+        self.classifier_dropout = classifier_dropout