PaddlePaddle · zhiqiu · Dec 26, 2024 · Dec 24, 2024 · Dec 25, 2024 · Dec 25, 2024
diff --git a/llm/auto_parallel/llama/README.md b/llm/auto_parallel/llama/README.md
@@ -32,3 +32,32 @@ cd ../../../slm/model_zoo/gpt-3/external_ops/ && python3 setup.py install && cd
 参考训练脚本 **run_pretrain_auto.sh**，并开启 `to_static=1`，运行8卡 dp2mp2pp2的并行策略。
 
 您可以参考 **run_pretrain_auto.sh**，按需求修改相关参数进行训练。
+
+## 4.推理
+推理流程包括：动态图推理 -> 动转静导出模型 -> 静态图推理。当前自动并行预训练保存的模型参数已支持用于动态图推理；动转静导出模型、静态图推理步骤请参考 [LLaMA 系列大模型运行文档](https://github.com/PaddlePaddle/PaddleNLP/blob/develop/llm/docs/predict/llama.md)。
+
+以动态图自动并行训练（dp2mp2pp2）为例。
+- 分布式 ckpt 合并为单卡模型参数：
+
+```python
+import paddle
+import paddle.distributed as dist
+
+ckpt_path='/path/for/dist_ckpt'
+# offload=1 将参数 offload 到 CPU，减少显存占用
+merged_state_dict = dist.checkpoint.load_state_dict.load_merged_state_dict(ckpt_path, offload=1)
+paddle.save(unsharded_state_dict, 'model_state.pdparams')
+
+# 上述合并的模型参数格式为Paddle原生格式，如需转换为unified_param格式(safetensors)，可继续执行如下代码：
+python PaddleNLP/llm/auto_parallel/utils/convert_to_safetensors.py --input_path input_path  [--output_path output_path] [--split_num split_num] [--offload offload]
+
+# 参数介绍
+--input_path: 输入的单卡模型参数路径
+--output_path: 可选，输出模型参数路径，默认为'./temp'
+--split_num: 可选，输出的模型参数分片数，默认为 1
+--offload: 可选，是否将参数 offload 到 CPU，默认为 false
+```
+
+- 动态图推理
+
+    [大模型推理教程](https://github.com/PaddlePaddle/PaddleNLP/blob/develop/llm/docs/predict/inference.md)
diff --git a/llm/auto_parallel/utils/convert_to_safetensors.py b/llm/auto_parallel/utils/convert_to_safetensors.py
@@ -0,0 +1,92 @@
+# Copyright (c) 2024 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.
+
+import json
+import os
+
+import paddle
+from safetensors.numpy import save_file as safe_save_file
+
+from paddlenlp.transformers.utils import dtype_byte_size
+from paddlenlp.utils.env import SAFE_WEIGHTS_INDEX_NAME
+
+
+def convert_to_unified_ckpt(path: str, output_dir: str = "./tmp", split_num: int = 1, offload: bool = False):
+    """
+    Convert a single card checkpoint to the unified format.
+
+    Args:
+        path (str): The path to the input checkpoint file.
+        output_dir (str, optional): The directory where the converted files will be saved. Defaults to ".".
+        split_num (int, optional): The number of shards to split the weights into output_dir. Defaults to 1.
+        offload (bool, optional): Whether to offload the weights to CPU memory before saving them. Defaults to False.
+    """
+
+    def get_sub_state_dict(sub_keys, state_dict, weight_filename, index_weight_file, total_size):
+        """
+        Get the sub-state dict and update the index weight file and total size.
+        Args:
+            sub_keys (list): A list of keys that belong to this sub-state dict.
+            state_dict (dict): The original state dict.
+            weight_filename (str): The filename of the corresponding weight file.
+            index_weight_file (dict): The dictionary containing the mapping from keys to their corresponding weight filenames.
+            total_size (int): The total size of the model so far.
+        """
+        sub_state_dict = {key: state_dict[key].numpy() for key in sub_keys}
+        for key in sub_keys:
+            index_weight_file[key] = weight_filename
+            total_size += state_dict[key].numel().item() * dtype_byte_size(state_dict[key].dtype)
+        return sub_state_dict, total_size
+
+    if offload:
+        paddle.set_device("cpu")
+    state_dict = paddle.load(path)
+    all_keys = list(state_dict.keys())
+    split_size = len(all_keys) // split_num
+    extra_keys = len(all_keys) % split_num
+    index_weight_file = {}
+    total_size = 0
+
+    os.makedirs(output_dir, exist_ok=True)
+
+    index = 0
+    for rank in range(split_num):
+        current_size = split_size + (1 if rank < extra_keys else 0)
+        sub_keys = all_keys[index : index + current_size]
+        index += current_size
+        weight_filename = f"model-{rank+1:04d}-of-{split_num:04d}.safetensors"
+        sub_state_dict, total_size = get_sub_state_dict(
+            sub_keys, state_dict, weight_filename, index_weight_file, total_size
+        )
+        safe_save_file(sub_state_dict, os.path.join(output_dir, weight_filename))
+    with open(os.path.join(output_dir, SAFE_WEIGHTS_INDEX_NAME), "w") as f:
+        json.dump({"metadata": {"total_size": total_size}, "weight_map": index_weight_file}, f, indent=4)
+
+
+if __name__ == "__main__":
+    import argparse
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--input_path", type=str, required=True, help="The path to the input checkpoint file.")
+    parser.add_argument(
+        "--output_dir", type=str, default="./tmp", help="The directory where the converted files will be saved."
+    )
+    parser.add_argument(
+        "--split_num", type=int, default=1, help="The number of shards to split the weights into output_dir."
+    )
+    parser.add_argument(
+        "--offload", type=bool, help="Whether to offload the weights to CPU memory before saving them."
+    )
+    args = parser.parse_args()
+    convert_to_unified_ckpt(args.input_path, args.output_dir, args.split_num, args.offload)