From 0c1060435d17a620a0eed5b07ab581055ca30c08 Mon Sep 17 00:00:00 2001 From: x54-729 <45304952+x54-729@users.noreply.github.com> Date: Mon, 17 Jul 2023 21:08:10 +0800 Subject: [PATCH] Use tempfile for convert2hf.py (#23) Fix https://github.com/InternLM/InternLM/issues/50 --- README-zh-Hans.md | 4 ++-- README.md | 4 ++-- tools/transformers/README-zh-Hans.md | 9 ++++----- tools/transformers/README.md | 11 +++++------ tools/transformers/convert2hf.py | 19 ++++++++++--------- tools/transformers/modeling_internlm.py | 2 +- 6 files changed, 24 insertions(+), 25 deletions(-) diff --git a/README-zh-Hans.md b/README-zh-Hans.md index c027896..c374344 100644 --- a/README-zh-Hans.md +++ b/README-zh-Hans.md @@ -145,10 +145,10 @@ streamlit run web_demo.py ### 转换为 Transformers 格式使用 -通过 InternLM 进行训练的模型可以很轻松地转换为 HuggingFace Transformers 格式，方便与社区各种开源项目无缝对接。借助 `tools/convert2hf.py` 可以将训练保存的权重一键转换为 transformers 格式 +通过 InternLM 进行训练的模型可以很轻松地转换为 HuggingFace Transformers 格式，方便与社区各种开源项目无缝对接。借助 `tools/transformers/convert2hf.py` 可以将训练保存的权重一键转换为 transformers 格式 ```bash -python convert2hf.py --src_folder origin_ckpt/ --tgt_folder hf_ckpt/ --tokenizer tokenizes/tokenizer.model +python tools/transformers/convert2hf.py --src_folder origin_ckpt/ --tgt_folder hf_ckpt/ --tokenizer ./tools/V7_sft.model ``` 转换之后可以通过以下的代码加载为 transformers diff --git a/README.md b/README.md index faa62bf..f1fa219 100644 --- a/README.md +++ b/README.md @@ -152,10 +152,10 @@ Please refer to [Usage Tutorial](./doc/en/usage.md) to start InternLM installati ### Convert to Transformers Format -The model trained by InternLM can be easily converted to HuggingFace Transformers format, which is convenient for seamless docking with various open source projects in the community. With the help of `tools/convert2hf.py`, the weights saved during training can be converted into transformers format with one command +The model trained by InternLM can be easily converted to HuggingFace Transformers format, which is convenient for seamless docking with various open source projects in the community. With the help of `tools/transformers/convert2hf.py`, the weights saved during training can be converted into transformers format with one command ```bash -python convert2hf.py --src_folder origin_ckpt/ --tgt_folder hf_ckpt/ --tokenizer tokenizes/tokenizer.model +python tools/transformers/convert2hf.py --src_folder origin_ckpt/ --tgt_folder hf_ckpt/ --tokenizer ./tools/V7_sft.model ``` After conversion, it can be loaded as transformers by the following code diff --git a/tools/transformers/README-zh-Hans.md b/tools/transformers/README-zh-Hans.md index 9b9a327..8bbdaf5 100644 --- a/tools/transformers/README-zh-Hans.md +++ b/tools/transformers/README-zh-Hans.md @@ -8,18 +8,17 @@ ## 权重转换 -`convert2hf.py` 可以将训练保存的权重一键转换为 transformers 格式。 +`convert2hf.py` 可以将训练保存的权重一键转换为 transformers 格式。在仓库根目录运行以下命令： ```bash -python convert2hf.py --src_folder origin_ckpt/ --tgt_folder hf_ckpt/ --tokenizer ../v7_sft.model +python tools/transformers/convert2hf.py --src_folder origin_ckpt/ --tgt_folder hf_ckpt/ --tokenizer ./tools/V7_sft.model ``` 然后可以使用 `from_pretrained` 接口加载： ```python -from modeling_internlm import InternLMForCausalLM - -model = InternForCausalLM.from_pretrained("hf_ckpt/") +>>> from transformers import AutoTokenizer, AutoModel +>>> model = AutoModel.from_pretrained("hf_ckpt/", trust_remote_code=True).cuda() ``` diff --git a/tools/transformers/README.md b/tools/transformers/README.md index 59d8ac8..4fe2a92 100644 --- a/tools/transformers/README.md +++ b/tools/transformers/README.md @@ -7,18 +7,17 @@ This folder contains the `InternLM` model in transformers format. ## Weight Conversion -`convert2hf.py` can convert saved training weights into the transformers format with a single command. +`convert2hf.py` can convert saved training weights into the transformers format with a single command. Execute the command in the root directory of repository: ```bash -python convert2hf.py --src_folder origin_ckpt/ --tgt_folder hf_ckpt/ --tokenizer ../v7_sft.model +python tools/transformers/convert2hf.py --src_folder origin_ckpt/ --tgt_folder hf_ckpt/ --tokenizer ./tools/V7_sft.model ``` Then, you can load it using the `from_pretrained` interface: ```python -from modeling_internlm import InternLMForCausalLM - -model = InternForCausalLM.from_pretrained("hf_ckpt/") +>>> from transformers import AutoTokenizer, AutoModel +>>> model = AutoModel.from_pretrained("hf_ckpt/", trust_remote_code=True).cuda() ``` -`intern_moss_example.py` demonstrates an example of how to use LoRA for fine-tuning on the `fnlp/moss-moon-002-sft` dataset. \ No newline at end of file +`intern_moss_example.py` demonstrates an example of how to use LoRA for fine-tuning on the `fnlp/moss-moon-002-sft` dataset. diff --git a/tools/transformers/convert2hf.py b/tools/transformers/convert2hf.py index 49ee1f6..181b68c 100644 --- a/tools/transformers/convert2hf.py +++ b/tools/transformers/convert2hf.py @@ -1,9 +1,9 @@ import argparse import math +import json import os -import random import re -import shutil +import tempfile import torch from modeling_internlm import InternLMConfig, InternLMForCausalLM @@ -15,10 +15,8 @@ NUM_SHARDS = { def convert2hf(model_config, states_tp_pps): - folder = f"/dev/shm/wait_to_upload_weight_tmp_{random.random()}/" - os.makedirs(folder, exist_ok=True) - try: + with tempfile.TemporaryDirectory() as folder: states = merge_pp(states_tp_pps)[0] if "embedding.word_embeddings.weight" in states: @@ -88,12 +86,9 @@ def convert2hf(model_config, states_tp_pps): config.save_pretrained(folder) torch.save(current_states, os.path.join(folder, "pytorch_model.bin")) - model = InternLMForCausalLM.from_pretrained(folder, torch_dtype=torch.float16, low_cpu_mem_usage=True) + model = InternLMForCausalLM.from_pretrained(folder, torch_dtype=torch.float16) del model.config._name_or_path - finally: - shutil.rmtree(folder) - return config, model @@ -169,6 +164,12 @@ if __name__ == "__main__": os.makedirs(target_folder, exist_ok=True) model.save_pretrained(target_folder, max_shard_size="20GB") + # TODO There should be a better way to add this. + with open(os.path.join(target_folder, "config.json")) as fp: + config_dict = json.load(fp) + config_dict["auto_map"]["AutoModel"] = "modeling_internlm.InternLMModel" + with open(os.path.join(target_folder, "config.json"), "w") as fp: + json.dump(config_dict, fp, indent=2) tokenizer = InternLMTokenizer(args.tokenizer) tokenizer.save_pretrained(target_folder) diff --git a/tools/transformers/modeling_internlm.py b/tools/transformers/modeling_internlm.py index 21eb7f5..df1e19f 100644 --- a/tools/transformers/modeling_internlm.py +++ b/tools/transformers/modeling_internlm.py @@ -31,7 +31,7 @@ from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutpu from transformers.modeling_utils import PreTrainedModel from transformers.generation.streamers import BaseStreamer from transformers.utils import add_start_docstrings, add_start_docstrings_to_model_forward, logging, replace_return_docstrings -from .configuration_internlm import InternLMConfig +from configuration_internlm import InternLMConfig logger = logging.get_logger(__name__)