快速体验LoRA微调Llama3-8B模型以及海光DCU推理加快（曙光超算互联网平台国 ...

王柳  金牌会员 | 2024-9-3 22:38:50 | 显示全部楼层 | 阅读模式
打印 上一主题 下一主题
楼主

主题 650|帖子 650|积分 1950

序言

本文以 LLaMA-Factory 为例，在超算互联网平台SCNet上利用异构加快卡AI 显存64GB PCIE，对 Llama3-8B-Instruct 模型进行 LoRA 微调、归并和推理。
一、参考资料

github仓库代码：LLaMA-Factory，利用最新的代码分支：v0.8.3。
超算互联网平台
异构加快卡AI 显存64GB PCIE
二、预备情况

1. 体系镜像

异构加快卡AI为国产加快卡（DCU），基于DTK软件栈（对标NVIDIA的CUDA），请选择 dtk24.04 版本的镜像情况。
以jupyterlab-pytorch:2.1.0-ubuntu20.04-dtk24.04-py310 镜像为例。
2. 软硬件依赖

特别注意：由 requirements.txt 文件可知，LLaMA-Factory项目要求最低版本 transformers>=4.41.2，vllm>=0.4.3 。
必须项至少推荐python3.83.11torch1.13.12.3.0transformers4.41.24.41.2datasets2.16.02.19.2accelerate0.30.10.30.1peft0.11.10.11.1trl0.8.60.9.4 可选项至少推荐CUDA11.612.2deepspeed0.10.00.14.0bitsandbytes0.39.00.43.1vllm0.4.30.4.3flash-attn2.3.02.5.9 3. 克隆base情况

  1. root@notebook-1813389960667746306-scnlbe5oi5-17811:/public/home/scnlbe5oi5/Downloads/models/LLaMA-Factory# conda create -n llama_factory_torch --clone base
  2. Source:      /opt/conda
  3. Destination: /opt/conda/envs/llama_factory_torch
  4. The following packages cannot be cloned out of the root environment:
  5. - https://repo.anaconda.com/pkgs/main/linux-64::conda-23.7.4-py310h06a4308_0
  6. Packages: 44
  7. Files: 53489
  9. Downloading and Extracting Packages
  12. Downloading and Extracting Packages
  14. Preparing transaction: done
  15. Verifying transaction: done
  16. Executing transaction: done
  17. #
  18. # To activate this environment, use
  19. #
  20. #     $ conda activate llama_factory_torch
  21. #
  22. # To deactivate an active environment, use
  23. #
  24. #     $ conda deactivate
复制代码
4. 安装 LLaMA Factory

  1. git clone --depth 1 https://github.com/hiyouga/LLaMA-Factory.git
  2. cd LLaMA-Factory
  3. pip install -e ".[torch,metrics]"
复制代码
  1. root@notebook-1813389960667746306-scnlbe5oi5-17811:/public/home/scnlbe5oi5/Downloads/models/LLaMA-Factory# source activate llama_factory_torch
  2. (llama_factory_torch) root@notebook-1813389960667746306-scnlbe5oi5-17811:/public/home/scnlbe5oi5/Downloads/m
  3. odels/LLaMA-Factory# pip install -e ".[torch,metrics]"
  4. Looking in indexes: https://pypi.tuna.tsinghua.edu.cn/simple
  5. Obtaining file:///public/home/scnlbe5oi5/Downloads/models/LLaMA-Factory
  6.   Installing build dependencies ... done
  7.   Checking if build backend supports build_editable ... done
  8.   Getting requirements to build editable ... done
  9.   Preparing editable metadata (pyproject.toml) ... done
  10. ...
  11. Checking if build backend supports build_editable ... done
  12. Building wheels for collected packages: llamafactory
  13.   Building editable for llamafactory (pyproject.toml) ... done
  14.   Created wheel for llamafactory: filename=llamafactory-0.8.4.dev0-0.editable-py3-none-any.whl size=20781 sha256=70c0480e2b648516e0eac3d39371d4100cbdaa1f277d87b657bf2adec9e0b2be
  15.   Stored in directory: /tmp/pip-ephem-wheel-cache-uhypmj_8/wheels/e9/b4/89/f13e921e37904ee0c839434aad2d7b2951c2c68e596667c7ef
  16. Successfully built llamafactory
  17. DEPRECATION: lmdeploy 0.1.0-git782048c.abi0.dtk2404.torch2.1. has a non-standard version number. pip 24.1 will enforce this behaviour change. A possible replacement is to upgrade to a newer version of lmdeploy or contact the author to suggest that they release a version with a conforming version number. Discussion can be found at https://github.com/pypa/pip/issues/12063
  18. DEPRECATION: mmcv 2.0.1-gitc0ccf15.abi0.dtk2404.torch2.1. has a non-standard version number. pip 24.1 will enforce this behaviour change. A possible replacement is to upgrade to a newer version of mmcv or contact the author to suggest that they release a version with a conforming version number. Discussion can be found at https://github.com/pypa/pip/issues/12063
  19. Installing collected packages: pydub, jieba, urllib3, tomlkit, shtab, semantic-version, scipy, ruff, rouge-chinese, joblib, importlib-resources, ffmpy, docstring-parser, aiofiles, nltk, tyro, sse-starlette, tokenizers, gradio-client, transformers, trl, peft, gradio, llamafactory
  20.   Attempting uninstall: urllib3
  21.     Found existing installation: urllib3 1.26.13
  22.     Uninstalling urllib3-1.26.13:
  23.       Successfully uninstalled urllib3-1.26.13
  24.   Attempting uninstall: tokenizers
  25.     Found existing installation: tokenizers 0.15.0
  26.     Uninstalling tokenizers-0.15.0:
  27.       Successfully uninstalled tokenizers-0.15.0
  28.   Attempting uninstall: transformers
  29.     Found existing installation: transformers 4.38.0
  30.     Uninstalling transformers-4.38.0:
  31.       Successfully uninstalled transformers-4.38.0
  32. ERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts.
  33. lmdeploy 0.1.0-git782048c.abi0.dtk2404.torch2.1. requires transformers==4.33.2, but you have transformers 4.43.3 which is incompatible.
  34. Successfully installed aiofiles-23.2.1 docstring-parser-0.16 ffmpy-0.4.0 gradio-4.40.0 gradio-client-1.2.0 importlib-resources-6.4.0 jieba-0.42.1 joblib-1.4.2 llamafactory-0.8.4.dev0 nltk-3.8.1 peft-0.12.0 pydub-0.25.1 rouge-chinese-1.0.3 ruff-0.5.5 scipy-1.14.0 semantic-version-2.10.0 shtab-1.7.1 sse-starlette-2.1.3 tokenizers-0.19.1 tomlkit-0.12.0 transformers-4.43.3 trl-0.9.6 tyro-0.8.5 urllib3-2.2.2
  35. WARNING: Running pip as the 'root' user can result in broken permissions and conflicting behaviour with the system package manager. It is recommended to use a virtual environment instead: https://pip.pypa.io/warnings/venv
  37. [notice] A new release of pip is available: 24.0 -> 24.2
  38. [notice] To update, run: pip install --upgrade pip
复制代码
5. 解决依赖包冲突

  1. (llama_factory_torch) root@notebook-1813389960667746306-scnlbe5oi5-17811:/public/home/scnlbe5oi5/Downloads/m
  2. odels/LLaMA-Factory# pip install --no-deps -e .
  3. Looking in indexes: https://pypi.tuna.tsinghua.edu.cn/simple
  4. Obtaining file:///public/home/scnlbe5oi5/Downloads/models/LLaMA-Factory
  5.   Installing build dependencies ... done
  6.   Checking if build backend supports build_editable ... done
  7.   Getting requirements to build editable ... done
  8.   Preparing editable metadata (pyproject.toml) ... done
  9. Building wheels for collected packages: llamafactory
  10.   Building editable for llamafactory (pyproject.toml) ... done
  11.   Created wheel for llamafactory: filename=llamafactory-0.8.4.dev0-0.editable-py3-none-any.whl size=20781 sha256=f874a791bc9fdca02075cda0459104b48a57d300a077eca00eee7221cde429c3
  12.   Stored in directory: /tmp/pip-ephem-wheel-cache-7vjiq3f3/wheels/e9/b4/89/f13e921e37904ee0c839434aad2d7b2951c2c68e596667c7ef
  13. Successfully built llamafactory
  14. Installing collected packages: llamafactory
  15.   Attempting uninstall: llamafactory
  16.     Found existing installation: llamafactory 0.8.4.dev0
  17.     Uninstalling llamafactory-0.8.4.dev0:
  18.       Successfully uninstalled llamafactory-0.8.4.dev0
  19. Successfully installed llamafactory-0.8.4.dev0
  20. WARNING: Running pip as the 'root' user can result in broken permissions and conflicting behaviour with the system package manager. It is recommended to use a virtual environment instead: https://pip.pypa.io/warnings/venv
  22. [notice] A new release of pip is available: 24.0 -> 24.2
  23. [notice] To update, run: pip install --upgrade pip
复制代码
6. 安装 vllm==0.4.3

  1. (llama_factory_torch) root@notebook-1813389960667746306-scnlbe5oi5-17811:/public/home/scnlbe5oi5/Downloads/m
  2. odels/LLaMA-Factory# pip list | grep llvm
  4. [notice] A new release of pip is available: 24.0 -> 24.2
  5. [notice] To update, run: pip install --upgrade pip
  6. (llama_factory_torch) root@notebook-1813389960667746306-scnlbe5oi5-17811:/public/home/scnlbe5oi5/Downloads/m
  7. odels/LLaMA-Factory# pip install --no-dependencies vllm==0.4.3
  8. Looking in indexes: https://pypi.tuna.tsinghua.edu.cn/simple
  9. Collecting vllm==0.4.3
  10.   Using cached https://pypi.tuna.tsinghua.edu.cn/packages/1a/1e/10bcb6566f4fa8b95ff85bddfd1675ff7db33ba861f59bd70aa3b92a46b7/vllm-0.4.3-cp310-cp310-manylinux1_x86_64.whl (131.1 MB)
  11. Installing collected packages: vllm
  12.   Attempting uninstall: vllm
  13.     Found existing installation: vllm 0.3.3+git3380931.abi0.dtk2404.torch2.1
  14.     Uninstalling vllm-0.3.3+git3380931.abi0.dtk2404.torch2.1:
  15.       Successfully uninstalled vllm-0.3.3+git3380931.abi0.dtk2404.torch2.1
  16. Successfully installed vllm-0.4.3
  17. WARNING: Running pip as the 'root' user can result in broken permissions and conflicting behaviour with the system package manager. It is recommended to use a virtual environment instead: https://pip.pypa.io/warnings/venv
  19. [notice] A new release of pip is available: 24.0 -> 24.2
  20. [notice] To update, run: pip install --upgrade pip
复制代码
三、关键步调

1. 获取Access Token

获取Access Token，并登录Hugging Face 账户。
详细步调，请参考另一篇博客：Hugging Face和ModelScope大模型/数据集的下载加快方法
  1. pip install --upgrade huggingface_hub
  2. huggingface-cli login
复制代码
2. llamafactory-cli 指令

利用 llamafactory-cli help 表现帮助信息。
  1. (llama_fct) root@notebook-1813389960667746306-scnlbe5oi5-17811:/public/home/scnlbe5oi5/Downloads/mod
  2. els/LLaMA-Factory# llamafactory-cli help
  3. No ROCm runtime is found, using ROCM_HOME='/opt/dtk'
  4. /opt/conda/envs/llama_fct/lib/python3.10/site-packages/torchvision/io/image.py:13: UserWarning: Fail                                     ed to load image Python extension: 'libc10_hip.so: cannot open shared object file: No such file or d                                     irectory'If you don't plan on using image functionality from `torchvision.io`, you can ignore this w                                     arning. Otherwise, there might be something wrong with your environment. Did you have `libjpeg` or `                                     libpng` installed before building `torchvision` from source?
  5.   warn(
  6. [2024-08-01 15:12:24,629] [INFO] [real_accelerator.py:158:get_accelerator] Setting ds_accelerator to                                      cuda (auto detect)
  7. ----------------------------------------------------------------------
  8. | Usage:                                                             |
  9. |   llamafactory-cli api -h: launch an OpenAI-style API server       |
  10. |   llamafactory-cli chat -h: launch a chat interface in CLI         |
  11. |   llamafactory-cli eval -h: evaluate models                        |
  12. |   llamafactory-cli export -h: merge LoRA adapters and export model |
  13. |   llamafactory-cli train -h: train models                          |
  14. |   llamafactory-cli webchat -h: launch a chat interface in Web UI   |
  15. |   llamafactory-cli webui: launch LlamaBoard                        |
  16. |   llamafactory-cli version: show version info                      |
  17. ----------------------------------------------------------------------
复制代码
3. 快速开始

下面三行命令分别对 Llama3-8B-Instruct 模型进行 LoRA 微调、归并和推理。
  1. llamafactory-cli train examples/train_lora/llama3_lora_sft.yaml
  3. llamafactory-cli export examples/merge_lora/llama3_lora_sft.yaml
  4. llamafactory-cli chat examples/inference/llama3_lora_sft.yaml
复制代码
3.1 LoRA 微调

模型微调训练是在DCU上进行的。
3.1.1 单卡情况

  1. (llama_factory_torch) root@notebook-1813389960667746306-scnlbe5oi5-17811:/public/home/scnlbe5oi5/Downloads/m
  2. odels/LLaMA-Factory# llamafactory-cli train examples/train_lora/llama3_lora_sft.yaml
  4. [2024-08-01 19:06:41,134] [INFO] [real_accelerator.py:158:get_accelerator] Setting ds_accelerator to cuda (a                                                                            uto detect)
  5. 08/01/2024 19:06:44 - INFO - llamafactory.hparams.parser - Process rank: 0, device: cuda:0, n_gpu: 1, distri                                                                            buted training: False, compute dtype: torch.bfloat16
  6. [INFO|tokenization_utils_base.py:2287] 2024-08-01 19:06:45,194 >> loading file tokenizer.json
  7. [INFO|tokenization_utils_base.py:2287] 2024-08-01 19:06:45,194 >> loading file added_tokens.json
  8. [INFO|tokenization_utils_base.py:2287] 2024-08-01 19:06:45,194 >> loading file special_tokens_map.json
  9. [INFO|tokenization_utils_base.py:2287] 2024-08-01 19:06:45,194 >> loading file tokenizer_config.json
  10. [INFO|tokenization_utils_base.py:2533] 2024-08-01 19:06:45,563 >> Special tokens have been added in the voca                                                                            bulary, make sure the associated word embeddings are fine-tuned or trained.
  11. 08/01/2024 19:06:45 - INFO - llamafactory.data.template - Replace eos token: <|eot_id|>
  12. 08/01/2024 19:06:45 - INFO - llamafactory.data.template - Add pad token: <|eot_id|>
  13. 08/01/2024 19:06:45 - INFO - llamafactory.data.loader - Loading dataset identity.json...
  14. Converting format of dataset (num_proc=16): 100%|███████████████████| 91/91 [00:00<00:00, 444.18 examples/s]
  15. 08/01/2024 19:06:47 - INFO - llamafactory.data.loader - Loading dataset alpaca_en_demo.json...
  16. Converting format of dataset (num_proc=16): 100%|██████████████| 1000/1000 [00:00<00:00, 4851.17 examples/s]
  17. Running tokenizer on dataset (num_proc=16): 100%|███████████████| 1091/1091 [00:02<00:00, 375.29 examples/s]
  18. training example:
  19. input_ids:
  20. [128000, 128006, 882, 128007, 271, 6151, 128009, 128006, 78191, 128007, 271, 9906, 0, 358, 1097, 5991, 609,                                                                             39254, 459, 15592, 18328, 8040, 555, 5991, 3170, 3500, 13, 2650, 649, 358, 7945, 499, 3432, 30, 128009]
  21. inputs:
  22. <|begin_of_text|><|start_header_id|>user<|end_header_id|>
  24. hi<|eot_id|><|start_header_id|>assistant<|end_header_id|>
  26. Hello! I am {{name}}, an AI assistant developed by {{author}}. How can I assist you today?<|eot_id|>
  27. label_ids:
  28. [-100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, 9906, 0, 358, 1097, 5991, 609, 39254, 459                                                                            , 15592, 18328, 8040, 555, 5991, 3170, 3500, 13, 2650, 649, 358, 7945, 499, 3432, 30, 128009]
  29. labels:
  30. Hello! I am {{name}}, an AI assistant developed by {{author}}. How can I assist you today?<|eot_id|>
  31. [INFO|configuration_utils.py:731] 2024-08-01 19:06:53,502 >> loading configuration file /root/.cache/modelsc                                                                            ope/hub/LLM-Research/Meta-Llama-3-8B-Instruct/config.json
  32. [INFO|configuration_utils.py:800] 2024-08-01 19:06:53,503 >> Model config LlamaConfig {
  33.   "_name_or_path": "/root/.cache/modelscope/hub/LLM-Research/Meta-Llama-3-8B-Instruct",
  34.   "architectures": [
  35.     "LlamaForCausalLM"
  36.   ],
  37.   "attention_bias": false,
  38.   "attention_dropout": 0.0,
  39.   "bos_token_id": 128000,
  40.   "eos_token_id": 128009,
  41.   "hidden_act": "silu",
  42.   "hidden_size": 4096,
  43.   "initializer_range": 0.02,
  44.   "intermediate_size": 14336,
  45.   "max_position_embeddings": 8192,
  46.   "mlp_bias": false,
  47.   "model_type": "llama",
  48.   "num_attention_heads": 32,
  49.   "num_hidden_layers": 32,
  50.   "num_key_value_heads": 8,
  51.   "pretraining_tp": 1,
  52.   "rms_norm_eps": 1e-05,
  53.   "rope_scaling": null,
  54.   "rope_theta": 500000.0,
  55.   "tie_word_embeddings": false,
  56.   "torch_dtype": "bfloat16",
  57.   "transformers_version": "4.43.3",
  58.   "use_cache": true,
  59.   "vocab_size": 128256
  60. }
  62. [INFO|modeling_utils.py:3631] 2024-08-01 19:06:53,534 >> loading weights file /root/.cache/modelscope/hub/LL                                                                            M-Research/Meta-Llama-3-8B-Instruct/model.safetensors.index.json
  63. [INFO|modeling_utils.py:1572] 2024-08-01 19:06:53,534 >> Instantiating LlamaForCausalLM model under default                                                                             dtype torch.bfloat16.
  64. [INFO|configuration_utils.py:1038] 2024-08-01 19:06:53,536 >> Generate config GenerationConfig {
  65.   "bos_token_id": 128000,
  66.   "eos_token_id": 128009
  67. }
  69. Loading checkpoint shards: 100%|██████████████████████████████████████████████| 4/4 [00:08<00:00,  2.04s/it]
  70. [INFO|modeling_utils.py:4463] 2024-08-01 19:07:01,775 >> All model checkpoint weights were used when initial                                                                            izing LlamaForCausalLM.
  72. [INFO|modeling_utils.py:4471] 2024-08-01 19:07:01,775 >> All the weights of LlamaForCausalLM were initialize                                                                            d from the model checkpoint at /root/.cache/modelscope/hub/LLM-Research/Meta-Llama-3-8B-Instruct.
  73. If your task is similar to the task the model of the checkpoint was trained on, you can already use LlamaFor                                                                            CausalLM for predictions without further training.
  74. [INFO|configuration_utils.py:991] 2024-08-01 19:07:01,779 >> loading configuration file /root/.cache/modelsc                                                                            ope/hub/LLM-Research/Meta-Llama-3-8B-Instruct/generation_config.json
  75. [INFO|configuration_utils.py:1038] 2024-08-01 19:07:01,780 >> Generate config GenerationConfig {
  76.   "bos_token_id": 128000,
  77.   "do_sample": true,
  78.   "eos_token_id": [
  79.     128001,
  80.     128009
  81.   ],
  82.   "max_length": 4096,
  83.   "temperature": 0.6,
  84.   "top_p": 0.9
  85. }
  87. 08/01/2024 19:07:01 - INFO - llamafactory.model.model_utils.checkpointing - Gradient checkpointing enabled.
  88. 08/01/2024 19:07:01 - INFO - llamafactory.model.model_utils.attention - Using vanilla attention implementati                                                                            on.
  89. 08/01/2024 19:07:01 - INFO - llamafactory.model.adapter - Upcasting trainable params to float32.
  90. 08/01/2024 19:07:01 - INFO - llamafactory.model.adapter - Fine-tuning method: LoRA
  91. 08/01/2024 19:07:01 - INFO - llamafactory.model.model_utils.misc - Found linear modules: q_proj,up_proj,v_pr                                                                            oj,down_proj,k_proj,o_proj,gate_proj
  92. 08/01/2024 19:07:04 - INFO - llamafactory.model.loader - trainable params: 20,971,520 || all params: 8,051,232,768 || trainable%: 0.2605
  93. Detected kernel version 3.10.0, which is below the recommended minimum of 5.5.0; this can cause the process to hang. It is recommended to upgrade the kernel to the minimum version or higher.
  94. [INFO|trainer.py:648] 2024-08-01 19:07:04,471 >> Using auto half precision backend
  95. [INFO|trainer.py:2134] 2024-08-01 19:07:04,831 >> ***** Running training *****
  96. [INFO|trainer.py:2135] 2024-08-01 19:07:04,831 >>   Num examples = 981
  97. [INFO|trainer.py:2136] 2024-08-01 19:07:04,831 >>   Num Epochs = 3
  98. [INFO|trainer.py:2137] 2024-08-01 19:07:04,832 >>   Instantaneous batch size per device = 1
  99. [INFO|trainer.py:2140] 2024-08-01 19:07:04,832 >>   Total train batch size (w. parallel, distributed & accumulation) = 8
  100. [INFO|trainer.py:2141] 2024-08-01 19:07:04,832 >>   Gradient Accumulation steps = 8
  101. [INFO|trainer.py:2142] 2024-08-01 19:07:04,832 >>   Total optimization steps = 366
  102. [INFO|trainer.py:2143] 2024-08-01 19:07:04,836 >>   Number of trainable parameters = 20,971,520
  103. {'loss': 1.5025, 'grad_norm': 1.3309401273727417, 'learning_rate': 2.702702702702703e-05, 'epoch': 0.08}
  104. {'loss': 1.3424, 'grad_norm': 1.8096668720245361, 'learning_rate': 5.405405405405406e-05, 'epoch': 0.16}
  105. {'loss': 1.1286, 'grad_norm': 1.2990491390228271, 'learning_rate': 8.108108108108109e-05, 'epoch': 0.24}
  106. {'loss': 0.9808, 'grad_norm': 1.1075998544692993, 'learning_rate': 9.997948550797227e-05, 'epoch': 0.33}
  107. {'loss': 0.9924, 'grad_norm': 1.8073676824569702, 'learning_rate': 9.961525153583327e-05, 'epoch': 0.41}
  108. {'loss': 1.0052, 'grad_norm': 1.2079122066497803, 'learning_rate': 9.879896064123961e-05, 'epoch': 0.49}
  109. {'loss': 0.9973, 'grad_norm': 1.7361079454421997, 'learning_rate': 9.753805025397779e-05, 'epoch': 0.57}
  110. {'loss': 0.8488, 'grad_norm': 1.1059085130691528, 'learning_rate': 9.584400884284545e-05, 'epoch': 0.65}
  111. {'loss': 0.9893, 'grad_norm': 0.8711654543876648, 'learning_rate': 9.373227124134888e-05, 'epoch': 0.73}
  112. {'loss': 0.9116, 'grad_norm': 1.3793599605560303, 'learning_rate': 9.122207801708802e-05, 'epoch': 0.82}
  113. {'loss': 1.0429, 'grad_norm': 1.3769993782043457, 'learning_rate': 8.833630016614976e-05, 'epoch': 0.9}
  114. {'loss': 0.9323, 'grad_norm': 1.2503643035888672, 'learning_rate': 8.510123072976239e-05, 'epoch': 0.98}
  115. {'loss': 0.9213, 'grad_norm': 2.449227809906006, 'learning_rate': 8.154634523184388e-05, 'epoch': 1.06}
  116. {'loss': 0.8386, 'grad_norm': 1.009852409362793, 'learning_rate': 7.770403312015721e-05, 'epoch': 1.14}
  117. 40%|███████████████████████████▌                                         | 146/366 [10:19<15:11,  4.14s/it]                                                                            {'loss': 0.856, 'grad_norm': 0.863474428653717, 'learning_rate': 7.360930265797935e-05, 'epoch': 1.22}
  118. {'loss': 0.838, 'grad_norm': 0.712546169757843, 'learning_rate': 6.929946195508932e-05, 'epoch': 1.3}
  119. {'loss': 0.8268, 'grad_norm': 1.6060960292816162, 'learning_rate': 6.481377904428171e-05, 'epoch': 1.39}
  120. {'loss': 0.7326, 'grad_norm': 0.7863644957542419, 'learning_rate': 6.019312410053286e-05, 'epoch': 1.47}
  121. {'loss': 0.7823, 'grad_norm': 0.8964634537696838, 'learning_rate': 5.547959706265068e-05, 'epoch': 1.55}
  122. {'loss': 0.7599, 'grad_norm': 0.5305138826370239, 'learning_rate': 5.0716144050239375e-05, 'epoch': 1.63}
  123. {'loss': 0.815, 'grad_norm': 0.8153926730155945, 'learning_rate': 4.594616607090028e-05, 'epoch': 1.71}
  124. {'loss': 0.8258, 'grad_norm': 1.3266267776489258, 'learning_rate': 4.121312358283463e-05, 'epoch': 1.79}
  125. {'loss': 0.7446, 'grad_norm': 1.8706341981887817, 'learning_rate': 3.656014051577713e-05, 'epoch': 1.88}
  126. {'loss': 0.7539, 'grad_norm': 1.5148639678955078, 'learning_rate': 3.202961135812437e-05, 'epoch': 1.96}
  127. {'loss': 0.7512, 'grad_norm': 1.3771291971206665, 'learning_rate': 2.7662814890184818e-05, 'epoch': 2.04}
  128. {'loss': 0.7128, 'grad_norm': 1.420331597328186, 'learning_rate': 2.3499538082923606e-05, 'epoch': 2.12}
  129. {'loss': 0.635, 'grad_norm': 0.9235875010490417, 'learning_rate': 1.9577713588953795e-05, 'epoch': 2.2}
  130. {'loss': 0.6628, 'grad_norm': 1.6558737754821777, 'learning_rate': 1.5933074128684332e-05, 'epoch': 2.28}
  131. {'loss': 0.681, 'grad_norm': 0.8138720393180847, 'learning_rate': 1.2598826920598772e-05, 'epoch': 2.36}
  132. {'loss': 0.6707, 'grad_norm': 1.0700312852859497, 'learning_rate': 9.605351122011309e-06, 'epoch': 2.45}
  133. {'loss': 0.6201, 'grad_norm': 1.3334729671478271, 'learning_rate': 6.979921036993042e-06, 'epoch': 2.53}
  134. {'loss': 0.6698, 'grad_norm': 1.440247893333435, 'learning_rate': 4.746457613389904e-06, 'epoch': 2.61}
  135. {'loss': 0.7072, 'grad_norm': 0.9171076416969299, 'learning_rate': 2.925310493105099e-06, 'epoch': 2.69}
  136. {'loss': 0.6871, 'grad_norm': 0.9809044003486633, 'learning_rate': 1.5330726014397668e-06, 'epoch': 2.77}
  137. {'loss': 0.5931, 'grad_norm': 1.7158288955688477, 'learning_rate': 5.824289648152126e-07, 'epoch': 2.85}
  138. {'loss': 0.6827, 'grad_norm': 1.3241132497787476, 'learning_rate': 8.204113433559201e-08, 'epoch': 2.94}
  139. 100%|█████████████████████████████████████████████████████████████████████| 366/366 [25:42<00:00,  4.02s/it]                                                                            [INFO|trainer.py:3503] 2024-08-01 19:32:47,527 >> Saving model checkpoint to saves/llama3-8b/lora/sft/checkp                                                                            oint-366
  140. [INFO|configuration_utils.py:731] 2024-08-01 19:32:47,556 >> loading configuration file /root/.cache/modelsc                                                                            ope/hub/LLM-Research/Meta-Llama-3-8B-Instruct/config.json
  141. [INFO|configuration_utils.py:800] 2024-08-01 19:32:47,557 >> Model config LlamaConfig {
  142.   "architectures": [
  143.     "LlamaForCausalLM"
  144.   ],
  145.   "attention_bias": false,
  146.   "attention_dropout": 0.0,
  147.   "bos_token_id": 128000,
  148.   "eos_token_id": 128009,
  149.   "hidden_act": "silu",
  150.   "hidden_size": 4096,
  151.   "initializer_range": 0.02,
  152.   "intermediate_size": 14336,
  153.   "max_position_embeddings": 8192,
  154.   "mlp_bias": false,
  155.   "model_type": "llama",
  156.   "num_attention_heads": 32,
  157.   "num_hidden_layers": 32,
  158.   "num_key_value_heads": 8,
  159.   "pretraining_tp": 1,
  160.   "rms_norm_eps": 1e-05,
  161.   "rope_scaling": null,
  162.   "rope_theta": 500000.0,
  163.   "tie_word_embeddings": false,
  164.   "torch_dtype": "bfloat16",
  165.   "transformers_version": "4.43.3",
  166.   "use_cache": true,
  167.   "vocab_size": 128256
  168. }
  170. [INFO|tokenization_utils_base.py:2702] 2024-08-01 19:32:47,675 >> tokenizer config file saved in saves/llama                                                                            3-8b/lora/sft/checkpoint-366/tokenizer_config.json
  171. [INFO|tokenization_utils_base.py:2711] 2024-08-01 19:32:47,677 >> Special tokens file saved in saves/llama3-                                                                            8b/lora/sft/checkpoint-366/special_tokens_map.json
  172. [INFO|trainer.py:2394] 2024-08-01 19:32:48,046 >>
  174. Training completed. Do not forget to share your model on huggingface.co/models =)
  177. {'train_runtime': 1543.2099, 'train_samples_per_second': 1.907, 'train_steps_per_second': 0.237, 'train_loss                                                                            ': 0.8416516305318947, 'epoch': 2.98}
  178. 100%|█████████████████████████████████████████████████████████████████████| 366/366 [25:43<00:00,  4.22s/it]
  179. [INFO|trainer.py:3503] 2024-08-01 19:32:48,050 >> Saving model checkpoint to saves/llama3-8b/lora/sft
  180. [INFO|configuration_utils.py:731] 2024-08-01 19:32:48,081 >> loading configuration file /root/.cache/modelsc                                                                            ope/hub/LLM-Research/Meta-Llama-3-8B-Instruct/config.json
  181. [INFO|configuration_utils.py:800] 2024-08-01 19:32:48,082 >> Model config LlamaConfig {
  182.   "architectures": [
  183.     "LlamaForCausalLM"
  184.   ],
  185.   "attention_bias": false,
  186.   "attention_dropout": 0.0,
  187.   "bos_token_id": 128000,
  188.   "eos_token_id": 128009,
  189.   "hidden_act": "silu",
  190.   "hidden_size": 4096,
  191.   "initializer_range": 0.02,
  192.   "intermediate_size": 14336,
  193.   "max_position_embeddings": 8192,
  194.   "mlp_bias": false,
  195.   "model_type": "llama",
  196.   "num_attention_heads": 32,
  197.   "num_hidden_layers": 32,
  198.   "num_key_value_heads": 8,
  199.   "pretraining_tp": 1,
  200.   "rms_norm_eps": 1e-05,
  201.   "rope_scaling": null,
  202.   "rope_theta": 500000.0,
  203.   "tie_word_embeddings": false,
  204.   "torch_dtype": "bfloat16",
  205.   "transformers_version": "4.43.3",
  206.   "use_cache": true,
  207.   "vocab_size": 128256
  208. }
  210. [INFO|tokenization_utils_base.py:2702] 2024-08-01 19:32:48,191 >> tokenizer config file saved in saves/llama                                                                            3-8b/lora/sft/tokenizer_config.json
  211. [INFO|tokenization_utils_base.py:2711] 2024-08-01 19:32:48,192 >> Special tokens file saved in saves/llama3-                                                                            8b/lora/sft/special_tokens_map.json
  212. ***** train metrics *****
  213.   epoch                    =     2.9847
  214.   total_flos               = 20619353GF
  215.   train_loss               =     0.8417
  216.   train_runtime            = 0:25:43.20
  217.   train_samples_per_second =      1.907
  218.   train_steps_per_second   =      0.237
  219. Figure saved at: saves/llama3-8b/lora/sft/training_loss.png
  220. 08/01/2024 19:32:48 - WARNING - llamafactory.extras.ploting - No metric eval_loss to plot.
  221. 08/01/2024 19:32:48 - WARNING - llamafactory.extras.ploting - No metric eval_accuracy to plot.
  222. [INFO|trainer.py:3819] 2024-08-01 19:32:48,529 >>
  223. ***** Running Evaluation *****
  224. [INFO|trainer.py:3821] 2024-08-01 19:32:48,529 >>   Num examples = 110
  225. [INFO|trainer.py:3824] 2024-08-01 19:32:48,529 >>   Batch size = 1
  226. 100%|█████████████████████████████████████████████████████████████████████| 110/110 [00:18<00:00,  6.07it/s]
  227. ***** eval metrics *****
  228.   epoch                   =     2.9847
  229.   eval_loss               =     0.9957
  230.   eval_runtime            = 0:00:18.23
  231.   eval_samples_per_second =      6.031
  232.   eval_steps_per_second   =      6.031
  233. [INFO|modelcard.py:449] 2024-08-01 19:33:06,773 >> Dropping the following result as it does not have all the                                                                             necessary fields:
  234. {'task': {'name': 'Causal Language Modeling', 'type': 'text-generation'}}
复制代码
输出效果
  1. root@notebook-1813389960667746306-scnlbe5oi5-17811:/public/home/scnlbe5oi5/Downloads/models# tree -L 6 LLaMA-Factory/saves/
  2. LLaMA-Factory/saves/
  3. `-- llama3-8b
  4.     `-- lora
  5.         `-- sft
  6.             |-- README.md
  7.             |-- adapter_config.json
  8.             |-- adapter_model.safetensors
  9.             |-- all_results.json
  10.             |-- checkpoint-366
  11.             |   |-- README.md
  12.             |   |-- adapter_config.json
  13.             |   |-- adapter_model.safetensors
  14.             |   |-- optimizer.pt
  15.             |   |-- rng_state.pth
  16.             |   |-- scheduler.pt
  17.             |   |-- special_tokens_map.json
  18.             |   |-- tokenizer.json
  19.             |   |-- tokenizer_config.json
  20.             |   |-- trainer_state.json
  21.             |   `-- training_args.bin
  22.             |-- eval_results.json
  23.             |-- special_tokens_map.json
  24.             |-- tokenizer.json
  25.             |-- tokenizer_config.json
  26.             |-- train_results.json
  27.             |-- trainer_log.jsonl
  28.             |-- trainer_state.json
  29.             |-- training_args.bin
  30.             `-- training_loss.png
复制代码
运行时的资源占用情况


3.1.2 多卡情况

  1. (llama_factory_torch) root@notebook-1819291427828183041-scnlbe5oi5-51898:/public/home/scnlbe5oi5/Downloads/models/LLaMA-Factory# CUDA_VISIBLE_DEVICES=0,1,2 llamafactory-cli train examples/train_lora/llama3_lora_sft.yaml
  3. [2024-08-02 18:08:58,775] [INFO] [real_accelerator.py:158:get_accelerator] Setting ds_accelerator to cuda (auto detect)
  4. 08/02/2024 18:09:01 - INFO - llamafactory.cli - Initializing distributed tasks at: 127.0.0.1:26472
  5. [2024-08-02 18:09:04,227] torch.distributed.run: [WARNING]
  6. [2024-08-02 18:09:04,227] torch.distributed.run: [WARNING] *****************************************
  7. [2024-08-02 18:09:04,227] torch.distributed.run: [WARNING] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed.
  8. [2024-08-02 18:09:04,227] torch.distributed.run: [WARNING] *****************************************
  9. [2024-08-02 18:09:09,155] [INFO] [real_accelerator.py:158:get_accelerator] Setting ds_accelerator to cuda (auto detect)
  10. [2024-08-02 18:09:09,269] [INFO] [real_accelerator.py:158:get_accelerator] Setting ds_accelerator to cuda (auto detect)
  11. [2024-08-02 18:09:09,457] [INFO] [real_accelerator.py:158:get_accelerator] Setting ds_accelerator to cuda (auto detect)
  12. WARNING: Logging before InitGoogleLogging() is written to STDERR
  13. I0802 18:09:12.353489 95618 ProcessGroupNCCL.cpp:686] [Rank 2] ProcessGroupNCCL initialization options:NCCL_ASYNC_ERROR_HANDLING: 1, NCCL_DESYNC_DEBUG: 0, NCCL_ENABLE_TIMING: 0, NCCL_BLOCKING_WAIT: 0, TIMEOUT(ms): 180000000000, USE_HIGH_PRIORITY_STREAM: 0, TORCH_DISTRIBUTED_DEBUG: OFF, NCCL_DEBUG: OFF, ID=353053408
  14. 08/02/2024 18:09:12 - WARNING - llamafactory.hparams.parser - `ddp_find_unused_parameters` needs to be set as False for LoRA in DDP training.
  15. 08/02/2024 18:09:12 - INFO - llamafactory.hparams.parser - Process rank: 2, device: cuda:2, n_gpu: 1, distributed training: True, compute dtype: torch.bfloat16
  16. WARNING: Logging before InitGoogleLogging() is written to STDERR
  17. I0802 18:09:12.555290 95617 ProcessGroupNCCL.cpp:686] [Rank 1] ProcessGroupNCCL initialization options:NCCL_ASYNC_ERROR_HANDLING: 1, NCCL_DESYNC_DEBUG: 0, NCCL_ENABLE_TIMING: 0, NCCL_BLOCKING_WAIT: 0, TIMEOUT(ms): 180000000000, USE_HIGH_PRIORITY_STREAM: 0, TORCH_DISTRIBUTED_DEBUG: OFF, NCCL_DEBUG: OFF, ID=369111936
  18. 08/02/2024 18:09:12 - WARNING - llamafactory.hparams.parser - `ddp_find_unused_parameters` needs to be set as False for LoRA in DDP training.
  19. 08/02/2024 18:09:12 - INFO - llamafactory.hparams.parser - Process rank: 1, device: cuda:1, n_gpu: 1, distributed training: True, compute dtype: torch.bfloat16
  20. WARNING: Logging before InitGoogleLogging() is written to STDERR
  21. I0802 18:09:13.120337 95616 ProcessGroupNCCL.cpp:686] [Rank 0] ProcessGroupNCCL initialization options:NCCL_ASYNC_ERROR_HANDLING: 1, NCCL_DESYNC_DEBUG: 0, NCCL_ENABLE_TIMING: 0, NCCL_BLOCKING_WAIT: 0, TIMEOUT(ms): 180000000000, USE_HIGH_PRIORITY_STREAM: 0, TORCH_DISTRIBUTED_DEBUG: OFF, NCCL_DEBUG: OFF, ID=359553664
  22. 08/02/2024 18:09:13 - WARNING - llamafactory.hparams.parser - `ddp_find_unused_parameters` needs to be set as False for LoRA in DDP training.
  23. 08/02/2024 18:09:13 - INFO - llamafactory.hparams.parser - Process rank: 0, device: cuda:0, n_gpu: 1, distributed training: True, compute dtype: torch.bfloat16
  24. 08/02/2024 18:09:13 - INFO - llamafactory.data.template - Replace eos token: <|eot_id|>
  25. 08/02/2024 18:09:13 - INFO - llamafactory.data.template - Add pad token: <|eot_id|>
  26. 08/02/2024 18:09:13 - INFO - llamafactory.data.template - Replace eos token: <|eot_id|>
  27. 08/02/2024 18:09:13 - INFO - llamafactory.data.template - Add pad token: <|eot_id|>
  28. I0802 18:09:14.158418 95618 ProcessGroupNCCL.cpp:2780] Rank 2 using GPU 2 to perform barrier as devices used by this process are currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect.Specify device_ids in barrier() to force use of a particular device.
  29. I0802 18:09:14.165846 95617 ProcessGroupNCCL.cpp:2780] Rank 1 using GPU 1 to perform barrier as devices used by this process are currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect.Specify device_ids in barrier() to force use of a particular device.
  30. [INFO|tokenization_utils_base.py:2287] 2024-08-02 18:09:14,276 >> loading file tokenizer.json
  31. [INFO|tokenization_utils_base.py:2287] 2024-08-02 18:09:14,276 >> loading file added_tokens.json
  32. [INFO|tokenization_utils_base.py:2287] 2024-08-02 18:09:14,276 >> loading file special_tokens_map.json
  33. [INFO|tokenization_utils_base.py:2287] 2024-08-02 18:09:14,276 >> loading file tokenizer_config.json
  34. [INFO|tokenization_utils_base.py:2533] 2024-08-02 18:09:14,684 >> Special tokens have been added in the vocabulary, make sure the associated word embeddings are fine-tuned or trained.
  35. 08/02/2024 18:09:14 - INFO - llamafactory.data.template - Replace eos token: <|eot_id|>
  36. 08/02/2024 18:09:14 - INFO - llamafactory.data.template - Add pad token: <|eot_id|>
  37. 08/02/2024 18:09:14 - INFO - llamafactory.data.loader - Loading dataset identity.json...
  38. Converting format of dataset (num_proc=16): 100%|████████████████████████████████████████████████████████████████████████████████████████████████████| 91/91 [00:00<00:00, 301.60 examples/s]
  39. 08/02/2024 18:09:16 - INFO - llamafactory.data.loader - Loading dataset alpaca_en_demo.json...
  40. Converting format of dataset (num_proc=16): 100%|███████████████████████████████████████████████████████████████████████████████████████████████| 1000/1000 [00:00<00:00, 3399.93 examples/s]
  41. I0802 18:09:18.295866 95616 ProcessGroupNCCL.cpp:2780] Rank 0 using GPU 0 to perform barrier as devices used by this process are currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect.Specify device_ids in barrier() to force use of a particular device.
  42. I0802 18:09:19.234498 95616 ProcessGroupNCCL.cpp:1340] NCCL_DEBUG: N/A
  43. 08/02/2024 18:09:19 - INFO - llamafactory.data.loader - Loading dataset identity.json...
  44. 08/02/2024 18:09:19 - INFO - llamafactory.data.loader - Loading dataset identity.json...
  45. Running tokenizer on dataset (num_proc=16):   0%|                                                                                                            | 0/1091 [00:00<?, ? examples/s]08/02/2024 18:09:20 - INFO - llamafactory.data.loader - Loading dataset alpaca_en_demo.json...
  46. 08/02/2024 18:09:20 - INFO - llamafactory.data.loader - Loading dataset alpaca_en_demo.json...
  47. Running tokenizer on dataset (num_proc=16): 100%|████████████████████████████████████████████████████████████████████████████████████████████████| 1091/1091 [00:03<00:00, 273.44 examples/s]
  48. training example:
  49. input_ids:
  50. [128000, 128006, 882, 128007, 271, 6151, 128009, 128006, 78191, 128007, 271, 9906, 0, 358, 1097, 5991, 609, 39254, 459, 15592, 18328, 8040, 555, 5991, 3170, 3500, 13, 2650, 649, 358, 7945, 499, 3432, 30, 128009]
  51. inputs:
  52. <|begin_of_text|><|start_header_id|>user<|end_header_id|>
  54. hi<|eot_id|><|start_header_id|>assistant<|end_header_id|>
  56. Hello! I am {{name}}, an AI assistant developed by {{author}}. How can I assist you today?<|eot_id|>
  57. label_ids:
  58. [-100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, 9906, 0, 358, 1097, 5991, 609, 39254, 459, 15592, 18328, 8040, 555, 5991, 3170, 3500, 13, 2650, 649, 358, 7945, 499, 3432, 30, 128009]
  59. labels:
  60. Hello! I am {{name}}, an AI assistant developed by {{author}}. How can I assist you today?<|eot_id|>
  61. [INFO|configuration_utils.py:731] 2024-08-02 18:09:24,080 >> loading configuration file /root/.cache/modelscope/hub/LLM-Research/Meta-Llama-3-8B-Instruct/config.json
  62. [INFO|configuration_utils.py:800] 2024-08-02 18:09:24,082 >> Model config LlamaConfig {
  63.   "_name_or_path": "/root/.cache/modelscope/hub/LLM-Research/Meta-Llama-3-8B-Instruct",
  64.   "architectures": [
  65.     "LlamaForCausalLM"
  66.   ],
  67.   "attention_bias": false,
  68.   "attention_dropout": 0.0,
  69.   "bos_token_id": 128000,
  70.   "eos_token_id": 128009,
  71.   "hidden_act": "silu",
  72.   "hidden_size": 4096,
  73.   "initializer_range": 0.02,
  74.   "intermediate_size": 14336,
  75.   "max_position_embeddings": 8192,
  76.   "mlp_bias": false,
  77.   "model_type": "llama",
  78.   "num_attention_heads": 32,
  79.   "num_hidden_layers": 32,
  80.   "num_key_value_heads": 8,
  81.   "pretraining_tp": 1,
  82.   "rms_norm_eps": 1e-05,
  83.   "rope_scaling": null,
  84.   "rope_theta": 500000.0,
  85.   "tie_word_embeddings": false,
  86.   "torch_dtype": "bfloat16",
  87.   "transformers_version": "4.43.3",
  88.   "use_cache": true,
  89.   "vocab_size": 128256
  90. }
  92. [INFO|modeling_utils.py:3631] 2024-08-02 18:09:24,119 >> loading weights file /root/.cache/modelscope/hub/LLM-Research/Meta-Llama-3-8B-Instruct/model.safetensors.index.json
  93. [INFO|modeling_utils.py:1572] 2024-08-02 18:09:24,119 >> Instantiating LlamaForCausalLM model under default dtype torch.bfloat16.
  94. [INFO|configuration_utils.py:1038] 2024-08-02 18:09:24,121 >> Generate config GenerationConfig {
  95.   "bos_token_id": 128000,
  96.   "eos_token_id": 128009
  97. }
  99. Loading checkpoint shards: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 4/4 [00:09<00:00,  2.49s/it]
  100. 08/02/2024 18:09:34 - INFO - llamafactory.model.model_utils.checkpointing - Gradient checkpointing enabled.
  101. 08/02/2024 18:09:34 - INFO - llamafactory.model.model_utils.attention - Using vanilla attention implementation.
  102. 08/02/2024 18:09:34 - INFO - llamafactory.model.adapter - Upcasting trainable params to float32.
  103. 08/02/2024 18:09:34 - INFO - llamafactory.model.adapter - Fine-tuning method: LoRA
  104. 08/02/2024 18:09:34 - INFO - llamafactory.model.model_utils.misc - Found linear modules: up_proj,gate_proj,o_proj,q_proj,k_proj,v_proj,down_proj
  105. Loading checkpoint shards: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 4/4 [00:10<00:00,  2.59s/it]
  106. [INFO|modeling_utils.py:4463] 2024-08-02 18:09:34,552 >> All model checkpoint weights were used when initializing LlamaForCausalLM.
  108. [INFO|modeling_utils.py:4471] 2024-08-02 18:09:34,552 >> All the weights of LlamaForCausalLM were initialized from the model checkpoint at /root/.cache/modelscope/hub/LLM-Research/Meta-Llama-3-8B-Instruct.
  109. If your task is similar to the task the model of the checkpoint was trained on, you can already use LlamaForCausalLM for predictions without further training.
  110. [INFO|configuration_utils.py:991] 2024-08-02 18:09:34,555 >> loading configuration file /root/.cache/modelscope/hub/LLM-Research/Meta-Llama-3-8B-Instruct/generation_config.json
  111. [INFO|configuration_utils.py:1038] 2024-08-02 18:09:34,555 >> Generate config GenerationConfig {
  112.   "bos_token_id": 128000,
  113.   "do_sample": true,
  114.   "eos_token_id": [
  115.     128001,
  116.     128009
  117.   ],
  118.   "max_length": 4096,
  119.   "temperature": 0.6,
  120.   "top_p": 0.9
  121. }
  123. 08/02/2024 18:09:34 - INFO - llamafactory.model.model_utils.checkpointing - Gradient checkpointing enabled.
  124. 08/02/2024 18:09:34 - INFO - llamafactory.model.model_utils.attention - Using vanilla attention implementation.
  125. 08/02/2024 18:09:34 - INFO - llamafactory.model.adapter - Upcasting trainable params to float32.
  126. 08/02/2024 18:09:34 - INFO - llamafactory.model.adapter - Fine-tuning method: LoRA
  127. 08/02/2024 18:09:34 - INFO - llamafactory.model.model_utils.misc - Found linear modules: k_proj,o_proj,v_proj,down_proj,q_proj,up_proj,gate_proj
  128. Loading checkpoint shards: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 4/4 [00:10<00:00,  2.52s/it]
  129. 08/02/2024 18:09:34 - INFO - llamafactory.model.model_utils.checkpointing - Gradient checkpointing enabled.
  130. 08/02/2024 18:09:34 - INFO - llamafactory.model.model_utils.attention - Using vanilla attention implementation.
  131. 08/02/2024 18:09:34 - INFO - llamafactory.model.adapter - Upcasting trainable params to float32.
  132. 08/02/2024 18:09:34 - INFO - llamafactory.model.adapter - Fine-tuning method: LoRA
  133. 08/02/2024 18:09:34 - INFO - llamafactory.model.model_utils.misc - Found linear modules: gate_proj,down_proj,q_proj,o_proj,up_proj,v_proj,k_proj
  134. 08/02/2024 18:09:34 - INFO - llamafactory.model.loader - trainable params: 20,971,520 || all params: 8,051,232,768 || trainable%: 0.2605
  135. 08/02/2024 18:09:34 - INFO - llamafactory.model.loader - trainable params: 20,971,520 || all params: 8,051,232,768 || trainable%: 0.2605
  136. Detected kernel version 3.10.0, which is below the recommended minimum of 5.5.0; this can cause the process to hang. It is recommended to upgrade the kernel to the minimum version or higher.
  137. [INFO|trainer.py:648] 2024-08-02 18:09:34,983 >> Using auto half precision backend
  138. 08/02/2024 18:09:35 - INFO - llamafactory.model.loader - trainable params: 20,971,520 || all params: 8,051,232,768 || trainable%: 0.2605
  139. [INFO|trainer.py:2134] 2024-08-02 18:09:37,114 >> ***** Running training *****
  140. [INFO|trainer.py:2135] 2024-08-02 18:09:37,114 >>   Num examples = 981
  141. [INFO|trainer.py:2136] 2024-08-02 18:09:37,114 >>   Num Epochs = 3
  142. [INFO|trainer.py:2137] 2024-08-02 18:09:37,114 >>   Instantaneous batch size per device = 1
  143. [INFO|trainer.py:2140] 2024-08-02 18:09:37,114 >>   Total train batch size (w. parallel, distributed & accumulation) = 24
  144. [INFO|trainer.py:2141] 2024-08-02 18:09:37,114 >>   Gradient Accumulation steps = 8
  145. [INFO|trainer.py:2142] 2024-08-02 18:09:37,114 >>   Total optimization steps = 120
  146. [INFO|trainer.py:2143] 2024-08-02 18:09:37,119 >>   Number of trainable parameters = 20,971,520
  147. {'loss': 1.4267, 'grad_norm': 1.401288628578186, 'learning_rate': 8.333333333333334e-05, 'epoch': 0.24}
  148. {'loss': 1.1319, 'grad_norm': 1.4780751466751099, 'learning_rate': 9.865224352899119e-05, 'epoch': 0.49}
  149. {'loss': 0.9963, 'grad_norm': 0.532632052898407, 'learning_rate': 9.330127018922194e-05, 'epoch': 0.73}
  150. {'loss': 0.9792, 'grad_norm': 0.7996620535850525, 'learning_rate': 8.43120818934367e-05, 'epoch': 0.98}
  151. {'loss': 0.937, 'grad_norm': 0.4041236639022827, 'learning_rate': 7.243995901002312e-05, 'epoch': 1.22}
  152. {'loss': 0.8805, 'grad_norm': 0.5675532221794128, 'learning_rate': 5.868240888334653e-05, 'epoch': 1.47}
  153. {'loss': 0.8467, 'grad_norm': 0.5038197636604309, 'learning_rate': 4.4195354293738484e-05, 'epoch': 1.71}
  154. {'loss': 0.8612, 'grad_norm': 0.7851077914237976, 'learning_rate': 3.019601169804216e-05, 'epoch': 1.96}
  155. {'loss': 0.818, 'grad_norm': 0.450968474149704, 'learning_rate': 1.7860619515673033e-05, 'epoch': 2.2}
  156. {'loss': 0.8308, 'grad_norm': 0.5961077809333801, 'learning_rate': 8.225609429353187e-06, 'epoch': 2.45}
  157. {'loss': 0.8071, 'grad_norm': 0.5323781371116638, 'learning_rate': 2.100524384225555e-06, 'epoch': 2.69}
  158. {'loss': 0.8061, 'grad_norm': 0.7563619017601013, 'learning_rate': 0.0, 'epoch': 2.94}
  159. 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 120/120 [09:46<00:00,  4.56s/it][INFO|trainer.py:3503] 2024-08-02 18:19:24,273 >> Saving model checkpoint to saves/llama3-8b/lora/sft/checkpoint-120
  160. [INFO|configuration_utils.py:731] 2024-08-02 18:19:24,304 >> loading configuration file /root/.cache/modelscope/hub/LLM-Research/Meta-Llama-3-8B-Instruct/config.json
  161. [INFO|configuration_utils.py:800] 2024-08-02 18:19:24,305 >> Model config LlamaConfig {
  162.   "architectures": [
  163.     "LlamaForCausalLM"
  164.   ],
  165.   "attention_bias": false,
  166.   "attention_dropout": 0.0,
  167.   "bos_token_id": 128000,
  168.   "eos_token_id": 128009,
  169.   "hidden_act": "silu",
  170.   "hidden_size": 4096,
  171.   "initializer_range": 0.02,
  172.   "intermediate_size": 14336,
  173.   "max_position_embeddings": 8192,
  174.   "mlp_bias": false,
  175.   "model_type": "llama",
  176.   "num_attention_heads": 32,
  177.   "num_hidden_layers": 32,
  178.   "num_key_value_heads": 8,
  179.   "pretraining_tp": 1,
  180.   "rms_norm_eps": 1e-05,
  181.   "rope_scaling": null,
  182.   "rope_theta": 500000.0,
  183.   "tie_word_embeddings": false,
  184.   "torch_dtype": "bfloat16",
  185.   "transformers_version": "4.43.3",
  186.   "use_cache": true,
  187.   "vocab_size": 128256
  188. }
  190. [INFO|tokenization_utils_base.py:2702] 2024-08-02 18:19:24,432 >> tokenizer config file saved in saves/llama3-8b/lora/sft/checkpoint-120/tokenizer_config.json
  191. [INFO|tokenization_utils_base.py:2711] 2024-08-02 18:19:24,434 >> Special tokens file saved in saves/llama3-8b/lora/sft/checkpoint-120/special_tokens_map.json
  192. [INFO|trainer.py:2394] 2024-08-02 18:19:24,832 >>
  194. Training completed. Do not forget to share your model on huggingface.co/models =)
  197. {'train_runtime': 587.7138, 'train_samples_per_second': 5.008, 'train_steps_per_second': 0.204, 'train_loss': 0.9434665679931641, 'epoch': 2.94}
  198. 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 120/120 [09:47<00:00,  4.90s/it]
  199. [INFO|trainer.py:3503] 2024-08-02 18:19:24,837 >> Saving model checkpoint to saves/llama3-8b/lora/sft
  200. [INFO|configuration_utils.py:731] 2024-08-02 18:19:24,907 >> loading configuration file /root/.cache/modelscope/hub/LLM-Research/Meta-Llama-3-8B-Instruct/config.json
  201. [INFO|configuration_utils.py:800] 2024-08-02 18:19:24,908 >> Model config LlamaConfig {
  202.   "architectures": [
  203.     "LlamaForCausalLM"
  204.   ],
  205.   "attention_bias": false,
  206.   "attention_dropout": 0.0,
  207.   "bos_token_id": 128000,
  208.   "eos_token_id": 128009,
  209.   "hidden_act": "silu",
  210.   "hidden_size": 4096,
  211.   "initializer_range": 0.02,
  212.   "intermediate_size": 14336,
  213.   "max_position_embeddings": 8192,
  214.   "mlp_bias": false,
  215.   "model_type": "llama",
  216.   "num_attention_heads": 32,
  217.   "num_hidden_layers": 32,
  218.   "num_key_value_heads": 8,
  219.   "pretraining_tp": 1,
  220.   "rms_norm_eps": 1e-05,
  221.   "rope_scaling": null,
  222.   "rope_theta": 500000.0,
  223.   "tie_word_embeddings": false,
  224.   "torch_dtype": "bfloat16",
  225.   "transformers_version": "4.43.3",
  226.   "use_cache": true,
  227.   "vocab_size": 128256
  228. }
  230. [INFO|tokenization_utils_base.py:2702] 2024-08-02 18:19:25,048 >> tokenizer config file saved in saves/llama3-8b/lora/sft/tokenizer_config.json
  231. [INFO|tokenization_utils_base.py:2711] 2024-08-02 18:19:25,055 >> Special tokens file saved in saves/llama3-8b/lora/sft/special_tokens_map.json
  232. ***** train metrics *****
  233.   epoch                    =     2.9358
  234.   total_flos               = 20332711GF
  235.   train_loss               =     0.9435
  236.   train_runtime            = 0:09:47.71
  237.   train_samples_per_second =      5.008
  238.   train_steps_per_second   =      0.204
  239. Figure saved at: saves/llama3-8b/lora/sft/training_loss.png
  240. 08/02/2024 18:19:25 - WARNING - llamafactory.extras.ploting - No metric eval_loss to plot.
  241. 08/02/2024 18:19:25 - WARNING - llamafactory.extras.ploting - No metric eval_accuracy to plot.
  242. [INFO|trainer.py:3819] 2024-08-02 18:19:25,357 >>
  243. ***** Running Evaluation *****
  244. [INFO|trainer.py:3821] 2024-08-02 18:19:25,357 >>   Num examples = 110
  245. [INFO|trainer.py:3824] 2024-08-02 18:19:25,357 >>   Batch size = 1
  246. 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 37/37 [00:08<00:00,  4.50it/s]
  247. ***** eval metrics *****
  248.   epoch                   =     2.9358
  249.   eval_loss               =     0.9702
  250.   eval_runtime            = 0:00:08.33
  251.   eval_samples_per_second =     13.193
  252.   eval_steps_per_second   =      4.438
  253. [INFO|modelcard.py:449] 2024-08-02 18:19:33,712 >> Dropping the following result as it does not have all the necessary fields:
  254. {'task': {'name': 'Causal Language Modeling', 'type': 'text-generation'}}
复制代码
运行时的资源占用情况


3.2 模型归并

模型归并实在CPU上进行的。
  1. (llama_factory_torch) root@notebook-1813389960667746306-scnlbe5oi5-17811:/public/home/scnlbe5oi5/Downloads/models/LLaMA-Factory# llamafactory-cli export examples/merge_lora/llama3_lora_sft.yaml
  2. [2024-08-01 21:34:37,394] [INFO] [real_accelerator.py:158:get_accelerator] Setting ds_accelerator to cuda (auto detect)
  3. [INFO|tokenization_utils_base.py:2287] 2024-08-01 21:34:41,664 >> loading file tokenizer.json
  4. [INFO|tokenization_utils_base.py:2287] 2024-08-01 21:34:41,664 >> loading file added_tokens.json
  5. [INFO|tokenization_utils_base.py:2287] 2024-08-01 21:34:41,664 >> loading file special_tokens_map.json
  6. [INFO|tokenization_utils_base.py:2287] 2024-08-01 21:34:41,664 >> loading file tokenizer_config.json
  7. [INFO|tokenization_utils_base.py:2533] 2024-08-01 21:34:42,030 >> Special tokens have been added in the vocabulary, make sure the associa                 ted word embeddings are fine-tuned or trained.
  8. 08/01/2024 21:34:42 - INFO - llamafactory.data.template - Replace eos token: <|eot_id|>
  9. 08/01/2024 21:34:42 - INFO - llamafactory.data.template - Add pad token: <|eot_id|>
  10. [INFO|configuration_utils.py:731] 2024-08-01 21:34:42,031 >> loading configuration file /root/.cache/modelscope/hub/LLM-Research/Meta-Lla                 ma-3-8B-Instruct/config.json
  11. [INFO|configuration_utils.py:800] 2024-08-01 21:34:42,032 >> Model config LlamaConfig {
  12.   "_name_or_path": "/root/.cache/modelscope/hub/LLM-Research/Meta-Llama-3-8B-Instruct",
  13.   "architectures": [
  14.     "LlamaForCausalLM"
  15.   ],
  16.   "attention_bias": false,
  17.   "attention_dropout": 0.0,
  18.   "bos_token_id": 128000,
  19.   "eos_token_id": 128009,
  20.   "hidden_act": "silu",
  21.   "hidden_size": 4096,
  22.   "initializer_range": 0.02,
  23.   "intermediate_size": 14336,
  24.   "max_position_embeddings": 8192,
  25.   "mlp_bias": false,
  26.   "model_type": "llama",
  27.   "num_attention_heads": 32,
  28.   "num_hidden_layers": 32,
  29.   "num_key_value_heads": 8,
  30.   "pretraining_tp": 1,
  31.   "rms_norm_eps": 1e-05,
  32.   "rope_scaling": null,
  33.   "rope_theta": 500000.0,
  34.   "tie_word_embeddings": false,
  35.   "torch_dtype": "bfloat16",
  36.   "transformers_version": "4.43.3",
  37.   "use_cache": true,
  38.   "vocab_size": 128256
  39. }
  41. 08/01/2024 21:34:42 - INFO - llamafactory.model.patcher - Using KV cache for faster generation.
  42. [INFO|modeling_utils.py:3631] 2024-08-01 21:34:42,058 >> loading weights file /root/.cache/modelscope/hub/LLM-Research/Meta-Llama-3-8B-In                 struct/model.safetensors.index.json
  43. [INFO|modeling_utils.py:1572] 2024-08-01 21:34:42,058 >> Instantiating LlamaForCausalLM model under default dtype torch.bfloat16.
  44. [INFO|configuration_utils.py:1038] 2024-08-01 21:34:42,059 >> Generate config GenerationConfig {
  45.   "bos_token_id": 128000,
  46.   "eos_token_id": 128009
  47. }
  49. Loading checkpoint shards: 100%|███████████████████████████████████████████████████████████████████████████| 4/4 [00:01<00:00,  3.40it/s]
  50. [INFO|modeling_utils.py:4463] 2024-08-01 21:34:43,324 >> All model checkpoint weights were used when initializing LlamaForCausalLM.
  52. [INFO|modeling_utils.py:4471] 2024-08-01 21:34:43,324 >> All the weights of LlamaForCausalLM were initialized from the model checkpoint a                 t /root/.cache/modelscope/hub/LLM-Research/Meta-Llama-3-8B-Instruct.
  53. If your task is similar to the task the model of the checkpoint was trained on, you can already use LlamaForCausalLM for predictions with                 out further training.
  54. [INFO|configuration_utils.py:991] 2024-08-01 21:34:43,327 >> loading configuration file /root/.cache/modelscope/hub/LLM-Research/Meta-Lla                 ma-3-8B-Instruct/generation_config.json
  55. [INFO|configuration_utils.py:1038] 2024-08-01 21:34:43,327 >> Generate config GenerationConfig {
  56.   "bos_token_id": 128000,
  57.   "do_sample": true,
  58.   "eos_token_id": [
  59.     128001,
  60.     128009
  61.   ],
  62.   "max_length": 4096,
  63.   "temperature": 0.6,
  64.   "top_p": 0.9
  65. }
  67. 08/01/2024 21:34:43 - INFO - llamafactory.model.model_utils.attention - Using vanilla attention implementation.
  68. 08/01/2024 21:40:34 - INFO - llamafactory.model.adapter - Merged 1 adapter(s).
  69. 08/01/2024 21:40:34 - INFO - llamafactory.model.adapter - Loaded adapter(s): saves/llama3-8b/lora/sft
  70. 08/01/2024 21:40:34 - INFO - llamafactory.model.loader - all params: 8,030,261,248
  71. 08/01/2024 21:40:34 - INFO - llamafactory.train.tuner - Convert model dtype to: torch.bfloat16.
  72. [INFO|configuration_utils.py:472] 2024-08-01 21:40:34,700 >> Configuration saved in models/llama3_lora_sft/config.json
  73. [INFO|configuration_utils.py:807] 2024-08-01 21:40:34,704 >> Configuration saved in models/llama3_lora_sft/generation_config.json
  74. [INFO|modeling_utils.py:2763] 2024-08-01 21:40:49,039 >> The model is bigger than the maximum size per checkpoint (2GB) and is going to be split in 9 checkpoint shards. You can find where each parameters has been saved in the index located at models/llama3_lora_sft/model.safetensors.index.json.
  75. [INFO|tokenization_utils_base.py:2702] 2024-08-01 21:40:49,046 >> tokenizer config file saved in models/llama3_lora_sft/tokenizer_config.json
  76. [INFO|tokenization_utils_base.py:2711] 2024-08-01 21:40:49,048 >> Special tokens file saved in models/llama3_lora_sft/special_tokens_map.json
复制代码
输出效果
  1. (llama_factory_torch) root@notebook-1813389960667746306-scnlbe5oi5-17811:/public/home/scnlbe5oi5/Downloads/models# tree -L 6 LLaMA-Factory/models/llama3_lora_sft/
  2. LLaMA-Factory/models/llama3_lora_sft/
  3. |-- config.json
  4. |-- generation_config.json
  5. |-- model-00001-of-00009.safetensors
  6. |-- model-00002-of-00009.safetensors
  7. |-- model-00003-of-00009.safetensors
  8. |-- model-00004-of-00009.safetensors
  9. |-- model-00005-of-00009.safetensors
  10. |-- model-00006-of-00009.safetensors
  11. |-- model-00007-of-00009.safetensors
  12. |-- model-00008-of-00009.safetensors
  13. |-- model-00009-of-00009.safetensors
  14. |-- model.safetensors.index.json
  15. |-- special_tokens_map.json
  16. |-- tokenizer.json
  17. `-- tokenizer_config.json
复制代码
运行时的资源占用情况

3.3 LoRA 推理

  1. (llama_factory_torch) root@notebook-1813389960667746306-scnlbe5oi5-20553:/public/home/scnlbe5oi5/Downloads/models/LLaMA-Factory# llamafactory-cli chat examples/inference/llama3_lora_sft.yaml
  2. [2024-08-02 22:08:48,070] [INFO] [real_accelerator.py:158:get_accelerator] Setting ds_accelerator to cuda (auto detect)
  3. 2024-08-02 22:08:52,267 - modelscope - WARNING - Using branch: master as version is unstable, use with caution
  4. [INFO|tokenization_utils_base.py:2287] 2024-08-02 22:08:52,535 >> loading file tokenizer.json
  5. [INFO|tokenization_utils_base.py:2287] 2024-08-02 22:08:52,535 >> loading file added_tokens.json
  6. [INFO|tokenization_utils_base.py:2287] 2024-08-02 22:08:52,535 >> loading file special_tokens_map.json
  7. [INFO|tokenization_utils_base.py:2287] 2024-08-02 22:08:52,535 >> loading file tokenizer_config.json
  8. [INFO|tokenization_utils_base.py:2533] 2024-08-02 22:08:52,818 >> Special tokens have been added in the vocabulary, make sure the associated word embeddings are fine-tuned or trained.
  9. 08/02/2024 22:08:52 - INFO - llamafactory.data.template - Replace eos token: <|eot_id|>
  10. 08/02/2024 22:08:52 - INFO - llamafactory.data.template - Add pad token: <|eot_id|>
  11. [INFO|configuration_utils.py:731] 2024-08-02 22:08:52,820 >> loading configuration file /root/.cache/modelscope/hub/LLM-Research/Meta-Llama-3-8B-Instruct/config.json
  12. [INFO|configuration_utils.py:800] 2024-08-02 22:08:52,821 >> Model config LlamaConfig {
  13.   "_name_or_path": "/root/.cache/modelscope/hub/LLM-Research/Meta-Llama-3-8B-Instruct",
  14.   "architectures": [
  15.     "LlamaForCausalLM"
  16.   ],
  17.   "attention_bias": false,
  18.   "attention_dropout": 0.0,
  19.   "bos_token_id": 128000,
  20.   "eos_token_id": 128009,
  21.   "hidden_act": "silu",
  22.   "hidden_size": 4096,
  23.   "initializer_range": 0.02,
  24.   "intermediate_size": 14336,
  25.   "max_position_embeddings": 8192,
  26.   "mlp_bias": false,
  27.   "model_type": "llama",
  28.   "num_attention_heads": 32,
  29.   "num_hidden_layers": 32,
  30.   "num_key_value_heads": 8,
  31.   "pretraining_tp": 1,
  32.   "rms_norm_eps": 1e-05,
  33.   "rope_scaling": null,
  34.   "rope_theta": 500000.0,
  35.   "tie_word_embeddings": false,
  36.   "torch_dtype": "bfloat16",
  37.   "transformers_version": "4.43.3",
  38.   "use_cache": true,
  39.   "vocab_size": 128256
  40. }
  42. 08/02/2024 22:08:52 - INFO - llamafactory.model.patcher - Using KV cache for faster generation.
  43. [INFO|modeling_utils.py:3631] 2024-08-02 22:08:52,847 >> loading weights file /root/.cache/modelscope/hub/LLM-Research/Meta-Llama-3-8B-Instruct/model.safetensors.index.json
  44. [INFO|modeling_utils.py:1572] 2024-08-02 22:08:52,847 >> Instantiating LlamaForCausalLM model under default dtype torch.bfloat16.
  45. [INFO|configuration_utils.py:1038] 2024-08-02 22:08:52,848 >> Generate config GenerationConfig {
  46.   "bos_token_id": 128000,
  47.   "eos_token_id": 128009
  48. }
  50. Loading checkpoint shards: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 4/4 [00:07<00:00,  1.98s/it]
  51. [INFO|modeling_utils.py:4463] 2024-08-02 22:09:01,148 >> All model checkpoint weights were used when initializing LlamaForCausalLM.
  53. [INFO|modeling_utils.py:4471] 2024-08-02 22:09:01,148 >> All the weights of LlamaForCausalLM were initialized from the model checkpoint at /root/.cache/modelscope/hub/LLM-Research/Meta-Llama-3-8B-Instruct.
  54. If your task is similar to the task the model of the checkpoint was trained on, you can already use LlamaForCausalLM for predictions without further training.
  55. [INFO|configuration_utils.py:991] 2024-08-02 22:09:01,151 >> loading configuration file /root/.cache/modelscope/hub/LLM-Research/Meta-Llama-3-8B-Instruct/generation_config.json
  56. [INFO|configuration_utils.py:1038] 2024-08-02 22:09:01,152 >> Generate config GenerationConfig {
  57.   "bos_token_id": 128000,
  58.   "do_sample": true,
  59.   "eos_token_id": [
  60.     128001,
  61.     128009
  62.   ],
  63.   "max_length": 4096,
  64.   "temperature": 0.6,
  65.   "top_p": 0.9
  66. }
  68. 08/02/2024 22:09:01 - INFO - llamafactory.model.model_utils.attention - Using vanilla attention implementation.
  69. 08/02/2024 22:09:06 - INFO - llamafactory.model.adapter - Merged 1 adapter(s).
  70. 08/02/2024 22:09:06 - INFO - llamafactory.model.adapter - Loaded adapter(s): saves/llama3-8b/lora/sft
  71. 08/02/2024 22:09:06 - INFO - llamafactory.model.loader - all params: 8,030,261,248
  72. Welcome to the CLI application, use `clear` to remove the history, use `exit` to exit the application.
  74. User: 中国深圳有哪些旅游景点
  75. Assistant: 深圳是一个非常有名的旅游城市，拥有许多名副其名的旅游景点。以下是一些主要的旅游景点：
  77. 1. Window of the World：这是一个规模宏大的主题公园，展示了世界各地的风土人情。
  79. 2. Splendid China Miniature Theme Park：这个公园展现了中国的历史和文化，拥有许多精致的模型和景观。
  81. 3. Dafen Oil Painting Village：这个村庄是中国最大的油画村，拥有数以万计的油画作品，展示了中国油画的技艺。
  83. 4. Dameisha Beach：这个沙滩是深圳最为人知的旅游景点之一，拥有洁洁的沙滩和清澈的海水，是一个非常适合休闲的场所。
  85. 5. Mangrove Forest Nature Reserve：这个自然保护区拥有广泛的 mangrove 森林，展示了中国的自然景观。
  87. 6. Shenzhen Museum：这个博物馆展现了深圳的历史和文化，拥有许多历史和艺术的收藏品。
  89. 7. Lianhua Mountain Park：这个公园是深圳最大的公园，拥有许多山路和景观，展示了中国的自然美景。
  91. 8. Shenzhen Bay Sports Center：这个体育中心拥有许多不同的运动场所，展示了中国的体育技艺。
  93. 9. OCT-LOFT：这个文化区拥有许多艺术和文化的项目，展示了中国的艺术和文化。
  95. 10. Fairy Lake Botanical Garden：这个植物园拥有许多不同的植物和花卉，展示了中国的自然美景。
  97. User: 中国广州有哪些旅游景点
  98. Assistant: 广州是一个非常有名的旅游城市，拥有许多名副其名的旅游景点。以下是一些主要的旅游景点：
  100. 1. Canton Tower：这是一个位于广州的超高建筑，拥有360度的观景台，展示了广州的全景。
  102. 2. Chimelong Paradise：这个主题公园拥有许多不同的游乐设施和景观，展示了中国的游乐技艺。
  104. 3. Baiyun Mountain：这个山区拥有许多不同的景观和游乐设施，展示了中国的自然美景。
  106. 4. Yuexiu Park：这个公园是广州最大的公园，拥有许多不同的景观和游乐设施，展示了中国的自然美景。
  108. 5. Temple of the Six Banyan Trees：这个寺庙拥有许多不同的文化和历史的收藏品，展示了中国的历史和文化。
  110. 6. Museum of the Chinese Revolution：这个博物馆展现了中国革命的历史和文化，拥有许多不同的收藏品和展品。
  112. 7. Guangzhou Tower：这个塔楼是广州最早的建筑，拥有许多不同的景观和游乐设施，展示了中国的历史和文化。
  114. 8. Guangzhou Museum：这个博物馆展现了广州的历史和文化，拥有许多不同的收藏品和展品。
  116. 9. Flower Street：这个街区拥有许多不同的花卉和景观，展示了中国的自然美景。
  118. 10. Shamian Island：这个岛区拥有许多不同的景观和游乐设施，展示了中国的自然美景和历史文化。
复制代码
四、FAQ

Q：OSError: You are trying to access a gated repo. Make sure to have access to it at https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct.

  1. (llama_fct) root@notebook-1813389960667746306-scnlbe5oi5-17811:/public/home/scnlbe5oi5/Downloads/mod
  2. els/LLaMA-Factory# llamafactory-cli train examples/train_lora/llama3_lora_sft.yaml
  4. No ROCm runtime is found, using ROCM_HOME='/opt/dtk'
  5. /opt/conda/envs/llama_fct/lib/python3.10/site-packages/torchvision/io/image.py:13: UserWarning: Fail                                     ed to load image Python extension: 'libc10_hip.so: cannot open shared object file: No such file or d                                     irectory'If you don't plan on using image functionality from `torchvision.io`, you can ignore this w                                     arning. Otherwise, there might be something wrong with your environment. Did you have `libjpeg` or `                                     libpng` installed before building `torchvision` from source?
  6.   warn(
  7. [2024-08-01 15:13:21,242] [INFO] [real_accelerator.py:158:get_accelerator] Setting ds_accelerator to                                      cuda (auto detect)
  8. 08/01/2024 15:13:24 - INFO - llamafactory.hparams.parser - Process rank: 0, device: cpu, n_gpu: 0, distributed training: False, compute dtype: torch.bfloat16
  9. [INFO|tokenization_auto.py:682] 2024-08-01 15:13:25,152 >> Could not locate the tokenizer configuration file, will try to use the model config instead.
  10. Traceback (most recent call last):
  11.   File "/opt/conda/envs/llama_fct/lib/python3.10/site-packages/huggingface_hub/utils/_errors.py", line 304, in hf_raise_for_status
  12.     response.raise_for_status()
  13.   File "/opt/conda/envs/llama_fct/lib/python3.10/site-packages/requests/models.py", line 1024, in raise_for_status
  14.     raise HTTPError(http_error_msg, response=self)
  15. requests.exceptions.HTTPError: 401 Client Error: Unauthorized for url: https://hf-mirror.com/meta-llama/Meta-Llama-3-8B-Instruct/resolve/main/config.json
  17. The above exception was the direct cause of the following exception:
  19. Traceback (most recent call last):
  20.   File "/opt/conda/envs/llama_fct/lib/python3.10/site-packages/transformers/utils/hub.py", line 402, in cached_file
  21.     resolved_file = hf_hub_download(
  22.   File "/opt/conda/envs/llama_fct/lib/python3.10/site-packages/huggingface_hub/utils/_deprecation.py", line 101, in inner_f
  23.     return f(*args, **kwargs)
  24.   File "/opt/conda/envs/llama_fct/lib/python3.10/site-packages/huggingface_hub/utils/_validators.py", line 114, in _inner_fn
  25.     return fn(*args, **kwargs)
  26.   File "/opt/conda/envs/llama_fct/lib/python3.10/site-packages/huggingface_hub/file_download.py", line 1240, in hf_hub_download
  27.     return _hf_hub_download_to_cache_dir(
  28.   File "/opt/conda/envs/llama_fct/lib/python3.10/site-packages/huggingface_hub/file_download.py", line 1347, in _hf_hub_download_to_cache_dir
  29.     _raise_on_head_call_error(head_call_error, force_download, local_files_only)
  30.   File "/opt/conda/envs/llama_fct/lib/python3.10/site-packages/huggingface_hub/file_download.py", line 1854, in _raise_on_head_call_error
  31.     raise head_call_error
  32.   File "/opt/conda/envs/llama_fct/lib/python3.10/site-packages/huggingface_hub/file_download.py", line 1751, in _get_metadata_or_catch_error
  33.     metadata = get_hf_file_metadata(
  34.   File "/opt/conda/envs/llama_fct/lib/python3.10/site-packages/huggingface_hub/utils/_validators.py", line 114, in _inner_fn
  35.     return fn(*args, **kwargs)
  36.   File "/opt/conda/envs/llama_fct/lib/python3.10/site-packages/huggingface_hub/file_download.py", line 1673, in get_hf_file_metadata
  37.     r = _request_wrapper(
  38.   File "/opt/conda/envs/llama_fct/lib/python3.10/site-packages/huggingface_hub/file_download.py", line 376, in _request_wrapper
  39.     response = _request_wrapper(
  40.   File "/opt/conda/envs/llama_fct/lib/python3.10/site-packages/huggingface_hub/file_download.py", line 400, in _request_wrapper
  41.     hf_raise_for_status(response)
  42.   File "/opt/conda/envs/llama_fct/lib/python3.10/site-packages/huggingface_hub/utils/_errors.py", line 321, in hf_raise_for_status
  43.     raise GatedRepoError(message, response) from e
  44. huggingface_hub.utils._errors.GatedRepoError: 401 Client Error. (Request ID: Root=1-66ab3595-53663c2f4d5cf81405b65b9e;080cfa15-3220-4ab1-b123-4a32ba31a03a)
  46. Cannot access gated repo for url https://hf-mirror.com/meta-llama/Meta-Llama-3-8B-Instruct/resolve/main/config.json.
  47. Access to model meta-llama/Meta-Llama-3-8B-Instruct is restricted. You must be authenticated to access it.
  49. The above exception was the direct cause of the following exception:
  51. Traceback (most recent call last):
  52.   File "/opt/conda/envs/llama_fct/bin/llamafactory-cli", line 8, in <module>
  53.     sys.exit(main())
  54.   File "/public/home/scnlbe5oi5/Downloads/models/LLaMA-Factory/src/llamafactory/cli.py", line 111, in main
  55.     run_exp()
  56.   File "/public/home/scnlbe5oi5/Downloads/models/LLaMA-Factory/src/llamafactory/train/tuner.py", line 50, in run_exp
  57.     run_sft(model_args, data_args, training_args, finetuning_args, generating_args, callbacks)
  58.   File "/public/home/scnlbe5oi5/Downloads/models/LLaMA-Factory/src/llamafactory/train/sft/workflow.py", line 44, in run_sft
  59.     tokenizer_module = load_tokenizer(model_args)
  60.   File "/public/home/scnlbe5oi5/Downloads/models/LLaMA-Factory/src/llamafactory/model/loader.py", line 69, in load_tokenizer
  61.     tokenizer = AutoTokenizer.from_pretrained(
  62.   File "/opt/conda/envs/llama_fct/lib/python3.10/site-packages/transformers/models/auto/tokenization_auto.py", line 853, in from_pretrained
  63.     config = AutoConfig.from_pretrained(
  64.   File "/opt/conda/envs/llama_fct/lib/python3.10/site-packages/transformers/models/auto/configuration_auto.py", line 972, in from_pretrained
  65.     config_dict, unused_kwargs = PretrainedConfig.get_config_dict(pretrained_model_name_or_path, **kwargs)
  66.   File "/opt/conda/envs/llama_fct/lib/python3.10/site-packages/transformers/configuration_utils.py", line 632, in get_config_dict
  67.     config_dict, kwargs = cls._get_config_dict(pretrained_model_name_or_path, **kwargs)
  68.   File "/opt/conda/envs/llama_fct/lib/python3.10/site-packages/transformers/configuration_utils.py", line 689, in _get_config_dict
  69.     resolved_config_file = cached_file(
  70.   File "/opt/conda/envs/llama_fct/lib/python3.10/site-packages/transformers/utils/hub.py", line 420, in cached_file
  71.     raise EnvironmentError(
  72. OSError: You are trying to access a gated repo.
  73. Make sure to have access to it at https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct.
  74. 401 Client Error. (Request ID: Root=1-66ab3595-53663c2f4d5cf81405b65b9e;080cfa15-3220-4ab1-b123-4a32ba31a03a)
  76. Cannot access gated repo for url https://hf-mirror.com/meta-llama/Meta-Llama-3-8B-Instruct/resolve/main/config.json.
  77. Access to model meta-llama/Meta-Llama-3-8B-Instruct is restricted. You must be authenticated to access it.
复制代码
错误缘故原由：默认是从Hugging Face中获取模型，由于Hugging Face 模型授权失败，导致获取模型失败。
解决方法：从modelscope下载模型。
  1. export USE_MODELSCOPE_HUB=1
  2. # Windows 使用 `set USE_MODELSCOPE_HUB=1`
复制代码
将 model_name_or_path 设置为模型 ID 来加载对应的模型。在魔搭社区查察全部可用的模型，例如 LLM-Research/Meta-Llama-3-8B-Instruct。
修改 llama3_lora_sft.yaml 文件：
  1. # model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
  2. 改为
  3. model_name_or_path: LLM-Research/Meta-Llama-3-8B-Instruct
复制代码
  1. llamafactory-cli train examples/train_lora/llama3_lora_sft.yaml
复制代码
Q：OSError: LLM-Research/Meta-Llama-3-8B-Instruct is not a local folder and is not a valid model identifier listed on 'https://huggingface.co/models'

  1. (llama_factory_torch) root@notebook-1813389960667746306-scnlbe5oi5-17811:/public/home/scnlbe5oi5/Downloads/models/LLaMA-Factory# llamafactory-cli chat examples/inference/llama3_lora_sft.yaml
  2. [2024-08-01 21:17:22,212] [INFO] [real_accelerator.py:158:get_accelerator] Setting ds_accelerator to cuda (auto detect)
  3. Traceback (most recent call last):
  4.   File "/opt/conda/envs/llama_factory_torch/lib/python3.10/site-packages/huggingface_hub/utils/_errors.py", line 304, in hf_raise_for_status
  5.     response.raise_for_status()
  6.   File "/opt/conda/envs/llama_factory_torch/lib/python3.10/site-packages/requests/models.py", line 1024, in raise_for_status
  7.     raise HTTPError(http_error_msg, response=self)
  8. requests.exceptions.HTTPError: 401 Client Error: Unauthorized for url: https://hf-mirror.com/LLM-Research/Meta-Llama-3-8B-Instruct/resolve/main/tokenizer_config.json
  10. The above exception was the direct cause of the following exception:
  12. Traceback (most recent call last):
  13.   File "/opt/conda/envs/llama_factory_torch/lib/python3.10/site-packages/transformers/utils/hub.py", line 402, in cached_file
  14.     resolved_file = hf_hub_download(
  15.   File "/opt/conda/envs/llama_factory_torch/lib/python3.10/site-packages/huggingface_hub/utils/_validators.py", line 114, in _inner_fn
  16.     return fn(*args, **kwargs)
  17.   File "/opt/conda/envs/llama_factory_torch/lib/python3.10/site-packages/huggingface_hub/file_download.py", line 1221, in hf_hub_download
  18.     return _hf_hub_download_to_cache_dir(
  19.   File "/opt/conda/envs/llama_factory_torch/lib/python3.10/site-packages/huggingface_hub/file_download.py", line 1325, in _hf_hub_download_to_cache_dir
  20.     _raise_on_head_call_error(head_call_error, force_download, local_files_only)
  21.   File "/opt/conda/envs/llama_factory_torch/lib/python3.10/site-packages/huggingface_hub/file_download.py", line 1823, in _raise_on_head_call_error
  22.     raise head_call_error
  23.   File "/opt/conda/envs/llama_factory_torch/lib/python3.10/site-packages/huggingface_hub/file_download.py", line 1722, in _get_metadata_or_catch_error
  24.     metadata = get_hf_file_metadata(url=url, proxies=proxies, timeout=etag_timeout, headers=headers)
  25.   File "/opt/conda/envs/llama_factory_torch/lib/python3.10/site-packages/huggingface_hub/utils/_validators.py", line 114, in _inner_fn
  26.     return fn(*args, **kwargs)
  27.   File "/opt/conda/envs/llama_factory_torch/lib/python3.10/site-packages/huggingface_hub/file_download.py", line 1645, in get_hf_file_metadata
  28.     r = _request_wrapper(
  29.   File "/opt/conda/envs/llama_factory_torch/lib/python3.10/site-packages/huggingface_hub/file_download.py", line 372, in _request_wrapper
  30.     response = _request_wrapper(
  31.   File "/opt/conda/envs/llama_factory_torch/lib/python3.10/site-packages/huggingface_hub/file_download.py", line 396, in _request_wrapper
  32.     hf_raise_for_status(response)
  33.   File "/opt/conda/envs/llama_factory_torch/lib/python3.10/site-packages/huggingface_hub/utils/_errors.py", line 352, in hf_raise_for_status
  34.     raise RepositoryNotFoundError(message, response) from e
  35. huggingface_hub.utils._errors.RepositoryNotFoundError: 401 Client Error. (Request ID: Root=1-66ab8ae6-4ed0547e1f86fcb201b723f8;acee559e-0676-48e4-8871-b6eb58e797ca)
  37. Repository Not Found for url: https://hf-mirror.com/LLM-Research/Meta-Llama-3-8B-Instruct/resolve/main/tokenizer_config.json.
  38. Please make sure you specified the correct `repo_id` and `repo_type`.
  39. If you are trying to access a private or gated repo, make sure you are authenticated.
  40. Invalid username or password.
  42. The above exception was the direct cause of the following exception:
  44. Traceback (most recent call last):
  45.   File "/opt/conda/envs/llama_factory_torch/bin/llamafactory-cli", line 8, in <module>
  46.     sys.exit(main())
  47.   File "/public/home/scnlbe5oi5/Downloads/models/LLaMA-Factory/src/llamafactory/cli.py", line 81, in main
  48.     run_chat()
  49.   File "/public/home/scnlbe5oi5/Downloads/models/LLaMA-Factory/src/llamafactory/chat/chat_model.py", line 125, in run_chat
  50.     chat_model = ChatModel()
  51.   File "/public/home/scnlbe5oi5/Downloads/models/LLaMA-Factory/src/llamafactory/chat/chat_model.py", line 44, in __init__
  52.     self.engine: "BaseEngine" = HuggingfaceEngine(model_args, data_args, finetuning_args, generating_args)
  53.   File "/public/home/scnlbe5oi5/Downloads/models/LLaMA-Factory/src/llamafactory/chat/hf_engine.py", line 53, in __init__
  54.     tokenizer_module = load_tokenizer(model_args)
  55.   File "/public/home/scnlbe5oi5/Downloads/models/LLaMA-Factory/src/llamafactory/model/loader.py", line 69, in load_tokenizer
  56.     tokenizer = AutoTokenizer.from_pretrained(
  57.   File "/opt/conda/envs/llama_factory_torch/lib/python3.10/site-packages/transformers/models/auto/tokenization_auto.py", line 833, in from_pretrained
  58.     tokenizer_config = get_tokenizer_config(pretrained_model_name_or_path, **kwargs)
  59.   File "/opt/conda/envs/llama_factory_torch/lib/python3.10/site-packages/transformers/models/auto/tokenization_auto.py", line 665, in get_tokenizer_config
  60.     resolved_config_file = cached_file(
  61.   File "/opt/conda/envs/llama_factory_torch/lib/python3.10/site-packages/transformers/utils/hub.py", line 425, in cached_file
  62.     raise EnvironmentError(
  63. OSError: LLM-Research/Meta-Llama-3-8B-Instruct is not a local folder and is not a valid model identifier listed on 'https://huggingface.co/models'
  64. If this is a private repository, make sure to pass a token having permission to this repo either by logging in with `huggingface-cli login` or by passing `token=<your_token>`
复制代码
错误缘故原由：找不到 LLM-Research/Meta-Llama-3-8B-Instruct模型。
解决方法：从modelscope下载模型。
  1. export USE_MODELSCOPE_HUB=1
复制代码
Q：ModuleNotFoundError: No module named 'modelscope'

  1. (llama_factory_torch) root@notebook-1813389960667746306-scnlbe5oi5-17811:/public/home/scnlbe5oi5/Downloads/models/LLaMA-Factory# llamafactory-cli train examples/train_lora/llama3_lora_sft.yaml
  2. [2024-08-01 19:05:15,320] [INFO] [real_accelerator.py:158:get_accelerator] Setting ds_accelerator to cuda (a                                                                            uto detect)08/01/2024 19:05:18 - INFO - llamafactory.hparams.parser - Process rank: 0, device: cuda:0, n_gpu: 1, distri                                                                            buted training: False, compute dtype: torch.bfloat16Traceback (most recent call last):  File "/public/home/scnlbe5oi5/Downloads/models/LLaMA-Factory/src/llamafactory/extras/misc.py", line 219, i                                                                            n try_download_model_from_ms    from modelscope import snapshot_downloadModuleNotFoundError: No module named 'modelscope'During handling of the above exception, another exception occurred:Traceback (most recent call last):  File "/opt/conda/envs/llama_factory_torch/bin/llamafactory-cli", line 8, in <module>    sys.exit(main())  File "/public/home/scnlbe5oi5/Downloads/models/LLaMA-Factory/src/llamafactory/cli.py", line 111, in main    run_exp()  File "/public/home/scnlbe5oi5/Downloads/models/LLaMA-Factory/src/llamafactory/train/tuner.py", line 50, in                                                                             run_exp    run_sft(model_args, data_args, training_args, finetuning_args, generating_args, callbacks)  File "/public/home/scnlbe5oi5/Downloads/models/LLaMA-Factory/src/llamafactory/train/sft/workflow.py", line                                                                             44, in run_sft    tokenizer_module = load_tokenizer(model_args)  File "/public/home/scnlbe5oi5/Downloads/models/LLaMA-Factory/src/llamafactory/model/loader.py", line 67, i                                                                            n load_tokenizer    init_kwargs = _get_init_kwargs(model_args)  File "/public/home/scnlbe5oi5/Downloads/models/LLaMA-Factory/src/llamafactory/model/loader.py", line 52, i                                                                            n _get_init_kwargs    model_args.model_name_or_path = try_download_model_from_ms(model_args)  File "/public/home/scnlbe5oi5/Downloads/models/LLaMA-Factory/src/llamafactory/extras/misc.py", line 224, i                                                                            n try_download_model_from_ms    raise ImportError("Please install modelscope via `pip install modelscope -U`")ImportError: Please install modelscope via `pip install modelscope -U`
复制代码
错误缘故原由：缺少modelscope依赖包。
解决方法：安装modelscope。
  1. (llama_factory_torch) root@notebook-1813389960667746306-scnlbe5oi5-17811:/public/home/scnlbe5oi5/Downloads/m
  2. odels/LLaMA-Factory# pip install --no-dependencies modelscope
  3. Looking in indexes: https://pypi.tuna.tsinghua.edu.cn/simple
  4. Collecting modelscope
  5.   Using cached https://pypi.tuna.tsinghua.edu.cn/packages/38/37/9fe505ebc67ba5e0345a69d6e8b2ee8630523975b484                                                                            d221691ef60182bd/modelscope-1.16.1-py3-none-any.whl (5.7 MB)
  6. Installing collected packages: modelscope
  7. Successfully installed modelscope-1.16.1
复制代码
Q：ImportError: /PATH/TO/site-packages/torch/lib/libtorch_hip.so: undefined symbol: ncclCommInitRankConfig

  1. (llama_fct_pytorch) root@notebook-1813389960667746306-scnlbe5oi5-17811:/public/home/scnlbe5oi5/Downloads/models/LLaMA-Factory# llamafactory-cli train examples/train_lora/llama3_lora_sft.yaml
  2. Traceback (most recent call last):  File "/opt/conda/envs/llama_fct_pytorch/bin/llamafactory-cli", line 5, in <module>    from llamafactory.cli import main  File "/public/home/scnlbe5oi5/Downloads/models/LLaMA-Factory/src/llamafactory/__init__.py", line 38, in <module>    from .cli import VERSION  File "/public/home/scnlbe5oi5/Downloads/models/LLaMA-Factory/src/llamafactory/cli.py", line 21, in <module>    from . import launcher  File "/public/home/scnlbe5oi5/Downloads/models/LLaMA-Factory/src/llamafactory/launcher.py", line 15, in <module>    from llamafactory.train.tuner import run_exp  File "/public/home/scnlbe5oi5/Downloads/models/LLaMA-Factory/src/llamafactory/train/tuner.py", line 19, in <module>    import torch  File "/opt/conda/envs/llama_fct_pytorch/lib/python3.10/site-packages/torch/__init__.py", line 237, in <module>    from torch._C import *  # noqa: F403ImportError: /opt/conda/envs/llama_fct_pytorch/lib/python3.10/site-packages/torch/lib/libtorch_hip.so: undefined symbol: ncclCommInitRankConfig
复制代码
  1. >>> import torch
  2. Traceback (most recent call last):
  3.   File "<stdin>", line 1, in <module>
  4.   File "/opt/conda/envs/llama_fct_pytorch/lib/python3.10/site-packages/torch/__init__.py", line 237, in <module>
  5.     from torch._C import *  # noqa: F403
  6. ImportError: /opt/conda/envs/llama_fct_pytorch/lib/python3.10/site-packages/torch/lib/libtorch_hip.so: undefined symbol: ncclCommInitRankConfig
复制代码
错误缘故原由：当前PyTorch版本不支持DCU。
该问题的解决方法，请参考下文的FAQ。
Q：PyTorch版本不支持DCU

  1. (llama_fct) root@notebook-1813389960667746306-scnlbe5oi5-17811:/public/home/scnlbe5oi5/Downloads/models/LLaMA-Factory# llamafactory-cli train examples/train_lora/llama3_lora_sft.yaml
  2. No ROCm runtime is found, using ROCM_HOME='/opt/dtk'/opt/conda/envs/llama_fct/lib/python3.10/site-packages/torchvision/io/image.py:13: UserWarning: Failed to lo                                                                            ad image Python extension: 'libc10_hip.so: cannot open shared object file: No such file or directory'If you                                                                             don't plan on using image functionality from `torchvision.io`, you can ignore this warning. Otherwise, there                                                                             might be something wrong with your environment. Did you have `libjpeg` or `libpng` installed before buildin                                                                            g `torchvision` from source?  warn([2024-08-01 17:49:08,805] [INFO] [real_accelerator.py:158:get_accelerator] Setting ds_accelerator to cuda (a                                                                            uto detect)08/01/2024 17:49:12 - INFO - llamafactory.hparams.parser - Process rank: 0, device: cpu, n_gpu: 0, distribut                                                                            ed training: False, compute dtype: torch.bfloat16Downloading: 100%|█████████████████████████████████████████████████████████| 654/654 [00:00<00:00, 2.56kB/s]Downloading: 100%|█████████████████████████████████████████████████████████| 48.0/48.0 [00:00<00:00, 183B/s]Downloading: 100%|███████████████████████████████████████████████████████████| 187/187 [00:00<00:00, 759B/s]Downloading: 100%|█████████████████████████████████████████████████████| 7.62k/7.62k [00:00<00:00, 29.9kB/s]Downloading: 100%|█████████████████████████████████████████████████████| 4.63G/4.63G [01:33<00:00, 53.4MB/s]Downloading: 100%|█████████████████████████████████████████████████████| 4.66G/4.66G [01:02<00:00, 79.9MB/s]Downloading: 100%|█████████████████████████████████████████████████████| 4.58G/4.58G [01:00<00:00, 81.7MB/s]Downloading: 100%|█████████████████████████████████████████████████████| 1.09G/1.09G [00:22<00:00, 51.6MB/s]Downloading: 100%|█████████████████████████████████████████████████████| 23.4k/23.4k [00:00<00:00, 53.6kB/s]Downloading: 100%|██████████████████████████████████████████████████████| 36.3k/36.3k [00:00<00:00, 125kB/s]Downloading: 100%|█████████████████████████████████████████████████████████| 73.0/73.0 [00:00<00:00, 293B/s]Downloading: 100%|█████████████████████████████████████████████████████| 8.66M/8.66M [00:00<00:00, 13.5MB/s]Downloading: 100%|█████████████████████████████████████████████████████| 49.8k/49.8k [00:00<00:00, 90.0kB/s]Downloading: 100%|█████████████████████████████████████████████████████| 4.59k/4.59k [00:00<00:00, 18.7kB/s][INFO|tokenization_utils_base.py:2287] 2024-08-01 17:53:53,510 >> loading file tokenizer.json[INFO|tokenization_utils_base.py:2287] 2024-08-01 17:53:53,511 >> loading file added_tokens.json[INFO|tokenization_utils_base.py:2287] 2024-08-01 17:53:53,511 >> loading file special_tokens_map.json[INFO|tokenization_utils_base.py:2287] 2024-08-01 17:53:53,511 >> loading file tokenizer_config.json[INFO|tokenization_utils_base.py:2533] 2024-08-01 17:53:53,854 >> Special tokens have been added in the voca                                                                            bulary, make sure the associated word embeddings are fine-tuned or trained.08/01/2024 17:53:53 - INFO - llamafactory.data.template - Replace eos token: <|eot_id|>08/01/2024 17:53:53 - INFO - llamafactory.data.template - Add pad token: <|eot_id|>08/01/2024 17:53:53 - INFO - llamafactory.data.loader - Loading dataset identity.json...Generating train split: 91 examples [00:00, 10580.81 examples/s]Converting format of dataset (num_proc=16): 100%|███████████████████| 91/91 [00:00<00:00, 427.78 examples/s]08/01/2024 17:53:56 - INFO - llamafactory.data.loader - Loading dataset alpaca_en_demo.json...Generating train split: 1000 examples [00:00, 66788.28 examples/s]Converting format of dataset (num_proc=16): 100%|██████████████████████████████████████████████████████████████████████████████████████████| 1000/1000 [00:00<00:00, 4688.60 examples/s]Running tokenizer on dataset (num_proc=16): 100%|███████████████████████████████████████████████████████████████████████████████████████████| 1091/1091 [00:03<00:00, 295.08 examples/s]training example:input_ids:[128000, 128006, 882, 128007, 271, 6151, 128009, 128006, 78191, 128007, 271, 9906, 0, 358, 1097, 5991, 609, 39254, 459, 15592, 18328, 8040, 555, 5991, 3170, 3500, 13, 2650, 649, 358, 7945, 499, 3432, 30, 128009]inputs:<|begin_of_text|><|start_header_id|>user<|end_header_id|>hi<|eot_id|><|start_header_id|>assistant<|end_header_id|>Hello! I am {{name}}, an AI assistant developed by {{author}}. How can I assist you today?<|eot_id|>label_ids:[-100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, 9906, 0, 358, 1097, 5991, 609, 39254, 459, 15592, 18328, 8040, 555, 5991, 3170, 3500, 13, 2650, 649, 358, 7945, 499, 3432, 30, 128009]labels:Hello! I am {{name}}, an AI assistant developed by {{author}}. How can I assist you today?<|eot_id|>[INFO|configuration_utils.py:731] 2024-08-01 17:54:02,106 >> loading configuration file /root/.cache/modelscope/hub/LLM-Research/Meta-Llama-3-8B-Instruct/config.json[INFO|configuration_utils.py:800] 2024-08-01 17:54:02,108 >> Model config LlamaConfig {  "_name_or_path": "/root/.cache/modelscope/hub/LLM-Research/Meta-Llama-3-8B-Instruct",  "architectures": [    "LlamaForCausalLM"  ],  "attention_bias": false,  "attention_dropout": 0.0,  "bos_token_id": 128000,  "eos_token_id": 128009,  "hidden_act": "silu",  "hidden_size": 4096,  "initializer_range": 0.02,  "intermediate_size": 14336,  "max_position_embeddings": 8192,  "mlp_bias": false,  "model_type": "llama",  "num_attention_heads": 32,  "num_hidden_layers": 32,  "num_key_value_heads": 8,  "pretraining_tp": 1,  "rms_norm_eps": 1e-05,  "rope_scaling": null,  "rope_theta": 500000.0,  "tie_word_embeddings": false,  "torch_dtype": "bfloat16",  "transformers_version": "4.43.3",  "use_cache": true,  "vocab_size": 128256}[INFO|modeling_utils.py:3631] 2024-08-01 17:54:02,139 >> loading weights file /root/.cache/modelscope/hub/LLM-Research/Meta-Llama-3-8B-Instruct/model.safetensors.index.json[INFO|modeling_utils.py:1572] 2024-08-01 17:54:02,140 >> Instantiating LlamaForCausalLM model under default dtype torch.bfloat16.[INFO|configuration_utils.py:1038] 2024-08-01 17:54:02,142 >> Generate config GenerationConfig {  "bos_token_id": 128000,  "eos_token_id": 128009}Loading checkpoint shards: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 4/4 [00:01<00:00,  2.68it/s][INFO|modeling_utils.py:4463] 2024-08-01 17:54:03,708 >> All model checkpoint weights were used when initializing LlamaForCausalLM.[INFO|modeling_utils.py:4471] 2024-08-01 17:54:03,709 >> All the weights of LlamaForCausalLM were initialized from the model checkpoint at /root/.cache/modelscope/hub/LLM-Research/Meta-Llama-3-8B-Instruct.If your task is similar to the task the model of the checkpoint was trained on, you can already use LlamaForCausalLM for predictions without further training.[INFO|configuration_utils.py:991] 2024-08-01 17:54:03,712 >> loading configuration file /root/.cache/modelscope/hub/LLM-Research/Meta-Llama-3-8B-Instruct/generation_config.json[INFO|configuration_utils.py:1038] 2024-08-01 17:54:03,713 >> Generate config GenerationConfig {  "bos_token_id": 128000,  "do_sample": true,  "eos_token_id": [    128001,    128009  ],  "max_length": 4096,  "temperature": 0.6,  "top_p": 0.9}08/01/2024 17:54:03 - INFO - llamafactory.model.model_utils.checkpointing - Gradient checkpointing enabled.08/01/2024 17:54:03 - INFO - llamafactory.model.model_utils.attention - Using torch SDPA for faster training and inference.08/01/2024 17:54:03 - INFO - llamafactory.model.adapter - Upcasting trainable params to float32.08/01/2024 17:54:03 - INFO - llamafactory.model.adapter - Fine-tuning method: LoRA08/01/2024 17:54:03 - INFO - llamafactory.model.model_utils.misc - Found linear modules: q_proj,down_proj,o_proj,k_proj,gate_proj,up_proj,v_proj08/01/2024 17:54:08 - INFO - llamafactory.model.loader - trainable params: 20,971,520 || all params: 8,051,232,768 || trainable%: 0.2605Detected kernel version 3.10.0, which is below the recommended minimum of 5.5.0; this can cause the process to hang. It is recommended to upgrade the kernel to the minimum version or higher.[INFO|trainer.py:648] 2024-08-01 17:54:08,091 >> Using cpu_amp half precision backend[INFO|trainer.py:2134] 2024-08-01 17:54:09,008 >> ***** Running training *****[INFO|trainer.py:2135] 2024-08-01 17:54:09,008 >>   Num examples = 981[INFO|trainer.py:2136] 2024-08-01 17:54:09,008 >>   Num Epochs = 3[INFO|trainer.py:2137] 2024-08-01 17:54:09,008 >>   Instantaneous batch size per device = 1[INFO|trainer.py:2140] 2024-08-01 17:54:09,008 >>   Total train batch size (w. parallel, distributed & accumulation) = 8[INFO|trainer.py:2141] 2024-08-01 17:54:09,008 >>   Gradient Accumulation steps = 8[INFO|trainer.py:2142] 2024-08-01 17:54:09,008 >>   Total optimization steps = 366[INFO|trainer.py:2143] 2024-08-01 17:54:09,012 >>   Number of trainable parameters = 20,971,520  0%|                                                                                                                                                           | 0/366 [00:00<?, ?it/s
复制代码
错误缘故原由：当前PyTorch不支持DCU，导致步伐卡住，模型无法微调训练。
解决方法：在光合社区中查询并下载安装PyTorch。以 torch-2.1.0+das1.1.git3ac1bdd.abi1.dtk2404-cp310-cp310-manylinux_2_31_x86_64 为例，尝试安装 torch-2.1.0。

免责声明：如果侵犯了您的权益，请联系站长，我们会及时删除侵权内容，谢谢合作！更多信息从访问主页：qidao123.com:ToB企服之家，中国第一个企服评测及商务社交产业平台。

本帖子中包含更多资源

您需要 登录 才可以下载或查看，没有账号？立即注册

x
回复

使用道具 举报

0 个回复

倒序浏览
返回列表

快速回复

高级模式
B Color Image Link Quote Code Smilies
您需要登录后才可以回帖 登录 or 立即注册

本版积分规则

王柳

金牌会员
这个人很懒什么都没写!

楼主热帖

标签云

挺好的 服务器
微信订阅号
微信服务号
微信客服
小程序
H5
关于我们 商务合作 网站地图
©Copyright 2022-2024 杭州祥升科技有限公司 版权所有 浙ICP备20004199号
快速回复 返回顶部 返回列表