From 1d0259b921395b8e541def087b476dee10f5399f Mon Sep 17 00:00:00 2001 From: Zhihao Lin <36994684+LZHgrla@users.noreply.github.com> Date: Wed, 6 Sep 2023 17:41:12 +0800 Subject: [PATCH] [Feature] Support Baichuan2 models (#102) * add baichuan2_7b_base * fix lm_head bug for Baichuan2 * Update README.md * Update README_zh-CN.md * Update README.md * remove infrequent configs * Update README.md * Update README_zh-CN.md * add baichuan2 chat template * Update README_zh-CN.md --- README.md | 24 +- README_zh-CN.md | 24 +- .../baichuan2_7b_base_qlora_alpaca_e3.py | 180 +++++++++++++++ .../baichuan2_7b_base_qlora_alpaca_enzh_e3.py | 198 ++++++++++++++++ ...an2_7b_base_qlora_alpaca_enzh_oasst1_e3.py | 211 +++++++++++++++++ .../baichuan2_7b_base_qlora_alpaca_zh_e3.py | 180 +++++++++++++++ ...ichuan2_7b_base_qlora_arxiv_gentitle_e3.py | 215 ++++++++++++++++++ .../baichuan2_7b_base_qlora_code_alpaca_e3.py | 184 +++++++++++++++ .../baichuan2_7b_base_qlora_colorist_e5.py | 180 +++++++++++++++ .../baichuan2_7b_base_qlora_lawyer_e3.py | 206 +++++++++++++++++ .../baichuan2_7b_base_qlora_oasst1_512_e3.py | 180 +++++++++++++++ .../baichuan2_7b_base_qlora_oasst1_e3.py | 180 +++++++++++++++ ...aichuan2_7b_base_qlora_open_platypus_e3.py | 180 +++++++++++++++ .../baichuan2_7b_base_qlora_sql_e3.py | 184 +++++++++++++++ xtuner/model/sft.py | 7 + xtuner/utils/templates.py | 3 + 16 files changed, 2314 insertions(+), 22 deletions(-) create mode 100644 xtuner/configs/baichuan/baichuan2_7b_base/baichuan2_7b_base_qlora_alpaca_e3.py create mode 100644 xtuner/configs/baichuan/baichuan2_7b_base/baichuan2_7b_base_qlora_alpaca_enzh_e3.py create mode 100644 xtuner/configs/baichuan/baichuan2_7b_base/baichuan2_7b_base_qlora_alpaca_enzh_oasst1_e3.py create mode 100644 xtuner/configs/baichuan/baichuan2_7b_base/baichuan2_7b_base_qlora_alpaca_zh_e3.py create mode 100644 xtuner/configs/baichuan/baichuan2_7b_base/baichuan2_7b_base_qlora_arxiv_gentitle_e3.py create mode 100644 xtuner/configs/baichuan/baichuan2_7b_base/baichuan2_7b_base_qlora_code_alpaca_e3.py create mode 100644 xtuner/configs/baichuan/baichuan2_7b_base/baichuan2_7b_base_qlora_colorist_e5.py create mode 100644 xtuner/configs/baichuan/baichuan2_7b_base/baichuan2_7b_base_qlora_lawyer_e3.py create mode 100644 xtuner/configs/baichuan/baichuan2_7b_base/baichuan2_7b_base_qlora_oasst1_512_e3.py create mode 100644 xtuner/configs/baichuan/baichuan2_7b_base/baichuan2_7b_base_qlora_oasst1_e3.py create mode 100644 xtuner/configs/baichuan/baichuan2_7b_base/baichuan2_7b_base_qlora_open_platypus_e3.py create mode 100644 xtuner/configs/baichuan/baichuan2_7b_base/baichuan2_7b_base_qlora_sql_e3.py diff --git a/README.md b/README.md index d28730113..c08a1eb3c 100644 --- a/README.md +++ b/README.md @@ -14,6 +14,7 @@ English | [简体中文](README_zh-CN.md) ## 🎉 News +- **\[2023.09.06\]** Support the training of [Baichuan2](https://huggingface.co/baichuan-inc) models! Try it out by `xtuner train baichuan2_7b_base_qlora_oasst1_e3`! - **\[2023.08.30\]** XTuner is released, with multiple fine-tuned adapters on [HuggingFace](https://huggingface.co/xtuner). ## 📖 Introduction @@ -21,7 +22,7 @@ English | [简体中文](README_zh-CN.md) XTuner is a toolkit for efficiently fine-tuning LLM, developed by the [MMRazor](https://github.com/open-mmlab/mmrazor) and [MMDeploy](https://github.com/open-mmlab/mmdeploy) teams. - **Efficiency**: Support LLM fine-tuning on consumer-grade GPUs. The minimum GPU memory required for 7B LLM fine-tuning is only **8GB**, indicating that users can use nearly any GPU (even the free resource, *e.g.*, Colab) to fine-tune custom LLMs. -- **Versatile**: Support various **LLMs** ([InternLM](https://github.com/InternLM/InternLM), [Llama2](https://github.com/facebookresearch/llama), [ChatGLM2](https://huggingface.co/THUDM/chatglm2-6b), [Qwen](https://github.com/QwenLM/Qwen-7B), [Baichuan](https://github.com/baichuan-inc), ...), **datasets** ([MOSS_003_SFT](https://huggingface.co/datasets/fnlp/moss-003-sft-data), [Alpaca](https://huggingface.co/datasets/tatsu-lab/alpaca), [WizardLM](https://huggingface.co/datasets/WizardLM/WizardLM_evol_instruct_V2_196k), [oasst1](https://huggingface.co/datasets/timdettmers/openassistant-guanaco), [Open-Platypus](https://huggingface.co/datasets/garage-bAInd/Open-Platypus), [Code Alpaca](https://huggingface.co/datasets/HuggingFaceH4/CodeAlpaca_20K), [Colorist](https://huggingface.co/datasets/burkelibbey/colors), ...) and **algorithms** ([QLoRA](http://arxiv.org/abs/2305.14314), [LoRA](http://arxiv.org/abs/2106.09685)), allowing users to choose the most suitable solution for their requirements. +- **Versatile**: Support various **LLMs** ([InternLM](https://huggingface.co/internlm), [Llama2](https://huggingface.co/meta-llama), [ChatGLM2](https://huggingface.co/THUDM/chatglm2-6b), [Qwen](https://huggingface.co/Qwen), [Baichuan2](https://huggingface.co/baichuan-inc), ...), **datasets** ([MOSS_003_SFT](https://huggingface.co/datasets/fnlp/moss-003-sft-data), [Alpaca](https://huggingface.co/datasets/tatsu-lab/alpaca), [WizardLM](https://huggingface.co/datasets/WizardLM/WizardLM_evol_instruct_V2_196k), [oasst1](https://huggingface.co/datasets/timdettmers/openassistant-guanaco), [Open-Platypus](https://huggingface.co/datasets/garage-bAInd/Open-Platypus), [Code Alpaca](https://huggingface.co/datasets/HuggingFaceH4/CodeAlpaca_20K), [Colorist](https://huggingface.co/datasets/burkelibbey/colors), ...) and **algorithms** ([QLoRA](http://arxiv.org/abs/2305.14314), [LoRA](http://arxiv.org/abs/2106.09685)), allowing users to choose the most suitable solution for their requirements. - **Compatibility**: Compatible with [DeepSpeed](https://github.com/microsoft/DeepSpeed) 🚀 and [HuggingFace](https://huggingface.co) 🤗 training pipeline, enabling effortless integration and utilization. ## 🌟 Demos @@ -70,17 +71,18 @@ XTuner is a toolkit for efficiently fine-tuning LLM, developed by the [MMRazor]( diff --git a/README_zh-CN.md b/README_zh-CN.md index 0ca037b7a..2bba9ad54 100644 --- a/README_zh-CN.md +++ b/README_zh-CN.md @@ -14,6 +14,7 @@ ## 🎉 更新 +- **\[2023.09.06\]** 支持 [Baichuan2](https://huggingface.co/baichuan-inc) 系列模型训练！快速体验：`xtuner train baichuan2_7b_base_qlora_oasst1_e3`！ - **\[2023.08.30\]** XTuner 正式发布！众多微调模型已上传至 [HuggingFace](https://huggingface.co/xtuner)！ ## 📖 介绍 @@ -21,7 +22,7 @@ XTuner 是一个轻量级微调大语言模型的工具库，由 [MMRazor](https://github.com/open-mmlab/mmrazor) 和 [MMDeploy](https://github.com/open-mmlab/mmdeploy) 团队联合开发。 - **轻量级**: 支持在消费级显卡上微调大语言模型。对于 7B 参数量，微调所需的最小显存仅为 **8GB**，这使得用户可以使用几乎任何显卡（甚至免费资源，例如Colab）来微调获得自定义大语言模型助手。 -- **多样性**: 支持多种**大语言模型**（[InternLM](https://github.com/InternLM/InternLM)、[Llama2](https://github.com/facebookresearch/llama)、[ChatGLM2](https://huggingface.co/THUDM/chatglm2-6b)、[Qwen](https://github.com/QwenLM/Qwen-7B)、[Baichuan](https://github.com/baichuan-inc), ...），**数据集**（[MOSS_003_SFT](https://huggingface.co/datasets/fnlp/moss-003-sft-data), [Alpaca](https://huggingface.co/datasets/tatsu-lab/alpaca), [WizardLM](https://huggingface.co/datasets/WizardLM/WizardLM_evol_instruct_V2_196k), [oasst1](https://huggingface.co/datasets/timdettmers/openassistant-guanaco), [Open-Platypus](https://huggingface.co/datasets/garage-bAInd/Open-Platypus), [Code Alpaca](https://huggingface.co/datasets/HuggingFaceH4/CodeAlpaca_20K), [Colorist](https://huggingface.co/datasets/burkelibbey/colors), ...）和**微调算法**（[QLoRA](http://arxiv.org/abs/2305.14314)、[LoRA](http://arxiv.org/abs/2106.09685)），支撑用户根据自身具体需求选择合适的解决方案。 +- **多样性**: 支持多种**大语言模型**（[InternLM](https://huggingface.co/internlm)、[Llama2](https://huggingface.co/meta-llama)、[ChatGLM2](https://huggingface.co/THUDM/chatglm2-6b)、[Qwen](https://huggingface.co/Qwen)、[Baichuan2](https://huggingface.co/baichuan-inc), ...），**数据集**（[MOSS_003_SFT](https://huggingface.co/datasets/fnlp/moss-003-sft-data), [Alpaca](https://huggingface.co/datasets/tatsu-lab/alpaca), [WizardLM](https://huggingface.co/datasets/WizardLM/WizardLM_evol_instruct_V2_196k), [oasst1](https://huggingface.co/datasets/timdettmers/openassistant-guanaco), [Open-Platypus](https://huggingface.co/datasets/garage-bAInd/Open-Platypus), [Code Alpaca](https://huggingface.co/datasets/HuggingFaceH4/CodeAlpaca_20K), [Colorist](https://huggingface.co/datasets/burkelibbey/colors), ...）和**微调算法**（[QLoRA](http://arxiv.org/abs/2305.14314)、[LoRA](http://arxiv.org/abs/2106.09685)），支撑用户根据自身具体需求选择合适的解决方案。 - **兼容性**: 兼容 [DeepSpeed](https://github.com/microsoft/DeepSpeed) 🚀 和 [HuggingFace](https://huggingface.co) 🤗 的训练流程，支撑用户无感式集成与使用。 ## 🌟 示例 @@ -70,17 +71,18 @@ XTuner 是一个轻量级微调大语言模型的工具库，由 [MMRazor](https diff --git a/xtuner/configs/baichuan/baichuan2_7b_base/baichuan2_7b_base_qlora_alpaca_e3.py b/xtuner/configs/baichuan/baichuan2_7b_base/baichuan2_7b_base_qlora_alpaca_e3.py new file mode 100644 index 000000000..3500806e3 --- /dev/null +++ b/xtuner/configs/baichuan/baichuan2_7b_base/baichuan2_7b_base_qlora_alpaca_e3.py @@ -0,0 +1,180 @@ +# Copyright (c) OpenMMLab. All rights reserved. +import torch +from bitsandbytes.optim import PagedAdamW32bit +from datasets import load_dataset +from mmengine.dataset import DefaultSampler +from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook, + LoggerHook, ParamSchedulerHook) +from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR +from peft import LoraConfig +from transformers import (AutoModelForCausalLM, AutoTokenizer, + BitsAndBytesConfig) + +from xtuner.dataset import process_hf_dataset +from xtuner.dataset.collate_fns import default_collate_fn +from xtuner.dataset.map_fns import alpaca_map_fn, template_map_fn_factory +from xtuner.engine import DatasetInfoHook, EvaluateChatHook +from xtuner.model import SupervisedFinetune +from xtuner.utils import PROMPT_TEMPLATE + +####################################################################### +# PART 1 Settings # +####################################################################### +# Model +pretrained_model_name_or_path = 'baichuan-inc/Baichuan2-7B-Base' + +# Data +alpaca_en_path = 'tatsu-lab/alpaca' +prompt_template = PROMPT_TEMPLATE.alpaca +max_length = 2048 +pack_to_max_length = True + +# Scheduler & Optimizer +batch_size = 1 # per_device +accumulative_counts = 16 +dataloader_num_workers = 0 +max_epochs = 3 +optim_type = PagedAdamW32bit +lr = 2e-4 +betas = (0.9, 0.999) +weight_decay = 0 +max_norm = 1 # grad clip + +# Evaluate the generation performance during the training +evaluation_freq = 500 +evaluation_inputs = [ + '请给我介绍五个上海的景点', 'Please tell me five scenic spots in Shanghai' +] + +####################################################################### +# PART 2 Model & Tokenizer # +####################################################################### +tokenizer = dict( + type=AutoTokenizer.from_pretrained, + pretrained_model_name_or_path=pretrained_model_name_or_path, + trust_remote_code=True, + padding_side='right') + +model = dict( + type=SupervisedFinetune, + llm=dict( + type=AutoModelForCausalLM.from_pretrained, + pretrained_model_name_or_path=pretrained_model_name_or_path, + trust_remote_code=True, + torch_dtype=torch.float16, + quantization_config=dict( + type=BitsAndBytesConfig, + load_in_4bit=True, + load_in_8bit=False, + llm_int8_threshold=6.0, + llm_int8_has_fp16_weight=False, + bnb_4bit_compute_dtype=torch.float16, + bnb_4bit_use_double_quant=True, + bnb_4bit_quant_type='nf4')), + lora=dict( + type=LoraConfig, + r=64, + lora_alpha=16, + lora_dropout=0.1, + bias='none', + task_type='CAUSAL_LM')) + +####################################################################### +# PART 3 Dataset & Dataloader # +####################################################################### +alpaca_en = dict( + type=process_hf_dataset, + dataset=dict(type=load_dataset, path=alpaca_en_path), + tokenizer=tokenizer, + max_length=max_length, + dataset_map_fn=alpaca_map_fn, + template_map_fn=dict( + type=template_map_fn_factory, template=prompt_template), + remove_unused_columns=True, + shuffle_before_pack=True, + pack_to_max_length=pack_to_max_length) + +train_dataloader = dict( + batch_size=batch_size, + num_workers=dataloader_num_workers, + dataset=alpaca_en, + sampler=dict(type=DefaultSampler, shuffle=True), + collate_fn=dict(type=default_collate_fn)) + +####################################################################### +# PART 4 Scheduler & Optimizer # +####################################################################### +# optimizer +optim_wrapper = dict( + type=AmpOptimWrapper, + optimizer=dict( + type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay), + clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False), + accumulative_counts=accumulative_counts, + loss_scale='dynamic', + dtype='float16') + +# learning policy +# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md # noqa: E501 +param_scheduler = dict( + type=CosineAnnealingLR, + eta_min=lr * 0.1, + by_epoch=True, + T_max=max_epochs, + convert_to_iter_based=True) + +# train, val, test setting +train_cfg = dict(by_epoch=True, max_epochs=max_epochs, val_interval=1) + +####################################################################### +# PART 5 Runtime # +####################################################################### +# Log the dialogue periodically during the training process, optional +custom_hooks = [ + dict(type=DatasetInfoHook, tokenizer=tokenizer), + dict( + type=EvaluateChatHook, + tokenizer=tokenizer, + every_n_iters=evaluation_freq, + evaluation_inputs=evaluation_inputs, + instruction=prompt_template.INSTRUCTION_START) +] + +# configure default hooks +default_hooks = dict( + # record the time of every iteration. + timer=dict(type=IterTimerHook), + # print log every 100 iterations. + logger=dict(type=LoggerHook, interval=10), + # enable the parameter scheduler. + param_scheduler=dict(type=ParamSchedulerHook), + # save checkpoint per epoch. + checkpoint=dict(type=CheckpointHook, interval=1), + # set sampler seed in distributed evrionment. + sampler_seed=dict(type=DistSamplerSeedHook), +) + +# configure environment +env_cfg = dict( + # whether to enable cudnn benchmark + cudnn_benchmark=False, + # set multi process parameters + mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0), + # set distributed parameters + dist_cfg=dict(backend='nccl'), +) + +# set visualizer +visualizer = None + +# set log level +log_level = 'INFO' + +# load from which checkpoint +load_from = None + +# whether to resume training from the loaded checkpoint +resume = False + +# Defaults to use random seed and disable `deterministic` +randomness = dict(seed=None, deterministic=False) diff --git a/xtuner/configs/baichuan/baichuan2_7b_base/baichuan2_7b_base_qlora_alpaca_enzh_e3.py b/xtuner/configs/baichuan/baichuan2_7b_base/baichuan2_7b_base_qlora_alpaca_enzh_e3.py new file mode 100644 index 000000000..2ec4464d3 --- /dev/null +++ b/xtuner/configs/baichuan/baichuan2_7b_base/baichuan2_7b_base_qlora_alpaca_enzh_e3.py @@ -0,0 +1,198 @@ +# Copyright (c) OpenMMLab. All rights reserved. +import torch +from bitsandbytes.optim import PagedAdamW32bit +from datasets import load_dataset +from mmengine.dataset import DefaultSampler +from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook, + LoggerHook, ParamSchedulerHook) +from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR +from peft import LoraConfig +from transformers import (AutoModelForCausalLM, AutoTokenizer, + BitsAndBytesConfig) + +from xtuner.dataset import ConcatDataset, process_hf_dataset +from xtuner.dataset.collate_fns import default_collate_fn +from xtuner.dataset.map_fns import (alpaca_map_fn, alpaca_zh_map_fn, + template_map_fn_factory) +from xtuner.engine import DatasetInfoHook, EvaluateChatHook +from xtuner.model import SupervisedFinetune +from xtuner.utils import PROMPT_TEMPLATE + +####################################################################### +# PART 1 Settings # +####################################################################### +# Model +pretrained_model_name_or_path = 'baichuan-inc/Baichuan2-7B-Base' + +# Data +alpaca_zh_path = 'silk-road/alpaca-data-gpt4-chinese' +alpaca_en_path = 'tatsu-lab/alpaca' +prompt_template = PROMPT_TEMPLATE.alpaca +max_length = 2048 +pack_to_max_length = True + +# Scheduler & Optimizer +batch_size = 1 # per_device +accumulative_counts = 16 +dataloader_num_workers = 0 +max_epochs = 3 +optim_type = PagedAdamW32bit +lr = 2e-4 +betas = (0.9, 0.999) +weight_decay = 0 +max_norm = 1 # grad clip + +# Evaluate the generation performance during the training +evaluation_freq = 500 +evaluation_inputs = [ + '请给我介绍五个上海的景点', 'Please tell me five scenic spots in Shanghai' +] + +####################################################################### +# PART 2 Model & Tokenizer # +####################################################################### +tokenizer = dict( + type=AutoTokenizer.from_pretrained, + pretrained_model_name_or_path=pretrained_model_name_or_path, + trust_remote_code=True, + padding_side='right') + +model = dict( + type=SupervisedFinetune, + llm=dict( + type=AutoModelForCausalLM.from_pretrained, + pretrained_model_name_or_path=pretrained_model_name_or_path, + trust_remote_code=True, + torch_dtype=torch.float16, + quantization_config=dict( + type=BitsAndBytesConfig, + load_in_4bit=True, + load_in_8bit=False, + llm_int8_threshold=6.0, + llm_int8_has_fp16_weight=False, + bnb_4bit_compute_dtype=torch.float16, + bnb_4bit_use_double_quant=True, + bnb_4bit_quant_type='nf4')), + lora=dict( + type=LoraConfig, + r=64, + lora_alpha=16, + lora_dropout=0.1, + bias='none', + task_type='CAUSAL_LM')) + +####################################################################### +# PART 3 Dataset & Dataloader # +####################################################################### +alpaca_en = dict( + type=process_hf_dataset, + dataset=dict(type=load_dataset, path=alpaca_en_path), + tokenizer=tokenizer, + max_length=max_length, + dataset_map_fn=alpaca_map_fn, + template_map_fn=dict( + type=template_map_fn_factory, template=prompt_template), + remove_unused_columns=True, + shuffle_before_pack=True, + pack_to_max_length=pack_to_max_length) + +alpaca_zh = dict( + type=process_hf_dataset, + dataset=dict(type=load_dataset, path=alpaca_zh_path), + tokenizer=tokenizer, + max_length=max_length, + dataset_map_fn=alpaca_zh_map_fn, + template_map_fn=dict( + type=template_map_fn_factory, template=prompt_template), + remove_unused_columns=True, + shuffle_before_pack=True, + pack_to_max_length=pack_to_max_length) + +train_dataset = dict( + type=ConcatDataset, + datasets_cfg=dict(alpaca_en=alpaca_en, alpaca_zh=alpaca_zh)) + +train_dataloader = dict( + batch_size=batch_size, + num_workers=dataloader_num_workers, + dataset=train_dataset, + sampler=dict(type=DefaultSampler, shuffle=True), + collate_fn=dict(type=default_collate_fn)) + +####################################################################### +# PART 4 Scheduler & Optimizer # +####################################################################### +# optimizer +optim_wrapper = dict( + type=AmpOptimWrapper, + optimizer=dict( + type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay), + clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False), + accumulative_counts=accumulative_counts, + loss_scale='dynamic', + dtype='float16') + +# learning policy +# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md # noqa: E501 +param_scheduler = dict( + type=CosineAnnealingLR, + eta_min=lr * 0.1, + by_epoch=True, + T_max=max_epochs, + convert_to_iter_based=True) + +# train, val, test setting +train_cfg = dict(by_epoch=True, max_epochs=max_epochs, val_interval=1) + +####################################################################### +# PART 5 Runtime # +####################################################################### +# Log the dialogue periodically during the training process, optional +custom_hooks = [ + dict(type=DatasetInfoHook, tokenizer=tokenizer), + dict( + type=EvaluateChatHook, + tokenizer=tokenizer, + every_n_iters=evaluation_freq, + evaluation_inputs=evaluation_inputs, + instruction=prompt_template.INSTRUCTION_START) +] + +# configure default hooks +default_hooks = dict( + # record the time of every iteration. + timer=dict(type=IterTimerHook), + # print log every 100 iterations. + logger=dict(type=LoggerHook, interval=10), + # enable the parameter scheduler. + param_scheduler=dict(type=ParamSchedulerHook), + # save checkpoint per epoch. + checkpoint=dict(type=CheckpointHook, interval=1), + # set sampler seed in distributed evrionment. + sampler_seed=dict(type=DistSamplerSeedHook), +) + +# configure environment +env_cfg = dict( + # whether to enable cudnn benchmark + cudnn_benchmark=False, + # set multi process parameters + mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0), + # set distributed parameters + dist_cfg=dict(backend='nccl'), +) + +# set visualizer +visualizer = None + +# set log level +log_level = 'INFO' + +# load from which checkpoint +load_from = None + +# whether to resume training from the loaded checkpoint +resume = False + +# Defaults to use random seed and disable `deterministic` +randomness = dict(seed=None, deterministic=False) diff --git a/xtuner/configs/baichuan/baichuan2_7b_base/baichuan2_7b_base_qlora_alpaca_enzh_oasst1_e3.py b/xtuner/configs/baichuan/baichuan2_7b_base/baichuan2_7b_base_qlora_alpaca_enzh_oasst1_e3.py new file mode 100644 index 000000000..e18bf308b --- /dev/null +++ b/xtuner/configs/baichuan/baichuan2_7b_base/baichuan2_7b_base_qlora_alpaca_enzh_oasst1_e3.py @@ -0,0 +1,211 @@ +# Copyright (c) OpenMMLab. All rights reserved. +import torch +from bitsandbytes.optim import PagedAdamW32bit +from datasets import load_dataset +from mmengine.dataset import DefaultSampler +from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook, + LoggerHook, ParamSchedulerHook) +from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR +from peft import LoraConfig +from transformers import (AutoModelForCausalLM, AutoTokenizer, + BitsAndBytesConfig) + +from xtuner.dataset import ConcatDataset, process_hf_dataset +from xtuner.dataset.collate_fns import default_collate_fn +from xtuner.dataset.map_fns import (alpaca_map_fn, alpaca_zh_map_fn, + oasst1_map_fn, template_map_fn_factory) +from xtuner.engine import DatasetInfoHook, EvaluateChatHook +from xtuner.model import SupervisedFinetune +from xtuner.utils import PROMPT_TEMPLATE + +####################################################################### +# PART 1 Settings # +####################################################################### +# Model +pretrained_model_name_or_path = 'baichuan-inc/Baichuan2-7B-Base' + +# Data +alpaca_zh_path = 'silk-road/alpaca-data-gpt4-chinese' +alpaca_en_path = 'tatsu-lab/alpaca' +oasst1_path = 'timdettmers/openassistant-guanaco' +prompt_template = PROMPT_TEMPLATE.alpaca +max_length = 2048 +pack_to_max_length = True + +# Scheduler & Optimizer +batch_size = 1 # per_device +accumulative_counts = 16 +dataloader_num_workers = 0 +max_epochs = 3 +optim_type = PagedAdamW32bit +lr = 2e-4 +betas = (0.9, 0.999) +weight_decay = 0 +max_norm = 1 # grad clip + +# Evaluate the generation performance during the training +evaluation_freq = 500 +evaluation_inputs = [ + '请给我介绍五个上海的景点', 'Please tell me five scenic spots in Shanghai' +] + +####################################################################### +# PART 2 Model & Tokenizer # +####################################################################### +tokenizer = dict( + type=AutoTokenizer.from_pretrained, + pretrained_model_name_or_path=pretrained_model_name_or_path, + trust_remote_code=True, + padding_side='right') + +model = dict( + type=SupervisedFinetune, + llm=dict( + type=AutoModelForCausalLM.from_pretrained, + pretrained_model_name_or_path=pretrained_model_name_or_path, + trust_remote_code=True, + torch_dtype=torch.float16, + quantization_config=dict( + type=BitsAndBytesConfig, + load_in_4bit=True, + load_in_8bit=False, + llm_int8_threshold=6.0, + llm_int8_has_fp16_weight=False, + bnb_4bit_compute_dtype=torch.float16, + bnb_4bit_use_double_quant=True, + bnb_4bit_quant_type='nf4')), + lora=dict( + type=LoraConfig, + r=64, + lora_alpha=16, + lora_dropout=0.1, + bias='none', + task_type='CAUSAL_LM')) + +####################################################################### +# PART 3 Dataset & Dataloader # +####################################################################### +alpaca_en = dict( + type=process_hf_dataset, + dataset=dict(type=load_dataset, path=alpaca_en_path), + tokenizer=tokenizer, + max_length=max_length, + dataset_map_fn=alpaca_map_fn, + template_map_fn=dict( + type=template_map_fn_factory, template=prompt_template), + remove_unused_columns=True, + shuffle_before_pack=True, + pack_to_max_length=pack_to_max_length) + +alpaca_zh = dict( + type=process_hf_dataset, + dataset=dict(type=load_dataset, path=alpaca_zh_path), + tokenizer=tokenizer, + max_length=max_length, + dataset_map_fn=alpaca_zh_map_fn, + template_map_fn=dict( + type=template_map_fn_factory, template=prompt_template), + remove_unused_columns=True, + shuffle_before_pack=True, + pack_to_max_length=pack_to_max_length) + +oasst1 = dict( + type=process_hf_dataset, + dataset=dict(type=load_dataset, path=oasst1_path), + tokenizer=tokenizer, + max_length=max_length, + dataset_map_fn=oasst1_map_fn, + template_map_fn=dict( + type=template_map_fn_factory, template=prompt_template), + remove_unused_columns=True, + shuffle_before_pack=True, + pack_to_max_length=pack_to_max_length) + +train_dataset = dict( + type=ConcatDataset, + datasets_cfg=dict(alpaca_en=alpaca_en, alpaca_zh=alpaca_zh, oasst1=oasst1)) + +train_dataloader = dict( + batch_size=batch_size, + num_workers=dataloader_num_workers, + dataset=train_dataset, + sampler=dict(type=DefaultSampler, shuffle=True), + collate_fn=dict(type=default_collate_fn)) + +####################################################################### +# PART 4 Scheduler & Optimizer # +####################################################################### +# optimizer +optim_wrapper = dict( + type=AmpOptimWrapper, + optimizer=dict( + type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay), + clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False), + accumulative_counts=accumulative_counts, + loss_scale='dynamic', + dtype='float16') + +# learning policy +# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md # noqa: E501 +param_scheduler = dict( + type=CosineAnnealingLR, + eta_min=lr * 0.1, + by_epoch=True, + T_max=max_epochs, + convert_to_iter_based=True) + +# train, val, test setting +train_cfg = dict(by_epoch=True, max_epochs=max_epochs, val_interval=1) + +####################################################################### +# PART 5 Runtime # +####################################################################### +# Log the dialogue periodically during the training process, optional +custom_hooks = [ + dict(type=DatasetInfoHook, tokenizer=tokenizer), + dict( + type=EvaluateChatHook, + tokenizer=tokenizer, + every_n_iters=evaluation_freq, + evaluation_inputs=evaluation_inputs, + instruction=prompt_template.INSTRUCTION_START) +] + +# configure default hooks +default_hooks = dict( + # record the time of every iteration. + timer=dict(type=IterTimerHook), + # print log every 100 iterations. + logger=dict(type=LoggerHook, interval=10), + # enable the parameter scheduler. + param_scheduler=dict(type=ParamSchedulerHook), + # save checkpoint per epoch. + checkpoint=dict(type=CheckpointHook, interval=1), + # set sampler seed in distributed evrionment. + sampler_seed=dict(type=DistSamplerSeedHook), +) + +# configure environment +env_cfg = dict( + # whether to enable cudnn benchmark + cudnn_benchmark=False, + # set multi process parameters + mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0), + # set distributed parameters + dist_cfg=dict(backend='nccl'), +) + +# set visualizer +visualizer = None + +# set log level +log_level = 'INFO' + +# load from which checkpoint +load_from = None + +# whether to resume training from the loaded checkpoint +resume = False + +# Defaults to use random seed and disable `deterministic` +randomness = dict(seed=None, deterministic=False) diff --git a/xtuner/configs/baichuan/baichuan2_7b_base/baichuan2_7b_base_qlora_alpaca_zh_e3.py b/xtuner/configs/baichuan/baichuan2_7b_base/baichuan2_7b_base_qlora_alpaca_zh_e3.py new file mode 100644 index 000000000..988f75b90 --- /dev/null +++ b/xtuner/configs/baichuan/baichuan2_7b_base/baichuan2_7b_base_qlora_alpaca_zh_e3.py @@ -0,0 +1,180 @@ +# Copyright (c) OpenMMLab. All rights reserved. +import torch +from bitsandbytes.optim import PagedAdamW32bit +from datasets import load_dataset +from mmengine.dataset import DefaultSampler +from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook, + LoggerHook, ParamSchedulerHook) +from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR +from peft import LoraConfig +from transformers import (AutoModelForCausalLM, AutoTokenizer, + BitsAndBytesConfig) + +from xtuner.dataset import process_hf_dataset +from xtuner.dataset.collate_fns import default_collate_fn +from xtuner.dataset.map_fns import alpaca_zh_map_fn, template_map_fn_factory +from xtuner.engine import DatasetInfoHook, EvaluateChatHook +from xtuner.model import SupervisedFinetune +from xtuner.utils import PROMPT_TEMPLATE + +####################################################################### +# PART 1 Settings # +####################################################################### +# Model +pretrained_model_name_or_path = 'baichuan-inc/Baichuan2-7B-Base' + +# Data +alpaca_zh_path = 'silk-road/alpaca-data-gpt4-chinese' +prompt_template = PROMPT_TEMPLATE.alpaca +max_length = 2048 +pack_to_max_length = True + +# Scheduler & Optimizer +batch_size = 1 # per_device +accumulative_counts = 16 +dataloader_num_workers = 0 +max_epochs = 3 +optim_type = PagedAdamW32bit +lr = 2e-4 +betas = (0.9, 0.999) +weight_decay = 0 +max_norm = 1 # grad clip + +# Evaluate the generation performance during the training +evaluation_freq = 500 +evaluation_inputs = [ + '请给我介绍五个上海的景点', 'Please tell me five scenic spots in Shanghai' +] + +####################################################################### +# PART 2 Model & Tokenizer # +####################################################################### +tokenizer = dict( + type=AutoTokenizer.from_pretrained, + pretrained_model_name_or_path=pretrained_model_name_or_path, + trust_remote_code=True, + padding_side='right') + +model = dict( + type=SupervisedFinetune, + llm=dict( + type=AutoModelForCausalLM.from_pretrained, + pretrained_model_name_or_path=pretrained_model_name_or_path, + trust_remote_code=True, + torch_dtype=torch.float16, + quantization_config=dict( + type=BitsAndBytesConfig, + load_in_4bit=True, + load_in_8bit=False, + llm_int8_threshold=6.0, + llm_int8_has_fp16_weight=False, + bnb_4bit_compute_dtype=torch.float16, + bnb_4bit_use_double_quant=True, + bnb_4bit_quant_type='nf4')), + lora=dict( + type=LoraConfig, + r=64, + lora_alpha=16, + lora_dropout=0.1, + bias='none', + task_type='CAUSAL_LM')) + +####################################################################### +# PART 3 Dataset & Dataloader # +####################################################################### +alpaca_zh = dict( + type=process_hf_dataset, + dataset=dict(type=load_dataset, path=alpaca_zh_path), + tokenizer=tokenizer, + max_length=max_length, + dataset_map_fn=alpaca_zh_map_fn, + template_map_fn=dict( + type=template_map_fn_factory, template=prompt_template), + remove_unused_columns=True, + shuffle_before_pack=True, + pack_to_max_length=pack_to_max_length) + +train_dataloader = dict( + batch_size=batch_size, + num_workers=dataloader_num_workers, + dataset=alpaca_zh, + sampler=dict(type=DefaultSampler, shuffle=True), + collate_fn=dict(type=default_collate_fn)) + +####################################################################### +# PART 4 Scheduler & Optimizer # +####################################################################### +# optimizer +optim_wrapper = dict( + type=AmpOptimWrapper, + optimizer=dict( + type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay), + clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False), + accumulative_counts=accumulative_counts, + loss_scale='dynamic', + dtype='float16') + +# learning policy +# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md # noqa: E501 +param_scheduler = dict( + type=CosineAnnealingLR, + eta_min=lr * 0.1, + by_epoch=True, + T_max=max_epochs, + convert_to_iter_based=True) + +# train, val, test setting +train_cfg = dict(by_epoch=True, max_epochs=max_epochs, val_interval=1) + +####################################################################### +# PART 5 Runtime # +####################################################################### +# Log the dialogue periodically during the training process, optional +custom_hooks = [ + dict(type=DatasetInfoHook, tokenizer=tokenizer), + dict( + type=EvaluateChatHook, + tokenizer=tokenizer, + every_n_iters=evaluation_freq, + evaluation_inputs=evaluation_inputs, + instruction=prompt_template.INSTRUCTION_START) +] + +# configure default hooks +default_hooks = dict( + # record the time of every iteration. + timer=dict(type=IterTimerHook), + # print log every 100 iterations. + logger=dict(type=LoggerHook, interval=10), + # enable the parameter scheduler. + param_scheduler=dict(type=ParamSchedulerHook), + # save checkpoint per epoch. + checkpoint=dict(type=CheckpointHook, interval=1), + # set sampler seed in distributed evrionment. + sampler_seed=dict(type=DistSamplerSeedHook), +) + +# configure environment +env_cfg = dict( + # whether to enable cudnn benchmark + cudnn_benchmark=False, + # set multi process parameters + mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0), + # set distributed parameters + dist_cfg=dict(backend='nccl'), +) + +# set visualizer +visualizer = None + +# set log level +log_level = 'INFO' + +# load from which checkpoint +load_from = None + +# whether to resume training from the loaded checkpoint +resume = False + +# Defaults to use random seed and disable `deterministic` +randomness = dict(seed=None, deterministic=False) diff --git a/xtuner/configs/baichuan/baichuan2_7b_base/baichuan2_7b_base_qlora_arxiv_gentitle_e3.py b/xtuner/configs/baichuan/baichuan2_7b_base/baichuan2_7b_base_qlora_arxiv_gentitle_e3.py new file mode 100644 index 000000000..1e1ad6924 --- /dev/null +++ b/xtuner/configs/baichuan/baichuan2_7b_base/baichuan2_7b_base_qlora_arxiv_gentitle_e3.py @@ -0,0 +1,215 @@ +# Copyright (c) OpenMMLab. All rights reserved. +import torch +from bitsandbytes.optim import PagedAdamW32bit +from datasets import load_dataset +from mmengine.dataset import DefaultSampler +from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook, + LoggerHook, ParamSchedulerHook) +from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR +from peft import LoraConfig +from transformers import (AutoModelForCausalLM, AutoTokenizer, + BitsAndBytesConfig) + +from xtuner.dataset import process_hf_dataset +from xtuner.dataset.collate_fns import default_collate_fn +from xtuner.dataset.map_fns import arxiv_map_fn, template_map_fn_factory +from xtuner.engine import DatasetInfoHook, EvaluateChatHook +from xtuner.model import SupervisedFinetune +from xtuner.utils import PROMPT_TEMPLATE + +####################################################################### +# PART 1 Settings # +####################################################################### +# Model +pretrained_model_name_or_path = 'baichuan-inc/Baichuan2-7B-Base' + +# Data +# 1. Download data from https://kaggle.com/datasets/Cornell-University/arxiv +# 2. Process data by `xtuner preprocess arxiv ${DOWNLOADED_DATA} ./data/arxiv_data.json [optional arguments]` # noqa: E501 +data_path = './data/arxiv_data.json' +prompt_template = PROMPT_TEMPLATE.title +max_length = 2048 +pack_to_max_length = True + +# Scheduler & Optimizer +batch_size = 1 # per_device +accumulative_counts = 16 +dataloader_num_workers = 0 +max_epochs = 3 +optim_type = PagedAdamW32bit +lr = 2e-4 +betas = (0.9, 0.999) +weight_decay = 0 +max_norm = 1 # grad clip + +# Evaluate the generation performance during the training +evaluation_freq = 500 +evaluation_inputs = [ + ('We present InternLM, a multilingual foundational language ' + 'model with 104B parameters. InternLM is pre-trained on a large ' + 'corpora with 1.6T tokens with a multi-phase progressive ' + 'process, and then fine-tuned to align with human preferences. ' + 'We also developed a training system called Uniscale-LLM for ' + 'efficient large language model training. The evaluation on a ' + 'number of benchmarks shows that InternLM achieves ' + 'state-of-the-art performance in multiple aspects, including ' + 'knowledge understanding, reading comprehension, mathematics, ' + 'and coding. With such well-rounded capabilities, InternLM ' + 'achieves outstanding performances on comprehensive exams, ' + 'including MMLU, AGIEval, C-Eval and GAOKAO-Bench, without ' + 'resorting to external tools. On these benchmarks, InternLM ' + 'not only significantly outperforms open-source models, but ' + 'also obtains superior performance compared to ChatGPT. Also, ' + 'InternLM demonstrates excellent capability of understanding ' + 'Chinese language and Chinese culture, which makes it a ' + 'suitable foundation model to support Chinese-oriented language ' + 'applications. This manuscript gives a detailed study of ' + 'our results, with benchmarks and examples across a diverse ' + 'set of knowledge domains and tasks.'), + ('In this work, we develop and release Llama 2, a collection of ' + 'pretrained and fine-tuned large language models (LLMs) ranging ' + 'in scale from 7 billion to 70 billion parameters.\nOur ' + 'fine-tuned LLMs, called LLAMA 2-CHAT, are optimized for ' + 'dialogue use cases. Our models outperform open-source chat ' + 'models on most benchmarks we tested, and based on our human ' + 'evaluations for helpfulness and safety, may be a suitable ' + 'substitute for closedsource models. We provide a detailed ' + 'description of our approach to fine-tuning and safety ' + 'improvements of LLAMA 2-CHAT in order to enable the community ' + 'to build on our work and contribute to the responsible ' + 'development of LLMs.') +] + +####################################################################### +# PART 2 Model & Tokenizer # +####################################################################### +tokenizer = dict( + type=AutoTokenizer.from_pretrained, + pretrained_model_name_or_path=pretrained_model_name_or_path, + trust_remote_code=True, + padding_side='right') + +model = dict( + type=SupervisedFinetune, + llm=dict( + type=AutoModelForCausalLM.from_pretrained, + pretrained_model_name_or_path=pretrained_model_name_or_path, + trust_remote_code=True, + torch_dtype=torch.float16, + quantization_config=dict( + type=BitsAndBytesConfig, + load_in_4bit=True, + load_in_8bit=False, + llm_int8_threshold=6.0, + llm_int8_has_fp16_weight=False, + bnb_4bit_compute_dtype=torch.float16, + bnb_4bit_use_double_quant=True, + bnb_4bit_quant_type='nf4')), + lora=dict( + type=LoraConfig, + r=64, + lora_alpha=16, + lora_dropout=0.1, + bias='none', + task_type='CAUSAL_LM')) + +####################################################################### +# PART 3 Dataset & Dataloader # +####################################################################### +train_dataset = dict( + type=process_hf_dataset, + dataset=dict( + type=load_dataset, path='json', data_files=dict(train=data_path)), + tokenizer=tokenizer, + max_length=max_length, + dataset_map_fn=arxiv_map_fn, + template_map_fn=dict( + type=template_map_fn_factory, template=prompt_template), + remove_unused_columns=True, + shuffle_before_pack=True, + pack_to_max_length=pack_to_max_length) + +train_dataloader = dict( + batch_size=batch_size, + num_workers=dataloader_num_workers, + dataset=train_dataset, + sampler=dict(type=DefaultSampler, shuffle=True), + collate_fn=dict(type=default_collate_fn)) + +####################################################################### +# PART 4 Scheduler & Optimizer # +####################################################################### +# optimizer +optim_wrapper = dict( + type=AmpOptimWrapper, + optimizer=dict( + type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay), + clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False), + accumulative_counts=accumulative_counts, + loss_scale='dynamic', + dtype='float16') + +# learning policy +# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md # noqa: E501 +param_scheduler = dict( + type=CosineAnnealingLR, + eta_min=lr * 0.1, + by_epoch=True, + T_max=max_epochs, + convert_to_iter_based=True) + +# train, val, test setting +train_cfg = dict(by_epoch=True, max_epochs=max_epochs, val_interval=1) + +####################################################################### +# PART 5 Runtime # +####################################################################### +# Log the dialogue periodically during the training process, optional +custom_hooks = [ + dict(type=DatasetInfoHook, tokenizer=tokenizer), + dict( + type=EvaluateChatHook, + tokenizer=tokenizer, + every_n_iters=evaluation_freq, + evaluation_inputs=evaluation_inputs, + instruction=prompt_template.INSTRUCTION_START) +] + +# configure default hooks +default_hooks = dict( + # record the time of every iteration. + timer=dict(type=IterTimerHook), + # print log every 100 iterations. + logger=dict(type=LoggerHook, interval=10), + # enable the parameter scheduler. + param_scheduler=dict(type=ParamSchedulerHook), + # save checkpoint per epoch. + checkpoint=dict(type=CheckpointHook, interval=1), + # set sampler seed in distributed evrionment. + sampler_seed=dict(type=DistSamplerSeedHook), +) + +# configure environment +env_cfg = dict( + # whether to enable cudnn benchmark + cudnn_benchmark=False, + # set multi process parameters + mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0), + # set distributed parameters + dist_cfg=dict(backend='nccl'), +) + +# set visualizer +visualizer = None + +# set log level +log_level = 'INFO' + +# load from which checkpoint +load_from = None + +# whether to resume training from the loaded checkpoint +resume = False + +# Defaults to use random seed and disable `deterministic` +randomness = dict(seed=None, deterministic=False) diff --git a/xtuner/configs/baichuan/baichuan2_7b_base/baichuan2_7b_base_qlora_code_alpaca_e3.py b/xtuner/configs/baichuan/baichuan2_7b_base/baichuan2_7b_base_qlora_code_alpaca_e3.py new file mode 100644 index 000000000..8c1d31968 --- /dev/null +++ b/xtuner/configs/baichuan/baichuan2_7b_base/baichuan2_7b_base_qlora_code_alpaca_e3.py @@ -0,0 +1,184 @@ +# Copyright (c) OpenMMLab. All rights reserved. +import torch +from bitsandbytes.optim import PagedAdamW32bit +from datasets import load_dataset +from mmengine.dataset import DefaultSampler +from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook, + LoggerHook, ParamSchedulerHook) +from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR +from peft import LoraConfig +from transformers import (AutoModelForCausalLM, AutoTokenizer, + BitsAndBytesConfig) + +from xtuner.dataset import process_hf_dataset +from xtuner.dataset.collate_fns import default_collate_fn +from xtuner.dataset.map_fns import code_alpaca_map_fn, template_map_fn_factory +from xtuner.engine import DatasetInfoHook, EvaluateChatHook +from xtuner.model import SupervisedFinetune +from xtuner.utils import PROMPT_TEMPLATE + +####################################################################### +# PART 1 Settings # +####################################################################### +# Model +pretrained_model_name_or_path = 'baichuan-inc/Baichuan2-7B-Base' + +# Data +data_path = 'HuggingFaceH4/CodeAlpaca_20K' +prompt_template = PROMPT_TEMPLATE.coder +max_length = 2048 +pack_to_max_length = True + +# Scheduler & Optimizer +batch_size = 1 # per_device +accumulative_counts = 16 +dataloader_num_workers = 0 +max_epochs = 3 +optim_type = PagedAdamW32bit +lr = 2e-4 +betas = (0.9, 0.999) +weight_decay = 0 +max_norm = 1 # grad clip + +# Evaluate the generation performance during the training +evaluation_freq = 100 +evaluation_inputs = [ + ('写一个Python函数，将十六进制颜色代码（如#0066ee）转换为对应的' + '红、绿、蓝（RGB）三个颜色分量值，并以元组的形式返回。'), + ('Write a Python function that takes a hexadecimal color code ' + '(e.g., #0066ee) as input and converts it into the corresponding ' + 'red, green, and blue (RGB) color component values.') +] + +####################################################################### +# PART 2 Model & Tokenizer # +####################################################################### +tokenizer = dict( + type=AutoTokenizer.from_pretrained, + pretrained_model_name_or_path=pretrained_model_name_or_path, + trust_remote_code=True, + padding_side='right') + +model = dict( + type=SupervisedFinetune, + llm=dict( + type=AutoModelForCausalLM.from_pretrained, + pretrained_model_name_or_path=pretrained_model_name_or_path, + trust_remote_code=True, + torch_dtype=torch.float16, + quantization_config=dict( + type=BitsAndBytesConfig, + load_in_4bit=True, + load_in_8bit=False, + llm_int8_threshold=6.0, + llm_int8_has_fp16_weight=False, + bnb_4bit_compute_dtype=torch.float16, + bnb_4bit_use_double_quant=True, + bnb_4bit_quant_type='nf4')), + lora=dict( + type=LoraConfig, + r=64, + lora_alpha=16, + lora_dropout=0.1, + bias='none', + task_type='CAUSAL_LM')) + +####################################################################### +# PART 3 Dataset & Dataloader # +####################################################################### +train_dataset = dict( + type=process_hf_dataset, + dataset=dict(type=load_dataset, path=data_path), + tokenizer=tokenizer, + max_length=max_length, + dataset_map_fn=code_alpaca_map_fn, + template_map_fn=dict( + type=template_map_fn_factory, template=prompt_template), + remove_unused_columns=True, + shuffle_before_pack=True, + pack_to_max_length=pack_to_max_length) + +train_dataloader = dict( + batch_size=batch_size, + num_workers=dataloader_num_workers, + dataset=train_dataset, + sampler=dict(type=DefaultSampler, shuffle=True), + collate_fn=dict(type=default_collate_fn)) + +####################################################################### +# PART 4 Scheduler & Optimizer # +####################################################################### +# optimizer +optim_wrapper = dict( + type=AmpOptimWrapper, + optimizer=dict( + type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay), + clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False), + accumulative_counts=accumulative_counts, + loss_scale='dynamic', + dtype='float16') + +# learning policy +# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md # noqa: E501 +param_scheduler = dict( + type=CosineAnnealingLR, + eta_min=lr * 0.1, + by_epoch=True, + T_max=max_epochs, + convert_to_iter_based=True) + +# train, val, test setting +train_cfg = dict(by_epoch=True, max_epochs=max_epochs, val_interval=1) + +####################################################################### +# PART 5 Runtime # +####################################################################### +# Log the dialogue periodically during the training process, optional +custom_hooks = [ + dict(type=DatasetInfoHook, tokenizer=tokenizer), + dict( + type=EvaluateChatHook, + tokenizer=tokenizer, + every_n_iters=evaluation_freq, + evaluation_inputs=evaluation_inputs, + instruction=prompt_template.INSTRUCTION_START) +] + +# configure default hooks +default_hooks = dict( + # record the time of every iteration. + timer=dict(type=IterTimerHook), + # print log every 100 iterations. + logger=dict(type=LoggerHook, interval=10), + # enable the parameter scheduler. + param_scheduler=dict(type=ParamSchedulerHook), + # save checkpoint per epoch. + checkpoint=dict(type=CheckpointHook, interval=1), + # set sampler seed in distributed evrionment. + sampler_seed=dict(type=DistSamplerSeedHook), +) + +# configure environment +env_cfg = dict( + # whether to enable cudnn benchmark + cudnn_benchmark=False, + # set multi process parameters + mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0), + # set distributed parameters + dist_cfg=dict(backend='nccl'), +) + +# set visualizer +visualizer = None + +# set log level +log_level = 'INFO' + +# load from which checkpoint +load_from = None + +# whether to resume training from the loaded checkpoint +resume = False + +# Defaults to use random seed and disable `deterministic` +randomness = dict(seed=None, deterministic=False) diff --git a/xtuner/configs/baichuan/baichuan2_7b_base/baichuan2_7b_base_qlora_colorist_e5.py b/xtuner/configs/baichuan/baichuan2_7b_base/baichuan2_7b_base_qlora_colorist_e5.py new file mode 100644 index 000000000..11075ad8b --- /dev/null +++ b/xtuner/configs/baichuan/baichuan2_7b_base/baichuan2_7b_base_qlora_colorist_e5.py @@ -0,0 +1,180 @@ +# Copyright (c) OpenMMLab. All rights reserved. +import torch +from bitsandbytes.optim import PagedAdamW32bit +from datasets import load_dataset +from mmengine.dataset import DefaultSampler +from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook, + LoggerHook, ParamSchedulerHook) +from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR +from peft import LoraConfig +from transformers import (AutoModelForCausalLM, AutoTokenizer, + BitsAndBytesConfig) + +from xtuner.dataset import process_hf_dataset +from xtuner.dataset.collate_fns import default_collate_fn +from xtuner.dataset.map_fns import colors_map_fn, template_map_fn_factory +from xtuner.engine import DatasetInfoHook, EvaluateChatHook +from xtuner.model import SupervisedFinetune +from xtuner.utils import PROMPT_TEMPLATE + +####################################################################### +# PART 1 Settings # +####################################################################### +# Model +pretrained_model_name_or_path = 'baichuan-inc/Baichuan2-7B-Base' + +# Data +data_path = 'burkelibbey/colors' +prompt_template = PROMPT_TEMPLATE.colorist +max_length = 2048 +pack_to_max_length = True + +# Scheduler & Optimizer +batch_size = 1 # per_device +accumulative_counts = 16 +dataloader_num_workers = 0 +max_epochs = 5 +optim_type = PagedAdamW32bit +lr = 2e-4 +betas = (0.9, 0.999) +weight_decay = 0 +max_norm = 1 # grad clip + +# Evaluate the generation performance during the training +evaluation_freq = 200 +evaluation_inputs = [ + '请给我一个像天空一样清澈透明的蓝色。', 'Please give me a clear blue like the sky.' +] + +####################################################################### +# PART 2 Model & Tokenizer # +####################################################################### +tokenizer = dict( + type=AutoTokenizer.from_pretrained, + pretrained_model_name_or_path=pretrained_model_name_or_path, + trust_remote_code=True, + padding_side='right') + +model = dict( + type=SupervisedFinetune, + llm=dict( + type=AutoModelForCausalLM.from_pretrained, + pretrained_model_name_or_path=pretrained_model_name_or_path, + trust_remote_code=True, + torch_dtype=torch.float16, + quantization_config=dict( + type=BitsAndBytesConfig, + load_in_4bit=True, + load_in_8bit=False, + llm_int8_threshold=6.0, + llm_int8_has_fp16_weight=False, + bnb_4bit_compute_dtype=torch.float16, + bnb_4bit_use_double_quant=True, + bnb_4bit_quant_type='nf4')), + lora=dict( + type=LoraConfig, + r=64, + lora_alpha=16, + lora_dropout=0.1, + bias='none', + task_type='CAUSAL_LM')) + +####################################################################### +# PART 3 Dataset & Dataloader # +####################################################################### +train_dataset = dict( + type=process_hf_dataset, + dataset=dict(type=load_dataset, path=data_path), + tokenizer=tokenizer, + max_length=max_length, + dataset_map_fn=colors_map_fn, + template_map_fn=dict( + type=template_map_fn_factory, template=prompt_template), + remove_unused_columns=True, + shuffle_before_pack=True, + pack_to_max_length=pack_to_max_length) + +train_dataloader = dict( + batch_size=batch_size, + num_workers=dataloader_num_workers, + dataset=train_dataset, + sampler=dict(type=DefaultSampler, shuffle=True), + collate_fn=dict(type=default_collate_fn)) + +####################################################################### +# PART 4 Scheduler & Optimizer # +####################################################################### +# optimizer +optim_wrapper = dict( + type=AmpOptimWrapper, + optimizer=dict( + type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay), + clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False), + accumulative_counts=accumulative_counts, + loss_scale='dynamic', + dtype='float16') + +# learning policy +# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md # noqa: E501 +param_scheduler = dict( + type=CosineAnnealingLR, + eta_min=lr * 0.1, + by_epoch=True, + T_max=max_epochs, + convert_to_iter_based=True) + +# train, val, test setting +train_cfg = dict(by_epoch=True, max_epochs=max_epochs, val_interval=1) + +####################################################################### +# PART 5 Runtime # +####################################################################### +# Log the dialogue periodically during the training process, optional +custom_hooks = [ + dict(type=DatasetInfoHook, tokenizer=tokenizer), + dict( + type=EvaluateChatHook, + tokenizer=tokenizer, + every_n_iters=evaluation_freq, + evaluation_inputs=evaluation_inputs, + instruction=prompt_template.INSTRUCTION_START) +] + +# configure default hooks +default_hooks = dict( + # record the time of every iteration. + timer=dict(type=IterTimerHook), + # print log every 100 iterations. + logger=dict(type=LoggerHook, interval=10), + # enable the parameter scheduler. + param_scheduler=dict(type=ParamSchedulerHook), + # save checkpoint per epoch. + checkpoint=dict(type=CheckpointHook, interval=1), + # set sampler seed in distributed evrionment. + sampler_seed=dict(type=DistSamplerSeedHook), +) + +# configure environment +env_cfg = dict( + # whether to enable cudnn benchmark + cudnn_benchmark=False, + # set multi process parameters + mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0), + # set distributed parameters + dist_cfg=dict(backend='nccl'), +) + +# set visualizer +visualizer = None + +# set log level +log_level = 'INFO' + +# load from which checkpoint +load_from = None + +# whether to resume training from the loaded checkpoint +resume = False + +# Defaults to use random seed and disable `deterministic` +randomness = dict(seed=None, deterministic=False) diff --git a/xtuner/configs/baichuan/baichuan2_7b_base/baichuan2_7b_base_qlora_lawyer_e3.py b/xtuner/configs/baichuan/baichuan2_7b_base/baichuan2_7b_base_qlora_lawyer_e3.py new file mode 100644 index 000000000..589747804 --- /dev/null +++ b/xtuner/configs/baichuan/baichuan2_7b_base/baichuan2_7b_base_qlora_lawyer_e3.py @@ -0,0 +1,206 @@ +# Copyright (c) OpenMMLab. All rights reserved. +import torch +from bitsandbytes.optim import PagedAdamW32bit +from datasets import load_dataset +from mmengine.dataset import DefaultSampler +from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook, + LoggerHook, ParamSchedulerHook) +from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR +from peft import LoraConfig +from transformers import (AutoModelForCausalLM, AutoTokenizer, + BitsAndBytesConfig) + +from xtuner.dataset import ConcatDataset, process_hf_dataset +from xtuner.dataset.collate_fns import default_collate_fn +from xtuner.dataset.map_fns import (crime_kg_assitant_map_fn, + law_reference_map_fn, + template_map_fn_factory) +from xtuner.engine import DatasetInfoHook, EvaluateChatHook +from xtuner.model import SupervisedFinetune +from xtuner.utils import PROMPT_TEMPLATE + +####################################################################### +# PART 1 Settings # +####################################################################### +# Model +pretrained_model_name_or_path = 'baichuan-inc/Baichuan2-7B-Base' + +# Data +# download data from https://github.com/LiuHC0428/LAW-GPT +crime_kg_assitant_path = './data/CrimeKgAssitant清洗后_52k.json' +law_reference_data_path = './data/训练数据_带法律依据_92k.json' +prompt_template = PROMPT_TEMPLATE.lawyer +max_length = 2048 +pack_to_max_length = True + +# Scheduler & Optimizer +batch_size = 1 # per_device +accumulative_counts = 16 +dataloader_num_workers = 0 +max_epochs = 3 +optim_type = PagedAdamW32bit +lr = 2e-4 +betas = (0.9, 0.999) +weight_decay = 0 +max_norm = 1 # grad clip + +# Evaluate the generation performance during the training +evaluation_freq = 500 +evaluation_inputs = ['请问离婚需要准备什么材料？', '销售鳄鱼皮包违法吗？'] + +####################################################################### +# PART 2 Model & Tokenizer # +####################################################################### +tokenizer = dict( + type=AutoTokenizer.from_pretrained, + pretrained_model_name_or_path=pretrained_model_name_or_path, + trust_remote_code=True, + padding_side='right') + +model = dict( + type=SupervisedFinetune, + llm=dict( + type=AutoModelForCausalLM.from_pretrained, + pretrained_model_name_or_path=pretrained_model_name_or_path, + trust_remote_code=True, + torch_dtype=torch.float16, + quantization_config=dict( + type=BitsAndBytesConfig, + load_in_4bit=True, + load_in_8bit=False, + llm_int8_threshold=6.0, + llm_int8_has_fp16_weight=False, + bnb_4bit_compute_dtype=torch.float16, + bnb_4bit_use_double_quant=True, + bnb_4bit_quant_type='nf4')), + lora=dict( + type=LoraConfig, + r=64, + lora_alpha=16, + lora_dropout=0.1, + bias='none', + task_type='CAUSAL_LM')) + +####################################################################### +# PART 3 Dataset & Dataloader # +####################################################################### +crime_kg_assitant = dict( + type=process_hf_dataset, + dataset=dict( + type=load_dataset, + path='json', + data_files=dict(train=crime_kg_assitant_path)), + tokenizer=tokenizer, + max_length=max_length, + dataset_map_fn=crime_kg_assitant_map_fn, + template_map_fn=dict( + type=template_map_fn_factory, template=prompt_template), + remove_unused_columns=True, + shuffle_before_pack=True, + pack_to_max_length=pack_to_max_length) + +law_reference_data = dict( + type=process_hf_dataset, + dataset=dict( + type=load_dataset, + path='json', + data_files=dict(train=law_reference_data_path)), + tokenizer=tokenizer, + max_length=max_length, + dataset_map_fn=law_reference_map_fn, + template_map_fn=dict( + type=template_map_fn_factory, template=prompt_template), + remove_unused_columns=True, + shuffle_before_pack=True, + pack_to_max_length=pack_to_max_length) + +train_dataset = dict( + type=ConcatDataset, + datasets_cfg=dict( + crime_kg_assitant=crime_kg_assitant, + law_reference_data=law_reference_data)) + +train_dataloader = dict( + batch_size=batch_size, + num_workers=dataloader_num_workers, + dataset=train_dataset, + sampler=dict(type=DefaultSampler, shuffle=True), + collate_fn=dict(type=default_collate_fn)) + +####################################################################### +# PART 4 Scheduler & Optimizer # +####################################################################### +# optimizer +optim_wrapper = dict( + type=AmpOptimWrapper, + optimizer=dict( + type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay), + clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False), + accumulative_counts=accumulative_counts, + loss_scale='dynamic', + dtype='float16') + +# learning policy +# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md # noqa: E501 +param_scheduler = dict( + type=CosineAnnealingLR, + eta_min=lr * 0.1, + by_epoch=True, + T_max=max_epochs, + convert_to_iter_based=True) + +# train, val, test setting +train_cfg = dict(by_epoch=True, max_epochs=max_epochs, val_interval=1) + +####################################################################### +# PART 5 Runtime # +####################################################################### +# Log the dialogue periodically during the training process, optional +custom_hooks = [ + dict(type=DatasetInfoHook, tokenizer=tokenizer), + dict( + type=EvaluateChatHook, + tokenizer=tokenizer, + every_n_iters=evaluation_freq, + evaluation_inputs=evaluation_inputs, + instruction=prompt_template.INSTRUCTION_START) +] + +# configure default hooks +default_hooks = dict( + # record the time of every iteration. + timer=dict(type=IterTimerHook), + # print log every 100 iterations. + logger=dict(type=LoggerHook, interval=10), + # enable the parameter scheduler. + param_scheduler=dict(type=ParamSchedulerHook), + # save checkpoint per epoch. + checkpoint=dict(type=CheckpointHook, interval=1), + # set sampler seed in distributed evrionment. + sampler_seed=dict(type=DistSamplerSeedHook), +) + +# configure environment +env_cfg = dict( + # whether to enable cudnn benchmark + cudnn_benchmark=False, + # set multi process parameters + mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0), + # set distributed parameters + dist_cfg=dict(backend='nccl'), +) + +# set visualizer +visualizer = None + +# set log level +log_level = 'INFO' + +# load from which checkpoint +load_from = None + +# whether to resume training from the loaded checkpoint +resume = False + +# Defaults to use random seed and disable `deterministic` +randomness = dict(seed=None, deterministic=False) diff --git a/xtuner/configs/baichuan/baichuan2_7b_base/baichuan2_7b_base_qlora_oasst1_512_e3.py b/xtuner/configs/baichuan/baichuan2_7b_base/baichuan2_7b_base_qlora_oasst1_512_e3.py new file mode 100644 index 000000000..2efef99c4 --- /dev/null +++ b/xtuner/configs/baichuan/baichuan2_7b_base/baichuan2_7b_base_qlora_oasst1_512_e3.py @@ -0,0 +1,180 @@ +# Copyright (c) OpenMMLab. All rights reserved. +import torch +from bitsandbytes.optim import PagedAdamW32bit +from datasets import load_dataset +from mmengine.dataset import DefaultSampler +from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook, + LoggerHook, ParamSchedulerHook) +from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR +from peft import LoraConfig +from transformers import (AutoModelForCausalLM, AutoTokenizer, + BitsAndBytesConfig) + +from xtuner.dataset import process_hf_dataset +from xtuner.dataset.collate_fns import default_collate_fn +from xtuner.dataset.map_fns import oasst1_map_fn, template_map_fn_factory +from xtuner.engine import DatasetInfoHook, EvaluateChatHook +from xtuner.model import SupervisedFinetune +from xtuner.utils import PROMPT_TEMPLATE + +####################################################################### +# PART 1 Settings # +####################################################################### +# Model +pretrained_model_name_or_path = 'baichuan-inc/Baichuan2-7B-Base' + +# Data +data_path = 'timdettmers/openassistant-guanaco' +prompt_template = PROMPT_TEMPLATE.openassistant +max_length = 512 +pack_to_max_length = False + +# Scheduler & Optimizer +batch_size = 1 # per_device +accumulative_counts = 16 +dataloader_num_workers = 0 +max_epochs = 3 +optim_type = PagedAdamW32bit +lr = 2e-4 +betas = (0.9, 0.999) +weight_decay = 0 +max_norm = 1 # grad clip + +# Evaluate the generation performance during the training +evaluation_freq = 500 +evaluation_inputs = [ + '请给我介绍五个上海的景点', 'Please tell me five scenic spots in Shanghai' +] + +####################################################################### +# PART 2 Model & Tokenizer # +####################################################################### +tokenizer = dict( + type=AutoTokenizer.from_pretrained, + pretrained_model_name_or_path=pretrained_model_name_or_path, + trust_remote_code=True, + padding_side='right') + +model = dict( + type=SupervisedFinetune, + llm=dict( + type=AutoModelForCausalLM.from_pretrained, + pretrained_model_name_or_path=pretrained_model_name_or_path, + trust_remote_code=True, + torch_dtype=torch.float16, + quantization_config=dict( + type=BitsAndBytesConfig, + load_in_4bit=True, + load_in_8bit=False, + llm_int8_threshold=6.0, + llm_int8_has_fp16_weight=False, + bnb_4bit_compute_dtype=torch.float16, + bnb_4bit_use_double_quant=True, + bnb_4bit_quant_type='nf4')), + lora=dict( + type=LoraConfig, + r=64, + lora_alpha=16, + lora_dropout=0.1, + bias='none', + task_type='CAUSAL_LM')) + +####################################################################### +# PART 3 Dataset & Dataloader # +####################################################################### +train_dataset = dict( + type=process_hf_dataset, + dataset=dict(type=load_dataset, path=data_path), + tokenizer=tokenizer, + max_length=max_length, + dataset_map_fn=oasst1_map_fn, + template_map_fn=dict( + type=template_map_fn_factory, template=prompt_template), + remove_unused_columns=True, + shuffle_before_pack=True, + pack_to_max_length=pack_to_max_length) + +train_dataloader = dict( + batch_size=batch_size, + num_workers=dataloader_num_workers, + dataset=train_dataset, + sampler=dict(type=DefaultSampler, shuffle=True), + collate_fn=dict(type=default_collate_fn)) + +####################################################################### +# PART 4 Scheduler & Optimizer # +####################################################################### +# optimizer +optim_wrapper = dict( + type=AmpOptimWrapper, + optimizer=dict( + type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay), + clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False), + accumulative_counts=accumulative_counts, + loss_scale='dynamic', + dtype='float16') + +# learning policy +# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md # noqa: E501 +param_scheduler = dict( + type=CosineAnnealingLR, + eta_min=lr * 0.1, + by_epoch=True, + T_max=max_epochs, + convert_to_iter_based=True) + +# train, val, test setting +train_cfg = dict(by_epoch=True, max_epochs=max_epochs, val_interval=1) + +####################################################################### +# PART 5 Runtime # +####################################################################### +# Log the dialogue periodically during the training process, optional +custom_hooks = [ + dict(type=DatasetInfoHook, tokenizer=tokenizer), + dict( + type=EvaluateChatHook, + tokenizer=tokenizer, + every_n_iters=evaluation_freq, + evaluation_inputs=evaluation_inputs, + instruction=prompt_template.INSTRUCTION_START) +] + +# configure default hooks +default_hooks = dict( + # record the time of every iteration. + timer=dict(type=IterTimerHook), + # print log every 100 iterations. + logger=dict(type=LoggerHook, interval=10), + # enable the parameter scheduler. + param_scheduler=dict(type=ParamSchedulerHook), + # save checkpoint per epoch. + checkpoint=dict(type=CheckpointHook, interval=1), + # set sampler seed in distributed evrionment. + sampler_seed=dict(type=DistSamplerSeedHook), +) + +# configure environment +env_cfg = dict( + # whether to enable cudnn benchmark + cudnn_benchmark=False, + # set multi process parameters + mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0), + # set distributed parameters + dist_cfg=dict(backend='nccl'), +) + +# set visualizer +visualizer = None + +# set log level +log_level = 'INFO' + +# load from which checkpoint +load_from = None + +# whether to resume training from the loaded checkpoint +resume = False + +# Defaults to use random seed and disable `deterministic` +randomness = dict(seed=None, deterministic=False) diff --git a/xtuner/configs/baichuan/baichuan2_7b_base/baichuan2_7b_base_qlora_oasst1_e3.py b/xtuner/configs/baichuan/baichuan2_7b_base/baichuan2_7b_base_qlora_oasst1_e3.py new file mode 100644 index 000000000..f1458f2b6 --- /dev/null +++ b/xtuner/configs/baichuan/baichuan2_7b_base/baichuan2_7b_base_qlora_oasst1_e3.py @@ -0,0 +1,180 @@ +# Copyright (c) OpenMMLab. All rights reserved. +import torch +from bitsandbytes.optim import PagedAdamW32bit +from datasets import load_dataset +from mmengine.dataset import DefaultSampler +from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook, + LoggerHook, ParamSchedulerHook) +from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR +from peft import LoraConfig +from transformers import (AutoModelForCausalLM, AutoTokenizer, + BitsAndBytesConfig) + +from xtuner.dataset import process_hf_dataset +from xtuner.dataset.collate_fns import default_collate_fn +from xtuner.dataset.map_fns import oasst1_map_fn, template_map_fn_factory +from xtuner.engine import DatasetInfoHook, EvaluateChatHook +from xtuner.model import SupervisedFinetune +from xtuner.utils import PROMPT_TEMPLATE + +####################################################################### +# PART 1 Settings # +####################################################################### +# Model +pretrained_model_name_or_path = 'baichuan-inc/Baichuan2-7B-Base' + +# Data +data_path = 'timdettmers/openassistant-guanaco' +prompt_template = PROMPT_TEMPLATE.openassistant +max_length = 2048 +pack_to_max_length = True + +# Scheduler & Optimizer +batch_size = 1 # per_device +accumulative_counts = 16 +dataloader_num_workers = 0 +max_epochs = 3 +optim_type = PagedAdamW32bit +lr = 2e-4 +betas = (0.9, 0.999) +weight_decay = 0 +max_norm = 1 # grad clip + +# Evaluate the generation performance during the training +evaluation_freq = 500 +evaluation_inputs = [ + '请给我介绍五个上海的景点', 'Please tell me five scenic spots in Shanghai' +] + +####################################################################### +# PART 2 Model & Tokenizer # +####################################################################### +tokenizer = dict( + type=AutoTokenizer.from_pretrained, + pretrained_model_name_or_path=pretrained_model_name_or_path, + trust_remote_code=True, + padding_side='right') + +model = dict( + type=SupervisedFinetune, + llm=dict( + type=AutoModelForCausalLM.from_pretrained, + pretrained_model_name_or_path=pretrained_model_name_or_path, + trust_remote_code=True, + torch_dtype=torch.float16, + quantization_config=dict( + type=BitsAndBytesConfig, + load_in_4bit=True, + load_in_8bit=False, + llm_int8_threshold=6.0, + llm_int8_has_fp16_weight=False, + bnb_4bit_compute_dtype=torch.float16, + bnb_4bit_use_double_quant=True, + bnb_4bit_quant_type='nf4')), + lora=dict( + type=LoraConfig, + r=64, + lora_alpha=16, + lora_dropout=0.1, + bias='none', + task_type='CAUSAL_LM')) + +####################################################################### +# PART 3 Dataset & Dataloader # +####################################################################### +train_dataset = dict( + type=process_hf_dataset, + dataset=dict(type=load_dataset, path=data_path), + tokenizer=tokenizer, + max_length=max_length, + dataset_map_fn=oasst1_map_fn, + template_map_fn=dict( + type=template_map_fn_factory, template=prompt_template), + remove_unused_columns=True, + shuffle_before_pack=True, + pack_to_max_length=pack_to_max_length) + +train_dataloader = dict( + batch_size=batch_size, + num_workers=dataloader_num_workers, + dataset=train_dataset, + sampler=dict(type=DefaultSampler, shuffle=True), + collate_fn=dict(type=default_collate_fn)) + +####################################################################### +# PART 4 Scheduler & Optimizer # +####################################################################### +# optimizer +optim_wrapper = dict( + type=AmpOptimWrapper, + optimizer=dict( + type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay), + clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False), + accumulative_counts=accumulative_counts, + loss_scale='dynamic', + dtype='float16') + +# learning policy +# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md # noqa: E501 +param_scheduler = dict( + type=CosineAnnealingLR, + eta_min=lr * 0.1, + by_epoch=True, + T_max=max_epochs, + convert_to_iter_based=True) + +# train, val, test setting +train_cfg = dict(by_epoch=True, max_epochs=max_epochs, val_interval=1) + +####################################################################### +# PART 5 Runtime # +####################################################################### +# Log the dialogue periodically during the training process, optional +custom_hooks = [ + dict(type=DatasetInfoHook, tokenizer=tokenizer), + dict( + type=EvaluateChatHook, + tokenizer=tokenizer, + every_n_iters=evaluation_freq, + evaluation_inputs=evaluation_inputs, + instruction=prompt_template.INSTRUCTION_START) +] + +# configure default hooks +default_hooks = dict( + # record the time of every iteration. + timer=dict(type=IterTimerHook), + # print log every 100 iterations. + logger=dict(type=LoggerHook, interval=10), + # enable the parameter scheduler. + param_scheduler=dict(type=ParamSchedulerHook), + # save checkpoint per epoch. + checkpoint=dict(type=CheckpointHook, interval=1), + # set sampler seed in distributed evrionment. + sampler_seed=dict(type=DistSamplerSeedHook), +) + +# configure environment +env_cfg = dict( + # whether to enable cudnn benchmark + cudnn_benchmark=False, + # set multi process parameters + mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0), + # set distributed parameters + dist_cfg=dict(backend='nccl'), +) + +# set visualizer +visualizer = None + +# set log level +log_level = 'INFO' + +# load from which checkpoint +load_from = None + +# whether to resume training from the loaded checkpoint +resume = False + +# Defaults to use random seed and disable `deterministic` +randomness = dict(seed=None, deterministic=False) diff --git a/xtuner/configs/baichuan/baichuan2_7b_base/baichuan2_7b_base_qlora_open_platypus_e3.py b/xtuner/configs/baichuan/baichuan2_7b_base/baichuan2_7b_base_qlora_open_platypus_e3.py new file mode 100644 index 000000000..7a9471bbd --- /dev/null +++ b/xtuner/configs/baichuan/baichuan2_7b_base/baichuan2_7b_base_qlora_open_platypus_e3.py @@ -0,0 +1,180 @@ +# Copyright (c) OpenMMLab. All rights reserved. +import torch +from bitsandbytes.optim import PagedAdamW32bit +from datasets import load_dataset +from mmengine.dataset import DefaultSampler +from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook, + LoggerHook, ParamSchedulerHook) +from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR +from peft import LoraConfig +from transformers import (AutoModelForCausalLM, AutoTokenizer, + BitsAndBytesConfig) + +from xtuner.dataset import process_hf_dataset +from xtuner.dataset.collate_fns import default_collate_fn +from xtuner.dataset.map_fns import alpaca_map_fn, template_map_fn_factory +from xtuner.engine import DatasetInfoHook, EvaluateChatHook +from xtuner.model import SupervisedFinetune +from xtuner.utils import PROMPT_TEMPLATE + +####################################################################### +# PART 1 Settings # +####################################################################### +# Model +pretrained_model_name_or_path = 'baichuan-inc/Baichuan2-7B-Base' + +# Data +data_path = 'garage-bAInd/Open-Platypus' +prompt_template = PROMPT_TEMPLATE.alpaca +max_length = 2048 +pack_to_max_length = True + +# Scheduler & Optimizer +batch_size = 1 # per_device +accumulative_counts = 16 +dataloader_num_workers = 0 +max_epochs = 3 +optim_type = PagedAdamW32bit +lr = 2e-4 +betas = (0.9, 0.999) +weight_decay = 0 +max_norm = 1 # grad clip + +# Evaluate the generation performance during the training +evaluation_freq = 500 +evaluation_inputs = [ + '请给我介绍五个上海的景点', 'Please tell me five scenic spots in Shanghai' +] + +####################################################################### +# PART 2 Model & Tokenizer # +####################################################################### +tokenizer = dict( + type=AutoTokenizer.from_pretrained, + pretrained_model_name_or_path=pretrained_model_name_or_path, + trust_remote_code=True, + padding_side='right') + +model = dict( + type=SupervisedFinetune, + llm=dict( + type=AutoModelForCausalLM.from_pretrained, + pretrained_model_name_or_path=pretrained_model_name_or_path, + trust_remote_code=True, + torch_dtype=torch.float16, + quantization_config=dict( + type=BitsAndBytesConfig, + load_in_4bit=True, + load_in_8bit=False, + llm_int8_threshold=6.0, + llm_int8_has_fp16_weight=False, + bnb_4bit_compute_dtype=torch.float16, + bnb_4bit_use_double_quant=True, + bnb_4bit_quant_type='nf4')), + lora=dict( + type=LoraConfig, + r=64, + lora_alpha=16, + lora_dropout=0.1, + bias='none', + task_type='CAUSAL_LM')) + +####################################################################### +# PART 3 Dataset & Dataloader # +####################################################################### +train_dataset = dict( + type=process_hf_dataset, + dataset=dict(type=load_dataset, path=data_path), + tokenizer=tokenizer, + max_length=max_length, + dataset_map_fn=alpaca_map_fn, + template_map_fn=dict( + type=template_map_fn_factory, template=prompt_template), + remove_unused_columns=True, + shuffle_before_pack=True, + pack_to_max_length=pack_to_max_length) + +train_dataloader = dict( + batch_size=batch_size, + num_workers=dataloader_num_workers, + dataset=train_dataset, + sampler=dict(type=DefaultSampler, shuffle=True), + collate_fn=dict(type=default_collate_fn)) + +####################################################################### +# PART 4 Scheduler & Optimizer # +####################################################################### +# optimizer +optim_wrapper = dict( + type=AmpOptimWrapper, + optimizer=dict( + type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay), + clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False), + accumulative_counts=accumulative_counts, + loss_scale='dynamic', + dtype='float16') + +# learning policy +# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md # noqa: E501 +param_scheduler = dict( + type=CosineAnnealingLR, + eta_min=lr * 0.1, + by_epoch=True, + T_max=max_epochs, + convert_to_iter_based=True) + +# train, val, test setting +train_cfg = dict(by_epoch=True, max_epochs=max_epochs, val_interval=1) + +####################################################################### +# PART 5 Runtime # +####################################################################### +# Log the dialogue periodically during the training process, optional +custom_hooks = [ + dict(type=DatasetInfoHook, tokenizer=tokenizer), + dict( + type=EvaluateChatHook, + tokenizer=tokenizer, + every_n_iters=evaluation_freq, + evaluation_inputs=evaluation_inputs, + instruction=prompt_template.INSTRUCTION_START) +] + +# configure default hooks +default_hooks = dict( + # record the time of every iteration. + timer=dict(type=IterTimerHook), + # print log every 100 iterations. + logger=dict(type=LoggerHook, interval=10), + # enable the parameter scheduler. + param_scheduler=dict(type=ParamSchedulerHook), + # save checkpoint per epoch. + checkpoint=dict(type=CheckpointHook, interval=1), + # set sampler seed in distributed evrionment. + sampler_seed=dict(type=DistSamplerSeedHook), +) + +# configure environment +env_cfg = dict( + # whether to enable cudnn benchmark + cudnn_benchmark=False, + # set multi process parameters + mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0), + # set distributed parameters + dist_cfg=dict(backend='nccl'), +) + +# set visualizer +visualizer = None + +# set log level +log_level = 'INFO' + +# load from which checkpoint +load_from = None + +# whether to resume training from the loaded checkpoint +resume = False + +# Defaults to use random seed and disable `deterministic` +randomness = dict(seed=None, deterministic=False) diff --git a/xtuner/configs/baichuan/baichuan2_7b_base/baichuan2_7b_base_qlora_sql_e3.py b/xtuner/configs/baichuan/baichuan2_7b_base/baichuan2_7b_base_qlora_sql_e3.py new file mode 100644 index 000000000..24862b83d --- /dev/null +++ b/xtuner/configs/baichuan/baichuan2_7b_base/baichuan2_7b_base_qlora_sql_e3.py @@ -0,0 +1,184 @@ +# Copyright (c) OpenMMLab. All rights reserved. +import torch +from bitsandbytes.optim import PagedAdamW32bit +from datasets import load_dataset +from mmengine.dataset import DefaultSampler +from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook, + LoggerHook, ParamSchedulerHook) +from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR +from peft import LoraConfig +from transformers import (AutoModelForCausalLM, AutoTokenizer, + BitsAndBytesConfig) + +from xtuner.dataset import process_hf_dataset +from xtuner.dataset.collate_fns import default_collate_fn +from xtuner.dataset.map_fns import sql_map_fn, template_map_fn_factory +from xtuner.engine import DatasetInfoHook, EvaluateChatHook +from xtuner.model import SupervisedFinetune +from xtuner.utils import PROMPT_TEMPLATE + +####################################################################### +# PART 1 Settings # +####################################################################### +# Model +pretrained_model_name_or_path = 'baichuan-inc/Baichuan2-7B-Base' + +# Data +data_path = 'b-mc2/sql-create-context' +prompt_template = PROMPT_TEMPLATE.sql +max_length = 2048 +pack_to_max_length = True + +# Scheduler & Optimizer +batch_size = 1 # per_device +accumulative_counts = 16 +dataloader_num_workers = 0 +max_epochs = 3 +optim_type = PagedAdamW32bit +lr = 2e-4 +betas = (0.9, 0.999) +weight_decay = 0 +max_norm = 1 # grad clip + +# Evaluate the generation performance during the training +evaluation_freq = 500 +evaluation_inputs = [ + ('CREATE TABLE station (name VARCHAR, lat VARCHAR, city VARCHAR)\n' + 'Find the name, latitude, and city of stations with latitude ' + 'above 50.'), + ('CREATE TABLE weather (zip_code VARCHAR, mean_visibility_miles ' + 'INTEGER)\n找到mean_visibility_miles最大的zip_code。') +] + +####################################################################### +# PART 2 Model & Tokenizer # +####################################################################### +tokenizer = dict( + type=AutoTokenizer.from_pretrained, + pretrained_model_name_or_path=pretrained_model_name_or_path, + trust_remote_code=True, + padding_side='right') + +model = dict( + type=SupervisedFinetune, + llm=dict( + type=AutoModelForCausalLM.from_pretrained, + pretrained_model_name_or_path=pretrained_model_name_or_path, + trust_remote_code=True, + torch_dtype=torch.float16, + quantization_config=dict( + type=BitsAndBytesConfig, + load_in_4bit=True, + load_in_8bit=False, + llm_int8_threshold=6.0, + llm_int8_has_fp16_weight=False, + bnb_4bit_compute_dtype=torch.float16, + bnb_4bit_use_double_quant=True, + bnb_4bit_quant_type='nf4')), + lora=dict( + type=LoraConfig, + r=64, + lora_alpha=16, + lora_dropout=0.1, + bias='none', + task_type='CAUSAL_LM')) + +####################################################################### +# PART 3 Dataset & Dataloader # +####################################################################### +train_dataset = dict( + type=process_hf_dataset, + dataset=dict(type=load_dataset, path=data_path), + tokenizer=tokenizer, + max_length=max_length, + dataset_map_fn=sql_map_fn, + template_map_fn=dict( + type=template_map_fn_factory, template=prompt_template), + remove_unused_columns=True, + shuffle_before_pack=True, + pack_to_max_length=pack_to_max_length) + +train_dataloader = dict( + batch_size=batch_size, + num_workers=dataloader_num_workers, + dataset=train_dataset, + sampler=dict(type=DefaultSampler, shuffle=True), + collate_fn=dict(type=default_collate_fn)) + +####################################################################### +# PART 4 Scheduler & Optimizer # +####################################################################### +# optimizer +optim_wrapper = dict( + type=AmpOptimWrapper, + optimizer=dict( + type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay), + clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False), + accumulative_counts=accumulative_counts, + loss_scale='dynamic', + dtype='float16') + +# learning policy +# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md # noqa: E501 +param_scheduler = dict( + type=CosineAnnealingLR, + eta_min=lr * 0.1, + by_epoch=True, + T_max=max_epochs, + convert_to_iter_based=True) + +# train, val, test setting +train_cfg = dict(by_epoch=True, max_epochs=max_epochs, val_interval=1) + +####################################################################### +# PART 5 Runtime # +####################################################################### +# Log the dialogue periodically during the training process, optional +custom_hooks = [ + dict(type=DatasetInfoHook, tokenizer=tokenizer), + dict( + type=EvaluateChatHook, + tokenizer=tokenizer, + every_n_iters=evaluation_freq, + evaluation_inputs=evaluation_inputs, + instruction=prompt_template.INSTRUCTION_START) +] + +# configure default hooks +default_hooks = dict( + # record the time of every iteration. + timer=dict(type=IterTimerHook), + # print log every 100 iterations. + logger=dict(type=LoggerHook, interval=10), + # enable the parameter scheduler. + param_scheduler=dict(type=ParamSchedulerHook), + # save checkpoint per epoch. + checkpoint=dict(type=CheckpointHook, interval=1), + # set sampler seed in distributed evrionment. + sampler_seed=dict(type=DistSamplerSeedHook), +) + +# configure environment +env_cfg = dict( + # whether to enable cudnn benchmark + cudnn_benchmark=False, + # set multi process parameters + mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0), + # set distributed parameters + dist_cfg=dict(backend='nccl'), +) + +# set visualizer +visualizer = None + +# set log level +log_level = 'INFO' + +# load from which checkpoint +load_from = None + +# whether to resume training from the loaded checkpoint +resume = False + +# Defaults to use random seed and disable `deterministic` +randomness = dict(seed=None, deterministic=False) diff --git a/xtuner/model/sft.py b/xtuner/model/sft.py index c15721fed..ea3a95679 100644 --- a/xtuner/model/sft.py +++ b/xtuner/model/sft.py @@ -33,6 +33,13 @@ def __init__(self, self.use_lora = lora is not None if self.use_lora: self._prepare_for_lora(peft_model, use_gradient_checkpointing) + try: + # for BaiChuan2, set first_flag to False to disable weight init + if self.llm.base_model.model.__class__.__name__.lower( + ) == 'BaichuanForCausalLM'.lower(): + self.llm.base_model.model.lm_head.first_flag = False + except Exception: + pass elif use_gradient_checkpointing: # For backward compatibility if hasattr(self.llm, 'enable_input_require_grads'): diff --git a/xtuner/utils/templates.py b/xtuner/utils/templates.py index 318cd38cc..21c3a1eee 100644 --- a/xtuner/utils/templates.py +++ b/xtuner/utils/templates.py @@ -82,6 +82,9 @@ baichuan_chat=dict( INSTRUCTION_START='{input}', INSTRUCTION='{input}'), + baichuan2_chat=dict( + INSTRUCTION_START='{input}', + INSTRUCTION='{input}'), wizardlm=dict( INSTRUCTION_START=('A chat between a curious user and an artificial ' 'intelligence assistant. The assistant gives '