ChatGLM + PEFT 进行finetune
kingglory/ChatGLM-Efficient-Tuning
一、前言
1.1 硬件需求
- 注：r 为LoRA 维数大小，p 为前缀词表大小，l 为微调层数，ex/s 为每秒训练的样本数。gradient_accumulation_steps 参数设置为 1。上述结果均来自于单个 Tesla V100 GPU，仅供参考。
1.2 微调方法
- LoRA：仅微调低秩适应器。
- P-Tuning V2：仅微调前缀编码器。
- Freeze：仅微调后几层的全连接层。
1.3 软件依赖
- Python 3.8+, PyTorch 1.13.1
- Transformers, Datasets, Accelerate, PEFT, TRL
- protobuf, cpm_kernels, sentencepiece
- jieba, rouge_chinese, nltk（用于评估）
- gradio, mdtex2html（用于网页端交互）
二、环境搭建
2.1 构建环境
conda create -n cet python=3.10 # 创建新环境
csource activate cet # 激活环境
2.2 下载代码
git clone https://github.com/hiyouga/ChatGLM-Efficient-Tuning.git
cd ChatGLM-Efficient-Tuning
切换到到 commit_id 为 13 版本
git reset --hard 13f773d958be6074dddf164b593e45cfe7260b77
2.3 安装依赖
pip install -r requirements.txt
三、资源准备
3.1 数据来源介绍
部分预置数据集简介：
- 注：BELLE 数据集是由 ChatGPT 产生的数据集，不保证数据准确性，所有类 GPT 模型产生的 self-instruction 数据集均不能保证其准确性。
部分数据集的使用需要确认，我们推荐使用下述命令登录您的 HuggingFace 账户。
pip install --upgrade huggingface_hub
huggingface-cli login
3.2 模型下载【建议下载到本地】
Huggingface 平台下载
- clone THUDM/chatglm-6b
git clone https://huggingface.co/THUDM/chatglm-6b
>>>
Cloning into 'chatglm-6b'...
remote: Enumerating objects: 372, done.
remote: Counting objects: 100% (369/369), done.
remote: Compressing objects: 100% (153/153), done.
remote: Total 372 (delta 229), reused 348 (delta 216), pack-reused 3
Receiving objects: 100% (372/372), 103.83 KiB | 2.00 MiB/s, done.
Resolving deltas: 100% (229/229), done.
Filtering content: 100% (9/9), 12.49 GiB | 734.00 KiB/s, done.
- 切换到到 commit_id 为 55 版本
git reset --hard 551a50efec3acc5a9b94de8ec46d33d0f81919f7
四、ChatGLM-6B 模型微调
4.1 ChatGLM-6B 模型微调 单 GPU 微调训练
运行以下指令进行微调：
- freeze 方式 微调
CUDA_VISIBLE_DEVICES=0 python src/finetune.py \
--do_train \
--dataset alpaca_gpt4_zh \
--finetuning_type freeze \ # none or lora or freeze or p_tuning
--output_dir path_to_checkpoint \
--per_device_train_batch_size 4 \
--gradient_accumulation_steps 4 \
--lr_scheduler_type cosine \
--logging_steps 10 \
--save_steps 1000 \
--learning_rate 5e-5 \
--num_train_epochs 1.0 \
--fp16 \
--quantization_bit=8
- p_tuning 方式 微调
CUDA_VISIBLE_DEVICES=0 python src/finetune.py \
--do_train \
--dataset alpaca_gpt4_zh \
--finetuning_type p_tuning \ # none or lora or freeze or p_tuning
--output_dir path_to_checkpoint \
--per_device_train_batch_size 4 \
--gradient_accumulation_steps 4 \
--lr_scheduler_type cosine \
--logging_steps 10 \
--save_steps 1000 \
--learning_rate 5e-5 \
--num_train_epochs 1.0 \
--quantization_bit=8
output
...
100%|█████████████████████████████████████████████████████████████████████| 12204/12204 [3:09:13<00:00, 1.14it/s][INFO|trainer.py:2039] 2023-04-27 13:13:17,316 >>
Training completed. Do not forget to share your model on huggingface.co/models =)
{'train_runtime': 11360.4803, 'train_samples_per_second': 4.297, 'train_steps_per_second': 1.074, 'train_loss': 2.4381695562642114, 'epoch': 1.0}
100%|█████████████████████████████████████████████████████████████████████| 12204/12204 [3:09:13<00:00, 1.07it/s]
***** train metrics *****
epoch = 1.0
train_loss = 2.4382
train_runtime = 3:09:20.48
train_samples_per_second = 4.297
train_steps_per_second = 1.074
04/27/2023 13:13:17 - INFO - utils.seq2seq - Saving model checkpoint to path_to_checkpoint
wandb: Waiting for W&B process to finish... (success).
wandb:
wandb:
wandb: Run history:
wandb: train/epoch ▁▁▁▁▂▂▂▂▂▃▃▃▃▃▃▄▄▄▄▄▅▅▅▅▅▅▆▆▆▆▆▇▇▇▇▇▇███
wandb: train/global_step ▁▁▁▁▂▂▂▂▂▃▃▃▃▃▃▄▄▄▄▄▅▅▅▅▅▅▆▆▆▆▆▇▇▇▇▇▇███
wandb: train/learning_rate ███████▇▇▇▇▇▆▆▆▆▆▅▅▅▄▄▄▄▃▃▃▃▂▂▂▂▂▁▁▁▁▁▁▁
wandb: train/loss ▆▅█▆▄▄▃▆▅▄▄▃▄▂▃▄▄▃▂▃▄▃▃▄▄▄▁▄▂▄▃▂▃▂▂▃▃▁▄▄
wandb: train/total_flos ▁
wandb: train/train_loss ▁
wandb: train/train_runtime ▁
wandb: train/train_samples_per_second ▁
wandb: train/train_steps_per_second ▁
wandb:
wandb: Run summary:
wandb: train/epoch 1.0
wandb: train/global_step 12204
wandb: train/learning_rate 0.0
wandb: train/loss 2.308
wandb: train/total_flos 1.5298043716141056e+17
wandb: train/train_loss 2.43817
wandb: train/train_runtime 11360.4803
wandb: train/train_samples_per_second 4.297
wandb: train/train_steps_per_second 1.074
如果是p_tuning 方式 微调，则不能使用 --fp16参数。
- lora 方式 微调
CUDA_VISIBLE_DEVICES=0 python src/finetune.py \
--do_train \
--dataset alpaca_gpt4_zh \
--finetuning_type lora \ # none or lora or freeze or p_tuning
--output_dir path_to_checkpoint \
--per_device_train_batch_size 4 \
--gradient_accumulation_steps 4 \
--lr_scheduler_type cosine \
--logging_steps 10 \
--save_steps 1000 \
--learning_rate 5e-5 \
--num_train_epochs 1.0 \
--fp16 \
--quantization_bit=8
output
...
100%|█████████████████████████████████████████████████████████████████████| 12204/12204 [4:43:25<00:00, 1.39s/it]
***** train metrics *****
epoch = 1.0
train_loss = 2.0096
train_runtime = 4:43:32.43
train_samples_per_second = 2.87
train_steps_per_second = 0.717
04/27/2023 08:58:11 - INFO - utils.seq2seq - Saving model checkpoint to path_to_checkpoint
wandb: Waiting for W&B process to finish... (success).
wandb:
wandb:
wandb: Run history:
wandb: train/epoch ▁▁▁▁▂▂▂▂▂▃▃▃▃▃▃▄▄▄▄▄▅▅▅▅▅▅▆▆▆▆▆▇▇▇▇▇▇███
wandb: train/global_step ▁▁▁▁▂▂▂▂▂▃▃▃▃▃▃▄▄▄▄▄▅▅▅▅▅▅▆▆▆▆▆▇▇▇▇▇▇███
wandb: train/learning_rate ███████▇▇▇▇▇▆▆▆▆▆▅▅▅▄▄▄▄▃▃▃▃▂▂▂▂▂▁▁▁▁▁▁▁
wandb: train/loss ▆▆▆▆▄▅▄█▇▄▄▄▆▃▆▅▇▄▄▅▇▆▄▆▆▇▁█▃▆▅▅▅▄▂▅▆▂█▇
wandb: train/total_flos ▁
wandb: train/train_loss ▁
wandb: train/train_runtime ▁
wandb: train/train_samples_per_second ▁
wandb: train/train_steps_per_second ▁
wandb:
wandb: Run summary:
wandb: train/epoch 1.0
wandb: train/global_step 12204
wandb: train/learning_rate 0.0
wandb: train/loss 1.9716
wandb: train/total_flos 3.059666492984525e+17
wandb: train/train_loss 2.00964
wandb: train/train_runtime 17012.4383
wandb: train/train_samples_per_second 2.87
wandb: train/train_steps_per_second 0.717
wandb:
wandb: Synced wild-paper-36: https://wandb.ai/13025232601/huggingface/runs/8kemyayq
wandb: Synced 6 W&B file(s), 0 media file(s), 0 artifact file(s) and 0 other file(s)
wandb: Find logs at: ./wandb/run-20230427_041440-8kemyayq/logs
4.2 ChatGLM-6B 模型 多GPU分布式微调
- 配置 分布式环境
accelerate config # 首先配置分布式环境
注：注意：若您使用 LoRA 方法进行微调，请指定以下参数 --ddp_find_unused_parameters False 来避免报错。
- 运行以下指令进行微调：
lora 方式 微调
accelerate launch src/finetune.py # 参数同上
五、ChatGLM-6B 评估预测
5.1 ChatGLM-6B 指标评估（BLEU分数和汉语ROUGE分数）
CUDA_VISIBLE_DEVICES=0 python src/finetune.py \
--do_eval \
--dataset alpaca_gpt4_zh \
--checkpoint_dir path_to_checkpoint \
--output_dir path_to_eval_result \
--per_device_eval_batch_size 8 \
--max_samples 50 \
--predict_with_generate
output
===================================BUG REPORT===================================
Welcome to bitsandbytes. For bug reports, please submit your error trace to: https://github.com/TimDettmers/bitsandbytes/issues
================================================================================
...
mon - Training/evaluation parameters Seq2SeqTrainingArguments(
_n_gpu=1,
adafactor=False,
adam_beta1=0.9,
adam_beta2=0.999,
adam_epsilon=1e-08,
auto_find_batch_size=False,
bf16=False,
bf16_full_eval=False,
data_seed=None,
dataloader_drop_last=False,
dataloader_num_workers=0,
dataloader_pin_memory=True,
ddp_bucket_cap_mb=None,
ddp_find_unused_parameters=None,
ddp_timeout=1800,
debug=[],
deepspeed=None,
disable_tqdm=False,
do_eval=True,
do_predict=False,
do_train=False,
eval_accumulation_steps=None,
eval_delay=0,
eval_steps=None,
evaluation_strategy=no,
fp16=False,
fp16_backend=auto,
fp16_full_eval=False,
fp16_opt_level=O1,
fsdp=[],
fsdp_config={'fsdp_min_num_params': 0, 'xla': False, 'xla_fsdp_grad_ckpt': False},
fsdp_min_num_params=0,
fsdp_transformer_layer_cls_to_wrap=None,
full_determinism=False,
generation_config=None,
generation_max_length=None,
generation_num_beams=None,
gradient_accumulation_steps=1,
gradient_checkpointing=False,
greater_is_better=None,
group_by_length=False,
half_precision_backend=auto,
hub_model_id=None,
hub_private_repo=False,
hub_strategy=every_save,
hub_token=,
ignore_data_skip=False,
include_inputs_for_metrics=False,
jit_mode_eval=False,
label_names=None,
label_smoothing_factor=0.0,
learning_rate=5e-05,
length_column_name=length,
load_best_model_at_end=False,
local_rank=-1,
log_level=passive,
log_level_replica=warning,
log_on_each_node=True,
logging_dir=path_to_eval_result/runs/May05_00-58-16_tgnet,
logging_first_step=False,
logging_nan_inf_filter=True,
logging_steps=500,
logging_strategy=steps,
lr_scheduler_type=linear,
max_grad_norm=1.0,
max_steps=-1,
metric_for_best_model=None,
mp_parameters=,
no_cuda=False,
num_train_epochs=3.0,
optim=adamw_torch,
optim_args=None,
output_dir=path_to_eval_result,
overwrite_output_dir=False,
past_index=-1,
per_device_eval_batch_size=1,
per_device_train_batch_size=8,
predict_with_generate=True,
prediction_loss_only=False,
push_to_hub=False,
push_to_hub_model_id=None,
push_to_hub_organization=None,
push_to_hub_token=,
ray_scope=last,
remove_unused_columns=True,
report_to=['tensorboard', 'wandb'],
resume_from_checkpoint=None,
run_name=path_to_eval_result,
save_on_each_node=False,
save_safetensors=False,
save_steps=500,
save_strategy=steps,
save_total_limit=None,
seed=42,
sharded_ddp=[],
skip_memory_metrics=True,
sortish_sampler=False,
tf32=None,
torch_compile=False,
torch_compile_backend=None,
torch_compile_mode=None,
torchdynamo=None,
tpu_metrics_debug=False,
tpu_num_cores=None,
use_ipex=False,
use_legacy_prediction_loop=False,
use_mps_device=False,
warmup_ratio=0.0,
warmup_steps=0,
weight_decay=0.0,
xpu_backend=None,
)
...
[INFO|configuration_utils.py:720] 2023-05-05 00:58:18,251 >> Model config ChatGLMConfig {
"_name_or_path": "THUDM/chatglm-6b",
"architectures": [
"ChatGLMModel"
],
"auto_map": {
"AutoConfig": "configuration_chatglm.ChatGLMConfig",
"AutoModel": "modeling_chatglm.ChatGLMForConditionalGeneration",
"AutoModelForSeq2SeqLM": "modeling_chatglm.ChatGLMForConditionalGeneration"
},
"bos_token_id": 130004,
"eos_token_id": 130005,
"gmask_token_id": 130001,
"hidden_size": 4096,
"inner_hidden_size": 16384,
"layernorm_epsilon": 1e-05,
"mask_token_id": 130000,
"max_sequence_length": 2048,
"model_type": "chatglm",
"num_attention_heads": 32,
"num_layers": 28,
"pad_token_id": 3,
"position_encoding_2d": true,
"pre_seq_len": null,
"prefix_projection": false,
"quantization_bit": 0,
"torch_dtype": "float16",
"transformers_version": "4.28.1",
"use_cache": true,
"vocab_size": 130528
}
[INFO|configuration_utils.py:575] 2023-05-05 00:58:18,291 >> Generate config GenerationConfig {
"_from_model_config": true,
"bos_token_id": 130004,
"eos_token_id": 130005,
"pad_token_id": 3,
"transformers_version": "4.28.1"
}
Loading checkpoint shards: 100%|████████████████████████████████████████████████████| 8/8 [00:10<00:00, 1.30s/it]
...
[INFO|modeling_utils.py:2839] 2023-05-05 00:58:29,280 >> Generation config file not found, using a generation config created from the model config.
05/05/2023 00:59:18 - INFO - utils.common - Quantized model to 4 bit.
05/05/2023 00:59:18 - INFO - utils.common - Fine-tuning method: P-Tuning V2
trainable params: 3670016 || all params: 3359416320 || trainable%: 0.1092
05/05/2023 00:59:18 - WARNING - datasets.arrow_dataset - Loading cached processed dataset at /.cache/huggingface/datasets/json/default-5c75ee3f92a08afd/0.0.0/fe5dd6ea2639a6df622901539cb550cf8797e5a6b2dd7af1cf934bed8e233e6e/cache-a21964d2ca8fe3cd.arrow
input_ids:
[3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 5, 112991, 80990, 66334, 63823, 130001, 130004]
inputs:
保持健康的三个提示。
label_ids:
[82235, 112991, 80990, 66334, 12, 4, 4, 9, 7, 5, 64442, 64310, 63987, 63823, 64354, 63912, 70666, 64310, 64287, 6, 63906, 71738, 63824, 70153, 63853, 68483, 6, 83231, 83242, 64176, 6, 65337, 66448, 65006, 6, 63885, 67623, 64651, 67266, 63823, 4, 4, 10, 7, 5, 71356, 65821, 63823, 64354, 65979, 73362, 66296, 63824, 66220, 63824, 64080, 89181, 63826, 100913, 64284, 94211, 65091, 6, 65073, 63905, 65044, 63824, 105241, 63826, 65521, 65060, 6, 63847, 112991, 108006, 63823, 4, 4, 13, 7, 5, 66625, 69769, 63823, 66625, 118143, 76038, 6, 73929, 64354, 64064, 64849, 5, 25, 11, 23, 5, 88081, 66625, 63823, 66584, 66625, 67623, 67455, 64700, 6, 64721, 64310, 65181, 6, 63885, 64299, 73066, 63826, 75991, 63823, 130001, 130004]
labels:
以下是保持健康的三个提示:
1. 保持身体活动。每天做适当的身体运动,如散步、跑步或游泳,能促进心血管健康,增强肌肉力量,并有助于减少体重。
2. 均衡饮食。每天食用新鲜的蔬菜、水果、全谷物和脂肪含量低的蛋白质食物,避免高糖、高脂肪和加工食品,以保持健康的饮食习惯。
3. 睡眠充足。睡眠对人体健康至关重要,成年人每天应保证 7-8 小时的睡眠。良好的睡眠有助于减轻压力,促进身体恢复,并提高注意力和记忆力。
[INFO|trainer.py:3129] 2023-05-05 00:59:19,444 >> ***** Running Evaluation *****
[INFO|trainer.py:3131] 2023-05-05 00:59:19,444 >> Num examples = 50
[INFO|trainer.py:3134] 2023-05-05 00:59:19,444 >> Batch size = 1
[INFO|configuration_utils.py:575] 2023-05-05 00:59:19,449 >> Generate config GenerationConfig {
"_from_model_config": true,
"bos_token_id": 130004,
"eos_token_id": 130005,
"pad_token_id": 3,
"transformers_version": "4.28.1"
}
100%|█████████████████████████████████████████████████████████████████████████████| 50/50 ...
100%|█████████████████████████████████████████████████████████████████████████████| 50/50 [07:41<00:00, 9.24s/it]
***** eval metrics *****
eval_bleu-4 = 13.0515
eval_rouge-1 = 33.0999
eval_rouge-2 = 13.6305
eval_rouge-l = 24.3066
eval_runtime = 0:07:43.40
eval_samples_per_second = 0.108
eval_steps_per_second = 0.108
wandb: Waiting for W&B process to finish... (success).
wandb:
wandb:
wandb: Run history:
wandb: eval/bleu-4 ▁
wandb: eval/rouge-1 ▁
wandb: eval/rouge-2 ▁
wandb: eval/rouge-l ▁
wandb: eval/runtime ▁
wandb: eval/samples_per_second ▁
wandb: eval/steps_per_second ▁
wandb: train/global_step ▁
wandb:
wandb: Run summary:
wandb: eval/bleu-4 13.05145
wandb: eval/rouge-1 33.09988
wandb: eval/rouge-2 13.63049
wandb: eval/rouge-l 24.30655
wandb: eval/runtime 463.4072
wandb: eval/samples_per_second 0.108
wandb: eval/steps_per_second 0.108
wandb: train/global_step 0
wandb:
wandb: Synced imperial-council-44: https://wandb.ai/13025232601/huggingface/runs/37i3uo74
wandb: Synced 6 W&B file(s), 0 media file(s), 0 artifact file(s) and 0 other file(s)
wandb: Find logs at: ./wandb/run-20230505_010703-37i3uo74/logs
5.2 ChatGLM-6B 模型预测
CUDA_VISIBLE_DEVICES=0 python src/finetune.py \
--do_predict \
--dataset alpaca_gpt4_zh \
--checkpoint_dir path_to_checkpoint \
--output_dir path_to_predict_result \
--per_device_eval_batch_size 8 \
--max_samples 50 \
--predict_with_generate
output
...
input_ids:
[3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 5, 112991, 80990, 66334, 63823, 130001, 130004]
inputs:
保持健康的三个提示。
label_ids:
[82235, 112991, 80990, 66334, 12, 4, 4, 9, 7, 5, 64442, 64310, 63987, 63823, 64354, 63912, 70666, 64310, 64287, 6, 63906, 71738, 63824, 70153, 63853, 68483, 6, 83231, 83242, 64176, 6, 65337, 66448, 65006, 6, 63885, 67623, 64651, 67266, 63823, 4, 4, 10, 7, 5, 71356, 65821, 63823, 64354, 65979, 73362, 66296, 63824, 66220, 63824, 64080, 89181, 63826, 100913, 64284, 94211, 65091, 6, 65073, 63905, 65044, 63824, 105241, 63826, 65521, 65060, 6, 63847, 112991, 108006, 63823, 4, 4, 13, 7, 5, 66625, 69769, 63823, 66625, 118143, 76038, 6, 73929, 64354, 64064, 64849, 5, 25, 11, 23, 5, 88081, 66625, 63823, 66584, 66625, 67623, 67455, 64700, 6, 64721, 64310, 65181, 6, 63885, 64299, 73066, 63826, 75991, 63823, 130001, 130004]
labels:
以下是保持健康的三个提示:
1. 保持身体活动。每天做适当的身体运动,如散步、跑步或游泳,能促进心血管健康,增强肌肉力量,并有助于减少体重。
2. 均衡饮食。每天食用新鲜的蔬菜、水果、全谷物和脂肪含量低的蛋白质食物,避免高糖、高脂肪和加工食品,以保持健康的饮食习惯。
3. 睡眠充足。睡眠对人体健康至关重要,成年人每天应保证 7-8 小时的睡眠。良好的睡眠有助于减轻压力,促进身体恢复,并提高注意力和记忆力。
[INFO|trainer.py:3129] 2023-05-05 04:42:02,531 >> ***** Running Prediction *****
[INFO|trainer.py:3131] 2023-05-05 04:42:02,531 >> Num examples = 50
[INFO|trainer.py:3134] 2023-05-05 04:42:02,531 >> Batch size = 8
[INFO|configuration_utils.py:575] 2023-05-05 04:42:02,539 >> Generate config GenerationConfig {
"_from_model_config": true,
"bos_token_id": 130004,
"eos_token_id": 130005,
"pad_token_id": 3,
"transformers_version": "4.28.1"
}
100%|████████████████████████████████████████████████████████████████████████████████| 7/7 [02:07<00:00, 19.63s/it]Building prefix dict from the default dictionary ...
05/05/2023 04:44:34 - DEBUG - jieba - Building prefix dict from the default dictionary ...
Loading model from cache /tmp/jieba.cache
05/05/2023 04:44:34 - DEBUG - jieba - Loading model from cache /tmp/jieba.cache
Loading model cost 0.221 seconds.
05/05/2023 04:44:34 - DEBUG - jieba - Loading model cost 0.221 seconds.
Prefix dict has been built successfully.
05/05/2023 04:44:34 - DEBUG - jieba - Prefix dict has been built successfully.
100%|████████████████████████████████████████████████████████████████████████████████| 7/7 [02:07<00:00, 18.26s/it]
***** predict metrics *****
predict_bleu-4 = 11.0719
predict_rouge-1 = 31.7603
predict_rouge-2 = 12.2967
predict_rouge-l = 23.4435
predict_runtime = 0:02:32.01
predict_samples_per_second = 0.329
predict_steps_per_second = 0.046
05/05/2023 04:44:34 - INFO - utils.seq2seq - Saving prediction results to path_to_predict_result/generated_predictions.txt
5.3 ChatGLM-6B 模型推理
CUDA_VISIBLE_DEVICES=0 python src/infer.py --checkpoint_dir path_to_checkpoint --quantization_bit=4
对话一
User: 政府可以采取哪些策略来减少空气污染？
ChatGLM-6B: 政府可以采取以下策略来减少空气污染：
1. 推广清洁能源：政府可以通过推广太阳能、风能、水能等清洁能源来减少空气污染。政府还可以鼓励人们使用电动汽车、混合动力汽车、节能灯具等环保产品。
2. 改善交通方式：政府可以改善交通方式，如减少汽车尾气排放，提高公共交通的使用效率，改善道路的通行条件等。
3. 控制工业污染：政府可以通过改善工业污染治理、加强环境监测来减少工业污染。
4. 加强环境教育：政府可以通过加强环境教育，提高公众的环境意识和环保意识，鼓励人们采取环保行动。
5. 实施环境法规：政府可以通过制定和执行环境法规，加强环境监管，打击环境违法行为，保证环境的可持续发展。
政府可以通过以上多种策略来减少空气污染，为人们创造更加健康、舒适的环境。
对话二
User: 自主机器人是计算机控制的机器，被编程执行特定任务而不需要任何人类输入。自主机器人在各个行业中被越来越广泛地应用，从制造业到医疗保健再到安全。
ChatGLM-6B: 自主机器人是指计算机控制的机器，被编程执行特定任务而不需要任何人类输入。这种技术在各个行业中都得到了越来越广泛的应用，从制造业到医疗保健再到安全。
在制造业中，自主机器人可以执行重复性任务，如装配线、包装、物流等。它们可以提高生产效率和质量，减少人工操作，提高安全性和可靠性。
在医疗保健领域，自主机器人可以执行手术、治疗、护理等任务。它们可以提高医疗效率和准确性，减少手术风险和错误。
在安全领域，自主机器人可以执行巡逻、监控、救援等任务。它们可以提高安全性和可靠性，减少人为错误和失误。
自主机器人技术在各个行业都有很多应用，可以提高生产效率、质量和安全性。随着技术的发展，自主机器人也会变得更加智能和人性化，成为人类的伙伴。
5.4 ChatGLM-6B 浏览器测试
CUDA_VISIBLE_DEVICES=0 python src/web_demo.py \
--checkpoint_dir path_to_checkpoint
六、踩坑笔记
6.1 第三步出现错误：RuntimeError: probability tensor contains either inf, nan or element < 0
- 问题描述
...
next_tokens = torch.multinomial(probs, num_samples=1).squeeze(1)
RuntimeError: probability tensor contains either inf, nan or element < 0
- 解决方法: 将 model.generate 中 设置 do_sample=False
参考/感谢
ChatGLM-Efficient-Tuning/README_zh.md
【LLMs 入门实战 —— 十一 】基于 PEFT 的高效 ChatGLM-6B 微调
ChatGLM-6B 小编填坑记
THUDM/ChatGLM-6B
huggingface.co/docs/peft
github.com/huggingface/peft/tuners