ChatGLM2

@article{zeng2022glm,
  title={Glm-130b: An open bilingual pre-trained model},
  author={Zeng, Aohan and Liu, Xiao and Du, Zhengxiao and Wang, Zihan and Lai, Hanyu and Ding, Ming and Yang, Zhuoyi and Xu, Yifan and Zheng, Wendi and Xia, Xiao and others},
  journal={arXiv preprint arXiv:2210.02414},
  year={2022}
}

# 运行如下命令，生成当前机器的RANK_TABLE_FILE的json文件
python ./mindformers/tools/hccl_tools.py --device_num "[0,8)"

{
    "version": "1.0",
    "server_count": "1",
    "server_list": [
        {
            "server_id": "xx.xx.xx.xx",
            "device": [
                {"device_id": "0","device_ip": "192.1.27.6","rank_id": "0"},
                {"device_id": "1","device_ip": "192.2.27.6","rank_id": "1"},
                {"device_id": "2","device_ip": "192.3.27.6","rank_id": "2"},
                {"device_id": "3","device_ip": "192.4.27.6","rank_id": "3"},
                {"device_id": "4","device_ip": "192.1.27.7","rank_id": "4"},
                {"device_id": "5","device_ip": "192.2.27.7","rank_id": "5"},
                {"device_id": "6","device_ip": "192.3.27.7","rank_id": "6"},
                {"device_id": "7","device_ip": "192.4.27.7","rank_id": "7"}],
             "host_nic_ip": "reserve"
        }
    ],
    "status": "completed"
}

# 运行如下命令，生成当前机器的RANK_TABLE_FILE的json文件
python ./mindformers/tools/hccl_tools.py --device_num "[0,8)" --server_ip xx.xx.xx.xx

# 运行如下命令，合并每个机器上的RANK_TABLE_FILE的json文件。
python ./mindformers/tools/merge_hccl.py hccl*.json

{
    "version": "1.0",
    "server_count": "2",
    "server_list": [
        {
            "server_id": "xx.xx.xx.xx",
            "device": [
                {
                    "device_id": "0", "device_ip": "192.168.0.0", "rank_id": "0"
                },
                {
                    "device_id": "1", "device_ip": "192.168.1.0", "rank_id": "1"
                },
                {
                    "device_id": "2", "device_ip": "192.168.2.0", "rank_id": "2"
                },
                {
                    "device_id": "3", "device_ip": "192.168.3.0", "rank_id": "3"
                },
                {
                    "device_id": "4", "device_ip": "192.168.0.1", "rank_id": "4"
                },
                {
                    "device_id": "5", "device_ip": "192.168.1.1", "rank_id": "5"
                },
                {
                    "device_id": "6", "device_ip": "192.168.2.1", "rank_id": "6"
                },
                {
                    "device_id": "7", "device_ip": "192.168.3.1", "rank_id": "7"
                }
            ],
            "host_nic_ip": "reserve"
        },
        {
            "server_id": "xx.xx.xx.xx",
            "device": [
                {
                    "device_id": "0", "device_ip": "192.168.0.1", "rank_id": "8"
                },
                {
                    "device_id": "1", "device_ip": "192.168.1.1", "rank_id": "9"
                },
                {
                    "device_id": "2", "device_ip": "192.168.2.1", "rank_id": "10"
                },
                {
                    "device_id": "3", "device_ip": "192.168.3.1", "rank_id": "11"
                },
                {
                    "device_id": "4", "device_ip": "192.168.0.2", "rank_id": "12"
                },
                {
                    "device_id": "5", "device_ip": "192.168.1.2", "rank_id": "13"
                },
                {
                    "device_id": "6", "device_ip": "192.168.2.2", "rank_id": "14"
                },
                {
                    "device_id": "7", "device_ip": "192.168.3.2", "rank_id": "15"
                }
            ],
            "host_nic_ip": "reserve"
        }
    ],
    "status": "completed"
}

python transform_ckpt.py \
--src_ckpt_strategy {path}/output/strategy/ \
--src_ckpt_dir {path}/output/checkpoint/ \
--dst_ckpt_dir {path}/target_checkpoint/ \
--prefix glm2_6b

# 参数说明
src_ckpt_strategy: 步骤1中的切分策略文件路径
src_ckpt_dir: 原切分权重文件夹
dst_ckpt_dir: 目标路径
prefix: ckpt文件前缀名

import mindspore
from mindformers import AutoConfig, AutoModel, AutoTokenizer

# 指定图模式，指定使用训练卡id
mindspore.set_context(mode=0, device_id=0)

tokenizer = AutoTokenizer.from_pretrained('glm2_6b')

# model的实例化有以下两种方式，选择其中一种进行实例化即可
# 1. 直接根据默认配置实例化
model = AutoModel.from_pretrained('glm2_6b')
# 2. 自定义修改配置后实例化
config = AutoConfig.from_pretrained('glm2_6b')
config.use_past = True                  # 此处修改默认配置，开启增量推理能够加速推理性能
# config.xxx = xxx                      # 根据需求自定义修改其余模型配置
model = AutoModel.from_config(config)   # 从自定义配置项中实例化模型

inputs = tokenizer("你好")["input_ids"]
# 首次调用model.generate()进行推理将包含图编译时间，推理性能显示不准确，多次重复调用以获取准确的推理性能
outputs = model.generate(inputs, max_new_tokens=20, do_sample=True, top_k=3)
response = tokenizer.decode(outputs)
print(response)
# ['你好，作为一名人工智能助手，我欢迎您随时向我提问。']

import mindspore
from mindformers.trainer import Trainer

# 指定图模式，指定使用训练卡id
mindspore.set_context(mode=0, device_id=0)

# 初始化预训练任务
trainer = Trainer(task='text_generation',
                  model='glm2_6b',
                  train_dataset='path/to/train_dataset',
                  eval_dataset='path/to/eval_dataset')

# 开启预训练
# 请参照多卡训练，glm2_6b不支持单卡启动训练
# trainer.train()

# 开启全量微调
# 请参照多卡微调，glm2_6b不支持单卡启动全量微调
# trainer.finetune()

# 开启评测
# 需要在configs/glm2/run_glm2_6b.yaml中将seq_length修改为256
trainer.evaluate()

# 开启推理
predict_result = trainer.predict(input_data="你好")
print(predict_result)
# [{'text_generation_text': ['你好，我是 ChatGLM2-6B， 一个人工智能助手。我背后使用的模型是 GLM2-6B， 是一种大型语言模型， 具有超过 2000 亿参数，支持多种任务。']}]

from mindformers import pipeline, TextGenerationPipeline
task_pipeline = pipeline(task='text_generation', model='glm2_6b', max_length=2048)
task_pipeline('你好')
# [{'text_generation_text': ['你好，我是 ChatGLM2-6B， 一个人工智能助手。我背后使用的模型是 GLM2-6B， 是一种大型语言模型， 具有超过 2000 亿参数，支持多种任务。']}]
pipeline = TextGenerationPipeline(model='glm2_6b', max_length=2048)
predict_result = pipeline("你好")
print(predict_result)
# [{'text_generation_text': ['你好，我是 ChatGLM2-6B， 一个人工智能助手。我背后使用的模型是 GLM2-6B， 是一种大型语言模型， 具有超过 2000 亿参数，支持多种任务。']}]

{
    "content": "类型#上衣*版型#宽松*版型#显瘦*图案#线条*衣样式#衬衫*衣袖型#泡泡袖*衣款式#抽绳",
    "summary": "这件衬衫的款式非常的宽松，利落的线条可以很好的隐藏身材上的小缺点，穿在身上有着很好的显瘦效果。领口装饰了一个可爱的抽绳，漂亮的绳结展现出了十足的个性，配合时尚的泡泡袖型，尽显女性甜美可爱的气息。"
}

AdvertiseGen
  ├── train.json
  └── dev.json

train_dataset: &train_dataset
  data_loader:
    type: ADGenDataLoader
    dataset_dir: "/path/to/AdvertiseGen/train.json"
    shuffle: True
    phase: "train"
    version: 2
    origin_columns: ["content", "summary"]
  tokenizer:
    type: ChatGLM2Tokenizer
    vocab_file: "/path/to/tokenizer.model"
  input_columns: ["input_ids", "labels"]
  max_source_length: 64
  max_target_length: 128
  ignore_pad_token_for_loss: True
  num_parallel_workers: 8
  python_multiprocessing: False
  drop_remainder: True
  batch_size: 1
  repeat: 1
  numa_enable: False
  prefetch_size: 1
  seed: 0

train_dataset_task:
  type: KeyWordGenDataset
  dataset_config: *train_dataset

eval_dataset: &eval_dataset
  data_loader:
    type: ADGenDataLoader
    dataset_dir: "/path/to/AdvertiseGen/dev.json"
    shuffle: False
    phase: "eval"
    version: 2
    origin_columns: ["content", "summary"]
  tokenizer:
    type: ChatGLM2Tokenizer
    vocab_file: "/path/to/tokenizer.model"
  max_source_length: 256
  max_target_length: 256
  ignore_pad_token_for_loss: True
  input_columns: ["input_ids", "labels"]
  num_parallel_workers: 8
  python_multiprocessing: False
  drop_remainder: True
  batch_size: 1
  repeat: 1
  numa_enable: False
  prefetch_size: 1
  seed: 0

eval_dataset_task:
  type: KeyWordGenDataset
  dataset_config: *eval_dataset

cd scripts
# Usage Help: bash run_distribute.sh [RANK_TABLE_FILE] [CONFIG_PATH] [DEVICE_RANGE] [RUN_STATUS]
bash run_distribute.sh /path/to/hccl_8p_01234567_127.0.1.1.json ../configs/glm2/run_glm2_6b_finetune*.yaml '[0,8]' finetune
# 将此处rank_table_file替换为实际路径

RANK_TABLE_FILE: 由mindformers/tools/hccl_tools.py生成的分布式json文件
CONFIG_PATH: 为configs文件夹下面的glm2/run_glm2_6b_finetune*.yaml配置文件
DEVICE_RANGE: 为单机分布式卡的范围，如 '[0,8]' 为8卡分布式，不包含8本身
RUN_STATUS: 为任务运行状态，支持关键字 train\finetune\eval\predict

server_count=12
device_num=8*$server_count
# launch ranks in the 0th server
cd scripts
bash run_distribute.sh $RANK_TABLE_FILE path/to/config.yaml [0,8] finetune $device_num

# launch ranks in the 1-11 server via ssh
for idx in {1..11}
do
    let rank_start=8*$idx
    let rank_end=$rank_start+8
    ssh ${IP_LIST[$idx]} "cd scripts; bash run_distribute.sh $RANK_TABLE_FILE path/to/config.yaml [$rank_start,$rank_end] finetune $device_num"
done

IP_LIST=("192.168.0.0", "192.168.0.1", ..., "192.168.0.11")

cd scripts
# Usage Help: bash run_stanalone.sh [CONFIG_PATH] [DEVICE_ID] [RUN_STATUS]
bash run_standalone.sh ../configs/glm2/run_glm2_6b_lora*.yaml 0 finetune

cd scripts
# Usage Help: bash run_distribute.sh [RANK_TABLE_FILE] [CONFIG_PATH] [DEVICE_RANGE] [RUN_STATUS]
bash run_distribute.sh /path/to/hccl_8p_01234567_127.0.1.1.json ../configs/glm2/run_glm2_6b_lora*.yaml '[0,8]' finetune
# 将此处rank_table_file替换为实际路径

server_count=12
device_num=8*$server_count
# launch ranks in the 0th server
cd scripts
bash run_distribute.sh $RANK_TABLE_FILE path/to/config_lora.yaml [0,8] finetune $device_num

# launch ranks in the 1-11 server via ssh
for idx in {1..11}
do
    let rank_start=8*$idx
    let rank_end=$rank_start+8
    ssh ${IP_LIST[$idx]} "cd scripts; bash run_distribute.sh $RANK_TABLE_FILE path/to/config_lora.yaml [$rank_start,$rank_end] finetune $device_num"
done

IP_LIST=("192.168.0.0", "192.168.0.1", ..., "192.168.0.11")

cd scripts
# Usage Help: bash run_stanalone.sh [CONFIG_PATH] [DEVICE_ID] [RUN_STATUS]
bash run_standalone.sh ../configs/glm2/run_glm2_6b_ptuning2.yaml 0 finetune

do_eval: True
eval_step_interval: 1788
eval_epoch_interval: -1

metric:
  type: ADGENMetric

model:
  model_config:
    seq_length: 192
train_dataset: &train_dataset
  max_source_length: 64
  max_target_length: 127
  batch_size: 8
eval_dataset: &eval_dataset
  max_source_length: 64
  max_target_length: 127
  batch_size: 8

do_eval: True
eval_step_interval: 1788
eval_epoch_interval: -1

metric:
  type: PerplexityMetric

model:
  model_config:
    seq_length: 192
train_dataset: &train_dataset
  data_loader:
    phase: "train"
  max_source_length: 64
  max_target_length: 127
  batch_size: 8
eval_dataset: &eval_dataset
  data_loader:
    phase: "train"
  max_source_length: 64
  max_target_length: 127
  batch_size: 8

model:
  model_config:
    seq_length: 256

python run_mindformer.py --config configs/glm2/run_glm2_6b_finetune_eval.yaml--run_mode eval --load_checkpoint /path/to/glm2_6b_finetune.ckpt --device_id 0 --use_parallel False

python run_mindformer.py --config configs/glm2/run_glm2_6b_lora_eval.yaml --run_mode eval --load_checkpoint /path/to/glm2_6b_lora.ckpt --device_id 0 --use_parallel False

from mindformers import AutoConfig, AutoModel, AutoTokenizer, AutoProcessor
import mindspore as ms

ms.set_context(mode=ms.GRAPH_MODE, device_target="Ascend", device_id=0)

# **注意** LoRA微调模型替换成 “glm2_6b_lora”,
# **注意** P-Tuning 微调模型替换成 “glm2_6b_ptuning2”
config = AutoConfig.from_pretrained("glm2_6b")
# 可以在此使用下行代码指定自定义权重进行推理，默认使用自动从obs上下载的预训练权重
# config.checkpoint_name_or_path = "/path/to/glm2_6b_finetune.ckpt"
config.use_past = True
model = AutoModel.from_config(config)

# 以下两种tokenizer实例化方式选其一即可
# 1. 在线加载方式
tokenizer = AutoTokenizer.from_pretrained("glm2_6b")
# 2. 本地加载方式
# tokenizer = AutoProcessor.from_pretrained("/path/to/your.yaml").tokenizer

batch_size = config.batch_size

inputs = tokenizer(tokenizer.build_prompt("你好"))["input_ids"]
outputs = model.generate([inputs]*batch_size, max_length=128)
print(tokenizer.decode(outputs))
# ['[Round 1]\n\n问：你好\n\n答： 你好👋！我是人工智能助手 ChatGLM2-6B，很高兴见到你，欢迎问我任何问题。']
inputs = tokenizer(tokenizer.build_prompt("请介绍一下华为"))["input_ids"]
outputs = model.generate([inputs]*batch_size, max_length=128)
print(tokenizer.decode(outputs))
# ['[Round 1]\n\n问：请介绍一下华为\n\n答： 华为是一家总部位于中国的全球知名科技公司,成立于1987年,是全球领先的信息与通信技术(ICT)解决方案供
# 应商之一。\n\n华为的业务范围涵盖了网络、终端、云计算、软件、芯片等多个领域,旗下的智能手机、电脑、平板电脑等消费电子产品在国内外市场上都享有较高
# 的声誉。此外,华为还在5G、人工智能、云计算等领域取得了重要的进展,为全球用户提供了更高效、更智能的科技体验。\n\n华为一直致力于技术创新,研发投入
# 占公司总收入的比例超过']
inputs = tokenizer(tokenizer.build_prompt("晚上睡不着应该怎么办"))["input_ids"]
outputs = model.generate([inputs]*batch_size, max_length=128)
print(tokenizer.decode(outputs))
# ['[Round 1]\n\n问：晚上睡不着应该怎么办\n\n答： 以下是一些有助于晚上睡觉的技巧:\n\n1. 创建一个规律的睡眠时间表:每天在相同的时间上床并尽量
# 在同一时间起床,有助于身体建立规律的生物钟。\n\n2. 创建一个舒适的睡眠环境:确保房间安静、黑暗、凉爽和舒适。如果需要,可以使用睡眠面罩、耳塞或空气
# 净化器来帮助创造一个更舒适的睡眠环境。\n\n3. 避免使用电子设备:在睡觉前一两个小时内避免使用电子设备,如']
inputs = tokenizer(tokenizer.build_prompt("类型#上衣*材质#牛仔布*颜色#白色*风格#简约*图案#刺绣*衣样式#外套*衣款式#破洞"))["input_ids"]
outputs = model.generate([inputs]*batch_size, max_length=128)
print(tokenizer.decode(outputs))
# ['[Round 1]\n\n问：类型#上衣*材质#牛仔布*颜色#白色*风格#简约*图案#刺绣*衣样式#外套*衣款式#破洞\n\n答： 上衣 材质:牛仔布 颜色:白色 风格:
# 简约 图案:刺绣 衣样式:外套 衣款式:破洞\n\n这件上衣由牛仔布制成,采用了简约的风格,图案为刺绣设计,衣样式为外套,衣款式为破洞。']

processor:
  tokenizer:
    vocab_file: "/path/to/tokenizer.model"

python run_mindformer.py --config path/to/config.yaml --run_mode predict --predict_data "[Round 1]\n\n问：你好\n\n答："
#  [{'text_generation_text': ['[Round 1]\n\n问：你好\n\n答： 你好👋！我是人工智能助手 ChatGLM2-6B，很高兴见到你，欢迎问我任何问题。']}]

# export
infer:
    prefill_model_path: "glm2_export/glm2_6b_prefill_seq512.mindir" # 保存mindir的位置
    increment_model_path: "glm2_export/glm2_6b_inc_seq512.mindir"   # 保存mindir的位置
    infer_seq_length: 512 # 需要保持跟 model-model_config-seq_length 一致

# ==== model config ====
model:
  model_config:
    seq_length: 512
    checkpoint_name_or_path: "/path/to/your/*.ckpt"

python mindformers/tools/export.py --config_path configs/glm2/export_glm2_6b.yaml

python run_infer_main.py --device_id 0 --model_name glm2_6b --prefill_model_path glm2_export/glm2_6b_prefill_seq512_graph.mindir --increment_model_path glm2_export/glm2_6b_inc_seq512_graph.mindir --config_path lite.ini --is_sample_acceleration False --seq_length 512 --add_special_tokens True

Please enter your predict data:

[Round 1]

问：你好。

答：

['[Round 1]\n\n问：你好。\n\n答： 你好👋！我是人工智能助手 ChatGLM2-6B，很高兴见到你，欢迎问我任何问题。']