推荐|4.1 低精度训练与大模型下载

1 背景介绍：

1.1 大模型训练的难点是什么？

1.2 模型训练的显存占用：

1.3 如何降低训练时的显存占用？

1 背景介绍：

1.1 大模型训练的难点是什么？

1.2 模型训练的显存占用：

1.3 如何降低训练时的显存占用？

1.4 模型本身的显存占用：

之前我们节约显存占用的优化手段都是针对降低训练时的显存占用，那么如何降低模型本身的显存占用呢？这就需要涉及低精度问题。

2 模型下载：

可以理解modelscope为一个更适合中国宝宝的transformers库，当然没有transformers那么完善，但用于下载模型还是可以的。

2.1 配置环境

pip install modelscope jupyterlab

2.2 从modelscope上下载


from modelscope.hub.snapshot_download import snapshot_download
 
# model_id 模型id
# cache_dir 本地缓存目录
# ignore_file_pattern 无需下载的文件
snapshot_download(
    model_id="Shanghai_AI_Laboratory/internlm-20b", cache_dir="F:\Modelscope_models"
)

3 模型加载（以chat-glm为例）：

因为transformers中没有chat_glm的官方实现，所以要设置trust_remote_code=True。

huggingface可以加载modelscope上的模型，但是不是所有模型都可以。


from transformers import AutoTokenizer, AutoModel
 
tokenizer = AutoTokenizer.from_pretrained("d:/Pretrained_models/ZhipuAI/chatglm2-6b/", trust_remote_code=True)
tokenizer

ChatGLMTokenizer(name_or_path='d:/Pretrained_models/ZhipuAI/chatglm2-6b/', vocab_size=64794, model_max_length=1000000000000000019884624838656, is_fast=False, padding_side='left', truncation_side='right', special_tokens={}, clean_up_tokenization_spaces=False)


model = AutoModel.from_pretrained("d:/Pretrained_models/ZhipuAI/chatglm2-6b/", trust_remote_code=True)
model


ChatGLMForConditionalGeneration(
  (transformer): ChatGLMModel(
    (embedding): Embedding(
      (word_embeddings): Embedding(65024, 4096)
    )
    (rotary_pos_emb): RotaryEmbedding()
    (encoder): GLMTransformer(
      (layers): ModuleList(
        (0): GLMBlock(
          (input_layernorm): RMSNorm()
          (self_attention): SelfAttention(
            (query_key_value): Linear(in_features=4096, out_features=4608, bias=True)
            (core_attention): CoreAttention(
              (attention_dropout): Dropout(p=0.0, inplace=False)
            )
            (dense): Linear(in_features=4096, out_features=4096, bias=False)
          )
          (post_attention_layernorm): RMSNorm()
          (mlp): MLP(
            (dense_h_to_4h): Linear(in_features=4096, out_features=27392, bias=False)
            (dense_4h_to_h): Linear(in_features=13696, out_features=4096, bias=False)
          )
        )
        (1): GLMBlock(
          (input_layernorm): RMSNorm()
          (self_attention): SelfAttention(
            (query_key_value): Linear(in_features=4096, out_features=4608, bias=True)
            (core_attention): CoreAttention(
              (attention_dropout): Dropout(p=0.0, inplace=False)
            )
            (dense): Linear(in_features=4096, out_features=4096, bias=False)
          )
          (post_attention_layernorm): RMSNorm()
          (mlp): MLP(
            (dense_h_to_4h): Linear(in_features=4096, out_features=27392, bias=False)
            (dense_4h_to_h): Linear(in_features=13696, out_features=4096, bias=False)
          )
        )
        (2): GLMBlock(
          (input_layernorm): RMSNorm()
          (self_attention): SelfAttention(
            (query_key_value): Linear(in_features=4096, out_features=4608, bias=True)
            (core_attention): CoreAttention(
              (attention_dropout): Dropout(p=0.0, inplace=False)
            )
            (dense): Linear(in_features=4096, out_features=4096, bias=False)
          )
          (post_attention_layernorm): RMSNorm()
          (mlp): MLP(
            (dense_h_to_4h): Linear(in_features=4096, out_features=27392, bias=False)
            (dense_4h_to_h): Linear(in_features=13696, out_features=4096, bias=False)
          )
        )
        (3): GLMBlock(
          (input_layernorm): RMSNorm()
          (self_attention): SelfAttention(
            (query_key_value): Linear(in_features=4096, out_features=4608, bias=True)
            (core_attention): CoreAttention(
              (attention_dropout): Dropout(p=0.0, inplace=False)
            )
            (dense): Linear(in_features=4096, out_features=4096, bias=False)
          )
          (post_attention_layernorm): RMSNorm()
          (mlp): MLP(
            (dense_h_to_4h): Linear(in_features=4096, out_features=27392, bias=False)
            (dense_4h_to_h): Linear(in_features=13696, out_features=4096, bias=False)
          )
        )
        (4): GLMBlock(
          (input_layernorm): RMSNorm()
          (self_attention): SelfAttention(
            (query_key_value): Linear(in_features=4096, out_features=4608, bias=True)
            (core_attention): CoreAttention(
              (attention_dropout): Dropout(p=0.0, inplace=False)
            )
            (dense): Linear(in_features=4096, out_features=4096, bias=False)
          )
          (post_attention_layernorm): RMSNorm()
          (mlp): MLP(
            (dense_h_to_4h): Linear(in_features=4096, out_features=27392, bias=False)
            (dense_4h_to_h): Linear(in_features=13696, out_features=4096, bias=False)
          )
        )
        (5): GLMBlock(
          (input_layernorm): RMSNorm()
          (self_attention): SelfAttention(
            (query_key_value): Linear(in_features=4096, out_features=4608, bias=True)
            (core_attention): CoreAttention(
              (attention_dropout): Dropout(p=0.0, inplace=False)
            )
            (dense): Linear(in_features=4096, out_features=4096, bias=False)
          )
          (post_attention_layernorm): RMSNorm()
          (mlp): MLP(
            (dense_h_to_4h): Linear(in_features=4096, out_features=27392, bias=False)
            (dense_4h_to_h): Linear(in_features=13696, out_features=4096, bias=False)
          )
        )
        (6): GLMBlock(
          (input_layernorm): RMSNorm()
          (self_attention): SelfAttention(
            (query_key_value): Linear(in_features=4096, out_features=4608, bias=True)
            (core_attention): CoreAttention(
              (attention_dropout): Dropout(p=0.0, inplace=False)
            )
            (dense): Linear(in_features=4096, out_features=4096, bias=False)
          )
          (post_attention_layernorm): RMSNorm()
          (mlp): MLP(
            (dense_h_to_4h): Linear(in_features=4096, out_features=27392, bias=False)
            (dense_4h_to_h): Linear(in_features=13696, out_features=4096, bias=False)
          )
        )
        (7): GLMBlock(
          (input_layernorm): RMSNorm()
          (self_attention): SelfAttention(
            (query_key_value): Linear(in_features=4096, out_features=4608, bias=True)
            (core_attention): CoreAttention(
              (attention_dropout): Dropout(p=0.0, inplace=False)
            )
            (dense): Linear(in_features=4096, out_features=4096, bias=False)
          )
          (post_attention_layernorm): RMSNorm()
          (mlp): MLP(
            (dense_h_to_4h): Linear(in_features=4096, out_features=27392, bias=False)
            (dense_4h_to_h): Linear(in_features=13696, out_features=4096, bias=False)
          )
        )
        (8): GLMBlock(
          (input_layernorm): RMSNorm()
          (self_attention): SelfAttention(
            (query_key_value): Linear(in_features=4096, out_features=4608, bias=True)
            (core_attention): CoreAttention(
              (attention_dropout): Dropout(p=0.0, inplace=False)
            )
            (dense): Linear(in_features=4096, out_features=4096, bias=False)
          )
          (post_attention_layernorm): RMSNorm()
          (mlp): MLP(
            (dense_h_to_4h): Linear(in_features=4096, out_features=27392, bias=False)
            (dense_4h_to_h): Linear(in_features=13696, out_features=4096, bias=False)
          )
        )
        (9): GLMBlock(
          (input_layernorm): RMSNorm()
          (self_attention): SelfAttention(
            (query_key_value): Linear(in_features=4096, out_features=4608, bias=True)
            (core_attention): CoreAttention(
              (attention_dropout): Dropout(p=0.0, inplace=False)
            )
            (dense): Linear(in_features=4096, out_features=4096, bias=False)
          )
          (post_attention_layernorm): RMSNorm()
          (mlp): MLP(
            (dense_h_to_4h): Linear(in_features=4096, out_features=27392, bias=False)
            (dense_4h_to_h): Linear(in_features=13696, out_features=4096, bias=False)
          )
        )
        (10): GLMBlock(
          (input_layernorm): RMSNorm()
          (self_attention): SelfAttention(
            (query_key_value): Linear(in_features=4096, out_features=4608, bias=True)
            (core_attention): CoreAttention(
              (attention_dropout): Dropout(p=0.0, inplace=False)
            )
            (dense): Linear(in_features=4096, out_features=4096, bias=False)
          )
          (post_attention_layernorm): RMSNorm()
          (mlp): MLP(
            (dense_h_to_4h): Linear(in_features=4096, out_features=27392, bias=False)
            (dense_4h_to_h): Linear(in_features=13696, out_features=4096, bias=False)
          )
        )
        (11): GLMBlock(
          (input_layernorm): RMSNorm()
          (self_attention): SelfAttention(
            (query_key_value): Linear(in_features=4096, out_features=4608, bias=True)
            (core_attention): CoreAttention(
              (attention_dropout): Dropout(p=0.0, inplace=False)
            )
            (dense): Linear(in_features=4096, out_features=4096, bias=False)
          )
          (post_attention_layernorm): RMSNorm()
          (mlp): MLP(
            (dense_h_to_4h): Linear(in_features=4096, out_features=27392, bias=False)
            (dense_4h_to_h): Linear(in_features=13696, out_features=4096, bias=False)
          )
        )
        (12): GLMBlock(
          (input_layernorm): RMSNorm()
          (self_attention): SelfAttention(
            (query_key_value): Linear(in_features=4096, out_features=4608, bias=True)
            (core_attention): CoreAttention(
              (attention_dropout): Dropout(p=0.0, inplace=False)
            )
            (dense): Linear(in_features=4096, out_features=4096, bias=False)
          )
          (post_attention_layernorm): RMSNorm()
          (mlp): MLP(
            (dense_h_to_4h): Linear(in_features=4096, out_features=27392, bias=False)
            (dense_4h_to_h): Linear(in_features=13696, out_features=4096, bias=False)
          )
        )
        (13): GLMBlock(
          (input_layernorm): RMSNorm()
          (self_attention): SelfAttention(
            (query_key_value): Linear(in_features=4096, out_features=4608, bias=True)
            (core_attention): CoreAttention(
              (attention_dropout): Dropout(p=0.0, inplace=False)
            )
            (dense): Linear(in_features=4096, out_features=4096, bias=False)
          )
          (post_attention_layernorm): RMSNorm()
          (mlp): MLP(
            (dense_h_to_4h): Linear(in_features=4096, out_features=27392, bias=False)
            (dense_4h_to_h): Linear(in_features=13696, out_features=4096, bias=False)
          )
        )
        (14): GLMBlock(
          (input_layernorm): RMSNorm()
          (self_attention): SelfAttention(
            (query_key_value): Linear(in_features=4096, out_features=4608, bias=True)
            (core_attention): CoreAttention(
              (attention_dropout): Dropout(p=0.0, inplace=False)
            )
            (dense): Linear(in_features=4096, out_features=4096, bias=False)
          )
          (post_attention_layernorm): RMSNorm()
          (mlp): MLP(
            (dense_h_to_4h): Linear(in_features=4096, out_features=27392, bias=False)
            (dense_4h_to_h): Linear(in_features=13696, out_features=4096, bias=False)
          )
        )
        (15): GLMBlock(
          (input_layernorm): RMSNorm()
          (self_attention): SelfAttention(
            (query_key_value): Linear(in_features=4096, out_features=4608, bias=True)
            (core_attention): CoreAttention(
              (attention_dropout): Dropout(p=0.0, inplace=False)
            )
            (dense): Linear(in_features=4096, out_features=4096, bias=False)
          )
          (post_attention_layernorm): RMSNorm()
          (mlp): MLP(
            (dense_h_to_4h): Linear(in_features=4096, out_features=27392, bias=False)
            (dense_4h_to_h): Linear(in_features=13696, out_features=4096, bias=False)
          )
        )
        (16): GLMBlock(
          (input_layernorm): RMSNorm()
          (self_attention): SelfAttention(
            (query_key_value): Linear(in_features=4096, out_features=4608, bias=True)
            (core_attention): CoreAttention(
              (attention_dropout): Dropout(p=0.0, inplace=False)
            )
            (dense): Linear(in_features=4096, out_features=4096, bias=False)
          )
          (post_attention_layernorm): RMSNorm()
          (mlp): MLP(
            (dense_h_to_4h): Linear(in_features=4096, out_features=27392, bias=False)
            (dense_4h_to_h): Linear(in_features=13696, out_features=4096, bias=False)
          )
        )
        (17): GLMBlock(
          (input_layernorm): RMSNorm()
          (self_attention): SelfAttention(
            (query_key_value): Linear(in_features=4096, out_features=4608, bias=True)
            (core_attention): CoreAttention(
              (attention_dropout): Dropout(p=0.0, inplace=False)
            )
            (dense): Linear(in_features=4096, out_features=4096, bias=False)
          )
          (post_attention_layernorm): RMSNorm()
          (mlp): MLP(
            (dense_h_to_4h): Linear(in_features=4096, out_features=27392, bias=False)
            (dense_4h_to_h): Linear(in_features=13696, out_features=4096, bias=False)
          )
        )
        (18): GLMBlock(
          (input_layernorm): RMSNorm()
          (self_attention): SelfAttention(
            (query_key_value): Linear(in_features=4096, out_features=4608, bias=True)
            (core_attention): CoreAttention(
              (attention_dropout): Dropout(p=0.0, inplace=False)
            )
            (dense): Linear(in_features=4096, out_features=4096, bias=False)
          )
          (post_attention_layernorm): RMSNorm()
          (mlp): MLP(
            (dense_h_to_4h): Linear(in_features=4096, out_features=27392, bias=False)
            (dense_4h_to_h): Linear(in_features=13696, out_features=4096, bias=False)
          )
        )
        (19): GLMBlock(
          (input_layernorm): RMSNorm()
          (self_attention): SelfAttention(
            (query_key_value): Linear(in_features=4096, out_features=4608, bias=True)
            (core_attention): CoreAttention(
              (attention_dropout): Dropout(p=0.0, inplace=False)
            )
            (dense): Linear(in_features=4096, out_features=4096, bias=False)
          )
          (post_attention_layernorm): RMSNorm()
          (mlp): MLP(
            (dense_h_to_4h): Linear(in_features=4096, out_features=27392, bias=False)
            (dense_4h_to_h): Linear(in_features=13696, out_features=4096, bias=False)
          )
        )
        (20): GLMBlock(
          (input_layernorm): RMSNorm()
          (self_attention): SelfAttention(
            (query_key_value): Linear(in_features=4096, out_features=4608, bias=True)
            (core_attention): CoreAttention(
              (attention_dropout): Dropout(p=0.0, inplace=False)
            )
            (dense): Linear(in_features=4096, out_features=4096, bias=False)
          )
          (post_attention_layernorm): RMSNorm()
          (mlp): MLP(
            (dense_h_to_4h): Linear(in_features=4096, out_features=27392, bias=False)
            (dense_4h_to_h): Linear(in_features=13696, out_features=4096, bias=False)
          )
        )
        (21): GLMBlock(
          (input_layernorm): RMSNorm()
          (self_attention): SelfAttention(
            (query_key_value): Linear(in_features=4096, out_features=4608, bias=True)
            (core_attention): CoreAttention(
              (attention_dropout): Dropout(p=0.0, inplace=False)
            )
            (dense): Linear(in_features=4096, out_features=4096, bias=False)
          )
          (post_attention_layernorm): RMSNorm()
          (mlp): MLP(
            (dense_h_to_4h): Linear(in_features=4096, out_features=27392, bias=False)
            (dense_4h_to_h): Linear(in_features=13696, out_features=4096, bias=False)
          )
        )
        (22): GLMBlock(
          (input_layernorm): RMSNorm()
          (self_attention): SelfAttention(
            (query_key_value): Linear(in_features=4096, out_features=4608, bias=True)
            (core_attention): CoreAttention(
              (attention_dropout): Dropout(p=0.0, inplace=False)
            )
            (dense): Linear(in_features=4096, out_features=4096, bias=False)
          )
          (post_attention_layernorm): RMSNorm()
          (mlp): MLP(
            (dense_h_to_4h): Linear(in_features=4096, out_features=27392, bias=False)
            (dense_4h_to_h): Linear(in_features=13696, out_features=4096, bias=False)
          )
        )
        (23): GLMBlock(
          (input_layernorm): RMSNorm()
          (self_attention): SelfAttention(
            (query_key_value): Linear(in_features=4096, out_features=4608, bias=True)
            (core_attention): CoreAttention(
              (attention_dropout): Dropout(p=0.0, inplace=False)
            )
            (dense): Linear(in_features=4096, out_features=4096, bias=False)
          )
          (post_attention_layernorm): RMSNorm()
          (mlp): MLP(
            (dense_h_to_4h): Linear(in_features=4096, out_features=27392, bias=False)
            (dense_4h_to_h): Linear(in_features=13696, out_features=4096, bias=False)
          )
        )
        (24): GLMBlock(
          (input_layernorm): RMSNorm()
          (self_attention): SelfAttention(
            (query_key_value): Linear(in_features=4096, out_features=4608, bias=True)
            (core_attention): CoreAttention(
              (attention_dropout): Dropout(p=0.0, inplace=False)
            )
            (dense): Linear(in_features=4096, out_features=4096, bias=False)
          )
          (post_attention_layernorm): RMSNorm()
          (mlp): MLP(
            (dense_h_to_4h): Linear(in_features=4096, out_features=27392, bias=False)
            (dense_4h_to_h): Linear(in_features=13696, out_features=4096, bias=False)
          )
        )
        (25): GLMBlock(
          (input_layernorm): RMSNorm()
          (self_attention): SelfAttention(
            (query_key_value): Linear(in_features=4096, out_features=4608, bias=True)
            (core_attention): CoreAttention(
              (attention_dropout): Dropout(p=0.0, inplace=False)
            )
            (dense): Linear(in_features=4096, out_features=4096, bias=False)
          )
          (post_attention_layernorm): RMSNorm()
          (mlp): MLP(
            (dense_h_to_4h): Linear(in_features=4096, out_features=27392, bias=False)
            (dense_4h_to_h): Linear(in_features=13696, out_features=4096, bias=False)
          )
        )
        (26): GLMBlock(
          (input_layernorm): RMSNorm()
          (self_attention): SelfAttention(
            (query_key_value): Linear(in_features=4096, out_features=4608, bias=True)
            (core_attention): CoreAttention(
              (attention_dropout): Dropout(p=0.0, inplace=False)
            )
            (dense): Linear(in_features=4096, out_features=4096, bias=False)
          )
          (post_attention_layernorm): RMSNorm()
          (mlp): MLP(
            (dense_h_to_4h): Linear(in_features=4096, out_features=27392, bias=False)
            (dense_4h_to_h): Linear(in_features=13696, out_features=4096, bias=False)
          )
        )
        (27): GLMBlock(
          (input_layernorm): RMSNorm()
          (self_attention): SelfAttention(
            (query_key_value): Linear(in_features=4096, out_features=4608, bias=True)
            (core_attention): CoreAttention(
              (attention_dropout): Dropout(p=0.0, inplace=False)
            )
            (dense): Linear(in_features=4096, out_features=4096, bias=False)
          )
          (post_attention_layernorm): RMSNorm()
          (mlp): MLP(
            (dense_h_to_4h): Linear(in_features=4096, out_features=27392, bias=False)
            (dense_4h_to_h): Linear(in_features=13696, out_features=4096, bias=False)
          )
        )
      )
      (final_layernorm): RMSNorm()
    )
    (output_layer): Linear(in_features=4096, out_features=65024, bias=False)
  )
)

1 背景介绍：

1.1 大模型训练的难点是什么？

1.2 模型训练的显存占用：

1.3 如何降低训练时的显存占用？

1.4 模型本身的显存占用：

2 模型下载：

2.1 配置环境

2.2 从modelscope上下载

3 模型加载（以chat-glm为例）：

评论记录：