0x0. 背景¶
尝试跑起来HuggingFace上release的DeepSeek V2，踩了几个坑，这里给出解决的方法。HuggingFace提供的开源DeepSeek V2 repo链接为：https://huggingface.co/deepseek-ai/DeepSeek-V2
0x1. 报错1: KeyError: 'sdpa'¶
这个问题社区也有人反馈了。https://huggingface.co/deepseek-ai/DeepSeek-V2/discussions/3
解决方法很简单，在工程里面的config.json最后加一句"_attn_implementation": "flash_attention_2"即可：
0x2. 报错2: 初始化阶段卡死¶
已经给accelerate提了一个pr解决这个问题。https://github.com/huggingface/accelerate/pull/2756
背景¶
当我尝试使用 transformers 库进行 deepseek-v2 模型推理时:
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM, GenerationConfig
model_name = "deepseek-ai/DeepSeek-V2"
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
# `max_memory` should be set based on your devices
max_memory = {i: "75GB" for i in range(8)}
# `device_map` cannot be set to `auto`
model = AutoModelForCausalLM.from_pretrained(model_name, trust_remote_code=True, device_map="sequential", torch_dtype=torch.bfloat16, max_memory=max_memory, attn_implementation="eager")
model.generation_config = GenerationConfig.from_pretrained(model_name)
model.generation_config.pad_token_id = model.generation_config.eos_token_id
text = "An attention function can be described as mapping a query and a set of key-value pairs to an output, where the query, keys, values, and output are all vectors. The output is"
inputs = tokenizer(text, return_tensors="pt")
outputs = model.generate(**inputs.to(model.device), max_new_tokens=100)
result = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(result)
我发现程序卡住了...
解决¶
根据栈信息，卡在 这里(https://github.com/huggingface/accelerate/blob/main/src/accelerate/utils/modeling.py#L1041) .
对于 DeepSeek V2,module_sizes 的长度为68185,而这里的代码复杂度为 O(N^2)(其中 N = module_sizes),需要非常长的时间来执行,给人一种卡住的错觉。这个 PR 优化了代码,使其复杂度变为 O(N),从而可以快速到达加载大模型的阶段。在一台8xA800机器上,经过这种优化后,推理结果也是正常的。
0x3. 单节点A800推理需要限制一下输出长度¶
由于模型的参数量有236B，用bf16来存储单节点8卡A800每张卡都已经占用了大约60G，如果输出长度太长，用HuggingFace直接推理的话KV Cache就顶不住了，如果你想在单节点跑通并且做一些调试的话，输入的prompt设短一点，然后max_new_tokens可以设置为64。实测是可以正常跑通的。我跑通的脚本为：
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM, GenerationConfig
model_name = "/mnt/data/zhaoliang/models/DeepSeek-V2/"
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
# `max_memory` should be set based on your devices
max_memory = {i: "75GB" for i in range(8)}
model = AutoModelForCausalLM.from_pretrained(model_name, trust_remote_code=True, device_map="auto", torch_dtype=torch.bfloat16, max_memory=max_memory)
model.generation_config = GenerationConfig.from_pretrained(model_name)
model.generation_config.pad_token_id = model.generation_config.eos_token_id
text = "An attention function can be described as mapping a query and a set of key-value pairs to an output, where the query, keys, values, and output are all vectors. The output is"
inputs = tokenizer(text, return_tensors="pt")
outputs = model.generate(**inputs.to(model.device), max_new_tokens=100)
result = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(result)
0x4. VLLM¶
如果想在单节点8卡A100/A800上加速推理并且输出更长长度的文本，目前可以使用vllm的实现，具体见这个pr：https://github.com/vllm-project/vllm/pull/4650
deepseek-v2的开源，技术报告，模型架构创新都很好，respect。但个人感觉上传开源代码的时候还是顺手做一下测试吧。
本文总阅读量次