怎样使用Accelerate库在多GPU上进行LLM推理呢？

大型语言模型(llm)已经彻底改变了自然语言处理领域。随着这些模型在规模和复杂性上的增长，推理的计算需求也显著增加。为了应对这一挑战利用多个gpu变得至关重要。

所以本文将在多个gpu上并行执行推理，主要包括：Accelerate库介绍，简单的方法与工作代码示例和使用多个gpu的性能基准测试。

本文将使用多个3090将llama2-7b的推理扩展在多个GPU上

基本示例

我们首先介绍一个简单的示例来演示使用Accelerate进行多gpu“消息传递”。

from accelerate import Accelerator
 from accelerate.utils import gather_object
 
 accelerator = Accelerator()
 
 # each GPU creates a string
 message=[ f"Hello this is GPU {accelerator.process_index}" ] 
 
 # collect the messages from all GPUs
 messages=gather_object(message)
 
 # output the messages only on the main process with accelerator.print() 
 accelerator.print(messages)

输出如下：

['Hello this is GPU 0', 
   'Hello this is GPU 1', 
   'Hello this is GPU 2', 
   'Hello this is GPU 3', 
   'Hello this is GPU 4']

多GPU推理

下面是一个简单的、非批处理的推理方法。代码很简单，因为Accelerate库已经帮我们做了很多工作，我们直接使用就可以：

from accelerate import Accelerator
 from accelerate.utils import gather_object
 from transformers import AutoModelForCausalLM, AutoTokenizer
 from statistics import mean
 import torch, time, json
 
 accelerator = Accelerator()
 
 # 10*10 Prompts. Source: https://www.penguin.co.uk/articles/2022/04/best-first-lines-in-books
 prompts_all=[
     "The King is dead. Long live the Queen.",
     "Once there were four children whose names were Peter, Susan, Edmund, and Lucy.",
     "The story so far: in the beginning, the universe was created.",
     "It was a bright cold day in April, and the clocks were striking thirteen.",
     "It is a truth universally acknowledged, that a single man in possession of a good fortune, must be in want of a wife.",
     "The sweat wis lashing oafay Sick Boy; he wis trembling.",
     "124 was spiteful. Full of Baby's venom.",
     "As Gregor Samsa awoke one morning from uneasy dreams he found himself transformed in his bed into a gigantic insect.",
     "I write this sitting in the kitchen sink.",
     "We were somewhere around Barstow on the edge of the desert when the drugs began to take hold.",
 ] * 10
 
 # load a base model and tokenizer
 model_path="models/llama2-7b"
 model = AutoModelForCausalLM.from_pretrained(
     model_path,    
     device_map={"": accelerator.process_index},
     torch_dtype=torch.bfloat16,
 )
 tokenizer = AutoTokenizer.from_pretrained(model_path)   
 
 # sync GPUs and start the timer
 accelerator.wait_for_everyone()
 start=time.time()
 
 # divide the prompt list onto the available GPUs 
 with accelerator.split_between_processes(prompts_all) as prompts:
     # store output of generations in dict
     results=dict(outputs=[], num_tokens=0)
 
     # have each GPU do inference, prompt by prompt
     for prompt in prompts:
         prompt_tokenized=tokenizer(prompt, return_tensors="pt").to("cuda")
         output_tokenized = model.generate(**prompt_tokenized, max_new_tokens=100)[0]
 
         # remove prompt from output 
         output_tokenized=output_tokenized[len(prompt_tokenized["input_ids"][0]):]
 
         # store outputs and number of tokens in result{}
         results["outputs"].append( tokenizer.decode(output_tokenized) )
         results["num_tokens"] += len(output_tokenized)
 
     results=[ results ] # transform to list, otherwise gather_object() will not collect correctly
 
 # collect results from all the GPUs
 results_gathered=gather_object(results)
 
 if accelerator.is_main_process:
     timediff=time.time()-start
     num_tokens=sum([r["num_tokens"] for r in results_gathered ])
 
     print(f"tokens/sec: {num_tokens//timediff}, time {timediff}, total tokens {num_tokens}, total prompts {len(prompts_all)}")

使用多个gpu会导致一些通信开销:性能在4个gpu时呈线性增长，然后在这种特定设置中趋于稳定。当然这里的性能取决于许多参数，如模型大小和量化、提示长度、生成的令牌数量和采样策略，所以我们只讨论一般的情况

1 GPU: 44个token /秒，时间:225.5s

2 gpu: 88个token /秒，时间:112.9s

3 gpu: 128个token /秒，时间:77.6s

4 gpu: 137个token /秒，时间:72.7s

5 gpu: 119个token /秒，时间:83.8s

在多GPU上进行批处理

现实世界中，我们可以使用批处理推理来加快速度。这会减少GPU之间的通讯，加快推理速度。我们只需要增加prepare_prompts函数将一批数据而不是单条数据输入到模型即可：

from accelerate import Accelerator
 from accelerate.utils import gather_object
 from transformers import AutoModelForCausalLM, AutoTokenizer
 from statistics import mean
 import torch, time, json
 
 accelerator = Accelerator()
 
 def write_pretty_json(file_path, data):
     import json
     with open(file_path, "w") as write_file:
         json.dump(data, write_file, indent=4)
 
 # 10*10 Prompts. Source: https://www.penguin.co.uk/articles/2022/04/best-first-lines-in-books
 prompts_all=[
     "The King is dead. Long live the Queen.",
     "Once there were four children whose names were Peter, Susan, Edmund, and Lucy.",
     "The story so far: in the beginning, the universe was created.",
     "It was a bright cold day in April, and the clocks were striking thirteen.",
     "It is a truth universally acknowledged, that a single man in possession of a good fortune, must be in want of a wife.",
     "The sweat wis lashing oafay Sick Boy; he wis trembling.",
     "124 was spiteful. Full of Baby's venom.",
     "As Gregor Samsa awoke one morning from uneasy dreams he found himself transformed in his bed into a gigantic insect.",
     "I write this sitting in the kitchen sink.",
     "We were somewhere around Barstow on the edge of the desert when the drugs began to take hold.",
 ] * 10
 
 # load a base model and tokenizer
 model_path="models/llama2-7b"
 model = AutoModelForCausalLM.from_pretrained(
     model_path,    
     device_map={"": accelerator.process_index},
     torch_dtype=torch.bfloat16,
 )
 tokenizer = AutoTokenizer.from_pretrained(model_path)   
 tokenizer.pad_token = tokenizer.eos_token
 
 # batch, left pad (for inference), and tokenize
 def prepare_prompts(prompts, tokenizer, batch_size=16):
     batches=[prompts[i:i + batch_size] for i in range(0, len(prompts), batch_size)]  
     batches_tok=[]
     tokenizer.padding_side="left"     
     for prompt_batch in batches:
         batches_tok.append(
             tokenizer(
                 prompt_batch, 
                 return_tensors="pt", 
                 padding='longest', 
                 truncation=False, 
                 pad_to_multiple_of=8,
                 add_special_tokens=False).to("cuda") 
             )
     tokenizer.padding_side="right"
     return batches_tok
 
 # sync GPUs and start the timer
 accelerator.wait_for_everyone()    
 start=time.time()
 
 # divide the prompt list onto the available GPUs 
 with accelerator.split_between_processes(prompts_all) as prompts:
     results=dict(outputs=[], num_tokens=0)
 
     # have each GPU do inference in batches
     prompt_batches=prepare_prompts(prompts, tokenizer, batch_size=16)
 
     for prompts_tokenized in prompt_batches:
         outputs_tokenized=model.generate(**prompts_tokenized, max_new_tokens=100)
 
         # remove prompt from gen. tokens
         outputs_tokenized=[ tok_out[len(tok_in):] 
             for tok_in, tok_out in zip(prompts_tokenized["input_ids"], outputs_tokenized) ] 
 
         # count and decode gen. tokens 
         num_tokens=sum([ len(t) for t in outputs_tokenized ])
         outputs=tokenizer.batch_decode(outputs_tokenized)
 
         # store in results{} to be gathered by accelerate
         results["outputs"].extend(outputs)
         results["num_tokens"] += num_tokens
 
     results=[ results ] # transform to list, otherwise gather_object() will not collect correctly
 
 # collect results from all the GPUs
 results_gathered=gather_object(results)
 
 if accelerator.is_main_process:
     timediff=time.time()-start
     num_tokens=sum([r["num_tokens"] for r in results_gathered ])
 
     print(f"tokens/sec: {num_tokens//timediff}, time elapsed: {timediff}, num_tokens {num_tokens}")

可以看到批处理会大大加快速度。

1 GPU: 520 token /sec，时间:19.2s

2 gpu: 900 token /sec，时间:11.1s

3 gpu: 1205个token /秒，时间:8.2s

4 gpu: 1655 token /sec，时间:6.0s

5 gpu: 1658 token /sec，时间:6.0s

总结

截止到本文为止，llama.cpp，ctransformer还不支持多GPU推理，好像llama.cpp在6月有个多GPU的merge，但是我没看到官方更新，所以这里暂时确定不支持多GPU。如果有小伙伴确认可以支持多GPU请留言。

huggingface的Accelerate包则为我们使用多GPU提供了一个很方便的选择，使用多个GPU推理可以显着提高性能，但gpu之间通信的开销随着gpu数量的增加而显著增加。