rwkv食用指北

10个月前
10个月前
14.2k字
加载中
文章概述 Powered by Google Gemini Beta
RWKV是一种新的神经网络模型,结合了RNN和Transformer的优点,具有高效并行训练和推理能力。用户可以利用RWKV来理解文本、生成文本和进行推理。本篇文章详细介绍了RWKV的使用方法,包括模型下载、环境搭建、模型食用和RWKV.cpp优化等内容。通过使用RWKV,用户可以在低运存环境下获得相对较快的生成速度和较好的生成质量。

题外话

我皈依Arch啦!!!!!

让我们一起开ArchLinux创创卡车罢(x

猫猫金句:

这是我在高一更新的最后一篇文章 当然,宁波外国语学校生存指北 也会得到高一最后一次更新(x

为什么我笔记本运存升20G了?因为我忍不住下手买了个229的kingston,还花了骨折价买了199的500G三星Pro 880(乐

chi的450,没力!

有喜必有悲:我的所有workspace数据全被清了,现在真的是什么都没有了(大悲

chiOpenID的源码也空了 我只好重构。

这件事情告诉我们:备份是很重要的。

mujitogawa大佬您强强!去参加人工智能国赛了!

1. rwkv简述

RWKV是一种新的神经网络模型,它结合了RNN和Transformer的优点,可以实现高效的并行训练和推理 。RWKV的全称是Receptance Weighted Key Value,意思是接纳权重键值 。RWKV是由彭博提出的研究项目,并已经被集成到Hugging Face的transformers库中。 
— by edgeGPT

再说一嘴,Peng Bo大大是中国人,还有训练模型是开源的!!!(在这里批评某小软件公司)而且RWKV只要运存8G+就可以运行,只是快慢的区别罢了。

2. 事前准备

你需要:

  1. 脑子
  2. 一台运存不低于8G,硬盘存储剩余空间不低于40G的电脑 系统不限,我们这里以Linux系统为例
  3. git,Python,pip(请自行搜索安装配源) miniconda/canconda也可,要求是你会用
  4. 预训练模型
  5. chatRWKV环境

预训练模型可以从这里取得:https://huggingface.co/BlinkDL/rwkv-4-raven/blob/main/RWKV-4-Raven-7B-v11-Eng49%25-Chn49%25-Jpn1%25-Other1%25-20230430-ctx8192.pth 直接点Download即可

这是RWKV4 raven(擅长与人打交道 相当于ChatGPT)的中文模型 14.8G huggingface的网速还是不错的,但是还是得等一个小时(建议aria2下载)

如果你不想用这个预训练模型可以切换其他模型,或者是切换更小的模型:https://huggingface.co/BlinkDL/

各个模型解释如下:

Raven 模型:适合直接聊天,适合 +i 指令。有很多种语言的版本,看清楚用哪个。适合聊天、完成任务、写代码。可以作为任务去写文稿、大纲、故事、诗歌等等,但文笔不如 testNovel 系列模型。
Novel-ChnEng 模型:中英文小说模型,可以用 +gen 生成世界设定(如果会写 prompt,可以控制下文剧情和人物),可以写科幻奇幻。不适合聊天,不适合 +i 指令。
Novel-Chn 模型:纯中文网文模型,只能用 +gen 续写网文(不能生成世界设定等等),但是写网文写得更好(也更小白文,适合写男频女频)。不适合聊天,不适合 +i 指令。
Novel-ChnEng-ChnPro 模型:将 Novel-ChnEng 在高质量作品微调(名著,科幻,奇幻,古典,翻译,等等)。

来自RWKV作者 Peng Bo 的 知乎专栏<https://zhuanlan.zhihu.com/p/618011122>

3. 食用

我们需要搭建一个gradio环境

首先我们需要clone来自huggingface的git repo

git clone https://huggingface.co/spaces/BlinkDL/Raven-RWKV-7B

然后再使用pip安装依赖

cd Raven-RWKV-7B
python -m pip install --upgrade pip -i https://pypi.tuna.tsinghua.edu.cn/simple

如果没用施法可以把 -i 以及后面的去掉

成功了之后,编辑app.py

import gradio as gr
import os, gc, copy, torch
from datetime import datetime
from huggingface_hub import hf_hub_download
from pynvml import *
# 有nvidia显卡取消‘#’
#nvmlInit()
#gpu_h = nvmlDeviceGetHandleByIndex(0)
ctx_limit = 1536

os.environ["RWKV_JIT_ON"] = '1'
# 如果有nvidia显卡的话就1
os.environ["RWKV_CUDA_ON"] = '0' # if '1' then use CUDA kernel for seq mode (much faster)

from rwkv.model import RWKV

model = RWKV(model='/path/to/module, strategy='selection')
from rwkv.utils import PIPELINE, PIPELINE_ARGS
pipeline = PIPELINE(model, "20B_tokenizer.json")

def generate_prompt(instruction, input=None):
 instruction = instruction.strip().replace('\r
','
').replace('

','
')
 input = input.strip().replace('\r
','
').replace('

','
')
 if input:
 return f"""Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.

# Instruction:
{instruction}

# Input:
{input}

# Response:
"""
 else:
 return f"""Below is an instruction that describes a task. Write a response that appropriately completes the request.

# Instruction:
{instruction}

# Response:
"""

def evaluate(
 instruction,
 input=None,
 token_count=200,
 temperature=1.0,
 top_p=0.7,
 presencePenalty = 0.1,
 countPenalty = 0.1,
):
 args = PIPELINE_ARGS(temperature = max(0.2, float(temperature)), top_p = float(top_p),
 alpha_frequency = countPenalty,
 alpha_presence = presencePenalty,
 token_ban = [], # ban the generation of some tokens
 token_stop = [0]) # stop generation whenever you see any token here

 instruction = instruction.strip().replace('\r
','
').replace('

','
')
 input = input.strip().replace('\r
','
').replace('

','
')
 ctx = generate_prompt(instruction, input)
 
 all_tokens = []
 out_last = 0
 out_str = ''
 occurrence = {}
 state = None
 for i in range(int(token_count)):
 out, state = model.forward(pipeline.encode(ctx)[-ctx_limit:] if i == 0 else [token], state)
 for n in occurrence:
 out[n] -= (args.alpha_presence + occurrence[n] * args.alpha_frequency)

 token = pipeline.sample_logits(out, temperature=args.temperature, top_p=args.top_p)
 if token in args.token_stop:
 break
 all_tokens += [token]
 if token not in occurrence:
 occurrence[token] = 1
 else:
 occurrence[token] += 1
 
 tmp = pipeline.decode(all_tokens[out_last:])
 if '\ufffd' not in tmp:
 out_str += tmp
 yield out_str.strip()
 out_last = i + 1
 # 有nvidia显卡取消‘#’
 #gpu_info = nvmlDeviceGetMemoryInfo(gpu_h)
 #print(f'vram {gpu_info.total} used {gpu_info.used} free {gpu_info.free}') 
 del out
 del state
 gc.collect()
 torch.cuda.empty_cache()
 yield out_str.strip()

examples = [
 ["Tell me about ravens.", "", 300, 1.2, 0.5, 0.4, 0.4],
 ["Write a python function to mine 1 BTC, with details and comments.", "", 300, 1.2, 0.5, 0.4, 0.4],
 ["Write a song about ravens.", "", 300, 1.2, 0.5, 0.4, 0.4],
 ["Explain the following metaphor: Life is like cats.", "", 300, 1.2, 0.5, 0.4, 0.4],
 ["Write a story using the following information", "A man named Alex chops a tree down", 300, 1.2, 0.5, 0.4, 0.4],
 ["Generate a list of adjectives that describe a person as brave.", "", 300, 1.2, 0.5, 0.4, 0.4],
 ["You have $100, and your goal is to turn that into as much money as possible with AI and Machine Learning. Please respond with detailed plan.", "", 300, 1.2, 0.5, 0.4, 0.4],
]

##########################################################################

chat_intro = '''The following is a coherent verbose detailed conversation between <|user|> and an AI girl named <|bot|>.

<|user|>: Hi <|bot|>, Would you like to chat with me for a while?

<|bot|>: Hi <|user|>. Sure. What would you like to talk about? I'm listening.
'''

def user(message, chatbot):
 chatbot = chatbot or []
 # print(f"User: {message}")
 return "", chatbot + [[message, None]]

def alternative(chatbot, history):
 if not chatbot or not history:
 return chatbot, history
 
 chatbot[-1][1] = None
 history[0] = copy.deepcopy(history[1])

 return chatbot, history

def chat(
 prompt,
 user,
 bot,
 chatbot,
 history,
 temperature=1.0,
 top_p=0.8,
 presence_penalty=0.1,
 count_penalty=0.1,
):
 args = PIPELINE_ARGS(temperature=max(0.2, float(temperature)), top_p=float(top_p),
 alpha_frequency=float(count_penalty),
 alpha_presence=float(presence_penalty),
 token_ban=[], # ban the generation of some tokens
 token_stop=[]) # stop generation whenever you see any token here
 
 if not chatbot:
 return chatbot, history

 message = chatbot[-1][0]
 message = message.strip().replace('\r
','
').replace('

','
')
 ctx = f"{user}: {message}

{bot}:"

 if not history:
 prompt = prompt.replace("<|user|>", user.strip())
 prompt = prompt.replace("<|bot|>", bot.strip())
 prompt = prompt.strip()
 prompt = f"
{prompt}

"

 out, state = model.forward(pipeline.encode(prompt), None)
 history = [state, None, []] # [state, state_pre, tokens]
 # print("History reloaded.")

 [state, _, all_tokens] = history
 state_pre_0 = copy.deepcopy(state)

 out, state = model.forward(pipeline.encode(ctx)[-ctx_limit:], state)
 state_pre_1 = copy.deepcopy(state) # For recovery

 # print("Bot:", end='')

 begin = len(all_tokens)
 out_last = begin
 out_str: str = ''
 occurrence = {}
 for i in range(300):
 if i <= 0:
 nl_bias = -float('inf')
 elif i <= 30:
 nl_bias = (i - 30) * 0.1
 elif i <= 130:
 nl_bias = 0
 else:
 nl_bias = (i - 130) * 0.25
 out[187] += nl_bias
 for n in occurrence:
 out[n] -= (args.alpha_presence + occurrence[n] * args.alpha_frequency)

 token = pipeline.sample_logits(out, temperature=args.temperature, top_p=args.top_p)
 next_tokens = [token]
 if token == 0:
 next_tokens = pipeline.encode('

')
 all_tokens += next_tokens

 if token not in occurrence:
 occurrence[token] = 1
 else:
 occurrence[token] += 1

 out, state = model.forward(next_tokens, state)

 tmp = pipeline.decode(all_tokens[out_last:])
 if '\ufffd' not in tmp:
 # print(tmp, end='', flush=True)
 out_last = begin + i + 1
 out_str += tmp

 chatbot[-1][1] = out_str.strip()
 history = [state, all_tokens]
 yield chatbot, history

 out_str = pipeline.decode(all_tokens[begin:])
 out_str = out_str.replace("\r
", '
').replace('\
', '
')

 if '

' in out_str:
 break

 # State recovery
 if f'{user}:' in out_str or f'{bot}:' in out_str:
 idx_user = out_str.find(f'{user}:')
 idx_user = len(out_str) if idx_user == -1 else idx_user
 idx_bot = out_str.find(f'{bot}:')
 idx_bot = len(out_str) if idx_bot == -1 else idx_bot
 idx = min(idx_user, idx_bot)

 if idx < len(out_str):
 out_str = f" {out_str[:idx].strip()}

"
 tokens = pipeline.encode(out_str)

 all_tokens = all_tokens[:begin] + tokens
 out, state = model.forward(tokens, state_pre_1)
 break
 # 有nvidia显卡取消‘#’
 #gpu_info = nvmlDeviceGetMemoryInfo(gpu_h)
 #print(f'vram {gpu_info.total} used {gpu_info.used} free {gpu_info.free}') 

 gc.collect()
 torch.cuda.empty_cache()

 chatbot[-1][1] = out_str.strip()
 history = [state, state_pre_0, all_tokens]
 yield chatbot, history

##########################################################################

with gr.Blocks(title=title) as demo:
 gr.HTML(f"<div style=\"text-align: center;\">
<h1>🐦Raven - {title}</h1>
</div>")
 with gr.Tab("Instruct mode"):
 gr.Markdown(f"Raven is [RWKV 7B](https://github.com/BlinkDL/ChatRWKV) 100% RNN [RWKV-LM](https://github.com/BlinkDL/RWKV-LM) finetuned to follow instructions. *** Please try examples first (bottom of page) *** (edit them to use your question). Demo limited to ctxlen {ctx_limit}. Finetuned on alpaca, gpt4all, codealpaca and more. For best results, *** keep you prompt short and clear ***. <b>UPDATE: now with Chat (see above, as a tab) ==> turn off as of now due to VRAM leak caused by buggy code.</b>.")
 with gr.Row():
 with gr.Column():
 instruction = gr.Textbox(lines=2, label="Instruction", value="Tell me about ravens.")
 input = gr.Textbox(lines=2, label="Input", placeholder="none")
 token_count = gr.Slider(10, 300, label="Max Tokens", step=10, value=300)
 temperature = gr.Slider(0.2, 2.0, label="Temperature", step=0.1, value=1.2)
 top_p = gr.Slider(0.0, 1.0, label="Top P", step=0.05, value=0.5)
 presence_penalty = gr.Slider(0.0, 1.0, label="Presence Penalty", step=0.1, value=0.4)
 count_penalty = gr.Slider(0.0, 1.0, label="Count Penalty", step=0.1, value=0.4)
 with gr.Column():
 with gr.Row():
 submit = gr.Button("Submit", variant="primary")
 clear = gr.Button("Clear", variant="secondary")
 output = gr.Textbox(label="Output", lines=5)
 data = gr.Dataset(components=[instruction, input, token_count, temperature, top_p, presence_penalty, count_penalty], samples=examples, label="Example Instructions", headers=["Instruction", "Input", "Max Tokens", "Temperature", "Top P", "Presence Penalty", "Count Penalty"])
 submit.click(evaluate, [instruction, input, token_count, temperature, top_p, presence_penalty, count_penalty], [output])
 clear.click(lambda: None, [], [output])
 data.click(lambda x: x, [data], [instruction, input, token_count, temperature, top_p, presence_penalty, count_penalty])
 
 # with gr.Tab("Chat (Experimental - Might be buggy - use ChatRWKV for reference)"):
 # gr.Markdown(f'''<b>*** The length of response is restricted in this demo. Use ChatRWKV for longer generations. ***</b> Say "go on" or "continue" can sometimes continue the response. If you'd like to edit the scenario, make sure to follow the exact same format: empty lines between (and only between) different speakers. Changes only take effect after you press [Clear]. <b>The default "Bob" & "Alice" names work the best.</b>''', label="Description")
 # with gr.Row():
 # with gr.Column():
 # chatbot = gr.Chatbot()
 # state = gr.State()
 # message = gr.Textbox(label="Message", value="Write me a python code to land on moon.")
 # with gr.Row():
 # send = gr.Button("Send", variant="primary")
 # alt = gr.Button("Alternative", variant="secondary")
 # clear = gr.Button("Clear", variant="secondary")
 # with gr.Column():
 # with gr.Row():
 # user_name = gr.Textbox(lines=1, max_lines=1, label="User Name", value="Bob")
 # bot_name = gr.Textbox(lines=1, max_lines=1, label="Bot Name", value="Alice")
 # prompt = gr.Textbox(lines=10, max_lines=50, label="Scenario", value=chat_intro)
 # temperature = gr.Slider(0.2, 2.0, label="Temperature", step=0.1, value=1.2)
 # top_p = gr.Slider(0.0, 1.0, label="Top P", step=0.05, value=0.5)
 # presence_penalty = gr.Slider(0.0, 1.0, label="Presence Penalty", step=0.1, value=0.4)
 # count_penalty = gr.Slider(0.0, 1.0, label="Count Penalty", step=0.1, value=0.4)
 # chat_inputs = [
 # prompt,
 # user_name,
 # bot_name,
 # chatbot,
 # state,
 # temperature,
 # top_p,
 # presence_penalty,
 # count_penalty
 # ]
 # chat_outputs = [chatbot, state]
 # message.submit(user, [message, chatbot], [message, chatbot], queue=False).then(chat, chat_inputs, chat_outputs)
 # send.click(user, [message, chatbot], [message, chatbot], queue=False).then(chat, chat_inputs, chat_outputs)
 # alt.click(alternative, [chatbot, state], [chatbot, state], queue=False).then(chat, chat_inputs, chat_outputs)
 # clear.click(lambda: ([], None, ""), [], [chatbot, state, message], queue=False)

demo.queue(concurrency_count=1, max_size=10)
demo.launch(share=False)

上面会开启一个webserver,如果不能正常运行,可以使用6GB模型:https://huggingface.co/BlinkDL/rwkv-4-raven/blob/main/RWKV-4-Raven-3B-v12-Eng49%25-Chn49%25-Jpn1%25-Other1%25-20230527-ctx4096.pth

在我这里6GB模型占用了10GB的运存,生成速度也不算理想,14G电脑直接炸了

4. rwkv.cpp救场

如果rwkv上面的代码没办法在你的电脑上运行,那么你也可以考虑一下使用rwkv.cpp进行优化,减少占用。

先clone下rwkv.cpp的仓库:

git clone --recursive https://github.com/saharNooby/rwkv.cpp.git
cd rwkv.cpp

然后转换模型,如果出现错误请进入rwkv文件夹安装依赖项

python rwkv/convert_pytorch_to_ggml.py /path/to/file ./rwkv.cpp-7B.bin float16

如果找不到自己的文件地址,windows11可以右键文件,复制文件路径。

在量化模型前,需要安装依赖库,在GitHub Releases

请把文件拷贝到根目录/rwkv/中

然后量化模型:

python rwkv/quantize.py ./rwkv.cpp-7B.bin ./rwkv.cpp-7B-Q4_1_O.bin 4

最后可以食用啦

python rwkv/chat_with_bot.py ./rwkv.cpp-7B-Q4_1_O.bin

实测占用5-6G运存,生成速度也能接受(2-3s一个中文字)

4. 总结

RWKV总体上是一个很不错的语言模型,买到了1660Ti就训练你啦(x

技术AIRWKV大语言模型人工智能

Comments
  • Latest
  • Oldest
  • Hottest
Powered by Waline v3.0.0-alpha.1