Llama2Chat
此筆記本示範如何使用 Llama2Chat 包裝器擴增 Llama-2 LLM，以支援 Llama-2 聊天提示格式。LangChain 中的數種 LLM 實作可用作 Llama-2 聊天模型的介面。這些包括 ChatHuggingFace、LlamaCpp、GPT4All，...，僅舉幾個例子。
Llama2Chat 是一個通用包裝器，實作 BaseChatModel，因此可以在應用程式中用作聊天模型。Llama2Chat 將訊息清單轉換為必要的聊天提示格式，並將格式化的提示作為 str 轉發到包裝的 LLM。
from langchain.chains import LLMChain
from langchain.memory import ConversationBufferMemory
from langchain_experimental.chat_models import Llama2Chat
對於以下的聊天應用程式範例，我們將使用以下聊天 prompt_template
from langchain_core.messages import SystemMessage
from langchain_core.prompts.chat import (
ChatPromptTemplate,
HumanMessagePromptTemplate,
MessagesPlaceholder,
)
template_messages = [
SystemMessage(content="You are a helpful assistant."),
MessagesPlaceholder(variable_name="chat_history"),
HumanMessagePromptTemplate.from_template("{text}"),
]
prompt_template = ChatPromptTemplate.from_messages(template_messages)
透過 HuggingFaceTextGenInference LLM 與 Llama-2 聊天
HuggingFaceTextGenInference LLM 封裝了對 text-generation-inference 伺服器的存取。在以下範例中，推論伺服器服務 meta-llama/Llama-2-13b-chat-hf 模型。它可以使用以下命令在本機啟動：
docker run \
--rm \
--gpus all \
--ipc=host \
-p 8080:80 \
-v ~/.cache/huggingface/hub:/data \
-e HF_API_TOKEN=${HF_API_TOKEN} \
ghcr.io/huggingface/text-generation-inference:0.9 \
--hostname 0.0.0.0 \
--model-id meta-llama/Llama-2-13b-chat-hf \
--quantize bitsandbytes \
--num-shard 4
例如，這適用於配備 4 個 RTX 3080ti 顯示卡的機器。調整 --num_shard 值以符合可用的 GPU 數量。HF_API_TOKEN 環境變數保存 Hugging Face API 權杖。
# !pip3 install text-generation
建立連接到本機推論伺服器的 HuggingFaceTextGenInference 執行個體，並將其包裝到 Llama2Chat 中。
from langchain_community.llms import HuggingFaceTextGenInference
llm = HuggingFaceTextGenInference(
inference_server_url="http://127.0.0.1:8080/",
max_new_tokens=512,
top_k=50,
temperature=0.1,
repetition_penalty=1.03,
)
model = Llama2Chat(llm=llm)
API 參考：HuggingFaceTextGenInference
然後，您就可以在 LLMChain 中一起使用聊天 model 以及 prompt_template 和對話 memory。
memory = ConversationBufferMemory(memory_key="chat_history", return_messages=True)
chain = LLMChain(llm=model, prompt=prompt_template, memory=memory)
print(
chain.run(
text="What can I see in Vienna? Propose a few locations. Names only, no details."
)
)
Sure, I'd be happy to help! Here are a few popular locations to consider visiting in Vienna:
1. Schönbrunn Palace
2. St. Stephen's Cathedral
3. Hofburg Palace
4. Belvedere Palace
5. Prater Park
6. Vienna State Opera
7. Albertina Museum
8. Museum of Natural History
9. Kunsthistorisches Museum
10. Ringstrasse
print(chain.run(text="Tell me more about #2."))
Certainly! St. Stephen's Cathedral (Stephansdom) is one of the most recognizable landmarks in Vienna and a must-see attraction for visitors. This stunning Gothic cathedral is located in the heart of the city and is known for its intricate stone carvings, colorful stained glass windows, and impressive dome.
The cathedral was built in the 12th century and has been the site of many important events throughout history, including the coronation of Holy Roman emperors and the funeral of Mozart. Today, it is still an active place of worship and offers guided tours, concerts, and special events. Visitors can climb up the south tower for panoramic views of the city or attend a service to experience the beautiful music and chanting.
透過 LlamaCPP LLM 與 Llama-2 聊天
若要將 Llama-2 聊天模型與 LlamaCPP LMM 搭配使用，請使用這些安裝指示安裝 llama-cpp-python 程式庫。以下範例使用量化的 llama-2-7b-chat.Q4_0.gguf 模型，該模型在本機儲存在 ~/Models/llama-2-7b-chat.Q4_0.gguf。
在建立 LlamaCpp 執行個體之後，llm 再次包裝到 Llama2Chat 中
from os.path import expanduser
from langchain_community.llms import LlamaCpp
model_path = expanduser("~/Models/llama-2-7b-chat.Q4_0.gguf")
llm = LlamaCpp(
model_path=model_path,
streaming=False,
)
model = Llama2Chat(llm=llm)
API 參考：LlamaCpp
並以與先前範例相同的方式使用。
memory = ConversationBufferMemory(memory_key="chat_history", return_messages=True)
chain = LLMChain(llm=model, prompt=prompt_template, memory=memory)
print(
chain.run(
text="What can I see in Vienna? Propose a few locations. Names only, no details."
)
)
Of course! Vienna is a beautiful city with a rich history and culture. Here are some of the top tourist attractions you might want to consider visiting:
1. Schönbrunn Palace
2. St. Stephen's Cathedral
3. Hofburg Palace
4. Belvedere Palace
5. Prater Park
6. MuseumsQuartier
7. Ringstrasse
8. Vienna State Opera
9. Kunsthistorisches Museum
10. Imperial Palace
These are just a few of the many amazing places to see in Vienna. Each one has its own unique history and charm, so I hope you enjoy exploring this beautiful city!
``````output
llama_print_timings: load time = 250.46 ms
llama_print_timings: sample time = 56.40 ms / 144 runs ( 0.39 ms per token, 2553.37 tokens per second)
llama_print_timings: prompt eval time = 1444.25 ms / 47 tokens ( 30.73 ms per token, 32.54 tokens per second)
llama_print_timings: eval time = 8832.02 ms / 143 runs ( 61.76 ms per token, 16.19 tokens per second)
llama_print_timings: total time = 10645.94 ms
print(chain.run(text="Tell me more about #2."))
Llama.generate: prefix-match hit
``````output
Of course! St. Stephen's Cathedral (also known as Stephansdom) is a stunning Gothic-style cathedral located in the heart of Vienna, Austria. It is one of the most recognizable landmarks in the city and is considered a symbol of Vienna.
Here are some interesting facts about St. Stephen's Cathedral:
1. History: The construction of St. Stephen's Cathedral began in the 12th century on the site of a former Romanesque church, and it took over 600 years to complete. The cathedral has been renovated and expanded several times throughout its history, with the most significant renovation taking place in the 19th century.
2. Architecture: St. Stephen's Cathedral is built in the Gothic style, characterized by its tall spires, pointed arches, and intricate stone carvings. The cathedral features a mix of Romanesque, Gothic, and Baroque elements, making it a unique blend of styles.
3. Design: The cathedral's design is based on the plan of a cross with a long nave and two shorter arms extending from it. The main altar is
``````output
llama_print_timings: load time = 250.46 ms
llama_print_timings: sample time = 100.60 ms / 256 runs ( 0.39 ms per token, 2544.73 tokens per second)
llama_print_timings: prompt eval time = 5128.71 ms / 160 tokens ( 32.05 ms per token, 31.20 tokens per second)
llama_print_timings: eval time = 16193.02 ms / 255 runs ( 63.50 ms per token, 15.75 tokens per second)
llama_print_timings: total time = 21988.57 ms