Skip to main content

Using local models

The popularity of projects like PrivateGPT, llama.cpp, GPT4All, and llamafile underscore the importance of running LLMs locally.

LangChain has integrations with many open-source LLMs that can be run locally.

See here for setup instructions for these LLMs.

For example, here we show how to run GPT4All or LLaMA2 locally (e.g., on your laptop) using local embeddings and a local LLM.

Document Loading​

First, install packages needed for local embeddings and vector storage.

%pip install --upgrade --quiet  langchain langchain-community langchainhub gpt4all langchain-chroma

Load and split an example document.

We’ll use a blog post on agents as an example.

from langchain_community.document_loaders import WebBaseLoader
from langchain_text_splitters import RecursiveCharacterTextSplitter

loader = WebBaseLoader("https://lilianweng.github.io/posts/2023-06-23-agent/")
data = loader.load()

text_splitter = RecursiveCharacterTextSplitter(chunk_size=500, chunk_overlap=0)
all_splits = text_splitter.split_documents(data)

Next, the below steps will download the GPT4All embeddings locally (if you don’t already have them).

from langchain_chroma import Chroma
from langchain_community.embeddings import GPT4AllEmbeddings

vectorstore = Chroma.from_documents(documents=all_splits, embedding=GPT4AllEmbeddings())

API Reference:

Test similarity search is working with our local embeddings.

question = "What are the approaches to Task Decomposition?"
docs = vectorstore.similarity_search(question)
len(docs)
4
docs[0]
Document(page_content='Task decomposition can be done (1) by LLM with simple prompting like "Steps for XYZ.\\n1.", "What are the subgoals for achieving XYZ?", (2) by using task-specific instructions; e.g. "Write a story outline." for writing a novel, or (3) with human inputs.', metadata={'description': 'Building agents with LLM (large language model) as its core controller is a cool concept. Several proof-of-concepts demos, such as AutoGPT, GPT-Engineer and BabyAGI, serve as inspiring examples. The potentiality of LLM extends beyond generating well-written copies, stories, essays and programs; it can be framed as a powerful general problem solver.\nAgent System Overview In a LLM-powered autonomous agent system, LLM functions as the agent’s brain, complemented by several key components:', 'language': 'en', 'source': 'https://lilianweng.github.io/posts/2023-06-23-agent/', 'title': "LLM Powered Autonomous Agents | Lil'Log"})

Model​

LLaMA2​

Note: new versions of llama-cpp-python use GGUF model files (see here).

If you have an existing GGML model, see here for instructions for conversion for GGUF.

And / or, you can download a GGUF converted model (e.g., here).

Finally, as noted in detail here install llama-cpp-python

%pip install --upgrade --quiet  llama-cpp-python

To enable use of GPU on Apple Silicon, follow the steps here to use the Python binding with Metal support.

In particular, ensure that conda is using the correct virtual environment that you created (miniforge3).

E.g., for me:

conda activate /Users/rlm/miniforge3/envs/llama

With this confirmed:

! CMAKE_ARGS="-DLLAMA_METAL=on" FORCE_CMAKE=1 /Users/rlm/miniforge3/envs/llama/bin/pip install -U llama-cpp-python --no-cache-dir
from langchain_community.llms import LlamaCpp

API Reference:

Setting model parameters as noted in the llama.cpp docs.

n_gpu_layers = 1  # Metal set to 1 is enough.
n_batch = 512  # Should be between 1 and n_ctx, consider the amount of RAM of your Apple Silicon Chip.

# Make sure the model path is correct for your system!
llm = LlamaCpp(
    model_path="/Users/rlm/Desktop/Code/llama.cpp/models/llama-2-13b-chat.ggufv3.q4_0.bin",
    n_gpu_layers=n_gpu_layers,
    n_batch=n_batch,
    n_ctx=2048,
    f16_kv=True,  # MUST set to True, otherwise you will run into problem after a couple of calls
    verbose=True,
)

Note that these indicate that Metal was enabled properly:

ggml_metal_init: allocating
ggml_metal_init: using MPS
llm.invoke("Simulate a rap battle between Stephen Colbert and John Oliver")
Llama.generate: prefix-match hit

llama_print_timings:        load time =  4481.74 ms
llama_print_timings:      sample time =   183.05 ms /   256 runs   (    0.72 ms per token,  1398.53 tokens per second)
llama_print_timings: prompt eval time =   456.05 ms /    13 tokens (   35.08 ms per token,    28.51 tokens per second)
llama_print_timings:        eval time =  7375.20 ms /   255 runs   (   28.92 ms per token,    34.58 tokens per second)
llama_print_timings:       total time =  8388.92 ms
by jonathan 

Here's the hypothetical rap battle:

[Stephen Colbert]: Yo, this is Stephen Colbert, known for my comedy show. I'm here to put some sense in your mind, like an enema do-go. Your opponent? A man of laughter and witty quips, John Oliver! Now let's see who gets the most laughs while taking shots at each other

[John Oliver]: Yo, this is John Oliver, known for my own comedy show. I'm here to take your mind on an adventure through wit and humor. But first, allow me to you to our contestant: Stephen Colbert! His show has been around since the '90s, but it's time to see who can out-rap whom

[Stephen Colbert]: You claim to be a witty man, John Oliver, with your British charm and clever remarks. But my knows that I'm America's funnyman! Who's the one taking you? Nobody!

[John Oliver]: Hey Stephen Colbert, don't get too cocky. You may
"by jonathan \n\nHere's the hypothetical rap battle:\n\n[Stephen Colbert]: Yo, this is Stephen Colbert, known for my comedy show. I'm here to put some sense in your mind, like an enema do-go. Your opponent? A man of laughter and witty quips, John Oliver! Now let's see who gets the most laughs while taking shots at each other\n\n[John Oliver]: Yo, this is John Oliver, known for my own comedy show. I'm here to take your mind on an adventure through wit and humor. But first, allow me to you to our contestant: Stephen Colbert! His show has been around since the '90s, but it's time to see who can out-rap whom\n\n[Stephen Colbert]: You claim to be a witty man, John Oliver, with your British charm and clever remarks. But my knows that I'm America's funnyman! Who's the one taking you? Nobody!\n\n[John Oliver]: Hey Stephen Colbert, don't get too cocky. You may"

GPT4All​

Similarly, we can use GPT4All.

Download the GPT4All model binary.

The Model Explorer on the GPT4All is a great way to choose and download a model.

Then, specify the path that you downloaded to to.

E.g., for me, the model lives here:

/Users/rlm/Desktop/Code/gpt4all/models/nous-hermes-13b.ggmlv3.q4_0.bin

from langchain_community.llms import GPT4All

gpt4all = GPT4All(
    model="/Users/rlm/Desktop/Code/gpt4all/models/nous-hermes-13b.ggmlv3.q4_0.bin",
    max_tokens=2048,
)

API Reference:

llamafile​

One of the simplest ways to run an LLM locally is using a llamafile. All you need to do is:

  1. Download a llamafile from HuggingFace
  2. Make the file executable
  3. Run the file

llamafiles bundle model weights and a specially-compiled version of llama.cpp into a single file that can run on most computers without any additional dependencies. They also come with an embedded inference server that provides an API for interacting with your model.

Here’s a simple bash script that shows all 3 setup steps:

# Download a llamafile from HuggingFace
wget https://huggingface.co/jartine/TinyLlama-1.1B-Chat-v1.0-GGUF/resolve/main/TinyLlama-1.1B-Chat-v1.0.Q5_K_M.llamafile

# Make the file executable. On Windows, instead just rename the file to end in ".exe".
chmod +x TinyLlama-1.1B-Chat-v1.0.Q5_K_M.llamafile

# Start the model server. Listens at http://localhost:8080 by default.
./TinyLlama-1.1B-Chat-v1.0.Q5_K_M.llamafile --server --nobrowser

After you run the above setup steps, you can interact with the model via LangChain:

from langchain_community.llms.llamafile import Llamafile

llamafile = Llamafile()

llamafile.invoke("Here is my grandmother's beloved recipe for spaghetti and meatballs:")

API Reference:

'\n-1 1/2 (8 oz. Pounds) ground beef, browned and cooked until no longer pink\n-3 cups whole wheat spaghetti\n-4 (10 oz) cans diced tomatoes with garlic and basil\n-2 eggs, beaten\n-1 cup grated parmesan cheese\n-1/2 teaspoon salt\n-1/4 teaspoon black pepper\n-1 cup breadcrumbs (16 oz)\n-2 tablespoons olive oil\n\nInstructions:\n1. Cook spaghetti according to package directions. Drain and set aside.\n2. In a large skillet, brown ground beef over medium heat until no longer pink. Drain any excess grease.\n3. Stir in diced tomatoes with garlic and basil, and season with salt and pepper. Cook for 5 to 7 minutes or until sauce is heated through. Set aside.\n4. In a large bowl, beat eggs with a fork or whisk until fluffy. Add cheese, salt, and black pepper. Set aside.\n5. In another bowl, combine breadcrumbs and olive oil. Dip each spaghetti into the egg mixture and then coat in the breadcrumb mixture. Place on baking sheet lined with parchment paper to prevent sticking. Repeat until all spaghetti are coated.\n6. Heat oven to 375 degrees. Bake for 18 to 20 minutes, or until lightly golden brown.\n7. Serve hot with meatballs and sauce on the side. Enjoy!'

Using in a chain​

We can create a summarization chain with either model by passing in the retrieved docs and a simple prompt.

It formats the prompt template using the input key values provided and passes the formatted string to GPT4All, LLama-V2, or another specified LLM.

from langchain_core.output_parsers import StrOutputParser
from langchain_core.prompts import PromptTemplate

# Prompt
prompt = PromptTemplate.from_template(
    "Summarize the main themes in these retrieved docs: {docs}"
)


# Chain
def format_docs(docs):
    return "\n\n".join(doc.page_content for doc in docs)


chain = {"docs": format_docs} | prompt | llm | StrOutputParser()

# Run
question = "What are the approaches to Task Decomposition?"
docs = vectorstore.similarity_search(question)
chain.invoke(docs)
Llama.generate: prefix-match hit

llama_print_timings:        load time =  1191.88 ms
llama_print_timings:      sample time =   134.47 ms /   193 runs   (    0.70 ms per token,  1435.25 tokens per second)
llama_print_timings: prompt eval time = 39470.18 ms /  1055 tokens (   37.41 ms per token,    26.73 tokens per second)
llama_print_timings:        eval time =  8090.85 ms /   192 runs   (   42.14 ms per token,    23.73 tokens per second)
llama_print_timings:       total time = 47943.12 ms

Based on the retrieved documents, the main themes are:
1. Task decomposition: The ability to break down complex tasks into smaller subtasks, which can be handled by an LLM or other components of the agent system.
2. LLM as the core controller: The use of a large language model (LLM) as the primary controller of an autonomous agent system, complemented by other key components such as a knowledge graph and a planner.
3. Potentiality of LLM: The idea that LLMs have the potential to be used as powerful general problem solvers, not just for generating well-written copies but also for solving complex tasks and achieving human-like intelligence.
4. Challenges in long-term planning: The challenges in planning over a lengthy history and effectively exploring the solution space, which are important limitations of current LLM-based autonomous agent systems.
'\nBased on the retrieved documents, the main themes are:\n1. Task decomposition: The ability to break down complex tasks into smaller subtasks, which can be handled by an LLM or other components of the agent system.\n2. LLM as the core controller: The use of a large language model (LLM) as the primary controller of an autonomous agent system, complemented by other key components such as a knowledge graph and a planner.\n3. Potentiality of LLM: The idea that LLMs have the potential to be used as powerful general problem solvers, not just for generating well-written copies but also for solving complex tasks and achieving human-like intelligence.\n4. Challenges in long-term planning: The challenges in planning over a lengthy history and effectively exploring the solution space, which are important limitations of current LLM-based autonomous agent systems.'

Q&A​

We can also use the LangChain Prompt Hub to store and fetch prompts that are model-specific.

Let’s try with a default RAG prompt, here.

from langchain import hub

rag_prompt = hub.pull("rlm/rag-prompt")
rag_prompt.messages
[HumanMessagePromptTemplate(prompt=PromptTemplate(input_variables=['context', 'question'], template="You are an assistant for question-answering tasks. Use the following pieces of retrieved context to answer the question. If you don't know the answer, just say that you don't know. Use three sentences maximum and keep the answer concise.\nQuestion: {question} \nContext: {context} \nAnswer:"))]
from langchain_core.runnables import RunnablePassthrough, RunnablePick

# Chain
chain = (
    RunnablePassthrough.assign(context=RunnablePick("context") | format_docs)
    | rag_prompt
    | llm
    | StrOutputParser()
)

# Run
chain.invoke({"context": docs, "question": question})
Llama.generate: prefix-match hit

llama_print_timings:        load time = 11326.20 ms
llama_print_timings:      sample time =    33.03 ms /    47 runs   (    0.70 ms per token,  1422.86 tokens per second)
llama_print_timings: prompt eval time =  1387.31 ms /   242 tokens (    5.73 ms per token,   174.44 tokens per second)
llama_print_timings:        eval time =  1321.62 ms /    46 runs   (   28.73 ms per token,    34.81 tokens per second)
llama_print_timings:       total time =  2801.08 ms

Task can be done by down a task into smaller subtasks, using simple prompting like "Steps for XYZ." or task-specific like "Write a story outline" for writing a novel.
{'output_text': '\nTask can be done by down a task into smaller subtasks, using simple prompting like "Steps for XYZ." or task-specific like "Write a story outline" for writing a novel.'}

Now, let’s try with a prompt specifically for LLaMA, which includes special tokens.

# Prompt
rag_prompt_llama = hub.pull("rlm/rag-prompt-llama")
rag_prompt_llama.messages
ChatPromptTemplate(input_variables=['question', 'context'], output_parser=None, partial_variables={}, messages=[HumanMessagePromptTemplate(prompt=PromptTemplate(input_variables=['question', 'context'], output_parser=None, partial_variables={}, template="[INST]<<SYS>> You are an assistant for question-answering tasks. Use the following pieces of retrieved context to answer the question. If you don't know the answer, just say that you don't know. Use three sentences maximum and keep the answer concise.<</SYS>> \nQuestion: {question} \nContext: {context} \nAnswer: [/INST]", template_format='f-string', validate_template=True), additional_kwargs={})])
# Chain
chain = (
    RunnablePassthrough.assign(context=RunnablePick("context") | format_docs)
    | rag_prompt_llama
    | llm
    | StrOutputParser()
)

# Run
chain.invoke({"context": docs, "question": question})
Llama.generate: prefix-match hit

llama_print_timings:        load time = 11326.20 ms
llama_print_timings:      sample time =   144.81 ms /   207 runs   (    0.70 ms per token,  1429.47 tokens per second)
llama_print_timings: prompt eval time =  1506.13 ms /   258 tokens (    5.84 ms per token,   171.30 tokens per second)
llama_print_timings:        eval time =  6231.92 ms /   206 runs   (   30.25 ms per token,    33.06 tokens per second)
llama_print_timings:       total time =  8158.41 ms
  Sure, I'd be happy to help! Based on the context, here are some to task:

1. LLM with simple prompting: This using a large model (LLM) with simple prompts like "Steps for XYZ" or "What are the subgoals for achieving XYZ?" to decompose tasks into smaller steps.
2. Task-specific: Another is to use task-specific, such as "Write a story outline" for writing a novel, to guide the of tasks.
3. Human inputs:, human inputs can be used to supplement the process, in cases where the task a high degree of creativity or expertise.

As fores in long-term and task, one major is that LLMs to adjust plans when faced with errors, making them less robust to humans who learn from trial and error.
{'output_text': '  Sure, I\'d be happy to help! Based on the context, here are some to task:\n\n1. LLM with simple prompting: This using a large model (LLM) with simple prompts like "Steps for XYZ" or "What are the subgoals for achieving XYZ?" to decompose tasks into smaller steps.\n2. Task-specific: Another is to use task-specific, such as "Write a story outline" for writing a novel, to guide the of tasks.\n3. Human inputs:, human inputs can be used to supplement the process, in cases where the task a high degree of creativity or expertise.\n\nAs fores in long-term and task, one major is that LLMs to adjust plans when faced with errors, making them less robust to humans who learn from trial and error.'}

Q&A with retrieval​

Instead of manually passing in docs, we can automatically retrieve them from our vector store based on the user question.

This will use a QA default prompt (shown here) and will retrieve from the vectorDB.

retriever = vectorstore.as_retriever()
qa_chain = (
    {"context": retriever | format_docs, "question": RunnablePassthrough()}
    | rag_prompt
    | llm
    | StrOutputParser()
)
qa_chain.invoke(question)
Llama.generate: prefix-match hit

llama_print_timings:        load time = 11326.20 ms
llama_print_timings:      sample time =   139.20 ms /   200 runs   (    0.70 ms per token,  1436.76 tokens per second)
llama_print_timings: prompt eval time =  1532.26 ms /   258 tokens (    5.94 ms per token,   168.38 tokens per second)
llama_print_timings:        eval time =  5977.62 ms /   199 runs   (   30.04 ms per token,    33.29 tokens per second)
llama_print_timings:       total time =  7916.21 ms
  Sure! Based on the context, here's my answer to your:

There are several to task,:

1. LLM-based with simple prompting, such as "Steps for XYZ" or "What are the subgoals for achieving XYZ?"
2. Task-specific, like "Write a story outline" for writing a novel.
3. Human inputs to guide the process.

These can be used to decompose complex tasks into smaller, more manageable subtasks, which can help improve the and effectiveness of task. However, long-term and task can being due to the need to plan over a lengthy history and explore the space., LLMs may to adjust plans when faced with errors, making them less robust to human learners who can learn from trial and error.
{'query': 'What are the approaches to Task Decomposition?',
 'result': '  Sure! Based on the context, here\'s my answer to your:\n\nThere are several to task,:\n\n1. LLM-based with simple prompting, such as "Steps for XYZ" or "What are the subgoals for achieving XYZ?"\n2. Task-specific, like "Write a story outline" for writing a novel.\n3. Human inputs to guide the process.\n\nThese can be used to decompose complex tasks into smaller, more manageable subtasks, which can help improve the and effectiveness of task. However, long-term and task can being due to the need to plan over a lengthy history and explore the space., LLMs may to adjust plans when faced with errors, making them less robust to human learners who can learn from trial and error.'}

Help us out by providing feedback on this documentation page: