L2M2: A Simple Python LLM Manager 💬👍

L2M2 ("LLM Manager" → "LLMM" → "L2M2") is a tiny and very simple LLM manager for Python that exposes lots of models through a unified API. This is useful for evaluation, demos, production applications etc. that need to easily be model-agnostic.

Features

• 31 supported models (see below) – regularly updated and with more on the way.
• Session chat memory – even across multiple models or with concurrent memory streams.
• JSON mode
• Prompt loading tools

Advantages

• Simple: Completely unified interface – just swap out the model name.
• Tiny: Only two external dependencies (httpx and typing_extensions).
• Fast: Fully asynchronous if concurrent calls are needed.

Supported Models

L2M2 currently supports the following models:

Model Name	Provider(s)	Model Version(s)
gpt-4o	OpenAI	gpt-4o-2024-11-20
gpt-4o-mini	OpenAI	gpt-4o-mini-2024-07-18
o1	OpenAI	o1
o1-preview	OpenAI	o1-preview
o1-mini	OpenAI	o1-mini
gpt-4-turbo	OpenAI	gpt-4-turbo-2024-04-09
gpt-3.5-turbo	OpenAI	gpt-3.5-turbo-0125
gemini-2.0-flash	Google	gemini-2.0-flash-exp
gemini-1.5-flash	Google	gemini-1.5-flash
gemini-1.5-flash-8b	Google	gemini-1.5-flash-8b
gemini-1.5-pro	Google	gemini-1.5-pro
claude-3.5-sonnet	Anthropic	claude-3-5-sonnet-latest
claude-3.5-haiku	Anthropic	claude-3-5-haiku-latest
claude-3-opus	Anthropic	claude-3-opus-20240229
claude-3-sonnet	Anthropic	claude-3-sonnet-20240229
claude-3-haiku	Anthropic	claude-3-haiku-20240307
command-r	Cohere	command-r
command-r-plus	Cohere	command-r-plus
mistral-large	Mistral	mistral-large-latest
ministral-3b	Mistral	ministral-3b-latest
ministral-8b	Mistral	ministral-8b-latest
mistral-small	Mistral	mistral-small-latest
mixtral-8x7b	Groq	mixtral-8x7b-32768
gemma-2-9b	Groq	gemma2-9b-it
llama-3-8b	Groq, Replicate	llama3-8b-8192, meta/meta-llama-3-8b-instruct
llama-3-70b	Groq, Replicate	llama3-70b-8192, meta/meta-llama-3-70b-instruct
llama-3.1-8b	Groq, Cerebras	llama-3.1-8b-instant, llama3.1-8b
llama-3.1-405b	Replicate	meta/meta-llama-3.1-405b-instruct
llama-3.2-1b	Groq	llama-3.2-1b-preview
llama-3.2-3b	Groq	llama-3.2-3b-preview
llama-3.3-70b	Groq, Cerebras	llama-3.3-70b-versatile, llama3.3-70b

Table of Contents

• Requirements
• Installation
• Usage
  • Basic Usage
  • Multi-Provider Models
  • Memory
    • Chat Memory
    • External Memory
  • Async Calls
  • Tools
    • JSON Mode
    • Prompt Loader
    • Other Capabilities
• Planned Features
• Contributing
• Contact

Requirements

• Python >= 3.9
• At least one valid API key for a supported provider

Installation

pip install l2m2

Usage

Import the LLM Client

from l2m2.client import LLMClient

client = LLMClient()

Activate Providers

To activate any of the providers, set the provider's API key in the corresponding environment variable shown below, and L2M2 will read it in to activate the provider.

Provider	Environment Variable
OpenAI	OPENAI_API_KEY
Anthropic	ANTHROPIC_API_KEY
Cohere	CO_API_KEY
Google	GOOGLE_API_KEY
Groq	GROQ_API_KEY
Replicate	REPLICATE_API_TOKEN
Mistral (La Plateforme)	MISTRAL_API_KEY
Cerebras	CEREBRAS_API_KEY

Additionally, you can activate providers programmatically as follows:

client = LLMClient(providers={
    "provider-a": "api-key-a",
    "provider-b": "api-key-b",
    ...
})

# Alternatively,
client.add_provider("provider-c", "api-key-c")

Call your LLM 💬👍

The call API is the same regardless of model or provider.

response = client.call(
    model="<model name>",
    prompt="<prompt>",
)

model and prompt are required, while system_prompt, temperature, and max_tokens are optional. When possible, L2M2 uses the provider's default model parameter values when they are not given.

response = client.call(
    model="<model name>",
    prompt="<prompt>",
    system_prompt="<system prompt>",
    temperature=<temperature>,
    max_tokens=<max tokens>,
)

Example

# example.py

from l2m2.client import LLMClient

client = LLMClient()

response = client.call(
    model="gpt-4o",
    prompt="How's the weather today?",
    system_prompt="Respond as if you were a pirate.",
)

print(response)

>> python3 example.py

Arrr, matey! The skies be clear as the Caribbean waters today, with the sun blazin' high 'bove us. A fine day fer settin' sail and huntin' fer treasure, it be. But keep yer eye on the horizon, for the weather can turn quicker than a sloop in a squall. Yarrr!

Multi-Provider Models

Some models are available from multiple providers, such as llama-3-70b from both Groq and Replicate. When multiple of such providers are active, you can use the parameter prefer_provider to specify which provider to use for a given inference.

response1 = client.call(
    model="llama-3-70b",
    prompt="Hello there",
    prefer_provider="groq",
) # Uses Groq

response2 = client.call(
    model="llama-3-70b",
    prompt="General Kenobi!",
    prefer_provider="replicate",
) # Uses Replicate

You can also set default preferred providers for the client using set_preferred_providers, to avoid having to specify prefer_provider for each call.

client.set_preferred_providers({
    "llama-3-70b": "groq",
    "llama-3-8b": "replicate",
})

response1 = client.call(model="llama-3-70b", prompt="Hello there") # Uses Groq
response2 = client.call(model="llama-3-8b", prompt="General Kenobi!") # Uses Replicate

Memory

L2M2 provides a simple memory system that allows you to maintain context and history across multiple calls and multiple models. There are two types of memory: ChatMemory, which natively hooks into models' conversation history, and ExternalMemory, which allows for custom memory implementations. Let's first take a look at ChatMemory.

from l2m2.client import LLMClient
from l2m2.memory import ChatMemory

client = LLMClient(memory=ChatMemory())

print(client.call(model="gpt-4o", prompt="My name is Pierce"))
print(client.call(model="claude-3-haiku", prompt="I am a software engineer."))
print(client.call(model="llama-3-8b", prompt="What's my name?"))
print(client.call(model="mixtral-8x7b", prompt="What's my job?"))

Hello, Pierce! How can I help you today?
A software engineer, you say? That's a noble profession.
Your name is Pierce.
You are a software engineer.

Chat memory is stored per session, with a sliding window of messages which defaults to the last 40 – this can be configured by passing memory_window_size to the client constructor.

For more control, you can instantiate a ChatMemory object on its own and manipulate it directly.

memory = ChatMemory()

memory.add_user_message("My favorite color is red.")
memory.add_user_message("My least favorite color is green.")
memory.add_agent_message("Ok, noted.")

client = LLMClient(memory=memory)
print(client.call(model="gpt-4o", prompt="What are my favorite and least favorite colors?"))
memory.clear()
print(client.call(model="gpt-4o", prompt="What are my favorite and least favorite colors?"))

Your favorite color is red, and your least favorite color is green.
I'm sorry, I don't have that information.

Caution

Some providers such as Anthropic enforce that chat messages in memory strictly alternate between one user and one agent message and will throw an error if this is not the case.

You can also load in alternate memory streams on the fly using the alt_memory parameter in call (This is especially useful for parallel memory streams – an example of this is shown in the Async Calls section).

m1 = ChatMemory()
m1.add_user_message("My favorite color is red.")
m1.add_user_message("My least favorite color is green.")
m1.add_agent_message("Ok, noted.")

m2 = ChatMemory()
m2.add_user_message("My favorite color is blue.")
m2.add_user_message("My least favorite color is yellow.")
m2.add_agent_message("Got it.")

client = LLMClient(memory=m1)
prompt = "What are my favorite and least favorite colors?"
print(client.call(model="gpt-4o", prompt=prompt)
print(client.call(model="gpt-4o", prompt=prompt, alt_memory=m2))

Your favorite color is red, and your least favorite color is green.
Your favorite color is blue, and your least favorite color is yellow.

Finally, memory can be bypassed for a single call by passing bypass_memory=True to call. This will cause the client to ignore previously stored memory and not write to it for the current call.

client = LLMClient(memory=ChatMemory())
client.call(model="gpt-4o", prompt="My name is Pierce")
client.call(model="gpt-4o", prompt="I am 25 years old")

print(client.call(model="gpt-4o", prompt="What is my name?"))
print(client.call(model="gpt-4o", prompt="What is my name?", bypass_memory=True))

client.call(model="gpt-4o", prompt="I am a software engineer", bypass_memory=True)
print(client.call(model="gpt-4o", prompt="What is my profession?"))

Your name is Pierce.
I'm sorry, but I don't have access to personal information, so I can't know your name.
You haven't mentioned your profession yet, Pierce.

External Memory

ExternalMemory is a simple but powerful memory mode that allows you to define your own memory implementation. This can be useful for more complex memory constructions (e.g., planning, reflecting) or for implementing custom persistence (e.g., saving memory to a database or a file). Its usage is much like ChatMemory, but unlike ChatMemory you must manage initializing and updating the memory yourself with get_contents and set_contents.

Here's a simple example of a custom memory implementation that has a description and a list of previous user/model message pairs:

# example_external_memory.py

from l2m2.client import LLMClient
from l2m2.memory import ExternalMemory

client = LLMClient(memory=ExternalMemory())

messages = [
    "My name is Pierce",
    "I am a software engineer",
    "What is my name?",
    "What is my profession?",
]

def update_memory(user_input, model_output):
    memory = client.get_memory() # ExternalMemory object
    contents = memory.get_contents()
    if contents == "":
        contents = "You are mid-conversation with me. Your memory of it is below:\n\n"
    contents += f"Me: {user_input}\nYou: {model_output}\n"
    memory.set_contents(contents)

for message in messages:
    response = client.call(model="gpt-4o", prompt=message)
    print(response)
    update_memory(message, response)

>> python3 example_external_memory.py

Nice to meet you, Pierce!
Nice! What kind of projects do you work on?
Your name is Pierce.
You are a software engineer.

By default, ExternalMemory contents are appended to the system prompt, or passed in as the system prompt if one is not given. Generally, models perform best when external memory is stored in the system prompt; however, you can configure the client to append the memory contents to the user prompt instead as follows:

from l2m2.memory import ExternalMemoryLoadingType

memory = ExternalMemory(loading_type=ExternalMemoryLoadingType.USER_PROMPT_APPEND)
client = LLMClient(memory=memory)

Similarly to ChatMemory, ExternalMemory can be passed into alt_memory and bypassed with bypass_memory.

Async Calls

L2M2 provides an asynchronous AsyncLLMClient in addition to the synchronous LLMClient. Its usage is identical to the synchronous client, but it's instantiated using async with and is called using await.

from l2m2.client import AsyncLLMClient

async def main():
    async with AsyncLLMClient() as client:
        response = await client.call(
            model="model",
            prompt="prompt",
            system_prompt="system prompt",
            # ...etc
        )

Under the hood, each AsyncLLMClient manages its own async http client, so calls are non-blocking. Here's an example of using the AsyncLLMClient to make concurrent calls to multiple models and measure the inference times:

# example_async.py

import os
import asyncio
import timeit
from l2m2.client import AsyncLLMClient

async def call_concurrent():
    async with AsyncLLMClient() as client:
        # Assumes no conflicts between active providers
        calls = [
            ("gpt-4o", "foo"),
            ("claude-3.5-sonnet", "bar"),
            ("gemini-1.5-pro", "baz"),
            ("command-r-plus", "qux"),
            ("llama-3-70b", "quux"),
            ("mixtral-8x7b", "corge"),
        ]
        system_prompt = "The secret word is {}"

        async def call_and_print(model, secret_word):
            start_time = timeit.default_timer()
            response = await client.call(
                model=model,
                prompt="What is the secret word? Respond briefly.",
                system_prompt=system_prompt.format(secret_word),
                temperature=0.2,
            )
            time = timeit.default_timer() - start_time
            print(f"{model}: {response} ({time:.2f}s)")

        await asyncio.gather(
            *[call_and_print(model, secret_word) for model, secret_word in calls]
        )

asyncio.run(call_concurrent())

>> python3 example_async.py

llama-3-70b: The secret word is quux. (0.21s)
mixtral-8x7b: The secret word is corge. (0.26s)
gpt-4o: foo (0.62s)
command-r-plus: The secret word is qux. (0.66s)
claude-3.5-sonnet: The secret word is bar. (0.70s)
gemini-1.5-pro: baz (0.73s)

As a general rule, I typically find it's best to use the synchronous LLMClient for research and demos, and AsyncLLMClient for apps.

Use Case: Parallel Memory Streams ⚡

One of the most powerful features of AsyncLLMClient is the ability to run maintain memory streams in parallel, such as in multi-agent systems with multiple interactions happening concurrently. Here's a simple example of how to easily achieve this using AsyncLLMClient and alt_memory.

# example_parallel_memory.py

import asyncio
from l2m2.client import AsyncLLMClient
from l2m2.memory import ChatMemory

async def call_concurrent_with_memory():
    m1 = ChatMemory()
    m2 = ChatMemory()

    calls1 = ["My name is Pierce", "My favorite color is red", "I am 25 years old"]
    calls2 = ["My name is Paul", "My favorite color is blue", "I am 60 years old"]
    question = "What is my name, favorite color, and age?"

    async with AsyncLLMClient() as client:
        client.set_preferred_providers({"mixtral-8x7b": "groq"})

        async def make_calls_1():
            for prompt in calls1:
                await client.call(model="mixtral-8x7b", prompt=prompt, alt_memory=m1)

        async def make_calls_2():
            for prompt in calls2:
                await client.call(model="mixtral-8x7b", prompt=prompt, alt_memory=m2)

        await asyncio.gather(make_calls_1(), make_calls_2())

        [res1, res2] = await asyncio.gather(
            client.call(model="mixtral-8x7b", prompt=question, alt_memory=m1),
            client.call(model="mixtral-8x7b", prompt=question, alt_memory=m2),
        )

        print("Memory 1:", res1)
        print("Memory 2:", res2)

asyncio.run(call_concurrent_with_memory())

>> python3 example_parallel_memory.py

Memory 1: Your name is Pierce, your favorite color is red, and you are 25 years old. I hope this information is helpful!
Memory 2: Your name is Paul, your favorite color is blue, and you are 60 years old. 😊

Tools: JSON Mode

L2M2 provides an optional json_mode flag that enforces JSON formatting on LLM responses. Importantly, this flag is applicable to all models and providers, whether or not they natively support JSON output enforcement. When JSON mode is not natively supported, json_mode will apply strategies to maximize the likelihood of valid JSON output.

# example_json_mode.py

response = client.call(
    model="gpt-4o",
    prompt="What are the capitals of each state of Australia?",
    system_prompt="Respond with the JSON format {'region': 'capital'}",
    json_mode=True,
)

print(response)

>> python3 example_json_mode.py

{
  "New South Wales": "Sydney",
  "Victoria": "Melbourne",
  "Queensland": "Brisbane",
  "South Australia": "Adelaide",
  "Western Australia": "Perth",
  "Tasmania": "Hobart",
  "Northern Territory": "Darwin",
  "Australian Capital Territory": "Canberra"
}

Important

Regardless of the model and even when json_mode is enabled, it's crucial to ensure that either the prompt or the system prompt mentions to return the output in JSON - and ideally, to specify the JSON format, as shown above.

The following models natively support JSON mode via the given provider:

• gpt-4o (via Openai)
• gpt-4o-mini (via Openai)
• gpt-4-turbo (via Openai)
• gpt-3.5-turbo (via Openai)
• gemini-2.0-flash (via Google)
• gemini-1.5-flash (via Google)
• gemini-1.5-flash-8b (via Google)
• gemini-1.5-pro (via Google)
• mistral-large (via Mistral)
• ministral-3b (via Mistral)
• ministral-8b (via Mistral)
• mistral-small (via Mistral)

JSON Mode Non-Native Strategies

For models that do not natively support JSON mode, L2M2 will attempt to enforce JSON formatting by applying one of the following two strategies under the hood:

1. Strip: This is usually the default strategy. It will attempt to extract the JSON from the response by searching for the first instance of { and the last instance of } in the response, and returning the between substring (inclusive). If no JSON is found, the response will be returned as-is.
2. Prepend: This strategy will attempt to enforce a valid JSON output by inserting a message ending with an opening { from the model into the conversation just after the user prompt and just before the model response, and re-prepending the opening { to the model response. By default this message is "Here is the JSON output:", but can be customized. More information is available on this strategy here. Importantly, the Prepend strategy is available whether or not memory is enabled, and will not interfere with memory.

If you'd like, you can specify a strategy by passing either JsonModeStrategy.strip() or JsonModeStrategy.prepend() to the json_mode_strategy parameter in call. If no strategy is given, L2M2 defaults to Strip for all models except for Anthropic's models, which will default to Prepend (more on this below).

# example_json_mode.py

from l2m2.client import LLMClient
from l2m2.tools import JsonModeStrategy

client = LLMClient()

response = client.call(
    model="claude-3-sonnet",
    prompt="What are the capitals of each Canadian province?",
    system_prompt="Respond with the JSON format {'region': 'capital'}",
    json_mode=True,
    json_mode_strategy=JsonModeStrategy.prepend(),
)

print(response)

>> python3 example_json_mode.py

{
  "Alberta": "Edmonton",
  "British Columbia": "Victoria",
  "Manitoba": "Winnipeg",
  "New Brunswick": "Fredericton",
  "Newfoundland and Labrador": "St. John's",
  "Nova Scotia": "Halifax",
  "Ontario": "Toronto",
  "Prince Edward Island": "Charlottetown",
  "Quebec": "Quebec City",
  "Saskatchewan": "Regina"
}

If using prepend, you can customize the message that gets prepended to the opening { by passing custom_prefix as follows:

response = client.call(
    model="claude-3-sonnet",
    prompt="What are the capitals of each Canadian province?",
    system_prompt="Respond with the JSON format {'region': 'capital'}",
    json_mode=True,
    json_mode_strategy=JsonModeStrategy.prepend(custom_prefix="Here is the JSON with provinces and capitals:"),
)

Ideally, this wouldn't change anything on the output – just under the hood – but this is useful for working with foreign languages, etc.

Tip

As mentioned above, L2M2 defaults to prepend for Anthropic models and strip for all others. I highly recommend sticking with these defaults, especially with Anthropic's models. From my personal testing, valid JSON is almost always produced when using prepend with Anthropic's models and almost never produced with strip, and vice versa for other models. I'll gather rigorous data on this eventually, but if anyone has any insights, please let me know!

Tools: Prompt Loader

L2M2 provides an optional prompt-loading utility that's useful for loading prompts with variables from a file. Usage is simple:

prompt.txt

Your name is {{name}} and you are a {{profession}}.

# example_prompt_loader.py

from l2m2.tools import PromptLoader

loader = PromptLoader()
prompt = loader.load_prompt(
    prompt_file="prompt.txt",
    variables={"name": "Pierce", "profession": "software engineer"},
)
print(prompt)

>> python3 example_prompt_loader.py

Your name is Pierce and you are a software engineer.

The default variable delimiters are {{ and }}. You can also optionally specify a prompt directory or customize the variable delimiters if needed.

path/to/prompts/prompt.txt

Your name is <<name>> and you are a <<profession>>.

# example_prompt_loader.py

from l2m2.tools import PromptLoader

loader = PromptLoader(
    prompts_base_dir="path/to/prompts",
    variable_delimiters=("<<", ">>"),
)
prompt = loader.load_prompt(
    prompt_file="prompt.txt",
    variables={"name": "Pierce", "profession": "software engineer"},
)
print(prompt)

>> python3 example_prompt_loader.py

Your name is Pierce and you are a software engineer.

Other Capabilities

Call Custom

If you'd like to call a language model from one of the supported providers that isn't officially supported by L2M2 (for example, older models such as gpt-4-0125-preview), you can similarly call_custom with the additional required parameter provider, and pass in the model name expected by the provider's API. Unlike call, call_custom doesn't guarantee correctness or well-defined behavior.

response = client.call_custom(
    provider="<provider name>",
    model_id="<model id for given provider>",
    prompt="<prompt>",
    ...
)

Extra Parameters

You can pass in extra parameters to the provider's API (For example, reasoning_effort on OpenAI's o1 series) by passing in the extra_params parameter to call or call_custom. These parameters are passed in as a dictionary of key-value pairs, where the values are of type str, int, or float. Similarly, using extra_params does not guarantee correctness or well-defined behavior, and you should refer to the provider's documentation for correct usage.

response = client.call(
    model="<model name>",
    prompt="<prompt>",
    extra_params={"foo": "bar", "baz": 123},
    ...
)

Planned Features

• Support for structured outputs where available (Just OpenAI as far as I know)
• Support for OSS and self-hosted (Hugging Face, Ollama, Gpt4all, etc.)
• Support for batch APIs where available (OpenAI, Anthropic, etc.)
• Basic (i.e., customizable & non-opinionated) agent & multi-agent system features
• Tools for common application workflows: RAG, prompt management, search, etc.
• Support for streaming responses
• ...etc.

Contributing

Contributions are welcome! Please see the below contribution guide.

• Requirements
  • Python >= 3.13
  • GNU Make
• Setup
  • Clone this repository and create a Python virtual environment.
  • Install dependencies: make init.
  • Create a feature branch and an issue with a description of the feature or bug fix.
• Develop
  • Run lint, typecheck and tests: make (make lint, make typecheck, and make test can also be run individually).
  • Generate test coverage: make coverage.
  • If you've updated the supported models, run make update-readme to reflect those changes in the README.
• Integration Test
  • cd into integration_tests.
  • Create a .env file with your API keys, and copy itests.example.py to itests.py.
  • Write your integration tests in itests.py.
  • Run locally with python itests.py -l.
    • Note: make sure to pass the -l flag or else it will look for an L2M2 distribution. Additionally, make sure l2m2 is not installed with pip when running the integration tests locally.
  • Once your changes are ready, from the top-level directory run make build to create the distribution and make itest to run your integration tests against the distribution.
    • Note: in order to ensure a clean test environment, make itest uninstalls all third-party Python packages before running the tests, so make sure to run make init when you're done working on integration tests.
• Contribute
  • Create a PR and ping me for a review.
  • Merge!

Contact

If you have requests, suggestions, or any other questions about l2m2 please shoot me a note at [email protected], open an issue on the Github repo, or DM me on the GenAI Collective Slack Channel.