A unified interface for interacting with multiple LLMs.
Project description
LLM Master
LLM Master is a powerful engine to boost your creativity by working with multiple generative AIs.
This is a Python library that provides a unified interface for interacting with multiple Large Language Models (LLMs) and multimedia generative AI models from different providers.
Features
- Support for various models from each provider
- Concurrent execution of multiple LLMs and multimedia generative AI models
- Thread-based execution for improved performance
- Customizable generation parameters
- Required parameters:
provider
andprompt
- Optional parameters:
model
and particular parameters for different model - Exception:
prompt
is not required for Audio-to-Text, Image-to-Image and Image-To-Video models, which is dependent on mode you select.
- Required parameters:
Supported LLM Providers and Models
LLM Master respects multi-modal approach. There are many generative AI models that can be used for different purposes. There are 2 categories: major data formats and 3D models.
1. Major Data Formats
The table below represents various conversion capabilities between different media types (text, image, audio/speech and video). Some conversions are available, some are pending or coming soon, and others are marked as not applicable (NA) at the moment.
Use highlighted word for provider
to make LLMMaster instance.
From \ To | Text | Image | Audio | Video |
---|---|---|---|---|
Text | openai , anthropic , google , groq , perplexity |
openai_tti , stable_diffusion_tti , adobe_firefly_tti (pending) |
openai_tts , elevenlabs_tts , google_tta (pending) |
(pending) |
Image | openai_itt , google_itt |
openai_iti , stable_diffusion_iti |
NA | stable_diffusion_itv |
Audio | openai_stt , google_stt |
NA | NA | NA |
Video | google_vtt |
NA | NA | NA |
And the list below represents the models that are supported by each provider. See each provider's documentation for full list.
Text-to-Text Models (typical)
- Anthropic (
claude-3-5-sonnet-20240620
) - Google (
gemini-1.5-flash
) - Groq (
llama3-70b-8192
) - OpenAI (
gpt-4o
) - Perplexity (
llama-3-sonar-large-32k-online
)
Text-to-Image Models
- OpenAI (
dall-e-3
,dall-e-2
) - Stable Diffusion (
core
,ultra
) - Adobe Firely (pending deployment)
Text-to-Audio (Speech) Models
- OpenAI (
tts-1
,tts-1-hd
) - ElevenLabs (
eleven_multilingual_v2
)
Image-to-Text Models (typical)
- OpenAI (
gpt-4o
) - Google (
gemini-1.5-flash
)
Image-to-Image Models
- OpenAI (
dall-e-2
) - Stable Diffusion (
v2beta
)
Image-to-Video Models
- Stable Diffusion (
v2beta
)
Audio(Speech)-to-Text Models (typical)
- OpenAI (
whisper-1
) - Google (
gemini-1.5-flash
)
Video-to-Text Models (typical)
- Google (
gemini-1.5-flash
)
Use highlighted word for model
to make LLMMaster instance.
The model
parameter is optional. If you do not specify the model, the default model defined in config.py
will be used.
2. 3D Models
From ver. 0.2.2, LLM Master supports 3D models thanks to Meshy API. Here are the list of supported models:
- MeshyTextToTexture:
meshy_tttx
- MeshyTextTo3D:
meshy_tt3d
- MeshyTextTo3DRefine:
meshy_tt3d_refine
- MeshyTextToVoxel:
meshy_ttvx
- MeshyImageTo3D:
meshy_it3d
More models will be supported soon!
Installation
To use LLM Master, you need to install the library. You can do this using pip:
pip install llmmaster
Relevant packages will also be installed.
Usage
Set API keys for your environment in advance
Set up your API keys as environment variables. Note that you do not have to set all of the API keys as shown below. You can set only the ones that you need.
Important: changed GEMINI_API_KEY
to GOOGLE_API_KEY
since ver. 0.1.4.
For Mac/Linux,
export ANTHROPIC_API_KEY="your_anthropic_key"
export GOOGLE_API_KEY="your_google_key"
export GROQ_API_KEY="your_groq_key"
export OPENAI_API_KEY="your_openai_key"
export PERPLEXITY_API_KEY="your_perplexity_key"
export STABLE_DIFFUSION_API_KEY="your_stable_diffusion_key"
export MESHY_API_KEY="your_meshy_key"
export ELEVENLABS_API_KEY="your_elevenlabs_key"
For Windows (cmd),
SET ANTHROPIC_API_KEY=your_anthropic_key
SET GOOGLE_API_KEY=your_google_key
SET GROQ_API_KEY=your_groq_key
SET OPENAI_API_KEY=your_openai_key
SET PERPLEXITY_API_KEY=your_perplexity_key
SET STABLE_DIFFUSION_API_KEY=your_stable_diffusion_key
SET MESHY_API_KEY=your_meshy_key
SET ELEVENLABS_API_KEY=your_elevenlabs_key
For Windows (PowerShell)
$env:ANTHROPIC_API_KEY="your_anthropic_key"
$env:GOOGLE_API_KEY="your_google_key"
$env:GROQ_API_KEY="your_groq_key"
$env:OPENAI_API_KEY="your_openai_key"
$env:PERPLEXITY_API_KEY="your_perplexity_key"
$env:STABLE_DIFFUSION_API_KEY="your_stable_diffusion_key"
$env:MESHY_API_KEY="your_meshy_key"
$env:ELEVENLABS_API_KEY="your_elevenlabs_key"
Use cases
You can find and download these use case codes in folder usecases.
Important: How to handle or save generated contents? LLM Master basically returns raw output defined by each provider. Some models return REST API Response object of requests library, while some other models return bytes object of media or specific class instance defined by provider. See RESULTSTYPE.md for brief description of what type of object is returned.
- Using single Text-to-Text LLM
This is the most basic usage of LLM Master.
openai_instance
is actually a unique label to manage multiple instances. You may set any string for it. The next case will be a multi-LLM example.
from llmmaster import LLMMaster
# Create an instance of LLMMaster
llmmaster = LLMMaster()
# Configure LLM instance
llmmaster.summon({
"openai_instance": llmmaster.pack_parameters(
provider="openai",
model="gpt-4o-mini",
prompt="Hello, how are you?",
max_tokens=100,
temperature=0.7
)
})
# Run LLM
print('Start running LLMMaster...')
llmmaster.run()
# Get results
results = llmmaster.results
print(f'OpenAI responded: {results["openai_instance"]}')
# Check elapsed time
print(f"Elapsed time: {llmmaster.elapsed_time} seconds")
# Clear instances
llmmaster.dismiss()
2 Using multiple Text-to-Text LLMs simultaneously
from llmmaster import LLMMaster
# Create an instance of LLMMaster
llmmaster = LLMMaster()
# Configure LLM instance
llmmaster.summon({
"openai_instance": llmmaster.pack_parameters(
provider="openai",
prompt="Summarize the main ideas of quantum computing."
),
"anthropic_instance": llmmaster.pack_parameters(
provider="anthropic",
prompt="Explain the concept of artificial general intelligence."
)
})
# You can check what parameters are set before run
print('Parameters set before run.')
for label, instance in llmmaster.instances.items():
print(f'{label} = {instance.parameters}')
# Run LLM
print('Start running LLMMaster...')
llmmaster.run()
# Get results
for instance, response in llmmaster.results.items():
print(f'{instance} responded: {response}')
# Check elapsed time
print(f"Elapsed time: {llmmaster.elapsed_time} seconds")
# Clear instances
llmmaster.dismiss()
- Using Text-to-Image Models
import requests
from llmmaster import LLMMaster
# Create an instance of LLMMaster
llmmaster = LLMMaster()
# Configure LLM instance
llmmaster.summon({
"openai_image": llmmaster.pack_parameters(
provider="openai_tti",
model="dall-e-3",
prompt="A futuristic cityscape with flying cars and holographic billboards",
size="1792x1024",
quality="hd"
),
"stable_diffusion_image": llmmaster.pack_parameters(
provider="stable_diffusion_tti",
model="core",
prompt="A serene landscape with a mountain lake at sunset",
aspect_ratio="16:9",
style_preset="cinematic"
)
})
# You can check what parameters are set before run
print('Parameters set before run.')
for label, instance in llmmaster.instances.items():
print(f'{label} = {instance.parameters}')
# Run LLM
print('Start running LLMMaster...')
llmmaster.run()
# Get results
response = llmmaster.results["openai_image"]
if hasattr(response, 'data'):
image_response = requests.get(response.data[0].url)
if image_response.status_code == 200:
with open("openai_image.png", 'wb') as f:
f.write(image_response.content)
print("openai_image.png saved")
response = llmmaster.results["stable_diffusion_image"]
if isinstance(response, bytes):
with open("stable_diffusion_image.png", 'wb') as f:
f.write(response)
print("stable_diffusion_image.png saved")
# Check elapsed time
print(f"Elapsed time: {llmmaster.elapsed_time} seconds")
# Clear instances
llmmaster.dismiss()
-
Using Text-to-Audio (Text-to-Speech) Models
4.1 OpenAI
from llmmaster import LLMMaster
from llmmaster.config import OPENAI_TTS_VOICE_OPTIONS
master = LLMMaster()
to_say = "Do not concentrate on the finger, or you will miss all that heavenly glory."
# try to generate all the voice patterns for a same saying
entries = []
for voice_pattern in OPENAI_TTS_VOICE_OPTIONS:
case = {'provider': 'openai_tts',
'prompt': to_say,
'voice': voice_pattern,
'response_format': 'mp3'}
entries.append({'name': f'openai_tts_{voice_pattern}', 'params': case})
for entry in entries:
master.summon({entry['name']: master.pack_parameters(**entry['params'])})
# You can check what parameters are set before run
print('Parameters set before run.')
for label, instance in master.instances.items():
print(f'{label} = {instance.parameters}')
# Run LLM
print('Start running LLMMaster...')
master.run()
# Get results
print('Responses')
for name, response in master.results.items():
save_as = f"{name}.mp3"
with open(save_as, 'wb') as f:
for chunk in response.iter_bytes():
f.write(chunk)
print(f'Saved as {save_as} for case {name}')
# Check elapsed time
print(f"Elapsed time: {master.elapsed_time} seconds")
master.dismiss()
4.2 ElevenLabs
import elevenlabs
from llmmaster import LLMMaster
master = LLMMaster()
to_say = "Do not concentrate on the finger, or you will miss all that heavenly glory."
test_case = master.pack_parameters(provider='elevenlabs_tts', prompt=to_say)
master.summon({'elevenlabs_tts': test_case})
# Run LLM
print('Start running LLMMaster...')
master.run()
# Get results
print('Responses')
if master.results['elevenlabs_tts']:
save_as = f"elevenlabs_tts.mp3"
elevenlabs.save(master.results['elevenlabs_tts'], save_as)
print(f'Saved as {save_as} for case {name}')
# Check elapsed time
print(f"Elapsed time: {master.elapsed_time} seconds")
master.dismiss()
- Using Image-to-Text Models
from llmmaster import LLMMaster
master = LLMMaster()
# Online images are preferred for OpenAI.
# Online images and local image paths are both supported for Google.
# Change "image_url" for your case
entries = [
{
'name': 'openai_itt_case',
'params': {
'provider': 'openai_itt',
'model': 'gpt-4o',
'prompt': 'Describe each image.',
'image_url': ['https://example.com/image-1.jpg',
'https://example.com/image-2.jpg']
}
},
{
'name': 'google_itt_case',
'params': {
'provider': 'google_itt',
'model': 'gemini-1.5-flash',
'prompt': 'What might be differences between these pictures?',
'image_url': ['https://example.com/image.png',
'/home/user/my_image.png']
}
}
]
for entry in entries:
master.summon({entry['name']: master.pack_parameters(**entry['params'])})
print('Start running LLMMaster...')
master.run()
for entry, result in master.results.items():
print(f'{entry} responded: {result}')
print(f'Elapsed time: {master.elapsed_time} seconds')
master.dismiss()
- Using Image-to-Image Models
import requests
from llmmaster import LLMMaster
master = LLMMaster()
test_image = '/home/user/test_image.png'
entries = [
{
'name': 'openai',
'params': {
'provider': 'openai_iti',
'mode': 'variations',
'image': test_image,
'n': 2
}
},
{
'name': 'stable_diffusion',
'params': {
'provider': 'stable_diffusion_iti',
'mode': 'remove_background',
'image': test_image,
}
}]
for entry in entries:
master.summon({entry['name']: master.pack_parameters(**entry['params'])})
print('Start running LLMMaster...')
master.run()
print('Results')
# Get results
response = master.results["openai"]
if hasattr(response, 'data'):
for i in range(len(response.data)):
image_response = requests.get(response.data[i].url)
if image_response.status_code == 200:
filename = f"openai_{i+1:02}.png"
with open(filename, 'wb') as f:
f.write(image_response.content)
print(f'Image No. {i+1:2} saved as {filename}')
else:
print(f'Image No. {i+1:2} failed to download')
response = master.results["stable_diffusion"]
if isinstance(response, bytes):
with open("stable_diffusion_result.png", 'wb') as f:
f.write(response)
print("stable_diffusion_result.png saved")
print(f"Elapsed time: {master.elapsed_time} seconds")
master.dismiss()
- Using Image-to-Video Model
from llmmaster import LLMMaster
master = LLMMaster()
# Set a path of local image file
# Change "image" for your case
entry = master.pack_parameters(
provider='stable_diffusion_itv',
image='/home/user/test-image.png')
master.summon({'sd_itv': entry})
print('Start running LLMMaster...')
master.run()
print('Results')
response = master.results["sd_itv"]
if isinstance(response, bytes):
with open("sd_itv.mp4", 'wb') as f:
f.write(response)
print("sd_itv.mp4 saved")
print(f"Elapsed time: {master.elapsed_time} seconds")
master.dismiss()
- Using Audio-to-Text (Speech-to-Text) Models
from llmmaster import LLMMaster
master = LLMMaster()
speech_file = '/home/user/sample-speech.mp3'
entries = [
{
'name': 'openai_stt_1',
'params': {
'provider': 'openai_stt',
'mode': 'translations',
'file': speech_file,
'response_format': 'json',
'temperature': 0.0
}
},
{
'name': 'openai_stt_2',
'params': {
'provider': 'openai_stt',
'mode': 'transcriptions',
'file': speech_file,
'response_format': 'text'
}
},
{
'name': 'openai_stt_3',
'params': {
'provider': 'openai_stt',
'mode': 'transcriptions',
'file': speech_file,
'response_format': 'verbose_json'
}
},
{
'name': 'google_stt',
'params': {
'provider': 'google_stt',
'prompt': 'What does the speaker imply?',
'audio_file': speech_file
}
}
]
for entry in entries:
master.summon({entry['name']: master.pack_parameters(**entry['params'])})
print(f'Start running LLMMaster...')
master.run()
# Different type of response given by different
# `response_format`, `transcriptions` or `translations` mode
# Handle with care for output
print('Results')
for name, response in master.results.items():
print(f'{name}: {response}')
print(f"Elapsed time: {master.elapsed_time} seconds")
master.dismiss()
- Using Video-to-Text Model
from llmmaster import LLMMaster
master = LLMMaster()
# Set a path of local video file
# Change "video_file" for your case
params = master.pack_parameters(
provider='google_vtt',
prompt='Describe attached video.',
video_file='/home/user/sample-video.mp4')
master.summon({'video_to_text': params})
print('Start running LLMMaster...')
master.run()
print(f'Answer = {master.results["video_to_text"]}')
print(f"Elapsed time: {master.elapsed_time} seconds")
master.dismiss()
- Using 3D Models (Meshy)
import requests
from requests.models import Response
from llmmaster import LLMMaster
REQUEST_OK = 200
TEST_OUTPUT_PATH = 'test-outputs'
master = LLMMaster()
# This is a case of generating Voxel model from text prompt
params = master.pack_parameters(provider='meshy_ttvx',
prompt='A cute robot boy.',
negative_prompt='ugly, low resolution')
master.summon({'meshy_ttvx_test': params})
# Run LLM
print('Start running LLMMaster...')
master.run()
# Get results
print('Responses')
if isinstance(master.results['meshy_ttvx_test'], Response):
json_result = master.results['meshy_ttvx_test'].json()
# Meshy returns several different model formats
for key, value in json_result['model_urls'].items():
if value:
res = requests.get(value)
if res.status_code == REQUEST_OK:
filename = f'meshy_ttvx_test.{key}'
filepath = os.path.join(TEST_OUTPUT_PATH, filename)
with open(filepath, 'wb') as f:
f.write(res.content)
print(f'Saved as {filepath}')
# Meshy also returns thumbnail png for preview
if 'thumbnail_url' in json_result:
res = requests.get(json_result['thumbnail_url'])
if res.status_code == REQUEST_OK:
filename = 'meshy_ttvx_test_thumbnail.png'
filepath = os.path.join(TEST_OUTPUT_PATH, filename)
with open(filepath, 'wb') as f:
f.write(res.content)
print(f'Saved as {filepath}')
# Check elapsed time
print(f"Elapsed time: {master.elapsed_time} seconds")
master.dismiss()
Applications
- Multi-AI brainstorming web app monju https://monju.ai. This app generates ideas from 3 different LLMs simultaneously, and then shows the result in mindmap.
Notes
- Please comply with the terms of service for each provider's API.
- Securely manage your API keys and be careful not to commit them to public repositories.
- There is a limit (32) to the number of LLM instances that can be created at once.
- If you apply a single model for multiple entrys, running simultaneously, each entry will start in every 1 second of interval. This is a strict limitation of providers for security reason.
- Input parameters are not strictly checked by the rules defined by each provider. You may face an error due to some wrong paramer or combination of parameters.
Customization
- You can easily adjust default models by updating the dictionary in
config.py
and creating a new thread class for the provider. - You can also run an individual provider Model directly without using LLMMaster. See each class definition for details.
Contributing
Contributions to LLM Master are welcome! Please feel free to submit a Pull Request and bug reports and feature requests through GitHub Issues.
License
This project is licensed under the MIT License.
This project is under development. New features and support for additional providers may be added. Please check the repository for the latest information.
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
Built Distribution
Hashes for llmmaster-0.2.3-py3-none-any.whl
Algorithm | Hash digest | |
---|---|---|
SHA256 | d9205fd3c39e672a5ac24f198fa7f5a9209beb163788b4fdf053292db7868ac4 |
|
MD5 | 45da8a3d6c5b26fab8a599376dc287bf |
|
BLAKE2b-256 | 9534b6665c00b08f7f25e29dbcacca1be1be94444d5c04e8c8e8ea1e95ff57d1 |