PrayogAditya_v2_29B

PrayogAdtiya_v2_29B is a Multimodel which is based on Next Generation Dynamic LLM modeling which can reduce the GPU cost by upto 50% and more without compromising the model parameters .

PrayogAdtiya_v2_29B have around 29 Billion Parameters and PrayogAdtiya_v2_29B have different pretrained models mentioned in modality_configuration.py which needs to be dynamically loaded when user do the prediction using this model for different modalities. All around 29 Billion Paremeters are from Pretrained Models only. PrayogAdtiya_v2_29B is not trained as a new LLM from Scratch or fine tuned any LLM instead it is build as an Interface to solve the most burning Problem in Artificial Intelligence Industry that is Computation Problem, as to run these LLM, GPU required which makes significant cost and GPU needs significatnt amount of electricity which may cause global warming. so if LLM can be run in a way computationally effective then it can save the cost and electricity both.

PrayogAdtiya_v2_29B is inspired from a vision that is-

Why One need to pay for information inside the LLM which is not required for the prediction one asked LLM to provide?

This Vision leads us to innovative solution which is based on Keras Subclass API and Trasnformers Subclass where One can orchestrate the multiple models inside a single model or single model graph. This orchestration have a top layer with Natural Language Inference (NLI) pretrained model based on the configuration mentioned in modality_configuration.py which helps to classify the input to be routed towards the selected pretrained model based on the configuration mentioned in modality_configuration.py file. The Key Features of Dynamic LLM Modeling Design used in this model is-

1. It does not load all the models into memory like Mixture of Expert Models it only loads and unloads the models based on the Input meaning Input first go to NLI Model then it classify which model to be loaded then NLI model unloaded then relevent model loaded for prediction. This way PrayogAdtiya_v2 make sure Optimal uses of compute resources like RAM and CPU/GPU.
2. Many Enterprise models and community models used for large scale customer serving where multi-processing is required, in that case for multiple instance of single model loaded can lead to Memory (RAM) overflow, that's why this model will not load any model when Initailize instead it loads the different model when user do the prediction and after prediction that model unloaded, that's how this Dynamic LLM Modeling saves the critical computes which is not available in others like Mixture of Experts.
3. This model is highly scalable because it does not have any fix pretrained models inside the modeling.py file which is core of this Dynamic LLM Model instead when user do the prediction then only This model dynamically load the pretrained models as mentioned in modality_configuration.py that's why this model can be scaled by just updating the different pretrained models model_id for the different modalities in modality_configuration.py file if required in any case.

There is no point to reinvent the wheels when Open Source Community has done the Great work that's why instead of training the New LLM from Scratch we have decided to first build a dynamic interface driven model which can provide user luxury to use already pretrained open source models.

Usage Example-

pip install diffusers, torch, trasnformers

from transformers import AutoModel

model = AutoModel.from_pretrained("Kratim-Budhimata/PrayogAdtiya_v2_29B", trust_remote_code=True, torch_dtype="auto")

model.predict("tell me something about Indian tourist places")

Note that this model downloads the models required at the time of prediction so if you want to use one time downloaded models then first download the models mentioned in the modality_configuration.py file, in the local file system and then update the local path of downloaded models in modality_configuration.py file accordingly.

License Agreement-

This model (including modeling.py file) is under MIT License and These models model_id mentioned below and in the modality_configuration.py (there is a dictionary for configuration in this file) file, will be loaded dynamically, when user do the prediction (call predict function) having their specifc license which user must need to agree for using these models when user do the prediction-

1. Orchestration Model or Routing Model (Nautal Language Inference Model) - facebook/bart-large-mnli - MIT License
2. Text Generation Model - databricks/dolly-v2-12b -MIT License
3. Image Model - HiDream-ai/HiDream-I1-Full - MIT License

Disclaimer-

The pretrained models are not build or finetuned by Kratim-Budhimata that's why Kratim-Budhimata is not responsible for any kind of response or error occured due to these pretrained models and User is solely responsible to use these models based on these models specific license terms and conditions and Kratim-Budhimata is not resposible for any kind of prediction done by pretrained models and any direct or indirect losses occurs to user or providers due to any kind of error or mistake occurs. User is solely responsible to use these community models based on user's own risk.

Feel Free to reach out to us if you have any query or suggestions-

Email Address- [email protected]