Monster API recently released its eponymous platform designed to democratize the use of the most advanced machine learning models (and techniques) available today.
This solution provides on-demand access to a pool of GPUs over a globally distributed computing network. This fact alone significantly lowers the barriers to entry (and cost) for training, improving, and deploying advanced statistical AI models, including large-scale language models (LLMs).
According to Monster API co-founder Gaurav VisiThe Monster API platform aims to “democratize AI models in general as the cost of GPU computing is prohibitively expensive.” So far, cloud availability has been limited to large enterprises. Our approach helps give developers and machine learning engineers access to GPU computing at a fraction of the cost. “
Backed by a set of GPUs at the platform’s disposal, the Monster API provides APIs to some of the most useful advanced machine learning models for natural language technology and computer vision applications. The platform also includes specific pipelines for training models and features for fine-tuning hyperparameters.
The optimization techniques involved, combined with cheap access to a large number of GPUs, greatly reduce traditional inhibitors for developers utilizing these models. The result is “democratized access to AI models and access to the compute underlying those models,” commented Saurabh Vij, CEO of Monster API.
GPU on demand
The foundation of Monster API’s platform is a distributed computing network of GPUs served by hundreds of data centers and individuals around the world. As a result, customers can access as many GPUs as they need from regions such as Europe, the US, and India. The network contains over 30,000 GPUs, ranging from traditional gaming to GPUs best suited for machine learning. “AirBnB has access to underutilized empty rooms, and just like Uber is doing it to cars, we’re doing it to GPUs,” Saurab said. clarified. The Monster API automatically scales to handle the workload, so the developer can spin up his GPU resources as needed.
Geographical concerns about where resources are spun up are abstracted from the user by the orchestrator that handles this task. Commercial GPUs designed for games are “prepared for AI models through packages that add containers, GPU drivers, libraries, and over 10 different frameworks to make them work with machine learning. ‘ he explained Saurabh. The platform itself is container-native and runs on Kubernetes. Security measures include five different encryption protocols, different access control models, container process isolation and data-level isolation, which states that “all processes and data volumes within a container are protected from other server-level processes. security,” says Gauruv. maintained.
Impressive optimization
The platform’s API gives users access to a number of cutting-edge AI models, including open-source options such as Whisper AI, Stable Diffusion, and more. In addition to optimizing consumer GPUs for machine learning, the platform incorporates optimization techniques to dramatically reduce the cost of training and deploying certain models. For specific use cases related to Whisper AI translation and speech-to-text transcription, “Improved throughput and reduced memory footprint” [of the model]” revealed Gaulf.
According to Gaurav, using Whisper AI out-of-the-box for translation and speech-to-text jobs on AWS would cost about $45,000. Using the Monster API’s optimized methods, the same job would cost him less than $3,000. “The model is also optimized, reducing the time required to process a request, which doubles the cost savings,” Gaurav noted. “And the cost of GPU infrastructure is already low.”
Model fine-tuning
The Monster API also supports the concept of model fine-tuning. This is another aspect of data science that can be costly if left unchecked. The platform incorporates a no-code fine-tuning solution, which involves users accessing machine learning models and datasets used to tune the models for their individual use cases. will be
“You can use pre-trained underlying models. You can grab datasets from free datasets like Hugging Face and immediately start fine-tuning these underlying models for custom datasets,” says Saurabh. explained Mr. “Instead of the hundreds of dollars we could spend fine-tuning these models, we can do it for less than $30-$40.”
Beyond cost savings
The reduction in cost of deploying advanced machine learning with the Monster API is remarkable. But the platform does more than just reduce the overhead of such work. The cost benefits of distributed GPU networks, optimization approaches, and model fine-tuning are matched by the overall accessibility provided to rapidly retrieve and curate these models for selected enterprise applications. increase.
So it affects not only garage developers but also executives, both of whom may benefit.
