The developer promises to reduce GPU and infrastructure operation expenses by up to 80%

Laboratoriya Chislitel has introduced a new edition of the Shturval container platform designed for machine learning and AI workloads. The Shturval ML/AI solution combines CPU and GPU computing within a single Kubernetes cluster, enabling the entire model lifecycle—from experimentation to launching production services.

The platform supports training and deploying large language models, including Qwen, DeepSeek, and Llama. The solution incorporates Kubeflow and MLflow tools for organizing experiments, training, and model versioning. KServe is used for model publishing, while language model serving is handled via vLLM.

The developers paid special attention to computational resource allocation. The platform allows graphics cards to be pooled together, allocated to specific teams, and monitored for utilization. NVIDIA Multi-Instance GPU (MIG) technology is applied to partition compatible accelerators. According to company estimates, this approach can reduce GPU and infrastructure operation expenses by up to 80%.

Updates to Kubernetes, drivers, and machine learning tools are performed jointly. The platform also provides centralized access management, vulnerability scanning for software images, and streaming of audit events to external security systems. The solution can be deployed in an air-gapped environment.

Shturval ML/AI supports a hybrid infrastructure where virtual machines are used for standard tasks and bare-metal servers with GPUs are used for training and running AI models. Supported accelerators include NVIDIA A100, H100, and L40S, while supported operating systems include Astra Linux, RED OS, Rosa, Alt Linux, Ubuntu, and Debian.