Models & Labs
Together AI Enhances GPU Cluster Reliability and Control
Together AI Bloghigh confidence
Why it matters
- →Enhancements improve the reliability of large-scale AI training jobs by quickly detecting and resolving hardware failures.
- →The new Slurm-on-Kubernetes stack addresses common operational issues, ensuring smoother and more efficient cluster management.
- →These updates enable users to manage AI workloads with less manual intervention, supporting scalability and robustness.
Together AI has rolled out updates to its GPU clusters, enhancing reliability and operational control. The updates include passive health checks and auto node repair to swiftly detect and address hardware failures, reducing downtime for training jobs. The Slurm-on-Kubernetes stack has been rebuilt to eliminate issues like zombie processes and inaccurate GPU states. These improvements allow users to manage clusters more efficiently, supporting robust and scalable AI workloads. The changes reflect a commitment to providing seamless and reliable infrastructure for AI operations.
Read originalMore from Together AI Blog
More in Models & Labs
Models & Labsmodels
llama.cpp b10043 release supports CUDA Virtual Devices
The b10043 release of llama.cpp marks a notable enhancement with the addition of CUDA Virtual Devices, which significantly improves GPU resource management. By removing the NCCL path when virtual devices are in use, the update fine-tunes performance for these specific setups. This release also includes a comprehensive code refactor and the implementation of GPUx2 server CI jobs, reflecting a commitment to better testing and deployment processes. While there are no new model architectures, the update enhances the platform's flexibility across various operating systems, making it more adaptable for developers working with a wide range of hardware configurations.
llama.cpp Releases
Models & Labsmodels
Llama.cpp b10051 Release Enhances SME Kernel Dispatch
The latest release of llama.cpp, b10051, addresses a critical issue in kernel dispatch by distinguishing between SME and SME2 capabilities. Previously, the integration treated SME as a single capability, leading to incorrect dispatch on SME(v1)-only hardware due to the use of SME2-specific instructions. This update introduces both build-time and runtime distinctions, ensuring that kernels are dispatched based on actual hardware support. This refinement enhances the accuracy and efficiency of operations on different hardware configurations, marking a significant improvement for developers working with these systems.
