Sagar Parmar·Nov 2Serving ResNet-50 on Kubernetes with Triton: How Dynamic Batching Delivered More ThroughputA GPU sitting at low utilization is the most expensive kind of idle there is.
Sagar Parmar·Sep 29Guide to LLM Serving with vLLMScaling an application is straightforward. Package it in a container, deploy a few replicas, add an autoscaler, and let Kubernetes handle…A response icon2A response icon2
Sagar Parmar·Sep 18LLM Inference Deep Dive: Prefill, Decode, KV Cache, FlashAttention, and Speculative DecodingLarge Language Models (LLMs) have transformed software engineering, powering intelligent applications across industries. However, serving…
Sagar Parmar·Sep 10Understanding LLM Precision and QuantizationIn traditional infrastructure, capacity planning is relatively straightforward. You size servers based on expected CPU, memory, storage and…A response icon1A response icon1
Sagar Parmar·Jun 22Beyond Native Kubernetes Scheduling: Why Volcano is the Missing Piece in Your AI InfrastructureFrom gang scheduling to topology aware placement. A deep dive into scaling distributed AI workloads on KubernetesA response icon2A response icon2
Sagar Parmar·Jun 9Demystifying JupyterHub: A Powerful Multi-User Platform for Data Science TeamsA practical guide to understanding JupyterHub architecture, deploying it on Kubernetes, and securing access with Azure AD SSO
Sagar Parmar·Mar 18Demystifying NVIDIA MPS: How Multi‑Process Service Improves GPU Sharing and PerformanceDemystifying NVIDIA MPS: The Key to Achieving Higher GPU Utilization and Faster Parallel WorkloadsA response icon1A response icon1
Sagar Parmar·Mar 15Beyond Partitioning: A Deep Dive into NVIDIA GPU Time-SlicingOptimizing Kubernetes GPU clusters: Beyond basic partitioning to high-throughput concurrency and hardware-level isolation
Sagar Parmar·Mar 3Connecting AI Agents to Your Kubernetes Cluster: Building an MCP Server in GoHow AI Agents Can Safely Inspect and Control Kubernetes Using GoA response icon1A response icon1
Sagar Parmar·Feb 20A Practical Guide to GPU Partitioning with MIG on On‑Prem Servers and KubernetesFrom hardware setup to Kubernetes deployment, everything you need to enable MIG and unlock reliable GPU sharing in on‑prem clusters.