Vensen·Oct 1Deterministic Collective Communication on ROCmIn multi-GPU training, AllReduce is often treated as infrastructure: call it, wait for it to complete, and move on. We rarely ask in what…
Zeenatriaz·Sep 28Why GPU Development Docker Images Balloon to 8+ GB for Simple HIP CompilationFixing dependency coupling, package mapping bottlenecks, and container cold-starts in AMD ROCm CI/CD engineering pipelines.
Zeenatriaz·Sep 28Ultimate Guide to MLOps: Optimizing LLM Inference with vLLM and AMD ROCmA production-ready blueprint for high-throughput machine learning systems engineering, advanced hardware acceleration, and zero-trust…
Shivank Chaudhary·Aug 26Deploying a CXI Fabric DRA Driver on Kubernetes: CDI, NRI, ResourceClaims, and Shared VNIs[learning…An end-to-end guide to exposing high-performance HPC NICs to pods with Dynamic Resource Allocation — the concepts, the commands, the actual…
Extremys·Aug 9Why I Kept Qwen3.5 9B at Q4 on a Radeon 780MThe smaller Q2 file generated faster tokens. It also completed none of my 15 local agent tasks.
Extremys·Aug 9How I Made a Radeon 780M Useful for Local LLM AgentsOn a GMKtec NucBox K6, the useful answer was not “get more tokens per second.” It was to spend a small integrated GPU budget where it…
Zecchi Zero·Jul 31Solución al error torch.cuda.is_available()Este artículo documenta cómo solucionar el error OSError: [Errno 28] No queda espacio en el dispositivo y el AttributeError: module 'torch'…
AlexHe99·Jun 9UniLab: CPU+GPU Heterogeneous Architecture for Robot RL, with Native AMD ROCm Support Across the…1. Introduction: Why Heterogeneous Computing Matters for Robot RL
Shaunak J·Jun 8Fine-Tuning LLMs on AMD ROCm: A Practical Axolotl Workflow for the MI300XBreaking the CUDA monopoly with ROCm, QLoRA, checkpointed training, and 192GB of VRAM.
Balaji Raghuram·May 22The Strix Halo Chronicles #1: Can WSL 2 Beat Native Windows for Qwen 3.6A hands-on engineering log documenting my attempt to run massive LLMs on AMD’s latest 8060S iGPU. This series tracks the setup, hurdles…
