PyTorch 2.11 on Linux: Installing, Configuring CUDA/ROCm, and Running Your First Model
PyTorch 2.11 changed how pip install torch behaves -- the default wheel now targets CUDA 13.0, not CPU-only, and Volta (V100) support was quietly dropped from the CUDA 12.8 builds. This guide covers every decision point: virtual environment setup, matching CUDA versions to your driver, GPU generation compatibility from Pascal to Blackwell, the complete AMD ROCm wheel path, and a verified inference example that confirms end-to-end GPU execution.
PyTorch 2.11 on Linux: Installing, Configuring CUDA/ROCm, and Running Your First Model
The critical PyPI default change in 2.11, Volta/V100 wheel compatibility, CUDA version matching, AMD ROCm pip installation, GPU verification, and a working ResNet-18 inference example. Includes interactive tools for install command generation and VRAM budget estimation.
AI / ML OpsRunning AI Workloads on Linux: A Beginner's Setup Guide
Start here before touching pip. Covers the hardware and OS prerequisites for running ML on Linux -- GPU driver requirements, CUDA vs ROCm decision framework, container-based inference options, and what to set up before installing any framework.
AI / ML OpsSetting Up a Local AI Stack on Linux: Ollama + Open WebUI + Docker
Ollama runs as a systemd service with full GPU access. Open WebUI connects to it from a Docker container and gives every model a ChatGPT-style interface with conversation history, RAG document uploads, and multi-user authentication. Pull additional models directly from the admin panel, add OpenAI-compatible cloud endpoints alongside your local ones, and update the entire stack with a single command without touching the named volume that holds all your data.
AI / ML OpsLinux GPU Acceleration for AI: CUDA vs ROCm — What You Need to Know in 2026
The install method changed in ROCm 7.2.1: AMD removed the amdgpu-install documentation from the primary flow and now leads with the native APT package manager. This guide covers what that means in practice, why AMD GPUs silently fall back to CPU without an error when ROCm rejects unsupported hardware, how HSA_OVERRIDE_GFX_VERSION fixes it, the updated GFX version table for every RDNA 2/3/4 card, and where PyTorch on ROCm still lags CUDA in 2026.
AI / ML OpsHow to Run LLMs Locally on Linux with Ollama (NVIDIA and AMD GPU Setup)
Ollama v0.20+ selects your GPU backend automatically -- CUDA for NVIDIA, ROCm v7 for supported AMD cards, and Vulkan as the fallback for Intel and older AMD hardware that ROCm rejects. This guide covers compute capability requirements, the HSA_OVERRIDE_GFX_VERSION fix for unofficially supported RDNA GPUs, GGUF quantization selection, KV cache tuning, and the verification commands that confirm your model is actually running on the GPU rather than silently falling back to CPU.
AI / ML OpsBest Linux Distros for AI and ML Work in 2026
The distro is not your AI stack -- it is the substrate beneath it. This guide covers what Ubuntu 24.04 LTS, Fedora 43, Pop!_OS 24.04 LTS, and Arch Linux each get right and wrong specifically for ML work: CUDA lazy module loading behavior, ROCm 7.2.1's hipBLASLt regression, the Blackwell open kernel module requirement, protecting a working driver stack from kernel updates, and why the uv package manager changes the Python freshness calculation between distros entirely.