KDnuggets Releases Ollama Cheat Sheet for Port 11434 Local LLM Management

Q: Why does one command—`ollama ps`—matter so much for spotting GPU limits, CPU fallback, and hidden slowdowns in local AI workflows?

`ollama ps` matters because it is the quickest reliable view into the real bottleneck in local inference: whether a model is fully on the GPU, split between GPU and CPU, or running on CPU alone. That distinction decides performance. A model that fits entirely in VRAM or Apple unified memory can generate at interactive speeds. If memory runs short, Ollama offloads part of the workload to system RAM and CPU. The model still works, but token generation can slow dramatically. `ollama ps` exposes that hidden shift in the `PROCESSOR` field, showing states such as `100% GPU`, `100% CPU`, or a split like `48%/52% CPU/GPU`. Without that readout, users may blame the model, prompt, or software version when the real problem is memory pressure. It also shows which models are still loaded and how much memory they occupy. That is important because local slowdowns often come from residency and context growth rather than installation problems. A long context window, multiple loaded models, or parallel requests can quietly push a system past GPU limits. `ollama ps` reveals those conditions immediately and helps users decide whether to unload a model with `ollama stop`, shrink context, choose a smaller quantization, or reduce concurrency. In practice, it is the operational dashboard for Ollama. The command turns vague complaints like “it feels slow” into a concrete diagnosis: GPU fit, CPU fallback, memory saturation, or stale models still sitting in memory.

1 articles · Updated · KDnuggets · Sep 30

KDnuggets published a new Ollama cheat sheet focused on running, configuring and optimizing local language models through Ollama’s localhost server and OpenAI-compatible interface.
Port 11434 sits at the center of that setup: Ollama serves models over HTTP, while ollama ps shows whether a model still fits in memory or has spilled from GPU to CPU, slowing generation.
The guide also flags a common macOS pitfall—desktop-launched Ollama ignores shell exports in .zshrc, so settings such as context length or model directory must be set with launchctl setenv.
Structured output is another focus: passing a JSON schema to format constrains replies to a fixed shape, while plain json only guarantees valid JSON without specific keys.
The cheat sheet rounds out local deployment basics with /api/chat, /api/embed, the /v1/ compatibility layer, disk-management commands, Modelfiles and environment variables controlling model residency and concurrency.