Is Your Local LLM Spilling Out of VRAM? How to Check

You load a model, ask it something, and the answer crawls out one word at a time. The model "works," but it feels broken. Often the cause is simple: the model didn't fit in your GPU's memory, and part of it quietly moved into system RAM.

The good news is you don't have to guess. Both of the popular local runners will tell you exactly where the model landed, if you know where to look.

Check it in Ollama with one command

Load your model, send it a prompt, then open a second terminal and run:

ollama ps

Look at the PROCESSOR column. The Ollama FAQ spells out what it means:

  • 100% GPU means the whole model is in video memory. This is what you want.
  • 100% CPU means it is running entirely from system memory.
  • A split like 48%/52% CPU/GPU means part of the model is on the GPU and part is in system RAM.

That split is the one that catches people. The model still answers, so it's easy to miss. But every token now has to touch the slower memory too, and generation speed drops hard.

Why a model that "should fit" doesn't

Most people size a model by its download size. That's a start, but it leaves things out. The rough rule I use is:

VRAM (GB) = parameters in billions × bytes per parameter × 1.2 + 2

At Q4_K_M, bytes per parameter is about 0.5. So an 8B model works out to 8 × 0.5 × 1.2 + 2, about 6.8 GB. A 14B model is about 10.4 GB. A 32B model is about 21.2 GB. The 1.2 covers overhead and a short context, and the extra 2 GB is for the system and anything else using the card, like your desktop and browser.

Run that math for a 14B model on an 8 GB card and you can see the problem right away. It comes to about 10.4 GB, more than the card has, so part of the model ends up on the CPU.

The context window is the sneaky part

Context also costs memory, and it grows as you raise it. Ollama's FAQ says its default context window is 4096 tokens, and you can raise it with the OLLAMA_CONTEXT_LENGTH environment variable or the num_ctx parameter. Raise it a lot and a model that ran at 100% GPU yesterday can show a CPU/GPU split today, with no other change.

So if ollama ps shows a split right after you bumped the context, that's your answer. Drop it back down and check again.

In llama.cpp, the flag is -ngl

If you run llama.cpp's llama-server directly, the setting that controls this is -ngl (also written --n-gpu-layers). Per the server README it sets the maximum number of layers to keep in VRAM, and it takes an exact number, auto, or all. The default is auto, and the server also has a --fit option, on by default, that adjusts unset settings to fit in device memory. If you set a layer count by hand and the model spills, lower it, or let auto do the work.

What to do when it spills

You have three honest options, and none of them is magic:

  1. Pick a smaller quant. Going from Q8 to Q4_K_M roughly halves the weight memory. For most chat use that's the right trade.
  2. Pick a smaller model. A 14B that fits fully on the GPU will usually feel better than a 32B that half lives in system RAM.
  3. Shrink the context. If you don't need a huge window, don't pay for one.

And if none of those works for what you want to run, that's the real signal it's time for a card with more memory.

Before you buy anything, run the numbers for the exact model and context you have in mind. I keep the full tables in the how much VRAM do I need guide, and the VRAM calculator does the same math with a context add-on.

Sources: Ollama FAQ (docs.ollama.com/faq), llama.cpp server README (github.com/ggml-org/llama.cpp, tools/server). Memory figures are formula estimates, not benchmarks.

Comments

Post a Comment

Popular posts from this blog

AMD RX 9070 XT for Local LLMs: ROCm Performance Guide (2026)

The AMD RX 9070 XT launched in early 2026 at $599 with 16GB GDDR6 and 648 GB/s memory bandwidth. For local LLM users who want a mid-range GPU without NVIDIA prices, this changes the calculus significantly. Here is what it actually delivers. RX 9070 XT Specs vs Competition GPU VRAM Bandwidth Price LLM Score RX 9070 XT 16 GB 648 GB/s $599 ~60 tok/s RTX 4070 Ti Super 16 GB 672 GB/s $799 ~62 tok/s RTX 4080 Super 16 GB 736 GB/s $999 ~68 tok/s RTX 4090 24 GB 1,008 GB/s ~$1,100 ~90 tok/s At $599 with 16GB VRAM, the RX 9070 XT is the best value GPU for local LLMs in the mid-range tier in 2026. See the full comparison at LLMHardware.io's AMD GPU guide . ROCm Setup on Linux (2026) # Ubuntu 22.04/24.04 sudo apt install rocm-hip-libraries hip-base # or use the AMD installer: wget https://repo.radeon.com/amdgpu-install/6.3/ubuntu/jammy/amdgpu-install_6.3.60300-1_all.deb sudo dpkg -i amdgpu-install*.deb sudo amdgpu-install --usecase=rocm,lrt,openclsdk # Verify ROCm sees your GPU r...
Read more

llama.cpp flash attention: how to enable and benchmark

Image
Flash attention computes attention with far less memory traffic. The math is identical; the memory movement is not. In llama.cpp it matters most in two situations: long context windows and tight video memory (VRAM). With flash attention enabled, the key-value cache shrinks substantially. A 32k-token context window becomes realistic on a 16GB card, and the prompt-processing phase speeds up, so long documents get ingested faster. The catch is compatibility: not every model and quantization combination supports it, and on some setups the speedup is modest. If you run short chats on a strong GPU, you may barely notice. Where it shines is retrieval-augmented generation (RAG) and long-document work on mid-range cards. Enable it, compare your tokens per second, and keep it if your workload benefits. For the common case of pasting a long PDF and asking questions, it is close to free performance. See how context length changes VRAM needs .
Read more

SillyTavern with Local LLMs: Complete Setup for AI Companions

Image
SillyTavern is the most feature-rich frontend for character-based AI conversations. What began as an AI roleplay tool has evolved into a powerful platform with persona management, memory systems, extensive prompt engineering tools, and connections to every major local inference backend. Here's how to set it up with local models. Quick connect: SillyTavern connects to Ollama, KoboldCpp, llama.cpp server, and TabbyAPI (ExLlamaV2) out of the box. Any of the hardware setups at LLMHardware.io provides a solid backend for SillyTavern. Installation Install via git: git clone https://github.com/SillyTavern/SillyTavern && cd SillyTavern && npm install then node server.js . Access at localhost:8000. For persistent installation, use the included start.sh script. SillyTavern uses Node.js and runs on any platform. The companion app (STMP) enables mobile access to your SillyTavern instance via Tailscale or local network. Character Cards and the TavernAI Format Chara...
Read more