The Zagitta/gpu-unlock repository shares setup notes and configuration tips for the Strata engine that powers large AI models running on local machines. It is aimed at developers and researchers who want private inference without sending prompts to cloud services. The notes focus on what your own GPU, drivers, and OS need to run the engine and how you can test it on your own system.
What it is and why it matters
The Zagitta/gpu-unlock guide explains the "gpu-unlock" concept: instructions and notes for enabling GPU acceleration for large models on consumer hardware under open-source tools like the Strata engine. Running models locally keeps data on your device and reduces the need for expensive cloud compute. The guide was created after the model became available publicly and shows the practical steps needed to get the engine working.
Hardware requirements
To run Strata with a large model on a consumer GPU you need a recent NVIDIA card (for example RTX 3060/3070/3080/3090 class) with at least 8 GB of fast VRAM. The repo recommends using drivers that support CUDA and keeping the OS and stack up to date. The notes stress checking that the correct CUDA toolkit version matches the driver before launching the engine.
| Item | Measured fact | Notes |
|---|---|---|
| Minimum VRAM for a 7B‑parameter model | 8 GB | Strata documentation and community notes point to this as the practical floor |
| Example GPU used in shared tests | [RTX 3080](/gpus/rtx-3080) (10 GB VRAM) | Shows the engine can run comfortably above the minimum |
| Power draw during inference | Roughly 150 W for a mid‑range card | Measured on a representative workload; actual usage varies by model and precision |
How to verify on your own machine
The guide tells you to confirm the setup yourself: check your GPU model and driver version, run `nvidia-smi` to see available VRAM, and test a small model with the Strata engine to ensure no errors occur. You can also time a simple prompt against a 7B model and compare the measured latency to what you expect for that hardware.
Core definition
"GPU unlock" refers to the set of steps and settings that enable a large model to run on a consumer graphics processor instead of only on CPUs or cloud GPUs.
Measuring performance yourself
The guide suggests measuring by running a short generation with a 7B model on your GPU, recording the time for the first token plus the average per‑token time for the next few tokens. Compare that to the baseline timing you would get on a CPU or a lower‑end GPU.
Cost estimate
Running a 7B model locally on a mid‑range card typically uses about 150 W of power while generating output. At a local electricity rate of $0.17 per kWh, a few hours of inference can cost well under a dollar, making local execution cheaper than renting cloud GPU time for occasional use.