# AI use cases for GPU Unlock (Zagitta/gpu-unlock) Repo: [Zagitta/gpu-unlock](https://github.com/Zagitta/gpu-unlock) | Use cases: 3 | Page: https://dedupedai.com/use-cases/zagitta-gpu-unlock/ gpu-unlock is a Linux kernel module that removes the half-rate cap NVIDIA puts on GeForce tensor cores for FP16, BF16 and TF32 math with FP32 accumulation. It loads at boot before the stock NVIDIA driver, with no patched driver or VBIOS flash. The README reports up to 2x tensor throughput and 25 to 55 percent faster LLM prefill on RTX 30 cards. This file holds every use case of this repo. Each one starts with a "## " heading and ends with its URL. - Speed up local LLM prompt processing on an RTX 30 card: https://dedupedai.com/use-cases/zagitta-gpu-unlock/speed-up-local-llm-prefill-on-an-rtx-3090/ - Run BF16 LoRA fine-tuning faster on a GeForce card: https://dedupedai.com/use-cases/zagitta-gpu-unlock/faster-bf16-lora-fine-tuning-on-a-geforce-card/ - Generate SDXL and ComfyUI images faster on RTX cards: https://dedupedai.com/use-cases/zagitta-gpu-unlock/faster-sdxl-and-comfyui-image-generation/ ## How to speed up local LLM prompt processing on an RTX 30 card with GPU Unlock Repo: [Zagitta/gpu-unlock](https://github.com/Zagitta/gpu-unlock) | Difficulty: Advanced | Time: 45 minutes | Basis: From the README Install the gpu_unlock module so FP16 and BF16 tensor math runs at full rate for vLLM, PyTorch and Transformers. **Who it is for:** A home lab user serving unquantized or AWQ models on an RTX 3060 to 3090 who waits on long prompts. **The problem:** GeForce cards run tensor core math with FP32 accumulation at half the rate of the same silicon sold as a pro card. **The outcome:** Faster prefill and time to first token on FP16, BF16, AWQ or GPTQ models, with bit-identical outputs. ### Steps 1. Run `git clone https://github.com/Zagitta/gpu-unlock`, `cd gpu-unlock` and `make install` 2. Append `gpu_unlock.targets=all` to your kernel command line (GRUB, systemd-boot or UKI, as the README shows) 3. On Debian or Ubuntu run `sudo update-grub`, then reboot 4. Rerun your usual vLLM or Transformers benchmark and compare prefill speed ### Prompt to paste into your agent ```text Walk me through installing Zagitta/gpu-unlock on this machine with an : build the module, add gpu_unlock.targets=all to the kernel command line for my bootloader, and give me a before and after benchmark to run. ``` ### Caveat This card is written from the Zagitta/gpu-unlock README, not from a test run by Deduped. Before you start, make sure you have Linux. Check the repo for newer instructions before you rely on these steps. Risk, system-change: This installs a kernel module that loads before the NVIDIA driver; keep a known good boot entry in case the system does not start. Risk, unofficial: This is not supported by NVIDIA. The README says harm is very unlikely but the card may run hotter. Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/ URL: https://dedupedai.com/use-cases/zagitta-gpu-unlock/speed-up-local-llm-prefill-on-an-rtx-3090/ ## How to run BF16 LoRA fine-tuning faster on a GeForce card with GPU Unlock Repo: [Zagitta/gpu-unlock](https://github.com/Zagitta/gpu-unlock) | Difficulty: Advanced | Time: 1 hour | Basis: From the README Unlock full tensor rate before a long LoRA or full fine-tune so training hours drop. **Who it is for:** A hobbyist or researcher fine-tuning models in BF16 on a single consumer GPU. **The problem:** Training runs on GeForce take far longer than the hardware could manage because of the tensor cap. **The outcome:** The README reports 25 to 50 percent faster LoRA and full fine-tuning in BF16 on supported cards. ### Steps 1. Check your card is in the Supported GPUs table 2. Install the module and add `gpu_unlock.targets=all` to the kernel command line 3. Reboot and confirm the unlock is active before starting training 4. Time a short training run before and after and post your numbers in the repo's compatibility reports if you like ### Prompt to paste into your agent ```text Help me use Zagitta/gpu-unlock before a BF16 LoRA run on my : confirm my card is supported, install it, and write a short timing script to compare steps per second before and after. ``` ### Caveat This card is written from the Zagitta/gpu-unlock README, not from a test run by Deduped. Before you start, make sure you have Linux. Check the repo for newer instructions before you rely on these steps. Risk, system-change: Kernel level change; laptops only stay unlocked if the dedicated GPU is never fully powered off. Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/ URL: https://dedupedai.com/use-cases/zagitta-gpu-unlock/faster-bf16-lora-fine-tuning-on-a-geforce-card/ ## How to generate SDXL and ComfyUI images faster on RTX cards with GPU Unlock Repo: [Zagitta/gpu-unlock](https://github.com/Zagitta/gpu-unlock) | Difficulty: Advanced | Time: 45 minutes | Basis: From the README Remove the FP16 tensor cap so diffusion models in ComfyUI render faster. **Who it is for:** An artist or tinkerer running SDXL, SD 1.5 or DiT models in ComfyUI on a GeForce card. **The problem:** Each image takes longer than it should because FP16 tensor math is held at half rate. **The outcome:** The README reports 30 to 55 percent faster FP16 or BF16 image generation, with identical outputs. ### Steps 1. Install GPU Unlock and set `gpu_unlock.targets=all` on the kernel command line 2. Reboot and open ComfyUI as usual 3. Render the same seed and workflow you used before and compare the time per image 4. Note quantized INT8, FP8 and INT4 models and gaming see no change ### Prompt to paste into your agent ```text Help me install Zagitta/gpu-unlock and measure ComfyUI SDXL render time on my before and after with the same seed and workflow. ``` ### Caveat This card is written from the Zagitta/gpu-unlock README, not from a test run by Deduped. Before you start, make sure you have Linux. Check the repo for newer instructions before you rely on these steps. Risk, system-change: Kernel module installed at boot; not supported by NVIDIA. Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/ URL: https://dedupedai.com/use-cases/zagitta-gpu-unlock/faster-sdxl-and-comfyui-image-generation/