From: Zagitta <github.com/Zagitta>Newsgroups: alt.freeweights.toolsSubject:

gpu-unlock

Release: noneCommit: 7eb51bcLicense: GPL-2.0

Your GeForce runs AI at half speed. This fixes it for 20/30/40 series cards. Up to 2× FP16 / BF16 / TF32 tensor throughput

inference engineno releasereleasegpu-unlockCGPL-2.0

FactsLicenseGPL-2.0LanguageCKindinference engineReleasenonePlatformsLinuxLinks to the projectcopyleft

$ git clone https://github.com/Zagitta/gpu-unlock && cd gpu-unlock
$ make install
$ Add gpu_unlock.targets=all to the kernel command line, then reboot
# use with Freeweights
$ Linux kernel module for GeForce RTX 20, 30 and 40 cards (not RTX 50 / Blackwell, not pro cards). It removes the consumer half-rate cap on FP16, BF16 and TF32 tensor math with FP32 accumulation, so local inference and fine-tuning of Freeweights models run faster. The project reports bit-identical outputs; Ampere is marked stable, Ada and Turing experimental. Laptops lose the unlock when the GPU powers off. Use at your own risk.

Overview Install Use with Freeweights License Provenance

Overview

Kernel module that lifts the GeForce half-rate tensor-core cap for FP16/BF16/TF32 (RTX 20/30/40, Linux).

Kind: inference engine
License: GPL-2.0
Language: C
Platforms: Linux
Latest release: no release
Repository: Zagitta/gpu-unlock

Freeweights links to the project and does not host tool binaries.

Install

$ git clone https://github.com/Zagitta/gpu-unlock && cd gpu-unlock
$ make install
$ Add gpu_unlock.targets=all to the kernel command line, then reboot
# use with Freeweights
$ Linux kernel module for GeForce RTX 20, 30 and 40 cards (not RTX 50 / Blackwell, not pro cards). It removes the consumer half-rate cap on FP16, BF16 and TF32 tensor math with FP32 accumulation, so local inference and fine-tuning of Freeweights models run faster. The project reports bit-identical outputs; Ampere is marked stable, Ada and Turing experimental. Laptops lose the unlock when the GPU powers off. Use at your own risk.

$ git clone https://github.com/Zagitta/gpu-unlock && cd gpu-unlock
$ make install
$ Add gpu_unlock.targets=all to the kernel command line, then reboot

Checked against the project's README on 2026-11-14. The project's own instructions win when they differ: install docs

Freeweights links to the project and does not host tool binaries.

Use with Freeweights

Run a catalog model

$ Linux kernel module for GeForce RTX 20, 30 and 40 cards (not RTX 50 / Blackwell, not pro cards). It removes the consumer half-rate cap on FP16, BF16 and TF32 tensor math with FP32 accumulation, so local inference and fine-tuning of Freeweights models run faster. The project reports bit-identical outputs; Ampere is marked stable, Ada and Turing experimental. Laptops lose the unlock when the GPU powers off. Use at your own risk.

Download the weights from Freeweights first (HF_ENDPOINT or fw.py), then point the tool at the folder.

Catalog models it can run (3852)

135M-lc-100k-rope-12B a2-rl-e2egit_large-50-32B AFM-4.5B AFM-4.5B-Base AFM-4.5B-ONNX AgentCPM-Report agentica-org_DeepCoder-14B-Preview-GGUF agentica-org_DeepScaleR-1.5B-Preview-GGUF agentica-org_DeepSWE-Preview-GGUF Agents-A1-4B AI21-Jamba-Reasoning-3B AI21-Jamba2-Mini ai21labs_AI21-Jamba-Reasoning-3B-GGUF ai21labs_AI21-Jamba2-Mini-GGUF ai9stars_G9v3-3B-GGUF alamios_Mistral-Small-3.1-DRAFT-0.5B-GGUF AlchemistCoder-L-7B AlchemistCoder-L-7B-GGUF Alexandria-Gemma-4-12B-it-Ornith9-Summaries-LoRA-r128 Alexandria-Gemma-4-12B-it-Ornith9-Summaries-LoRA-r64 Alexandria-Gemma-4-12B-it-Qwen27-Knowledge-Units-LoRA-r128 Alexandria-Gemma-4-12B-it-Qwen27-Summaries-LoRA-r128 Alexandria-Gemma-4-12B-it-Qwen27-Summaries-LoRA-r64 Alexandria-Gemma-4-E4B-it-Qwen27-Knowledge-Units-LoRA-r128 Alibaba-NLP_Tongyi-DeepResearch-30B-A3B-GGUF ALLaM-7B-Instruct-preview allenai_Olmo-3-7B-Instruct-GGUF allenai_Olmo-3-7B-Think-GGUF allenai_Olmo-3.1-32B-Instruct-GGUF allenai_Olmo-3.1-32B-Think-GGUF AM-Thinking-v1 AM-Thinking-v1-GGUF AMD-OLMo-1B AMD-OLMo-1B AMD-OLMo-1B-SFT AMD-OLMo-1B-SFT AMD-OLMo-1B-SFT-DPO AMD-OLMo-1B-SFT-DPO AMD-OLMo-1B-SFT-DPO-GGUF AMD-OLMo-1B-SFT-GGUF Apertus-70B-2509 Apertus-70B-Instruct-2509 Apertus-70B-Instruct-2509-GGUF Apertus-8B-2509 Apertus-8B-Instruct-2509 Apertus-8B-Instruct-2509-GGUF Apertus-8B-Instruct-2509-ONNX Apertus-8B-Instruct-2509-unsloth-bnb-4bit Apertus-v1.1-0.5B Apertus-v1.1-0.5B-Instruct Apertus-v1.1-0.5B-Instruct-ONNX Apertus-v1.1-0.5B-Instruct-QAD-INT4-ONNX Apertus-v1.1-1.5B Apertus-v1.1-1.5B-Instruct Apertus-v1.1-1.5B-Instruct-ONNX Apertus-v1.1-1.5B-Instruct-QAD-INT4-ONNX Apertus-v1.1-4B Apertus-v1.1-4B-Instruct Apertus-v1.1-4B-Instruct-ONNX Apertus-v1.1-4B-Instruct-QAD-INT4-ONNX

and 3792 more.

License

License: GPL-2.0
Class: Copyleft (GPL)

GPL-2.0 text is in licenses/GPL-2.0.txt (no root LICENSE file, so GitHub shows none). The build downloads NVIDIA and CachyOS inputs over HTTPS and generates firmware on your machine; check NVIDIA's terms for your use. Freeweights links to the source and hosts none of it.

Read the license

Freeweights links to the project and does not host tool binaries.

Provenance

From: Zagitta <github.com/Zagitta>Newsgroups: alt.freeweights.toolsSubject: gpu-unlockRelease: noneCommit: 7eb51bcLicense: GPL-2.0

Repository: Zagitta/gpu-unlock
Pinned release: none
Release commit: not known
Default branch: main 7eb51bccae0a0b7e599e084d971170e9a1e82d0c
Fetched: 2026-11-14
Takedowns: None.

Freeweights links to the project and does not host tool binaries.

Get it

Install Project repository

Works with

3852 catalog models. See the list