llama-cpp-python
llama-cpp-python copied to clipboard
abetlen
CUDA wheel installs, but GPU is never used on Windows 11 (Python 3.11, CUDA 12.1, torch finds GPU)
On Windows 11, with Python 3.11 and a CUDA 12.1-compatible NVIDIA GPU, I can successfully install llama-cpp-python via pip from the cu121 wheel, but no matter what, all model layers are always assigned to CPU only.
torch works and finds the GPU (torch.cuda.is_available() == True).
All paths and DLLs appear correct.
No errors, just silent fallback to CPU for every layer, even with n_gpu_layers set. Tried -1 and 32
I have confirmed this across multiple clean venvs and installs.
System Info
OS: Windows 11 Pro
GPU: NVIDIA RTX 4060 Ti, 16 GB VRAM
NVIDIA Driver Version: NVIDIA-SMI, 581.57
CUDA Toolkit Version: 12.1 (from PATH and NVIDIA tools)
Python Version: 3.11.x
llama-cpp-python Version: 0.3.16 (from pip show llama-cpp-python)
Installation Command:
pip install llama-cpp-python --extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cu121
Model Used: Mistral 7B Q5_K GGUF (4.7GB)
What I Tried
Fresh venv (python -m venv envgpu311)
Confirmed correct Python version: where python → E:\envgpu311\Scripts\python.exe
pip list shows torch 2.5.1+cu121, llama-cpp-python 0.3.16
System PATH contains correct CUDA 12.1 bin directory
cudart64_121.dll is present in CUDA bin
Test script:
from llama_cpp import Llama llm = Llama(model_path="E:/Python/Models/Mistral/lucian_model.gguf", n_ctx=512, n_gpu_layers=32, verbose=True) print(llm("Hello, world!"))
Logs never mention CUDA or GPU, only assigned to device CPU for every layer.
torch tests pass:
import torch print(torch.version) # 2.5.1+cu121 print(torch.cuda.is_available()) # True
Tried uninstalling all other venvs, cleaning PATH, and purging pip cache.
Also tried n_gpu_layers=-1 and 32 as stated above
What Happens
Model loads, runs, and produces output—but all layers go to CPU, no GPU assignment or CUDA logs, and performance matches CPU.
What I Expected
Layers assigned to GPU, CUDA logs in output, and at least some portion of the model running on GPU
Other Details
No error messages, just silent fallback to CPU.
Issue persists across all wheels and source builds tested.
pip show llama-cpp-python:
Name: llama_cpp_python Version: 0.3.16 Location: E:\envgpu311\Lib\site-packages Requires: diskcache, jinja2, numpy, typing-extensions
torch and other CUDA tools/apps detect GPU fine.
Latest NVIDIA driver installed 10/15/2025
Use a fork that provides prebuilt Windows wheels, such as JamePeng/llama-cpp-python
pip install https://github.com/JamePeng/llama-cpp-python/releases/download/v0.3.16-cu128-AVX2-win-20251007/llama_cpp_python-0.3.16-cp311-cp311-win_amd64.whl
The original llama-cpp-python repository does not provide Windows wheels.
You can confirm this on the releases page abetlen/llama-cpp-python/releases or wheels page - the file names include the Python version and OS.
To verify the installation
python -c "import llama_cpp; print(llama_cpp.llama_print_system_info())"
# Example output for CPU support:
# CPU : SSE3 = 1 | SSSE3 = 1 | ...
# Example output for CUDA support:
# ggml_cuda_init: found 1 CUDA devices: ...
# CUDA : ARCHS = 700,750,800,860,890,900 | FORCE_MMQ = 1 ...
its
pip install https://github.com/JamePeng/llama-cpp-python/releases/download/v0.3.16-cu128-AVX2-win-20251023/llama_cpp_python-0.3.16-cp311-cp311-win_amd64.whl
It seems it also wrongly states (see below) that pre-built wheels for CUDA 12.5 exist when this actually shows 404: https://abetlen.github.io/llama-cpp-python/whl/cu125/llama-cpp-python/
Latest pre-built wheel available only for CUDA 12.4: https://abetlen.github.io/llama-cpp-python/whl/cu124/llama-cpp-python/
