Date: 11.01.2026

GLM4 9B CUDA PyTorch Test

Test environment

• Workstation 40 GB RAM, 500GB SSD, 750W Power supply
• Ubuntu 24.04 LTS HWE Kernel
• Install python 3.12

My test environment: HP Z440 + NVIDIA RTX 3090

Ubuntu preparation

sudo apt-get install --install-recommends linux-generic-hwe-24.04
hwe-support-status --verbose
sudo apt dist-upgrade
sudo reboot

Driver setup

• Install drivers nvidia-driver-570

sudo apt install nvidia-driver-570 clinfo
sudo reboot

• Check installation

nvidia-smi
clinfo

Dry-run GLM4 9B

• Preapre python environment for CUDA:

mkdir -p ~/llm && cd ~/llm
python3 -m venv .venv_llm_glm4
source ./.venv_llm_glm4/bin/activate
python -m pip install --upgrade pip
pip install "torch==2.7.1" "torchvision==0.22.1" "torchaudio==2.7.1" --index-url https://download.pytorch.org/whl/cu128
pip install transformers accelerate

• Get the GLM4 9B model:

git lfs install
git clone https://huggingface.co/zai-org/GLM-4-9B-0414 glm4

• Create script test_cuda_glm4.py:

from transformers import AutoTokenizer, AutoModelForCausalLM, set_seed
import torch

print("GPU available:", torch.cuda.is_available())
print("GPU name:", torch.cuda.get_device_name(0))

model_path = "/home/sysadmin/llm/glm4"

seed = torch.seed() % (2**32)
print(f"Using seed: {seed}")
set_seed(seed)

tokenizer = AutoTokenizer.from_pretrained(model_path)
model     = AutoModelForCausalLM.from_pretrained(
    model_path,
    torch_dtype=torch.float16,
).to("cuda")

prompt = """People love to know how fast (or slow) \
their computers are. We have always \
been interested in speed; it is human \
nature. To help us with this quest, \
we can use various benchmark test \
programs to measure aspects of processor \
and system performance. Although no \
single numerical measurement can \
completely describe the performance \
of a complex device such as a processor \
or a complete PC, benchmarks can be \
useful tools for comparing different \
components and systems."""

inputs = tokenizer(
    prompt, 
    return_tensors="pt").to("cuda")

output_ids = model.generate(
    **inputs,
    max_new_tokens= 30000, #32768,
    temperature=0.7,
    top_p=0.9,
    do_sample=True,
    # repetition loop prevention
    repetition_penalty=1.1,
    no_repeat_ngram_size=4
)

print(tokenizer.decode(output_ids[0], skip_special_tokens=True))

• Run test

python test_cuda_glm4.py

Enjoy the result!