Instructions to use unsloth/GLM-5.1-GGUF with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use unsloth/GLM-5.1-GGUF with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="unsloth/GLM-5.1-GGUF")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("unsloth/GLM-5.1-GGUF", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

llama.cpp

How to use unsloth/GLM-5.1-GGUF with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf unsloth/GLM-5.1-GGUF:UD-Q4_K_M
# Run inference directly in the terminal:
llama cli -hf unsloth/GLM-5.1-GGUF:UD-Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf unsloth/GLM-5.1-GGUF:UD-Q4_K_M
# Run inference directly in the terminal:
llama cli -hf unsloth/GLM-5.1-GGUF:UD-Q4_K_M

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf unsloth/GLM-5.1-GGUF:UD-Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf unsloth/GLM-5.1-GGUF:UD-Q4_K_M

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf unsloth/GLM-5.1-GGUF:UD-Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf unsloth/GLM-5.1-GGUF:UD-Q4_K_M

Use Docker

docker model run hf.co/unsloth/GLM-5.1-GGUF:UD-Q4_K_M

LM Studio
Jan

vLLM

How to use unsloth/GLM-5.1-GGUF with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "unsloth/GLM-5.1-GGUF"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "unsloth/GLM-5.1-GGUF",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/unsloth/GLM-5.1-GGUF:UD-Q4_K_M

SGLang

How to use unsloth/GLM-5.1-GGUF with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "unsloth/GLM-5.1-GGUF" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "unsloth/GLM-5.1-GGUF",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "unsloth/GLM-5.1-GGUF" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "unsloth/GLM-5.1-GGUF",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Ollama
How to use unsloth/GLM-5.1-GGUF with Ollama:
```
ollama run hf.co/unsloth/GLM-5.1-GGUF:UD-Q4_K_M
```
Unsloth Desktop

How to use unsloth/GLM-5.1-GGUF with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf unsloth/GLM-5.1-GGUF:UD-Q4_K_M

Configure the model in Pi

# Install Pi:
npm install -g @earendil-works/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "unsloth/GLM-5.1-GGUF:UD-Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use unsloth/GLM-5.1-GGUF with Docker Model Runner:
```
docker model run hf.co/unsloth/GLM-5.1-GGUF:UD-Q4_K_M
```

Lemonade

How to use unsloth/GLM-5.1-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull unsloth/GLM-5.1-GGUF:UD-Q4_K_M

Run and chat with the model

lemonade run user.GLM-5.1-GGUF-UD-Q4_K_M

List all available models

lemonade list

Hermes Agent

How to use unsloth/GLM-5.1-GGUF with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf unsloth/GLM-5.1-GGUF:UD-Q4_K_M

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default unsloth/GLM-5.1-GGUF:UD-Q4_K_M

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use unsloth/GLM-5.1-GGUF with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf unsloth/GLM-5.1-GGUF:UD-Q4_K_M

Configure OpenClaw

# Install OpenClaw:
npm install -g openclaw@latest
# Register the local server and set it as the default model:
openclaw onboard --non-interactive --mode local \
  --auth-choice custom-api-key \
  --custom-base-url http://127.0.0.1:8080/v1 \
  --custom-model-id "unsloth/GLM-5.1-GGUF:UD-Q4_K_M" \
  --custom-provider-id llama-cpp \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

openclaw agent --local --agent main --message "Hello from Hugging Face"

[Solved in b8808] IQ4_NL Gibberish in llama.cpp

by jpsequeira - opened Apr 8

Discussion

jpsequeira

Apr 8

Hey,

This with the command:

CUDA_VISIBLE_DEVICES=0,1,2,3,4 llama-server -m ~/.cache/huggingface/hub/models--unsloth--GLM-5.1-GGUF/snapshots/3238253553497e969f3144fda297dac98b99dbbe/UD-IQ4_NL/GLM-5.1-UD-IQ4_NL-00001-of-00009.gguf --temp 1.0 --top-p 0.95  --host 0.0.0.0 --alias "GLM 5.1"

In Ik_llama with:

CUDA_VISIBLE_DEVICES=0,1,2,3,4 ik-cli -m ~/.cache/huggingface/hub/models--unsloth--GLM-5.1-GGUF/snapshots/3238253553497e969f3144fda297dac98b99dbbe/UD-IQ4_NL/GLM-5.1-UD-IQ4_NL-00001-of-00009.gguf --temp 1.0 --top-p 0.95   --merge-qkv -ctk q8_0 -ctv q4_0 -mla 3 --no-mmap --jinja -b 8192 -ub 8192 -amb 512 -mqkv -ngl 99 --host 0.0.0.0 -khad -vhad --alias "GLM 5.1"  -cnv

very coherant.

jpsequeira

Apr 8

version: 8709 (85d482e6b)
built with GNU 15.2.1 for Linux x86_64

alpine3

Apr 8

•

edited Apr 8

The Q4_K_XL quantization keeps outputting question marks on my machine.

jpsequeira

Apr 9

@alpine3 have you tried running it in ik_llama.cpp? For me it works with the command above

DrRos

Apr 9

IQ_4XS broken for me in mainline llama.cpp

shimmyshimmer

Unsloth AI org Apr 9

version: 8709 (85d482e6b)
built with GNU 15.2.1 for Linux x86_64

The Q4_K_XL quantization keeps outputting question marks on my machine.

IQ_4XS broken for me in mainline llama.cpp

What CUDA version are you guys using?

DrRos

Apr 9

•

edited Apr 9

What CUDA version are you guys using?

13.2

danielhanchen

Unsloth AI org Apr 9

•

edited Apr 9

What CUDA version are you guys using?

13.2

That's why, at the moment CUDA 13.2 doesn't work for any GGUF and it spits out gibberish no matter the uploader. It's a known issue and we told NVIDIA to fix it but it looks like they are not done yet.

https://github.com/unslothai/unsloth/issues/4849

jpsequeira

Apr 9

in the meantime it works on ik_llama. for thinking we need to enable it --chat-template-kwargs '{"enable_thinking":true}'

DrRos

Apr 9

What CUDA version are you guys using?

13.2

That's why, at the moment CUDA 13.2 doesn't work for any GGUF and it spits out GGUF no matter the uploader. It's a known issue and we told NVIDIA to fix it but it looks like they are not done yet....

https://github.com/unslothai/unsloth/issues/4849

This is actually a rather strange bug. Some requests do work fine, actually, and some doesn't.
here is some samples:
Bad one (request in Russian, about 5000 tokens) - gibberish:

Short request in Russian - perfectly fine coherent answer:

Short request in English - also works fine.

Long summarization request (about 8000 tokens) - gibberish:

Here is some startup logs:

srv          load: spawning server instance with name=GLM-5-1-IQ4-XS:default on port 37067
srv          load: spawning server instance with args:
srv          load:   /home/drros/llama.cpp/build/bin/llama-server
srv          load:   --host
srv          load:   127.0.0.1
srv          load:   --mlock
srv          load:   --no-mmap
srv          load:   --no-mmproj-offload
srv          load:   --port
srv          load:   37067
srv          load:   --temperature
srv          load:   1.0
srv          load:   --top-p
srv          load:   0.95
srv          load:   --webui-mcp-proxy
srv          load:   --alias
srv          load:   GLM-5-1-IQ4-XS:default
srv          load:   --ctx-size
srv          load:   102400
srv          load:   --cache-type-k
srv          load:   q8_0
srv          load:   --cache-type-v
srv          load:   q8_0
srv          load:   --fit-target
srv          load:   1024,1024,1024
srv          load:   --kv-unified
srv          load:   --model
srv          load:   /mnt/ds1nfs/codellamaweights/glm-5-1-iq4-xxs/GLM-5.1-UD-IQ4_XS-00001-of-00009.gguf
srv          load:   --parallel
srv          load:   4
srv          load:   --reasoning
srv          load:   on
srv          load:   --ubatch-size
srv          load:   2048
srv  log_server_r: done request: POST /models/load 192.168.0.61 200
[37067] ggml_cuda_init: found 3 CUDA devices (Total VRAM: 71963 MiB):
[37067]   Device 0: NVIDIA RTX PRO 4000 Blackwell, compute capability 12.0, VMM: yes, VRAM: 23987 MiB
[37067]   Device 1: NVIDIA RTX PRO 4000 Blackwell, compute capability 12.0, VMM: yes, VRAM: 23987 MiB
[37067]   Device 2: NVIDIA RTX PRO 4000 Blackwell, compute capability 12.0, VMM: yes, VRAM: 23987 MiB
[37067] build_info: b8720-d12cc3d1c

danielhanchen

Unsloth AI org Apr 9

•

edited Apr 9

in the meantime it works on ik_llama. for thinking we need to enable it --chat-template-kwargs '{"enable_thinking":true}'

What CUDA version are you guys using?

13.2

That's why, at the moment CUDA 13.2 doesn't work for any GGUF and it spits out GGUF no matter the uploader. It's a known issue and we told NVIDIA to fix it but it looks like they are not done yet....

https://github.com/unslothai/unsloth/issues/4849

This is actually a rather strange bug. Some requests do work fine, actually, and some doesn't.
here is some samples:
Bad one (request in Russian, about 5000 tokens) - gibberish:

Short request in Russian - perfectly fine coherent answer:

Short request in English - also works fine.

Long summarization request (about 8000 tokens) - gibberish:

Here is some startup logs:

srv          load: spawning server instance with name=GLM-5-1-IQ4-XS:default on port 37067
srv          load: spawning server instance with args:
srv          load:   /home/drros/llama.cpp/build/bin/llama-server
srv          load:   --host
srv          load:   127.0.0.1
srv          load:   --mlock
srv          load:   --no-mmap
srv          load:   --no-mmproj-offload
srv          load:   --port
srv          load:   37067
srv          load:   --temperature
srv          load:   1.0
srv          load:   --top-p
srv          load:   0.95
srv          load:   --webui-mcp-proxy
srv          load:   --alias
srv          load:   GLM-5-1-IQ4-XS:default
srv          load:   --ctx-size
srv          load:   102400
srv          load:   --cache-type-k
srv          load:   q8_0
srv          load:   --cache-type-v
srv          load:   q8_0
srv          load:   --fit-target
srv          load:   1024,1024,1024
srv          load:   --kv-unified
srv          load:   --model
srv          load:   /mnt/ds1nfs/codellamaweights/glm-5-1-iq4-xxs/GLM-5.1-UD-IQ4_XS-00001-of-00009.gguf
srv          load:   --parallel
srv          load:   4
srv          load:   --reasoning
srv          load:   on
srv          load:   --ubatch-size
srv          load:   2048
srv  log_server_r: done request: POST /models/load 192.168.0.61 200
[37067] ggml_cuda_init: found 3 CUDA devices (Total VRAM: 71963 MiB):
[37067]   Device 0: NVIDIA RTX PRO 4000 Blackwell, compute capability 12.0, VMM: yes, VRAM: 23987 MiB
[37067]   Device 1: NVIDIA RTX PRO 4000 Blackwell, compute capability 12.0, VMM: yes, VRAM: 23987 MiB
[37067]   Device 2: NVIDIA RTX PRO 4000 Blackwell, compute capability 12.0, VMM: yes, VRAM: 23987 MiB
[37067] build_info: b8720-d12cc3d1c

Using any version lower than CUDA 13.2 will solve your issue. Unsloth Studio for example uses 13.0 so there is no issue with generation

I think we should make a PSA and let everyone know not to use CUDA 13.2 because it seems like it's more problematic than originally anticipated

jpsequeira

Apr 9

How come ik_llama works but llama.cpp doesn't? If the problem lies in CUDA 13.2...

danielhanchen

Unsloth AI org Apr 9

How come ik_llama works but llama.cpp doesn't? If the problem lies in CUDA 13.2...

If ik_llama works then continue using that, but still would not recommend using CUDA 13.2. The issue has been confirmed by us and 30+ users.

WohthaN

Apr 14

•

edited Apr 14

I have a similar problem. llama.cpp-b8733.tar.gz works well; following versions (didn't pinpoint the exact breaking change version yet) give gibberish output most of the times.
This might give some insight on why ik_llama works thogh?

I see this problem with CUDA 13.0 (Toolkit version 13.0.2, nvcc V13.0.88) Driver Version: 595.58.03

WohthaN

Apr 16

b8808 seems to work fine now! I believe there was a problem with FA in some previous release that got fixed.

danielhanchen

Unsloth AI org Apr 16

b8808 seems to work fine now! I believe there was a problem with FA in some previous release that got fixed.

Awesome thanks for sharing! @jpsequeira could you edit your post if possible incase other people also encoutner the problem? Thanks so much! :)

jpsequeira changed discussion title from IQ4_NL Gibberish in llama.cpp to [Solved in b8808] IQ4_NL Gibberish in llama.cpp Apr 16

jpsequeira

Apr 16

@danielhanchen perhaps updating the do not use cuda post would be a good as well

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment