Skip to main content
Deepseek-R1 is an advanced large language model designed to handle a wide range of conversational and generative tasks. It has proven capabilities in various benchmarks and excels in complex reasoning. Deepseek R1 was trained in fp8 precision and has 671B parameters. In fp8 precision each parameter is quantized to 8 bits bringing the memory required to run the DeepSeek to ~700GB. 700GB worth of GPU RAM can only be achieved on a single node of 8xH200 GPUs. Even 8xH100 GPU nodes have a combined GPU memory of 640 GB. Hence, GPU memory becomes a bottleneck for deploying the entire model. Our friends at Unsloth AI released GGUF quants of the DeepSeek Model which require only 140 GB of GPU memory. Their quants have an average quantisation of 1.58 bit (some parameters are 4-bit, some are 8 bit , some are 1 bit). This allows us to deploy our model on variety of GPU combinations that can fit on a single node. In this guide we will walk you through deploying GGUF quants of DeepSeek R1 671B on your cloud account using Tensorfuse. We will deploy the 140 GB model on a combination of 4xL40S GPUs. This combination has a GPU memory of 192 GB giving us sufficient space to run the model, and also have some headroom for KV cache and long enough context length.
You can deploy other GGUF quant models by modifying the entrypoint.sh script below. You can also tinker with the number of GPUs and GPU type to deploy on other GPU combinations. If you have more GPU memory than 192 GB, I would also recommend playing around with the --ctx-size parameter.

​
Prerequisites

Before you begin, ensure you have configured Tensorfuse on your AWS account. If you haven’t done that yet, follow the Getting Started guide.

​
Deploying Deepseek-R1-671B with Tensorfuse

Each Tensorkube deployment requires:
  1. Your environment (as a Dockerfile).
  2. Your code (in this example, the entrypoint.sh script).
  3. A deployment configuration (deployment.yaml).

​
Step 1: Prepare the Dockerfile

We will use the official llama.cpp image as our base image. This image comes with all the necessary dependencies to run llama.cpp. The image is present on the Github Registry as gerganov/llama.cpp. We will then set the environment variables required to run the model and install the necessary huggingface dependencies to download the model. We will then copy our code and set the permissions for the entrypoint script.
Dockerfile
FROM ghcr.io/ggerganov/llama.cpp:full-cuda

# Set environment variables
ENV CUDA_VISIBLE_DEVICES=0,1,2,3
ENV GGML_CUDA_MAX_STREAMS=16
ENV GGML_CUDA_MMQ_Y=1
ENV HF_HUB_ENABLE_HF_TRANSFER=1
WORKDIR /app

# Install dependencies
RUN apt-get update && \
    apt-get install -y python3-pip && \
    pip3 install huggingface_hub hf-transfer

# Copy and set permissions
COPY entrypoint.sh .
RUN chmod +x /app/entrypoint.sh

EXPOSE 8080

ENTRYPOINT ["/app/entrypoint.sh"]
We’ve configured the vLLM server with numerous CLI flags tailored to our specific use case. A comprehensive list of all other vLLM flags is available for further reference, and if you have questions about selecting flags for production, the Tensorfuse Community is an excellent place to seek guidance.

​
Step 2: Prepare the entrypoint script

In this step, we first download the GGUF model using snapshot_download from huggingface_hub. We then start the llama server with the necessary flags to run the model.
You can deploy other GGUF quant models by modifying the entrypoint.sh script below. You will have to change the repo_id and local_dir flag in the snapshot_download parameters and change the --model flag in the llama-server command.
entrypoint.sh
#!/bin/bash
set -e

# Download model shards if missing
if [ ! -d "/app/DeepSeek-R1-GGUF" ]; then
  echo "Downloading model..."
  python3 -c "
from huggingface_hub import snapshot_download
snapshot_download(
  repo_id='unsloth/DeepSeek-R1-GGUF',
  local_dir='DeepSeek-R1-GGUF',
  allow_patterns=['*UD-IQ1_S*']
)"
fi

echo "Downloading model finished. Now waiting to start the llama server with optimisations for one batch latency"

# Start server with single-request optimizations
./llama-server \
  --model DeepSeek-R1-GGUF/DeepSeek-R1-UD-IQ1_S/DeepSeek-R1-UD-IQ1_S-00001-of-00003.gguf\
  --host 0.0.0.0 \
  --port 8080 \
  --n-gpu-layers 62 \
  --parallel 4 \
  --ctx-size 5120 \
  --mlock \
  --threads 42 \
  --tensor-split 1,1,1,1 \
  --no-mmap \
  --rope-freq-base 1000000 \
  --rope-freq-scale 0.25 \
  --metrics
We’ve configured the vLLM server with numerous CLI flags tailored to our specific use case. A comprehensive list of all other llama-server flags is available for further reference, and if you have questions about selecting flags for production, the Tensorfuse Community is an excellent place to seek guidance.

​
Step 3: Deployment config

Although you can deploy tensorfuse apps using command line, it is always recommended to have a config file so that you can follow a GitOps approach to deployment.
deployment.yaml
# deployment.yaml for Deepseek R1 GGUf quants

gpus: 4
gpu_type: l40s
port: 8080
readiness:
    httpGet:
        path: /health
        port: 8080
Don’t forget the readiness endpoint in your config. Tensorfuse uses this endpoint to ensure that your service is healthy. llama-server exposes readiness by default on the /health endpoint. Remember that we have set port to 8080 in deployment.yaml as llama-server runs on that port and we have exposed 8080 in the Dockerfile.
If no readiness endpoint is configured, Tensorfuse tries the /readiness path on port 80 by default which can cause issues if your app is not listening on that path.
Now you can deploy your service using the following command: 
tensorkube deploy --config-file ./deployment.yaml

​
Step 4: Accessing the deployed app

Voila! Your autoscaling production llama.cpp service is ready. Once the deployment is successful, you can see the status of your app by running: 
tensorkube deployment list
And that’s it! You have successfully deployed the world’s strongest Open Source Reasoning Model in a quantised format.
Remember to configure a TLS endpoint with a custom domain before going to production.
To test it out, replace YOUR_APP_URL with the endpoint shown in the output of the above command and run: 
curl --request POST \
  --url YOUR_APP_URL/v1/completions \
  --header 'Content-Type: application/json' \
  --data '{
    "model": "deepseek-ai/DeepSeek-R1",
    "prompt": "Earth to Robotland. What's up?",
    "max_tokens": 200
}'
Because llama-server is compatible with the OpenAI API, you can use OpenAI’s client libraries as well. Here’s a sample snippet using Python: 
import openai

# Replace with your actual URL and token
base_url = "YOUR_APP_URL/v1"

openai.api_base = base_url

response = openai.Completion.create(
    model="deepseek-ai/DeepSeek-R1",
    prompt="Hello, Deepseek R1! How are you today?",
    max_tokens=200
)

print(response)
Click here to get started with Tensorfuse. You can also directly use the Deepseek-R1 GitHub repository for more details and updates on these Dockerfiles.