The best way to Run A number of LLMs Domestically Utilizing Llama-Swap on a Single Server

Picture by Creator | Ideogram

Working a number of giant language fashions will be helpful, whether or not for evaluating mannequin outputs, organising a fallback in case one fails, or customizing habits (like utilizing one mannequin for coding and one other for technical writing). That is how we frequently use LLMs in follow. There are apps like poe.com that supply this sort of setup. It’s a single platform the place you possibly can run a number of LLMs. However what if you wish to do all of it regionally, save on API prices, and preserve your knowledge personal?

Nicely, that’s the place the true downside reveals up. Setting this up normally means juggling totally different ports, operating separate processes, and switching between them manually. Not preferrred.

That’s precisely the ache Llama-Swap solves. It’s an open-source proxy server that’s tremendous light-weight (only a single binary), and it allows you to change between a number of native LLMs simply. In easy phrases, it listens for OpenAI-style API calls in your machine and routinely begins or stops the suitable mannequin server primarily based on the mannequin you request. Let’s break down the way it works and stroll by a step-by-step setup to get it operating in your native machine.

# How Llama-Swap Works

 
Conceptually, Llama-Swap sits in entrance of your LLM servers as a sensible router. When an API request arrives (e.g., a POST /v1/chat/completions name), it seems on the "mannequin" subject within the JSON payload. It then masses the suitable server course of for that mannequin, shutting down every other mannequin if wanted. For instance, should you first request mannequin "A" after which request mannequin "B", Llama-Swap will routinely cease the server for “A” and begin the server for “B” so that every request is served by the proper mannequin. This dynamic swapping occurs transparently, so purchasers see the anticipated response with out worrying concerning the underlying processes.

By default, Llama-Swap permits just one mannequin to run at a time (it unloads others when switching). Nevertheless, its Teams characteristic allows you to change this habits. A gaggle can listing a number of fashions and management their swap habits. For instance, setting swap: false in a bunch means all group members can run collectively with out unloading. In follow, you would possibly use one group for heavyweight fashions (just one energetic at a time) and one other “parallel” group for small fashions you need operating concurrently. This offers you full management over useful resource utilization and concurrency on a single server.

# Conditions

 
Earlier than getting began, guarantee your system has the next:

• Python 3 (>=3.8): Wanted for fundamental scripting and tooling.
• Homebrew (on macOS): Makes putting in LLM runtimes straightforward. For instance, you possibly can set up the llama.cpp server with:

This offers the llama-server binary for internet hosting fashions regionally.

• llama.cpp (llama-server): The OpenAI-compatible server binary (put in by way of Homebrew above, or constructed from supply) that truly runs the LLM mannequin.
• Hugging Face CLI: For downloading fashions on to your native machine with out logging into the positioning or manually navigating mannequin pages. Set up it utilizing:

pip set up -U "huggingface_hub[cli]"

• {Hardware}: Any trendy CPU will work. For quicker inference, a GPU is beneficial. (On Apple Silicon Macs, you possibly can run on the CPU or attempt PyTorch’s MPS backend for supported fashions. On Linux/Home windows with NVIDIA GPUs, you should utilize Docker/CUDA containers for acceleration.)
• Docker (Non-compulsory): To run the pre-built Docker photographs. Nevertheless, I selected to not use this for this information as a result of these photographs are designed primarily for x86 (Intel/AMD) programs and don’t work reliably on Apple Silicon (M1/M2) Macs. As an alternative, I used the bare-metal set up technique, which works straight on macOS with none container overhead.

In abstract, you’ll want a Python setting and a neighborhood LLM server (just like the `llama.cpp` server). We’ll use these to host two instance fashions on one machine.

# Step-by-Step Directions

// 1. Putting in Llama-Swap

Obtain the newest Llama-Swap launch in your OS from the GitHub releases web page. For instance, I might see v126 as the newest launch. Run the next instructions:

# Step 1: Obtain the proper file
curl -L -o llama-swap.tar.gz 
  https://github.com/mostlygeek/llama-swap/releases/obtain/v126/llama-swap_126_darwin_arm64.tar.gz

Output:
  % Complete    % Obtained % Xferd  Common Pace   Time    Time     Time  Present
                                 Dload  Add   Complete   Spent    Left  Pace
  0     0    0     0    0     0      0      0 --:--:-- --:--:-- --:--:--     0
100 3445k  100 3445k    0     0  1283k      0  0:00:02  0:00:02 --:--:-- 5417k

Now, extract the file, make it executable, and take a look at it by checking the model:

# Step 2: Extract it
tar -xzf llama-swap.tar.gz

# Step 3: Make it executable
chmod +x llama-swap

# Step 4: Check it
./llama-swap --version

Output:
model: 126 (591a9cdf4d3314fe4b3906e939a17e76402e1655), constructed at 2025-06-16T23:53:50Z

// 2. Downloading and Getting ready Two or Extra LLMs

Select two instance fashions to run. We’ll use Qwen2.5-0.5B and SmolLM2-135M (small fashions) from Hugging Face. You want the mannequin recordsdata (in GGUF or comparable format) in your machine. For instance, utilizing the Hugging Face CLI:

mkdir -p ~/llm-models

huggingface-cli obtain bartowski/SmolLM2-135M-Instruct-GGUF 
  --include "SmolLM2-135M-Instruct-Q4_K_M.gguf" --local-dir ~/llm-models

huggingface-cli obtain bartowski/Qwen2.5-0.5B-Instruct-GGUF 
  --include "Qwen2.5-0.5B-Instruct-Q4_K_M.gguf" --local-dir ~/llm-models

It will:

• Create the listing llm-models in your person’s house folder
• Obtain the GGUF mannequin recordsdata safely into that folder. After obtain, you possibly can verify it’s there:

Output:

SmolLM2-135M-Instruct-Q4_K_M.gguf
Qwen2.5-0.5B-Instruct-Q4_K_M.gguf

// 3. Making a Llama-Swap Configuration

Llama-Swap makes use of a single YAML file to outline fashions and server instructions. Create a config.yaml file with contents like this:

fashions:
  "smollm2":
    cmd: |
      llama-server
      --model /path/to/fashions/llm-models/SmolLM2-135M-Instruct-Q4_K_M.gguf
      --port ${PORT}

  "qwen2.5":
    cmd: |
      llama-server
      --model /path/to/fashions/llm-models/Qwen2.5-0.5B-Instruct-Q4_K_M.gguf
      --port ${PORT}

Substitute /path/to/fashions/ along with your precise native path. Every entry beneath fashions: provides an ID (like "qwen2.5") and a shell cmd: to run its server. We use llama-server (from llama.cpp) with --model pointing to the GGUF file and --port ${PORT}. The ${PORT} macro tells Llama-Swap to assign a free port to every mannequin routinely. The teams part is elective. I’ve omitted it for this instance, so by default, Llama-Swap will solely run one mannequin at a time. You may customise many choices per mannequin (aliases, timeouts, and so forth.) on this configuration. For extra particulars on obtainable choices, see the Full Configuration Instance File.

// 4. Working Llama-Swap

With the binary and config.yaml prepared, begin Llama-Swap pointing to your config:

./llama-swap --config config.yaml --listen 127.0.0.1:8080

This launches the proxy server on localhost:8080. It can learn config.yaml and (at first) load no fashions till the primary request arrives. Llama-Swap will now deal with API requests on port 8080, forwarding them to the suitable underlying llama-server course of primarily based on the "mannequin" parameter.

// 5. Interacting with Your Fashions

Now you can also make OpenAI-style API calls to check every mannequin. Set up jq should you don’t have it earlier than operating the instructions beneath:

// Utilizing Qwen2.5

curl -s http://localhost:8080/v1/completions 
  -H "Content material-Kind: utility/json" 
  -H "Authorization: Bearer no-key" 
  -d '{
        "mannequin": "qwen2.5",
        "immediate": "Consumer: What's Python?nAssistant:",
        "max_tokens": 100
      }' | jq '.decisions[0].textual content'

Output:
"Python is a well-liked general-purpose programming language. It's straightforward to be taught, has a big customary library, and is appropriate with many working programs. Python is used for net growth, knowledge evaluation, scientific computing, and machine studying.nPython is a language that's standard for net growth on account of its simplicity, versatility and its use of contemporary options. It's utilized in a variety of purposes together with net growth, knowledge evaluation, scientific computing, machine studying and extra. Python is a well-liked language within the"

// Utilizing SmolLM2

curl -s http://localhost:8080/v1/completions 
  -H "Content material-Kind: utility/json" 
  -H "Authorization: Bearer no-key" 
  -d '{
        "mannequin": "smollm2",
        "immediate": "Consumer: What's Python?nAssistant:",
        "max_tokens": 100
      }' | jq '.decisions[0].textual content'

Output:
"Python is a high-level programming language designed for simplicity and effectivity. It is identified for its readability, syntax, and flexibility, making it a preferred selection for newbies and builders alike.nnWhat is Python?"

Every mannequin will reply based on its coaching. The fantastic thing about Llama-Swap is you don’t should restart something manually — simply change the "mannequin" subject, and it handles the remaining. As proven within the examples above, you may see:

• qwen2.5: a extra verbose, technical response
• smollm2: a less complicated, extra concise reply

That confirms Llama-Swap is routing requests to the proper mannequin!

# Conclusion

 
Congratulations! You have arrange Llama-Swap to run two LLMs on one machine, and now you can change between them on the fly by way of API calls. We put in a proxy, ready a YAML configuration with two fashions, and noticed how Llama-Swap routes requests to the proper backend.

Subsequent steps: You may develop this to incorporate:

• Bigger fashions (like TinyLlama, Phi-2, Mistral)
• Teams for concurrent serving
• Integration with LangChain, FastAPI, or different frontends

Have enjoyable exploring totally different fashions and configurations!
 
 

Kanwal Mehreen is a machine studying engineer and a technical author with a profound ardour for knowledge science and the intersection of AI with drugs. She co-authored the e-book “Maximizing Productiveness with ChatGPT”. As a Google Technology Scholar 2022 for APAC, she champions range and tutorial excellence. She’s additionally acknowledged as a Teradata Range in Tech Scholar, Mitacs Globalink Analysis Scholar, and Harvard WeCode Scholar. Kanwal is an ardent advocate for change, having based FEMCodes to empower ladies in STEM fields.