Easy methods to Run A number of LLMs Regionally Utilizing Llama-Swap on a Single Server

Picture by Writer | Ideogram

Operating a number of massive language fashions could be helpful, whether or not for evaluating mannequin outputs, organising a fallback in case one fails, or customizing conduct (like utilizing one mannequin for coding and one other for technical writing). That is how we regularly use LLMs in follow. There are apps like poe.com that provide this sort of setup. It’s a single platform the place you may run a number of LLMs. However what if you wish to do all of it domestically, save on API prices, and hold your knowledge non-public?

Nicely, that’s the place the actual drawback reveals up. Setting this up normally means juggling totally different ports, working separate processes, and switching between them manually. Not ultimate.

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 correct mannequin server based mostly on the mannequin you request. Let’s break down the way it works and stroll by means of a step-by-step setup to get it working 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 to be on the "mannequin" discipline within the JSON payload. It then hundreds the suitable server course of for that mannequin, shutting down some other mannequin if wanted. For instance, in case 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 in regards to 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 conduct. A bunch can checklist a number of fashions and management their swap conduct. For instance, setting swap: false in a gaggle 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 working concurrently. This provides 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 primary scripting and tooling.
• Homebrew (on macOS): Makes putting in LLM runtimes straightforward. For instance, you may set up the llama.cpp server with:

This supplies the llama-server binary for internet hosting fashions domestically.

• llama.cpp (llama-server): The OpenAI-compatible server binary (put in through 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 location or manually navigating mannequin pages. Set up it utilizing:

pip set up -U "huggingface_hub[cli]"

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

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

# Step-by-Step Directions

// 1. Putting in Llama-Swap

Obtain the most recent Llama-Swap launch on your OS from the GitHub releases web page. For instance, I may see v126 as the most recent 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 check 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 related 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

This can:

• Create the listing llm-models in your person’s house folder
• Obtain the GGUF mannequin recordsdata safely into that folder. After obtain, you may 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/ together with your precise native path. Every entry below fashions: offers 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 optionally available. I’ve omitted it for this instance, so by default, Llama-Swap will solely run one mannequin at a time. You possibly can customise many choices per mannequin (aliases, timeouts, and so on.) on this configuration. For extra particulars on out there choices, see the Full Configuration Instance File.

// 4. Operating 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 is going to 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 based mostly on the "mannequin" parameter.

// 5. Interacting with Your Fashions

Now you may make OpenAI-style API calls to check every mannequin. Set up jq in case you don’t have it earlier than working the instructions under:

// Utilizing Qwen2.5

curl -s http://localhost:8080/v1/completions 
  -H "Content material-Kind: software/json" 
  -H "Authorization: Bearer no-key" 
  -d '{
        "mannequin": "qwen2.5",
        "immediate": "Person: What's Python?nAssistant:",
        "max_tokens": 100
      }' | jq '.selections[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 suitable with many working methods. Python is used for internet improvement, knowledge evaluation, scientific computing, and machine studying.nPython is a language that's widespread for internet improvement because of its simplicity, versatility and its use of recent options. It's utilized in a variety of functions together with internet improvement, 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: software/json" 
  -H "Authorization: Bearer no-key" 
  -d '{
        "mannequin": "smollm2",
        "immediate": "Person: What's Python?nAssistant:",
        "max_tokens": 100
      }' | jq '.selections[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 alternative for learners and builders alike.nnWhat is Python?"

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

• qwen2.5: a extra verbose, technical response
• smollm2: an easier, extra concise reply

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

# Conclusion

 
Congratulations! You’ve got arrange Llama-Swap to run two LLMs on one machine, and now you can change between them on the fly through 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 possibly can increase 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 medication. She co-authored the book “Maximizing Productiveness with ChatGPT”. As a Google Technology Scholar 2022 for APAC, she champions range and educational 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.