Distributed Llama

Tensor parallelism is all you need. Run LLMs on weak devices or make powerful devices even more powerful by distributing the workload and dividing the RAM usage. This project proves that it's possible split the workload of LLMs across multiple devices and achieve a significant speedup. Distributed Llama allows you to run huge LLMs in-house. The project uses TCP sockets to synchronize the state. You can easily configure your AI cluster by using a home router.

_{^{Distributed Llama running Llama 2 70B on 8 Raspberry Pi 4B devices}}

🔥 Setup Root Node by Single Command

Python 3 and C++ compiler required. The command will download the model and the tokenizer.

Model	Purpose	Size	Command
TinyLlama 1.1B 3T Q40	Benchmark	844 MB	`python launch.py tinyllama_1_1b_3t_q40`
Llama 3 8B Q40	Benchmark	6.32 GB	`python launch.py llama3_8b_q40`
Llama 3 8B Instruct Q40	Chat, API	6.32 GB	`python launch.py llama3_8b_instruct_q40`
Llama 3.1 8B Instruct Q40	Chat, API	6.32 GB	`python launch.py llama3_1_8b_instruct_q40`

🛠️ Convert Model Manually

Supported architectures: Llama, Mixtral, Grok

🚧 Known Limitations

You can run Distributed Llama only on 1, 2, 4... 2^n nodes.
The maximum number of nodes is equal to the number of KV heads in the model #70.
CPU support only, GPU support is planned, optimized for (weights format × buffer format):
- ARM CPUs
  - ✅ F32 × F32
  - ❌ F16 × F32
  - ✅ Q40 × F32
  - ✅ Q40 × Q80
- x86_64 AVX2 CPUs
  - ✅ F32 × F32
  - ❌ F16 × F32
  - ✅ Q40 × F32
  - ✅ Q40 × Q80

👷 Architecture

The project is split up into two parts:

Root node - it's responsible for loading the model and weights and forward them to workers. Also, it synchronizes the state of the neural network. The root node is also a worker, it processes own slice of the neural network.
Worker node - it processes own slice of the neural network. It doesn't require any configuration related to the model.

You always need the root node and you can add 2^n - 1 worker nodes to speed up the inference. The RAM usage of the neural network is split up across all nodes. The root node requires a bit more RAM than worker nodes.

🎹 Commands

dllama inference - run the inference with a simple benchmark,
dllama chat - run the CLI chat,
dllama worker - run the worker node,
dllama-api - run the API server.

Inference, Chat, API

Argument	Description	Example
`--model <path>`	Path to model.	`dllama_model_meta-llama-3-8b_q40.m`
`--tokenizer <path>`	Tokenizer to model.	`dllama_tokenizer_llama3.t`
`--buffer-float-type <type>`	Float precision of synchronization.	`q80`
`--workers <workers>`	Addresses of workers (ip:port), separated by space.	`10.0.0.1:9991 10.0.0.2:9991`

Inference, Chat, Worker, API

Argument	Description	Example
`--nthreads <n>`	Amount of threads. Don't set a higher value than number of CPU cores.	`4`

Worker, API

Argument	Description	Example
`--port <port>`	Binding port.	`9999`

Inference

Argument	Description	Example
`--prompt <prompt>`	Initial prompt.	`"Hello World"`
`--steps <steps>`	Number of tokens to generate.	`256`

📊 Measurements

Average Token Generation Time

I - inference time of the root node, T - network transfer time of the root node.

Raspberry Pi 5 8GB

_{^{Weights = Q40, Buffer = Q80, nSamples = 16, switch = TP-Link LS1008G, tested on 0.3.1 version}}

Model	1 x RasPi 5 8 GB	2 x RasPi 5 8 GB	4 x RasPi 5 8 GB
Llama 2 7B	441.09 ms, 2.26 t/s _{^{I: 434.84 ms, T: 5.25 ms}}	341.46 ms, 2.92 t/s _{^{I: 257.78 ms, T: 83.27 ms}}	219.08 ms, 4.56 t/s 🔥 _{^{I: 163.42 ms, T: 55.25 ms}}
Llama 3 8B	564.31 ms, 1.77 t/s _{^{I: 556.67 ms, T: 6.17 ms}}	444.27 ms, 2.25 t/s _{^{I: 362.73 ms, T: 80.11 ms}}	331.47 ms, 3.01 t/s 🔥 _{^{I: 267.62 ms, T: 62.34 ms}}

Raspberry Pi 4B 8 GB

_{^{Weights = Q40, Buffer = Q80, nSamples = 16, switch = TP-Link LS1008G, tested on 0.1.0 version}}

_{^{8 x Raspberry Pi 4B 8GB}}

Model	1 x RasPi 4B 8 GB	2 x RasPi 4B 8 GB	4 x RasPi 4B 8 GB	8 x RasPi 4B 8 GB
Llama 2 7B	1312.50 ms _{^{I: 1307.94 ms, T: 1.81 ms}}	793.69 ms _{^{I: 739.00 ms, T: 52.50 ms}}	494.00 ms 🔥 _{^{I: 458.81 ms, T: 34.06 ms}}	588.19 ms _{^{I: 296.69 ms, T: 289.75 ms}}
Llama 2 13B	_{^{Not enough RAM}}	1497.19 ms _{^{I: 1465.06 ms, T: 30.88 ms}}	848.19 ms 🔥 _{^{I: 746.88 ms, T: 99.50 ms}}	1114.88 ms _{^{I: 460.8 ms, T: 652.88 ms}}
Llama 2 70B	_{^{Not enough RAM}}	_{^{Not enough RAM}}	_{^{Not enough RAM}}	4842.81 ms 🔥 _{^{I: 2121.94 ms, T: 2719.62 ms}}

x86_64 CPU Cloud Server

_{^{Weights = Q40, Buffer = Q80, nSamples = 16, VMs = c3d-highcpu-30, tested on 0.1.0 version}}

Model	1 x VM	2 x VM	4 x VM
Llama 2 7B	101.81 ms _{^{I: 101.06 ms, T: 0.19 ms}}	69.69 ms _{^{I: 61.50 ms, T: 7.62 ms}}	53.69 ms 🔥 _{^{I: 40.25 ms, T: 12.81 ms}}
Llama 2 13B	184.19 ms _{^{I: 182.88 ms, T: 0.69 ms}}	115.38 ms _{^{I: 107.12 ms, T: 7.81 ms}}	86.81 ms 🔥 _{^{I: 66.25 ms, T: 19.94 ms}}
Llama 2 70B	909.69 ms _{^{I: 907.25 ms, T: 1.75 ms}}	501.38 ms _{^{I: 475.50 ms, T: 25.00 ms}}	293.06 ms 🔥 _{^{I: 264.00 ms, T: 28.50 ms}}

Network Transfer for Generating Token

F32 Buffer

Model	2 devices	4 devices	8 devices
Llama 3 8B	2048 kB	6144 kB	14336 kB

Q80 Buffer

Model	2 devices	4 devices	8 devices
Llama 3 8B	544 kB	1632 kB	3808 kB

📟 Setup Raspberry Pi Devices

Install Raspberry Pi OS Lite (64 bit) on your Raspberry Pi devices. This OS doesn't have desktop environment.
Connect all devices to your switch or router.
Connect to all devices via SSH.

ssh user@raspberrypi1.local
ssh user@raspberrypi2.local

Install Git:

sudo apt install git

Clone this repository and compile Distributed Llama on all devices:

git clone https://github.com/b4rtaz/distributed-llama.git
make dllama
make dllama-api

Transfer weights and the tokenizer file to the root device.
Optional: assign static IP addresses.

sudo ip addr add 10.0.0.1/24 dev eth0 # 1th device
sudo ip addr add 10.0.0.2/24 dev eth0 # 2th device

Run worker nodes on worker devices:

sudo nice -n -20 ./dllama worker --port 9998 --nthreads 4

Run root node on the root device:

sudo nice -n -20 ./dllama inference --model dllama_model_meta-llama-3-8b_q40.m --tokenizer dllama_tokenizer_llama3.t --buffer-float-type q80 --prompt "Hello world" --steps 16 --nthreads 4 --workers 10.0.0.2:9998

To add more worker nodes, just add more addresses to the --workers argument.

./dllama inference ... --workers 10.0.0.2:9998 10.0.0.3:9998 10.0.0.4:9998

💻 Setup computers with MacOS, Linux, or Windows

You need x86_64 AVX2 CPUs or ARM CPUs. Different devices may have different CPUs.

MacOS or Linux

The below instructions are for Debian-based distributions but you can easily adapt them to your distribution, macOS.

Install Git and GCC:

sudo apt install git build-essential

Clone this repository and compile Distributed Llama on all computers:

git clone https://github.com/b4rtaz/distributed-llama.git
make dllama
make dllama-api

Continue to point 3.

Windows

Install Git and Mingw (via Chocolatey):

choco install mingw

Clone this repository and compile Distributed Llama on all computers:

git clone https://github.com/b4rtaz/distributed-llama.git
make dllama
make dllama-api