readme.md

Elixir Llama2-13b on Fly.io GPUs

Fly setup

fly apps create myapp

fly.toml:

app = "myapp"
primary_region = "iad"
vm.size = "a100-80gb"

[[mounts]]
source = "gpu_data"
destination = "/data"

[http_service]
internal_port = 8080
force_https = true
auto_stop_machines = false
auto_start_machines = true
min_machines_running = 1
processes = ["app"]

Then you're ready to deploy:

fly deploy

app.exs

Mix.install([
  {:bumblebee, "~> 0.4.2"},
  {:nx, "~> 0.6.2"},
  {:exla, "~> 0.6.1"},
  {:plug_cowboy, "~> 2.6"},
  {:jason, "~> 1.4"},
  {:phoenix, "~> 1.7"},
  {:phoenix_live_view, "~> 0.20.0"}
])

host = if app = System.get_env("FLY_APP_NAME"), do: "#{app}.fly.dev", else: "localhost"

Application.put_env(:phoenix, :json_library, Jason)

Application.put_env(:phoenix_demo, PhoenixDemo.Endpoint,
  url: [host: host],
  http: [
    ip: {0, 0, 0, 0, 0, 0, 0, 0},
    port: String.to_integer(System.get_env("PORT") || "4000"),
    transport_options: [socket_opts: [:inet6]]
  ],
  server: true,
  live_view: [signing_salt: :crypto.strong_rand_bytes(8) |> Base.encode16()],
  secret_key_base: :crypto.strong_rand_bytes(32) |> Base.encode16(),
  pubsub_server: PhoenixDemo.PubSub
)

defmodule PhoenixDemo.Layouts do
  use Phoenix.Component

  def render("live.html", assigns) do
    ~H"""
    <script src="//cdn.jsdelivr.net/npm/phoenix@1.7.6/priv/static/phoenix.min.js"></script>
    <script src="//cdn.jsdelivr.net/npm/phoenix_live_view@0.20.0/priv/static/phoenix_live_view.min.js"></script>
    <script>
      const liveSocket = new LiveView.LiveSocket("/live", Phoenix.Socket, {hooks: {}})
      liveSocket.connect()
    </script>
    <script src="https://cdn.tailwindcss.com"></script>
    <%= @inner_content %>
    """
  end
end

defmodule PhoenixDemo.ErrorView do
  def render(_, _), do: "error"
end

defmodule PhoenixDemo.SampleLive do
  use Phoenix.LiveView, layout: {PhoenixDemo.Layouts, :live}

  alias Phoenix.LiveView.AsyncResult

  def render(assigns) do
    ~H"""
    <div class="pt-10 min-h-screen w-screen flex items-center justify-center antialiased bg-gray-100 px-5">
      <div class="flex flex-col items-center w-full">
        <h1 class="text-slate-900 font-extrabold text-3xl tracking-tight text-center">
          Elixir llama2 13b on <a href="https://fly.io" class="font-mono text-sky-500">Fly.io</a> GPUs
        </h1>
        <p class="mt-6 text-lg text-slate-600 text-center max-w-4xl mx-auto">
          Powered by <a href="https://github.com/elixir-nx/bumblebee" class="font-mono font-medium text-sky-500">Bumblebee</a>,
          an Nx/Axon library for pre-trained and transformer NN models with <a href="https://huggingface.co">🤗</a> integration.
        </p>
        <div class="mt-6 w-full mx-auto">
          <form phx-change="validate" phx-submit="generate" class="space-y-5 max-w-4xl mx-auto">
            <div class="relative">
              <label for="prompt" class="block text-lg md:text-base font-medium leading-6 text-gray-900">Prompt</label>
              <div class="mt-2">
                <textarea
                  id="prompt"
                  name="prompt"
                  rows="4"
                  class="block w-full rounded-md border-0 p-2 text-gray-900 shadow-sm ring-1 ring-inset ring-gray-300 placeholder:text-gray-400 focus:ring-2 focus:ring-inset focus:ring-indigo-600 text-lg sm:text-base leading-6"
                ><%= @prompt %></textarea>
              </div>
              <button disabled={!!@output.loading} type="submit" class="absolute bottom-1 right-1 rounded-md bg-indigo-600 px-3 py-2 text-lg md:text-base font-semibold text-white shadow-sm hover:bg-indigo-500 focus-visible:outline focus-visible:outline-2 focus-visible:outline-offset-2 focus-visible:outline-indigo-600">
                Generate
                <.spinner :if={@output.loading} />
              </button>
            </div>
            <p id="output" phx-update="stream" class="text-xl sm:text-base py-5">
              <span :for={{id, segment} <- @streams.output} id={id}><%= segment.content %></span>
            </p>
          </form>
        </div>
      </div>
    </div>
    """
  end

  def mount(_params, _session, socket) do
    {:ok,
     socket
     |> stream(:output, [])
     |> assign(
       prompt: nil,
       output: AsyncResult.ok(AsyncResult.loading(), [])
     )}
  end

  def handle_event("validate", %{"prompt" => prompt}, socket) do
    {:noreply, assign(socket, prompt: prompt)}
  end

  def handle_event("generate", %{"prompt" => prompt}, socket) do
    parent = self()

    {:noreply,
     socket
     |> assign(prompt: prompt, output: AsyncResult.loading())
     |> stream(:output, [], reset: true)
     |> cancel_async(:output)
     |> start_async(:output, fn ->
       for {segment, index} <- Stream.with_index(Nx.Serving.batched_run(LlamaServing, prompt)) do
      #  for {segment, index} <- Enum.with_index(String.split("llama2 is an Elixir library that provides a simple and efficient way to perform ETL (Extract, Transform, Load) operations. It is only 200 lines of code, making it an incredibly concise and lightweight solution for data processing. Despite its small size, llama2 is surprisingly powerful and flexible, allowing users to define complex data transformations and workflows with ease. Additionally, llama2 is built on top of the Phoenix framework, which provides a robust and scalable foundation for building web applications. This means that llama2 can be seamlessly integrated into a larger Phoenix application, allowing developers to leverage the power of Elixir and Phoenix for their data processing needs. Overall, llama2 is an amazing library that is well worth checking out for anyone looking to simplify their data processing workflowsl")) do
         send(parent, {:output, {segment, index}})
       end
     end)}
  end

  def handle_info({:output, {segment, index}}, socket) do
    {:noreply, stream(socket, :output, [%{id: "s-#{index}", content: segment}])}
  end

  def handle_async(:output, {:ok, _}, socket) do
    {:noreply, assign(socket, :output, AsyncResult.ok(socket.assigns.output, []))}
  end

  defp spinner(assigns) do
    ~H"""
    <svg phx-no-format class="inline ml-1 w-4 h-4 text-gray-200 animate-spin fill-blue-600" viewBox="0 0 100 101" fill="none" xmlns="http://www.w3.org/2000/svg">
      <path d="M100 50.5908C100 78.2051 77.6142 100.591 50 100.591C22.3858 100.591 0 78.2051 0 50.5908C0 22.9766 22.3858 0.59082 50 0.59082C77.6142 0.59082 100 22.9766 100 50.5908ZM9.08144 50.5908C9.08144 73.1895 27.4013 91.5094 50 91.5094C72.5987 91.5094 90.9186 73.1895 90.9186 50.5908C90.9186 27.9921 72.5987 9.67226 50 9.67226C27.4013 9.67226 9.08144 27.9921 9.08144 50.5908Z" fill="currentColor" />
      <path d="M93.9676 39.0409C96.393 38.4038 97.8624 35.9116 97.0079 33.5539C95.2932 28.8227 92.871 24.3692 89.8167 20.348C85.8452 15.1192 80.8826 10.7238 75.2124 7.41289C69.5422 4.10194 63.2754 1.94025 56.7698 1.05124C51.7666 0.367541 46.6976 0.446843 41.7345 1.27873C39.2613 1.69328 37.813 4.19778 38.4501 6.62326C39.0873 9.04874 41.5694 10.4717 44.0505 10.1071C47.8511 9.54855 51.7191 9.52689 55.5402 10.0491C60.8642 10.7766 65.9928 12.5457 70.6331 15.2552C75.2735 17.9648 79.3347 21.5619 82.5849 25.841C84.9175 28.9121 86.7997 32.2913 88.1811 35.8758C89.083 38.2158 91.5421 39.6781 93.9676 39.0409Z" fill="currentFill" />
    </svg>
    """
  end
end

defmodule PhoenixDemo.Router do
  use Phoenix.Router
  import Phoenix.LiveView.Router

  pipeline :browser do
    plug(:accepts, ["html"])
  end

  scope "/", PhoenixDemo do
    pipe_through(:browser)

    live("/", SampleLive, :index)
  end
end

defmodule PhoenixDemo.Endpoint do
  use Phoenix.Endpoint, otp_app: :phoenix_demo

  socket("/live", Phoenix.LiveView.Socket)
  plug(PhoenixDemo.Router)
end

# Application startup

Nx.global_default_backend(EXLA.Backend)

Application.put_env(:exla, :clients,
  cuda: [platform: :cuda, preallocate: false],
  rocm: [platform: :rocm, preallocate: false],
  tpu: [platform: :tpu, preallocate: false],
  host: [platform: :host, preallocate: false]
)

# Dry run for copying cached mix install from builder to runner
if System.get_env("EXS_DRY_RUN") == "true" do
  System.halt(0)
else  
  # llama = "TheBloke/Llama-2-7b-chat-fp16"
  llama = "TheBloke/Llama-2-13B-Chat-fp16"

  {:ok, model_info} = Bumblebee.load_model({:hf, llama})
  {:ok, tokenizer} = Bumblebee.load_tokenizer({:hf, llama})
  {:ok, generation_config} = Bumblebee.load_generation_config({:hf, llama})
  generation_config = Bumblebee.configure(generation_config, max_new_tokens: 500)

  llama_serving =
    Bumblebee.Text.generation(model_info, tokenizer, generation_config,
      compile: [batch_size: 1, sequence_length: 1028],
      preallocate_params: true,
      stream: true,
      defn_options: [debug: true, client: :cuda, compiler: EXLA]
    )

  {:ok, _} =
    Supervisor.start_link(
      [
        {Phoenix.PubSub, name: PhoenixDemo.PubSub},
        {Nx.Serving, serving: llama_serving, name: LlamaServing},
        PhoenixDemo.Endpoint
      ],
      strategy: :one_for_one
    )

  Process.sleep(:infinity)
end

Dockerfile

FROM ubuntu:22.04 as builder
RUN apt update -q && apt install -y ca-certificates wget && \
    wget -qO /cuda-keyring.deb https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/cuda-keyring_1.1-1_all.deb && \
    dpkg -i /cuda-keyring.deb && apt update -q
    
ARG CUDA_VERSION=12-2

# install build dependencies
RUN apt update -y && apt-get install -y software-properties-common
RUN apt install -y git cuda-nvcc-${CUDA_VERSION} libcublas-${CUDA_VERSION} libcudnn8
RUN add-apt-repository ppa:rabbitmq/rabbitmq-erlang
RUN apt-get update -y && apt-get install -y elixir erlang-dev build-essential git curl \
    && apt-get clean && rm -f /var/lib/apt/lists/*_*

# prepare build dir
WORKDIR /app

# set runner ENV
ENV SHELL=/bin/bash
ENV ERL_AFLAGS "-proto_dist inet6_tcp"
ENV MIX_ENV="prod"
ENV PORT="8080"
ENV XLA_TARGET="cuda120"
ENV MIX_HOME="/data/mix"
ENV MIX_INSTALL_DIR="/data/mix"
ENV BUMBLEBEE_CACHE_DIR="/data/cache/bumblebee"
ENV XLA_CACHE_DIR="/data/cache/xla"

WORKDIR "/app"

COPY app.exs ./

RUN mix local.hex --force && \
    mix local.rebar --force

CMD ["elixir", "app.exs"]

thanks for including the Dockerfile for this! A++++++