ci: Fix tag detection for 'latest'

Drop armv7 builds

.dockerignore

@@ -0,0 +1 @@
+models/*.bin

.env

@@ -0,0 +1 @@
+THREADS=14

.github/workflows/image.yml

@@ -2,6 +2,7 @@
 name: 'build container images'
 
 on:
+  pull_request:
   push:
     branches:
       - master
@@ -12,68 +13,42 @@ jobs:
   docker:
     runs-on: ubuntu-latest
     steps:
-      - name: Release space from worker
-        run: |
-          echo "Listing top largest packages"
-          pkgs=$(dpkg-query -Wf '${Installed-Size}\t${Package}\t${Status}\n' | awk '$NF == "installed"{print $1 "\t" $2}' | sort -nr)
-          head -n 30 <<< "${pkgs}"
-          echo
-          df -h
-          echo
-          sudo apt-get remove -y '^llvm-.*|^libllvm.*' || true
-          sudo apt-get remove --auto-remove android-sdk-platform-tools || true
-          sudo apt-get purge --auto-remove android-sdk-platform-tools || true
-          sudo rm -rf /usr/local/lib/android
-          sudo apt-get remove -y '^dotnet-.*|^aspnetcore-.*' || true
-          sudo rm -rf /usr/share/dotnet
-          sudo apt-get remove -y '^mono-.*' || true
-          sudo apt-get remove -y '^ghc-.*' || true
-          sudo apt-get remove -y '.*jdk.*|.*jre.*' || true
-          sudo apt-get remove -y 'php.*' || true
-          sudo apt-get remove -y hhvm powershell firefox monodoc-manual msbuild || true
-          sudo apt-get remove -y '^google-.*' || true
-          sudo apt-get remove -y azure-cli || true
-          sudo apt-get remove -y '^mongo.*-.*|^postgresql-.*|^mysql-.*|^mssql-.*' || true
-          sudo apt-get remove -y '^gfortran-.*' || true
-          sudo apt-get autoremove -y
-          sudo apt-get clean
-          echo
-          echo "Listing top largest packages"
-          pkgs=$(dpkg-query -Wf '${Installed-Size}\t${Package}\t${Status}\n' | awk '$NF == "installed"{print $1 "\t" $2}' | sort -nr)
-          head -n 30 <<< "${pkgs}"
-          echo
-          sudo rm -rfv build || true
-          df -h
       - name: Checkout
         uses: actions/checkout@v3
+
       - name: Prepare
         id: prep
         run: |
           DOCKER_IMAGE=quay.io/go-skynet/llama-cli
-          VERSION=latest
+          VERSION=master
           SHORTREF=${GITHUB_SHA::8}
+
           # If this is git tag, use the tag name as a docker tag
           if [[ $GITHUB_REF == refs/tags/* ]]; then
             VERSION=${GITHUB_REF#refs/tags/}
           fi
           TAGS="${DOCKER_IMAGE}:${VERSION},${DOCKER_IMAGE}:${SHORTREF}"
+
           # If the VERSION looks like a version number, assume that
           # this is the most recent version of the image and also
           # tag it 'latest'.
-          if [[ $VERSION =~ ^[0-9]{1,3}\.[0-9]{1,3}\.[0-9]{1,3}$ ]]; then
+          if [[ $VERSION =~ ^v[0-9]{1,3}\.[0-9]{1,3}\.[0-9]{1,3}$ ]]; then
             TAGS="$TAGS,${DOCKER_IMAGE}:latest"
           fi
+
           # Set output parameters.
           echo ::set-output name=tags::${TAGS}
           echo ::set-output name=docker_image::${DOCKER_IMAGE}
-          echo ::set-output name=image::${DOCKER_IMAGE}:${VERSION}
+
       - name: Set up QEMU
         uses: docker/setup-qemu-action@master
         with:
           platforms: all
+
       - name: Set up Docker Buildx
         id: buildx
         uses: docker/setup-buildx-action@master
+
       - name: Login to DockerHub
         if: github.event_name != 'pull_request'
         uses: docker/login-action@v2
@@ -81,9 +56,23 @@ jobs:
           registry: quay.io
           username: ${{ secrets.QUAY_USERNAME }}
           password: ${{ secrets.QUAY_PASSWORD }}
-      - uses: earthly/actions/setup-earthly@v1
       - name: Build
-        run: |
-            earthly config "global.conversion_parallelism" "1"
-            earthly config "global.buildkit_max_parallelism" "1"
-            earthly --push +image-all --IMAGE=${{ steps.prep.outputs.image }}
+        if: github.event_name != 'pull_request'
+        uses: docker/build-push-action@v4
+        with:
+          builder: ${{ steps.buildx.outputs.name }}
+          context: .
+          file: ./Dockerfile
+          platforms: linux/amd64,linux/arm64
+          push: true
+          tags: ${{ steps.prep.outputs.tags }}
+      - name: Build PRs
+        if: github.event_name == 'pull_request'
+        uses: docker/build-push-action@v4
+        with:
+          builder: ${{ steps.buildx.outputs.name }}
+          context: .
+          file: ./Dockerfile
+          platforms: linux/amd64
+          push: false
+          tags: ${{ steps.prep.outputs.tags }}

.gitignore

@@ -0,0 +1,2 @@
+llama-cli
+models/*.bin

Dockerfile

@@ -0,0 +1,19 @@
+ARG GO_VERSION=1.20
+ARG DEBIAN_VERSION=11
+
+FROM golang:$GO_VERSION as builder
+
+WORKDIR /build
+RUN git clone --recurse-submodules https://github.com/go-skynet/go-llama.cpp
+RUN cd go-llama.cpp && make libbinding.a
+COPY go.mod ./
+COPY go.sum ./
+RUN go mod download
+RUN apt-get update
+COPY . .
+RUN go mod edit -replace github.com/go-skynet/go-llama.cpp=/build/go-llama.cpp
+RUN C_INCLUDE_PATH=/build/go-llama.cpp LIBRARY_PATH=/build/go-llama.cpp go build -o llama-cli ./
+
+FROM debian:$DEBIAN_VERSION
+COPY --from=builder /build/llama-cli /usr/bin/llama-cli
+ENTRYPOINT [ "/usr/bin/llama-cli" ]

Earthfile

@@ -1,31 +1,5 @@
 VERSION 0.7
 
-go-deps:
-    ARG GO_VERSION=1.20
-    FROM golang:$GO_VERSION
-    WORKDIR /build
-    COPY go.mod ./
-    COPY go.sum ./
-    RUN go mod download
-    RUN apt-get update
-    SAVE ARTIFACT go.mod AS LOCAL go.mod
-    SAVE ARTIFACT go.sum AS LOCAL go.sum
-
 build:
-    FROM +go-deps
-    WORKDIR /build
-    RUN git clone https://github.com/go-skynet/llama
-    RUN cd llama && make libllama.a
-    COPY . .
-    RUN C_INCLUDE_PATH=/build/llama LIBRARY_PATH=/build/llama go build -o llama-cli ./
-    SAVE ARTIFACT llama-cli AS LOCAL llama-cli
-
-image:
-    FROM +go-deps
-    ARG IMAGE=alpaca-cli-nomodel
-    COPY +build/llama-cli /llama-cli
-    ENTRYPOINT [ "/llama-cli" ]
-    SAVE IMAGE --push $IMAGE
-
-image-all:
-    BUILD --platform=linux/amd64 --platform=linux/arm64 +image
+    FROM DOCKERFILE -f Dockerfile .
+    SAVE ARTIFACT /usr/bin/llama-cli AS LOCAL llama-cli

README.md

@@ -1,16 +1,70 @@
 ## :camel: llama-cli
 
 
-llama-cli is a straightforward golang CLI interface for [llama.cpp](https://github.com/ggerganov/llama.cpp), providing a simple API and a command line interface that allows text generation using a GPT-based model like llama directly from the terminal. It is also compatible with [gpt4all](https://github.com/nomic-ai/gpt4all) and [alpaca](https://github.com/tatsu-lab/stanford_alpaca).
+llama-cli is a straightforward golang CLI interface and API compatible with OpenAI for [llama.cpp](https://github.com/ggerganov/llama.cpp), it supports multiple-models and also provides a simple command line interface that allows text generation using a GPT-based model like llama directly from the terminal. 
 
-`llama-cli` uses https://github.com/go-skynet/llama, which is a fork of [llama.cpp](https://github.com/ggerganov/llama.cpp) providing golang binding.
+It is compatible with the models supported by `llama.cpp`. You might need to convert older models to the new format, see [here](https://github.com/ggerganov/llama.cpp#using-gpt4all) for instance to run `gpt4all`.
+
+`llama-cli` doesn't shell-out, it uses https://github.com/go-skynet/go-llama.cpp, which is a golang binding of [llama.cpp](https://github.com/ggerganov/llama.cpp).
+
+## Usage
+
+You can use `docker-compose`:
+
+```bash
+
+git clone https://github.com/go-skynet/llama-cli
+cd llama-cli
+
+# copy your models to models/
+cp your-model.bin models/
+
+# (optional) Edit the .env file to set the number of concurrent threads used for inference
+# echo "THREADS=14" > .env
+
+# start with docker-compose
+docker compose up -d --build
+
+# Now API is accessible at localhost:8080
+curl http://localhost:8080/v1/models
+# {"object":"list","data":[{"id":"your-model.bin","object":"model"}]}
+curl http://localhost:8080/v1/completions -H "Content-Type: application/json" -d '{
+     "model": "your-model.bin",            
+     "prompt": "A long time ago in a galaxy far, far away",
+     "temperature": 0.7
+   }'
+
+
+```
+
+Note: You can use a use a default template for every model in your model path, by creating a corresponding file with the `.tmpl` suffix next to your model. For instance, if the model is called `foo.bin`, you can create a sibiling file, `foo.bin.tmpl` which will be used as a default prompt, for instance this can be used with alpaca:
+
+```
+Below is an instruction that describes a task. Write a response that appropriately completes the request.
+
+### Instruction:
+{{.Input}}
+
+### Response:
+```
 
 ## Container images
 
+`llama-cli` comes by default as a container image. You can check out all the available images with corresponding tags [here](https://quay.io/repository/go-skynet/llama-cli?tab=tags&tag=latest)
+
 To begin, run:
 
 ```
-docker run -ti --rm quay.io/go-skynet/llama-cli:v0.4  --instruction "What's an alpaca?" --topk 10000 --model ...
+docker run -ti --rm quay.io/go-skynet/llama-cli:latest  --instruction "What's an alpaca?" --topk 10000 --model ...
+```
+
+Where `--model` is the path of the model you want to use. 
+
+Note: you need to mount a volume to the docker container in order to load a model, for instance:
+
+```
+# assuming your model is in /path/to/your/models/foo.bin
+docker run -v /path/to/your/models:/models -ti --rm quay.io/go-skynet/llama-cli:latest  --instruction "What's an alpaca?" --topk 10000 --model /models/foo.bin
 ```
 
 You will receive a response like the following:
@@ -39,8 +93,6 @@ llama-cli --model <model_path> --instruction <instruction> [--input <input>] [--
 | top_p        | TOP_P                | 0.85          | The cumulative probability for top-p sampling. |
 | top_k        | TOP_K                | 20            | The number of top-k tokens to consider for text generation.  |
 | context-size | CONTEXT_SIZE         | 512           | Default token context size. |
-| alpaca       | ALPACA               | true          | Set to true for alpaca models. |
-| gpt4all       | GPT4ALL               | false          | Set to true for gpt4all models. |
 
 Here's an example of using `llama-cli`:
 
@@ -50,14 +102,14 @@ llama-cli --model ~/ggml-alpaca-7b-q4.bin --instruction "What's an alpaca?"
 
 This will generate text based on the given model and instruction.
 
-## Advanced usage
+## API
 
-`llama-cli` also provides an API for running text generation as a service. The model will be pre-loaded and kept in memory.
+`llama-cli` also provides an API for running text generation as a service. The models once loaded the first time will be kept in memory.
 
 Example of starting the API with `docker`:
 
 ```bash
-docker run -p 8080:8080 -ti --rm quay.io/go-skynet/llama-cli:v0.4 api --context-size 700 --threads 4
+docker run -p 8080:8080 -ti --rm quay.io/go-skynet/llama-cli:latest api --models-path /path/to/models --context-size 700 --threads 4
 ```
 
 And you'll see:
@@ -72,36 +124,68 @@ And you'll see:
 └───────────────────────────────────────────────────┘ 
 ```
 
+Note: Models have to end up with `.bin`.
+
 You can control the API server options with command line arguments:
 
 ```
-llama-cli api --model <model_path> [--address <address>] [--threads <num_threads>]
+llama-cli api --models-path <model_path> [--address <address>] [--threads <num_threads>]
 ```
 
 The API takes takes the following:
 
 | Parameter    | Environment Variable | Default Value | Description                            |
 | ------------ | -------------------- | ------------- | -------------------------------------- |
-| model        | MODEL_PATH           |               | The path to the pre-trained GPT-based model.      |
+| models-path        | MODELS_PATH           |               | The path where you have models (ending with `.bin`).      |
 | threads      | THREADS              | CPU cores     | The number of threads to use for text generation. |
 | address      | ADDRESS              | :8080         | The address and port to listen on. |
 | context-size | CONTEXT_SIZE         | 512           | Default token context size. |
-| alpaca       | ALPACA               | true          | Set to true for alpaca models. |
-| gpt4all       | GPT4ALL               | false          | Set to true for gpt4all models. |
 
+Once the server is running, you can start making requests to it using HTTP, using the OpenAI API. 
 
-Once the server is running, you can start making requests to it using HTTP. For example, to generate text based on an instruction, you can send a POST request to the `/predict` endpoint with the instruction as the request body:
+### Supported OpenAI API endpoints
+
+You can check out the [OpenAI API reference](https://platform.openai.com/docs/api-reference/chat/create). 
+
+Following the list of endpoints/parameters supported.
+
+#### Chat completions
+
+For example, to generate a chat completion, you can send a POST request to the `/v1/chat/completions` endpoint with the instruction as the request body:
 
 ```
-curl --location --request POST 'http://localhost:8080/predict' --header 'Content-Type: application/json' --data-raw '{
-    "text": "What is an alpaca?",
-    "topP": 0.8,
-    "topK": 50,
-    "temperature": 0.7,
-    "tokens": 100
-}'
+curl http://localhost:8080/v1/chat/completions -H "Content-Type: application/json" -d '{
+     "model": "ggml-koala-7b-model-q4_0-r2.bin",
+     "messages": [{"role": "user", "content": "Say this is a test!"}],
+     "temperature": 0.7
+   }'
 ```
 
+Available additional parameters: `top_p`, `top_k`, `max_tokens`
+
+#### Completions
+
+For example, to generate a comletion, you can send a POST request to the `/v1/completions` endpoint with the instruction as the request body:
+```
+curl http://localhost:8080/v1/completions -H "Content-Type: application/json" -d '{
+     "model": "ggml-koala-7b-model-q4_0-r2.bin",
+     "prompt": "A long time ago in a galaxy far, far away",
+     "temperature": 0.7
+   }'
+```
+
+Available additional parameters: `top_p`, `top_k`, `max_tokens`
+
+#### List models
+
+You can list all the models available with:
+
+```
+curl http://localhost:8080/v1/models
+```
+
+## Web interface
+
 There is also available a simple web interface (for instance, http://localhost:8080/) which can be used as a playground.
 
 Note: The API doesn't inject a template for talking to the instance, while the CLI does. You have to use a prompt similar to what's described in the standford-alpaca docs: https://github.com/tatsu-lab/stanford_alpaca#data-release, for instance:
@@ -115,18 +199,9 @@ Below is an instruction that describes a task. Write a response that appropriate
 ### Response:
 ```
 
+
 ## Using other models
 
-You can specify a model binary to be used for inference with `--model`.
-
-13B and 30B alpaca models are known to work:
-
-```
-# Download the model image, extract the model
-# Use the model with llama-cli
-docker run -v $PWD:/models -p 8080:8080 -ti --rm quay.io/go-skynet/llama-cli:v0.4 api --model /models/model.bin
-```
-
 gpt4all (https://github.com/nomic-ai/gpt4all) works as well, however the original model needs to be converted (same applies for old alpaca models, too):
 
 ```bash
@@ -154,7 +229,7 @@ import (
 
 func main() {
 
-	cli := client.NewClient("http://ip:30007")
+	cli := client.NewClient("http://ip:port")
 
 	out, err := cli.Predict("What's an alpaca?")
 	if err != nil {
@@ -185,9 +260,8 @@ In order to build the `llama-cli` container image locally you can use `docker`:
 
 ```
 # build the image as "alpaca-image"
-docker run --privileged -v /var/run/docker.sock:/var/run/docker.sock --rm -t -v "$(pwd)":/workspace -v earthly-tmp:/tmp/earthly:rw earthly/earthly:v0.7.2 +image --IMAGE=alpaca-image
-# run the image
-docker run alpaca-image --instruction "What's an alpaca?"
+docker build -t llama-cli .
+docker run llama-cli --instruction "What's an alpaca?"
 ```
 
 Or build the binary with:
@@ -201,11 +275,11 @@ docker run --privileged -v /var/run/docker.sock:/var/run/docker.sock --rm -t -v
 
 ## Short-term roadmap
 
-- Mimic OpenAI API (https://github.com/go-skynet/llama-cli/issues/10)
+- [x] Mimic OpenAI API (https://github.com/go-skynet/llama-cli/issues/10)
 - Binary releases (https://github.com/go-skynet/llama-cli/issues/6)
 - Upstream our golang bindings to llama.cpp (https://github.com/ggerganov/llama.cpp/issues/351)
-- Multi-model support
-- Full Deployment and compatibility with https://github.com/mckaywrigley/chatbot-ui
+- [x] Multi-model support
+- Have a webUI!
 
 ## License
 

api.go

@@ -1,91 +0,0 @@
-package main
-
-import (
-	"embed"
-	"net/http"
-	"strconv"
-	"sync"
-
-	llama "github.com/go-skynet/llama/go"
-	"github.com/gofiber/fiber/v2"
-	"github.com/gofiber/fiber/v2/middleware/filesystem"
-)
-
-//go:embed index.html
-var indexHTML embed.FS
-
-func api(l *llama.LLama, listenAddr string, threads int) error {
-	app := fiber.New()
-	app.Use("/", filesystem.New(filesystem.Config{
-		Root:         http.FS(indexHTML),
-		NotFoundFile: "index.html",
-	}))
-	/*
-		curl --location --request POST 'http://localhost:8080/predict' --header 'Content-Type: application/json' --data-raw '{
-		    "text": "What is an alpaca?",
-		    "topP": 0.8,
-		    "topK": 50,
-		    "temperature": 0.7,
-		    "tokens": 100
-		}'
-	*/
-	var mutex = &sync.Mutex{}
-
-	// Endpoint to generate the prediction
-	app.Post("/predict", func(c *fiber.Ctx) error {
-		mutex.Lock()
-		defer mutex.Unlock()
-		// Get input data from the request body
-		input := new(struct {
-			Text string `json:"text"`
-		})
-		if err := c.BodyParser(input); err != nil {
-			return err
-		}
-
-		// Set the parameters for the language model prediction
-		topP, err := strconv.ParseFloat(c.Query("topP", "0.9"), 64) // Default value of topP is 0.9
-		if err != nil {
-			return err
-		}
-
-		topK, err := strconv.Atoi(c.Query("topK", "40")) // Default value of topK is 40
-		if err != nil {
-			return err
-		}
-
-		temperature, err := strconv.ParseFloat(c.Query("temperature", "0.5"), 64) // Default value of temperature is 0.5
-		if err != nil {
-			return err
-		}
-
-		tokens, err := strconv.Atoi(c.Query("tokens", "128")) // Default value of tokens is 128
-		if err != nil {
-			return err
-		}
-
-		// Generate the prediction using the language model
-		prediction, err := l.Predict(
-			input.Text,
-			llama.SetTemperature(temperature),
-			llama.SetTopP(topP),
-			llama.SetTopK(topK),
-			llama.SetTokens(tokens),
-			llama.SetThreads(threads),
-		)
-		if err != nil {
-			return err
-		}
-
-		// Return the prediction in the response body
-		return c.JSON(struct {
-			Prediction string `json:"prediction"`
-		}{
-			Prediction: prediction,
-		})
-	})
-
-	// Start the server
-	app.Listen(":8080")
-	return nil
-}

api/api.go

@@ -0,0 +1,275 @@
+package api
+
+import (
+	"embed"
+	"fmt"
+	"net/http"
+	"strconv"
+	"strings"
+	"sync"
+
+	model "github.com/go-skynet/llama-cli/pkg/model"
+
+	llama "github.com/go-skynet/go-llama.cpp"
+	"github.com/gofiber/fiber/v2"
+	"github.com/gofiber/fiber/v2/middleware/cors"
+	"github.com/gofiber/fiber/v2/middleware/filesystem"
+	"github.com/gofiber/fiber/v2/middleware/recover"
+)
+
+type OpenAIResponse struct {
+	Created int      `json:"created,omitempty"`
+	Object  string   `json:"chat.completion,omitempty"`
+	ID      string   `json:"id,omitempty"`
+	Model   string   `json:"model,omitempty"`
+	Choices []Choice `json:"choices,omitempty"`
+}
+
+type Choice struct {
+	Index        int     `json:"index,omitempty"`
+	FinishReason string  `json:"finish_reason,omitempty"`
+	Message      Message `json:"message,omitempty"`
+	Text         string  `json:"text,omitempty"`
+}
+
+type Message struct {
+	Role    string `json:"role,omitempty"`
+	Content string `json:"content,omitempty"`
+}
+
+type OpenAIModel struct {
+	ID     string `json:"id"`
+	Object string `json:"object"`
+}
+
+type OpenAIRequest struct {
+	Model string `json:"model"`
+
+	// Prompt is read only by completion API calls
+	Prompt string `json:"prompt"`
+	// Messages is readh only by chat/completion API calls
+	Messages []Message `json:"messages"`
+
+	// Common options between all the API calls
+	TopP        float64 `json:"top_p"`
+	TopK        int     `json:"top_k"`
+	Temperature float64 `json:"temperature"`
+	Maxtokens   int     `json:"max_tokens"`
+}
+
+//go:embed index.html
+var indexHTML embed.FS
+
+func openAIEndpoint(chat bool, defaultModel *llama.LLama, loader *model.ModelLoader, threads int, defaultMutex *sync.Mutex, mutexMap *sync.Mutex, mutexes map[string]*sync.Mutex) func(c *fiber.Ctx) error {
+	return func(c *fiber.Ctx) error {
+		var err error
+		var model *llama.LLama
+
+		input := new(OpenAIRequest)
+		// Get input data from the request body
+		if err := c.BodyParser(input); err != nil {
+			return err
+		}
+
+		if input.Model == "" {
+			if defaultModel == nil {
+				return fmt.Errorf("no default model loaded, and no model specified")
+			}
+			model = defaultModel
+		} else {
+			model, err = loader.LoadModel(input.Model)
+			if err != nil {
+				return err
+			}
+		}
+
+		// This is still needed, see: https://github.com/ggerganov/llama.cpp/discussions/784
+		if input.Model != "" {
+			mutexMap.Lock()
+			l, ok := mutexes[input.Model]
+			if !ok {
+				m := &sync.Mutex{}
+				mutexes[input.Model] = m
+				l = m
+			}
+			mutexMap.Unlock()
+			l.Lock()
+			defer l.Unlock()
+		} else {
+			defaultMutex.Lock()
+			defer defaultMutex.Unlock()
+		}
+
+		// Set the parameters for the language model prediction
+		topP := input.TopP
+		if topP == 0 {
+			topP = 0.7
+		}
+		topK := input.TopK
+		if topK == 0 {
+			topK = 80
+		}
+
+		temperature := input.Temperature
+		if temperature == 0 {
+			temperature = 0.9
+		}
+
+		tokens := input.Maxtokens
+		if tokens == 0 {
+			tokens = 512
+		}
+
+		predInput := input.Prompt
+		if chat {
+			mess := []string{}
+			for _, i := range input.Messages {
+				mess = append(mess, i.Content)
+			}
+
+			predInput = strings.Join(mess, "\n")
+		}
+
+		// A model can have a "file.bin.tmpl" file associated with a prompt template prefix
+		templatedInput, err := loader.TemplatePrefix(input.Model, struct {
+			Input string
+		}{Input: predInput})
+		if err == nil {
+			predInput = templatedInput
+		}
+
+		// Generate the prediction using the language model
+		prediction, err := model.Predict(
+			predInput,
+			llama.SetTemperature(temperature),
+			llama.SetTopP(topP),
+			llama.SetTopK(topK),
+			llama.SetTokens(tokens),
+			llama.SetThreads(threads),
+		)
+		if err != nil {
+			return err
+		}
+
+		if chat {
+			// Return the chat prediction in the response body
+			return c.JSON(OpenAIResponse{
+				Model:   input.Model,
+				Choices: []Choice{{Message: Message{Role: "assistant", Content: prediction}}},
+			})
+		}
+
+		// Return the prediction in the response body
+		return c.JSON(OpenAIResponse{
+			Model:   input.Model,
+			Choices: []Choice{{Text: prediction}},
+		})
+	}
+}
+
+func Start(defaultModel *llama.LLama, loader *model.ModelLoader, listenAddr string, threads int) error {
+	app := fiber.New()
+
+	// Default middleware config
+	app.Use(recover.New())
+	app.Use(cors.New())
+
+	// This is still needed, see: https://github.com/ggerganov/llama.cpp/discussions/784
+	var mutex = &sync.Mutex{}
+	mu := map[string]*sync.Mutex{}
+	var mumutex = &sync.Mutex{}
+
+	// openAI compatible API endpoint
+	app.Post("/v1/chat/completions", openAIEndpoint(true, defaultModel, loader, threads, mutex, mumutex, mu))
+	app.Post("/v1/completions", openAIEndpoint(false, defaultModel, loader, threads, mutex, mumutex, mu))
+	app.Get("/v1/models", func(c *fiber.Ctx) error {
+		models, err := loader.ListModels()
+		if err != nil {
+			return err
+		}
+
+		dataModels := []OpenAIModel{}
+		for _, m := range models {
+			dataModels = append(dataModels, OpenAIModel{ID: m, Object: "model"})
+		}
+		return c.JSON(struct {
+			Object string        `json:"object"`
+			Data   []OpenAIModel `json:"data"`
+		}{
+			Object: "list",
+			Data:   dataModels,
+		})
+	})
+
+	app.Use("/", filesystem.New(filesystem.Config{
+		Root:         http.FS(indexHTML),
+		NotFoundFile: "index.html",
+	}))
+
+	/*
+		curl --location --request POST 'http://localhost:8080/predict' --header 'Content-Type: application/json' --data-raw '{
+		    "text": "What is an alpaca?",
+		    "topP": 0.8,
+		    "topK": 50,
+		    "temperature": 0.7,
+		    "tokens": 100
+		}'
+	*/
+	// Endpoint to generate the prediction
+	app.Post("/predict", func(c *fiber.Ctx) error {
+		mutex.Lock()
+		defer mutex.Unlock()
+		// Get input data from the request body
+		input := new(struct {
+			Text string `json:"text"`
+		})
+		if err := c.BodyParser(input); err != nil {
+			return err
+		}
+
+		// Set the parameters for the language model prediction
+		topP, err := strconv.ParseFloat(c.Query("topP", "0.9"), 64) // Default value of topP is 0.9
+		if err != nil {
+			return err
+		}
+
+		topK, err := strconv.Atoi(c.Query("topK", "40")) // Default value of topK is 40
+		if err != nil {
+			return err
+		}
+
+		temperature, err := strconv.ParseFloat(c.Query("temperature", "0.5"), 64) // Default value of temperature is 0.5
+		if err != nil {
+			return err
+		}
+
+		tokens, err := strconv.Atoi(c.Query("tokens", "128")) // Default value of tokens is 128
+		if err != nil {
+			return err
+		}
+
+		// Generate the prediction using the language model
+		prediction, err := defaultModel.Predict(
+			input.Text,
+			llama.SetTemperature(temperature),
+			llama.SetTopP(topP),
+			llama.SetTopK(topK),
+			llama.SetTokens(tokens),
+			llama.SetThreads(threads),
+		)
+		if err != nil {
+			return err
+		}
+
+		// Return the prediction in the response body
+		return c.JSON(struct {
+			Prediction string `json:"prediction"`
+		}{
+			Prediction: prediction,
+		})
+	})
+
+	// Start the server
+	app.Listen(listenAddr)
+	return nil
+}

docker-compose.yaml

@@ -0,0 +1,15 @@
+version: '3.6'
+
+services:
+  api:
+    image: quay.io/go-skynet/llama-cli:latest
+    build: .
+    volumes:
+      - ./models:/models
+    ports:
+      - 8080:8080
+    environment:
+      - MODELS_PATH=/models
+      - CONTEXT_SIZE=700
+      - THREADS=$THREADS
+    command: api

go.mod

@@ -17,6 +17,7 @@ require (
 	github.com/aymanbagabas/go-osc52/v2 v2.0.1 // indirect
 	github.com/containerd/console v1.0.3 // indirect
 	github.com/cpuguy83/go-md2man/v2 v2.0.2 // indirect
+	github.com/go-skynet/go-llama.cpp v0.0.0-20230405204601-5429d2339021 // indirect
 	github.com/google/uuid v1.3.0 // indirect
 	github.com/klauspost/compress v1.15.9 // indirect
 	github.com/lucasb-eyer/go-colorful v1.2.0 // indirect

go.sum

@@ -19,6 +19,10 @@ github.com/containerd/console v1.0.3 h1:lIr7SlA5PxZyMV30bDW0MGbiOPXwc63yRuCP0ARu
 github.com/containerd/console v1.0.3/go.mod h1:7LqA/THxQ86k76b8c/EMSiaJ3h1eZkMkXar0TQ1gf3U=
 github.com/cpuguy83/go-md2man/v2 v2.0.2 h1:p1EgwI/C7NhT0JmVkwCD2ZBK8j4aeHQX2pMHHBfMQ6w=
 github.com/cpuguy83/go-md2man/v2 v2.0.2/go.mod h1:tgQtvFlXSQOSOSIRvRPT7W67SCa46tRHOmNcaadrF8o=
+github.com/go-skynet/go-llama.cpp v0.0.0-20230404185816-24b85a924f09 h1:WPUWvw7DOv3WUuhtNfv+xJVE2CCTGa1op1PKGcNk2Bk=
+github.com/go-skynet/go-llama.cpp v0.0.0-20230404185816-24b85a924f09/go.mod h1:yD5HHNAHPReBlvWGWUr9OcMeE5BJH3xOUDtKCwjxdEQ=
+github.com/go-skynet/go-llama.cpp v0.0.0-20230405204601-5429d2339021 h1:SsUkTjdCCAJjULfspizf99Sfw8Fx9OAHF30kp3i6cxc=
+github.com/go-skynet/go-llama.cpp v0.0.0-20230405204601-5429d2339021/go.mod h1:yD5HHNAHPReBlvWGWUr9OcMeE5BJH3xOUDtKCwjxdEQ=
 github.com/go-skynet/llama v0.0.0-20230321172246-7be5326e18cc h1:NcmO8mA7iRZIX0Qy2SjcsSaV14+g87MiTey1neUJaFQ=
 github.com/go-skynet/llama v0.0.0-20230321172246-7be5326e18cc/go.mod h1:ZtYsAIud4cvP9VTTI9uhdgR1uCwaO/gGKnZZ95h9i7w=
 github.com/go-skynet/llama v0.0.0-20230325223742-a3563a2690ba h1:u6OhAqlWFHsTjfWKePdK2kP4/mTyXX5vsmKwrK5QX6o=

interactive.go

@@ -1,142 +0,0 @@
-package main
-
-// A simple program demonstrating the text area component from the Bubbles
-// component library.
-
-import (
-	"fmt"
-	"strings"
-
-	"github.com/charmbracelet/bubbles/textarea"
-	"github.com/charmbracelet/bubbles/viewport"
-	tea "github.com/charmbracelet/bubbletea"
-	"github.com/charmbracelet/lipgloss"
-	llama "github.com/go-skynet/llama/go"
-)
-
-func startInteractive(l *llama.LLama, opts ...llama.PredictOption) error {
-	p := tea.NewProgram(initialModel(l, opts...))
-
-	_, err := p.Run()
-	return err
-}
-
-type (
-	errMsg error
-)
-
-type model struct {
-	viewport    viewport.Model
-	messages    *[]string
-	textarea    textarea.Model
-	senderStyle lipgloss.Style
-	err         error
-	l           *llama.LLama
-	opts        []llama.PredictOption
-
-	predictC chan string
-}
-
-func initialModel(l *llama.LLama, opts ...llama.PredictOption) model {
-	ta := textarea.New()
-	ta.Placeholder = "Send a message..."
-	ta.Focus()
-
-	ta.Prompt = "┃ "
-	ta.CharLimit = 280
-
-	ta.SetWidth(200)
-	ta.SetHeight(3)
-
-	// Remove cursor line styling
-	ta.FocusedStyle.CursorLine = lipgloss.NewStyle()
-
-	ta.ShowLineNumbers = false
-
-	vp := viewport.New(200, 5)
-	vp.SetContent(`Welcome to llama-cli. Type a message and press Enter to send. Alpaca doesn't keep context of the whole chat (yet).`)
-
-	ta.KeyMap.InsertNewline.SetEnabled(false)
-
-	predictChannel := make(chan string)
-	messages := []string{}
-	m := model{
-		textarea:    ta,
-		messages:    &messages,
-		viewport:    vp,
-		senderStyle: lipgloss.NewStyle().Foreground(lipgloss.Color("5")),
-		err:         nil,
-		l:           l,
-		opts:        opts,
-		predictC:    predictChannel,
-	}
-	go func() {
-		for p := range predictChannel {
-			str, _ := templateString(emptyInput, struct {
-				Instruction string
-				Input       string
-			}{Instruction: p})
-			res, _ := l.Predict(
-				str,
-				opts...,
-			)
-
-			mm := *m.messages
-			*m.messages = mm[:len(mm)-1]
-			*m.messages = append(*m.messages, m.senderStyle.Render("llama: ")+res)
-			m.viewport.SetContent(strings.Join(*m.messages, "\n"))
-			ta.Reset()
-			m.viewport.GotoBottom()
-		}
-	}()
-
-	return m
-}
-
-func (m model) Init() tea.Cmd {
-	return textarea.Blink
-}
-
-func (m model) Update(msg tea.Msg) (tea.Model, tea.Cmd) {
-	var (
-		tiCmd tea.Cmd
-		vpCmd tea.Cmd
-	)
-
-	m.textarea, tiCmd = m.textarea.Update(msg)
-	m.viewport, vpCmd = m.viewport.Update(msg)
-
-	switch msg := msg.(type) {
-	case tea.WindowSizeMsg:
-
-	//	m.viewport.Width = msg.Width
-	//	m.viewport.Height = msg.Height
-	case tea.KeyMsg:
-		switch msg.Type {
-		case tea.KeyCtrlC, tea.KeyEsc:
-			fmt.Println(m.textarea.Value())
-			return m, tea.Quit
-		case tea.KeyEnter:
-			*m.messages = append(*m.messages, m.senderStyle.Render("You: ")+m.textarea.Value(), m.senderStyle.Render("Loading response..."))
-			m.predictC <- m.textarea.Value()
-			m.viewport.SetContent(strings.Join(*m.messages, "\n"))
-			m.textarea.Reset()
-			m.viewport.GotoBottom()
-		}
-
-	// We handle errors just like any other message
-	case errMsg:
-		m.err = msg
-		return m, nil
-	}
-
-	return m, tea.Batch(tiCmd, vpCmd)
-}
-
-func (m model) View() string {
-	return fmt.Sprintf(
-		"%s\n\n%s",
-		m.viewport.View(),
-		m.textarea.View(),
-	) + "\n\n"
-}

kubernetes/deployment.yaml

@@ -25,7 +25,7 @@ spec:
         - name: llama
           args:
           - api
-          image: quay.io/go-skynet/llama-cli:v0.3
+          image: quay.io/go-skynet/llama-cli:latest
 ---
 apiVersion: v1
 kind: Service

main.go

@@ -8,7 +8,10 @@ import (
 	"runtime"
 	"text/template"
 
-	llama "github.com/go-skynet/llama/go"
+	llama "github.com/go-skynet/go-llama.cpp"
+	api "github.com/go-skynet/llama-cli/api"
+	model "github.com/go-skynet/llama-cli/pkg/model"
+
 	"github.com/urfave/cli/v2"
 )
 
@@ -33,12 +36,6 @@ var nonEmptyInput string = `Below is an instruction that describes a task, paire
 
 func llamaFromOptions(ctx *cli.Context) (*llama.LLama, error) {
 	opts := []llama.ModelOption{llama.SetContext(ctx.Int("context-size"))}
-	if ctx.Bool("alpaca") {
-		opts = append(opts, llama.EnableAlpaca)
-	}
-	if ctx.Bool("gpt4all") {
-		opts = append(opts, llama.EnableGPT4All)
-	}
 	return llama.New(ctx.String("model"), opts...)
 }
 
@@ -92,16 +89,6 @@ var modelFlags = []cli.Flag{
 		EnvVars: []string{"TOP_K"},
 		Value:   20,
 	},
-	&cli.BoolFlag{
-		Name:    "alpaca",
-		EnvVars: []string{"ALPACA"},
-		Value:   true,
-	},
-	&cli.BoolFlag{
-		Name:    "gpt4all",
-		EnvVars: []string{"GPT4ALL"},
-		Value:   false,
-	},
 }
 
 func main() {
@@ -134,24 +121,6 @@ echo "An Alpaca (Vicugna pacos) is a domesticated species of South American came
 `,
 		Copyright: "go-skynet authors",
 		Commands: []*cli.Command{
-			{
-				Flags: modelFlags,
-				Name:  "interactive",
-				Action: func(ctx *cli.Context) error {
-
-					l, err := llamaFromOptions(ctx)
-					if err != nil {
-						fmt.Println("Loading the model failed:", err.Error())
-						os.Exit(1)
-					}
-
-					return startInteractive(l, llama.SetTemperature(ctx.Float64("temperature")),
-						llama.SetTopP(ctx.Float64("topp")),
-						llama.SetTopK(ctx.Int("topk")),
-						llama.SetTokens(ctx.Int("tokens")),
-						llama.SetThreads(ctx.Int("threads")))
-				},
-			},
 			{
 
 				Name: "api",
@@ -162,24 +131,18 @@ echo "An Alpaca (Vicugna pacos) is a domesticated species of South American came
 						Value:   runtime.NumCPU(),
 					},
 					&cli.StringFlag{
-						Name:    "model",
-						EnvVars: []string{"MODEL_PATH"},
+						Name:    "models-path",
+						EnvVars: []string{"MODELS_PATH"},
+					},
+					&cli.StringFlag{
+						Name:    "default-model",
+						EnvVars: []string{"default-model"},
 					},
 					&cli.StringFlag{
 						Name:    "address",
 						EnvVars: []string{"ADDRESS"},
 						Value:   ":8080",
 					},
-					&cli.BoolFlag{
-						Name:    "alpaca",
-						EnvVars: []string{"ALPACA"},
-						Value:   true,
-					},
-					&cli.BoolFlag{
-						Name:    "gpt4all",
-						EnvVars: []string{"GPT4ALL"},
-						Value:   false,
-					},
 					&cli.IntFlag{
 						Name:    "context-size",
 						EnvVars: []string{"CONTEXT_SIZE"},
@@ -187,13 +150,19 @@ echo "An Alpaca (Vicugna pacos) is a domesticated species of South American came
 					},
 				},
 				Action: func(ctx *cli.Context) error {
-					l, err := llamaFromOptions(ctx)
-					if err != nil {
-						fmt.Println("Loading the model failed:", err.Error())
-						os.Exit(1)
+
+					var defaultModel *llama.LLama
+					defModel := ctx.String("default-model")
+					if defModel != "" {
+						opts := []llama.ModelOption{llama.SetContext(ctx.Int("context-size"))}
+						var err error
+						defaultModel, err = llama.New(ctx.String("default-model"), opts...)
+						if err != nil {
+							return err
+						}
 					}
 
-					return api(l, ctx.String("address"), ctx.Int("threads"))
+					return api.Start(defaultModel, model.NewModelLoader(ctx.String("models-path")), ctx.String("address"), ctx.Int("threads"))
 				},
 			},
 		},

pkg/model/loader.go

@@ -0,0 +1,114 @@
+package model
+
+import (
+	"bytes"
+	"fmt"
+	"io/ioutil"
+	"os"
+	"path/filepath"
+	"strings"
+	"sync"
+	"text/template"
+
+	llama "github.com/go-skynet/go-llama.cpp"
+)
+
+type ModelLoader struct {
+	modelPath        string
+	mu               sync.Mutex
+	models           map[string]*llama.LLama
+	promptsTemplates map[string]*template.Template
+}
+
+func NewModelLoader(modelPath string) *ModelLoader {
+	return &ModelLoader{modelPath: modelPath, models: make(map[string]*llama.LLama), promptsTemplates: make(map[string]*template.Template)}
+}
+
+func (ml *ModelLoader) ListModels() ([]string, error) {
+	files, err := ioutil.ReadDir(ml.modelPath)
+	if err != nil {
+		return []string{}, err
+	}
+
+	models := []string{}
+	for _, file := range files {
+		if strings.HasSuffix(file.Name(), ".bin") {
+			models = append(models, strings.TrimRight(file.Name(), ".bin"))
+		}
+	}
+
+	return models, nil
+}
+
+func (ml *ModelLoader) TemplatePrefix(modelName string, in interface{}) (string, error) {
+	ml.mu.Lock()
+	defer ml.mu.Unlock()
+
+	m, ok := ml.promptsTemplates[modelName]
+	if !ok {
+		// try to find a s.bin
+		modelBin := fmt.Sprintf("%s.bin", modelName)
+		m, ok = ml.promptsTemplates[modelBin]
+		if !ok {
+			return "", fmt.Errorf("no prompt template available")
+		}
+	}
+
+	var buf bytes.Buffer
+
+	if err := m.Execute(&buf, in); err != nil {
+		return "", err
+	}
+	return buf.String(), nil
+}
+
+func (ml *ModelLoader) LoadModel(modelName string, opts ...llama.ModelOption) (*llama.LLama, error) {
+	ml.mu.Lock()
+	defer ml.mu.Unlock()
+
+	// Check if we already have a loaded model
+	modelFile := filepath.Join(ml.modelPath, modelName)
+
+	if m, ok := ml.models[modelFile]; ok {
+		return m, nil
+	}
+
+	// Check if the model path exists
+	if _, err := os.Stat(modelFile); os.IsNotExist(err) {
+		// try to find a s.bin
+		modelBin := fmt.Sprintf("%s.bin", modelFile)
+		if _, err := os.Stat(modelBin); os.IsNotExist(err) {
+			return nil, err
+		} else {
+			modelName = fmt.Sprintf("%s.bin", modelName)
+			modelFile = modelBin
+		}
+	}
+
+	// Load the model and keep it in memory for later use
+	model, err := llama.New(modelFile, opts...)
+	if err != nil {
+		return nil, err
+	}
+
+	// If there is a prompt template, load it
+
+	modelTemplateFile := fmt.Sprintf("%s.tmpl", modelFile)
+	// Check if the model path exists
+	if _, err := os.Stat(modelTemplateFile); err == nil {
+		dat, err := os.ReadFile(modelTemplateFile)
+		if err != nil {
+			return nil, err
+		}
+
+		// Parse the template
+		tmpl, err := template.New("prompt").Parse(string(dat))
+		if err != nil {
+			return nil, err
+		}
+		ml.promptsTemplates[modelName] = tmpl
+	}
+
+	ml.models[modelFile] = model
+	return model, err
+}