wip

added python rag news summary

.dockerignore

@@ -1,4 +1,8 @@
 .vscode
 ollama
 app
+dist
+scripts
 llm/llama.cpp/ggml
+llm/llama.cpp/gguf
+.env

.gitmodules

@@ -1,4 +1,10 @@
 [submodule "llm/llama.cpp/ggml"]
-	path = llm/llama.cpp/ggml
-	url = https://github.com/ggerganov/llama.cpp.git
-	ignore = dirty
+    path = llm/llama.cpp/ggml
+    url = https://github.com/ggerganov/llama.cpp.git
+    ignore = dirty
+    shallow = true
+[submodule "llm/llama.cpp/gguf"]
+    path = llm/llama.cpp/gguf
+    url = https://github.com/ggerganov/llama.cpp.git
+    ignore = dirty
+    shallow = true

Dockerfile

@@ -1,21 +1,23 @@
-FROM golang:alpine
+FROM nvidia/cuda:11.8.0-devel-ubuntu22.04
+
+ARG TARGETARCH
+ARG GOFLAGS="'-ldflags=-w -s'"
 
 WORKDIR /go/src/github.com/jmorganca/ollama
-RUN apk add --no-cache git build-base cmake
+RUN apt-get update && apt-get install -y git build-essential cmake
+ADD https://dl.google.com/go/go1.21.3.linux-$TARGETARCH.tar.gz /tmp/go1.21.3.tar.gz
+RUN mkdir -p /usr/local && tar xz -C /usr/local </tmp/go1.21.3.tar.gz
 
 COPY . .
-RUN go generate ./... && go build -ldflags '-linkmode external -extldflags "-static"' .
-
-FROM alpine
-ENV OLLAMA_HOST 0.0.0.0
-RUN apk add --no-cache libstdc++
-
-ARG USER=ollama
-ARG GROUP=ollama
-RUN addgroup $GROUP && adduser -D -G $GROUP $USER
+ENV GOARCH=$TARGETARCH
+ENV GOFLAGS=$GOFLAGS
+RUN /usr/local/go/bin/go generate ./... \
+    && /usr/local/go/bin/go build .
 
+FROM ubuntu:22.04
+RUN apt-get update && apt-get install -y ca-certificates
 COPY --from=0 /go/src/github.com/jmorganca/ollama/ollama /bin/ollama
-
-USER $USER:$GROUP
+EXPOSE 11434
+ENV OLLAMA_HOST 0.0.0.0
 ENTRYPOINT ["/bin/ollama"]
 CMD ["serve"]

Dockerfile.build

@@ -0,0 +1,31 @@
+# centos7 amd64 dependencies
+FROM --platform=linux/amd64 nvidia/cuda:11.3.1-devel-centos7 AS base-amd64
+RUN yum install -y https://repo.ius.io/ius-release-el7.rpm centos-release-scl && \
+    yum update -y && \
+    yum install -y devtoolset-10-gcc devtoolset-10-gcc-c++ git236 wget
+RUN wget "https://github.com/Kitware/CMake/releases/download/v3.27.6/cmake-3.27.6-linux-x86_64.sh" -O cmake-installer.sh && chmod +x cmake-installer.sh && ./cmake-installer.sh --skip-license --prefix=/usr/local
+ENV PATH /opt/rh/devtoolset-10/root/usr/bin:$PATH
+
+# centos8 arm64 dependencies
+FROM --platform=linux/arm64 nvidia/cuda-arm64:11.3.1-devel-centos8 AS base-arm64
+RUN sed -i -e 's/mirrorlist/#mirrorlist/g' -e 's|#baseurl=http://mirror.centos.org|baseurl=http://vault.centos.org|g' /etc/yum.repos.d/CentOS-*
+RUN yum install -y git cmake
+
+FROM base-${TARGETARCH}
+ARG TARGETARCH
+ARG GOFLAGS="'-ldflags -w -s'"
+
+# install go
+ADD https://dl.google.com/go/go1.21.3.linux-$TARGETARCH.tar.gz /tmp/go1.21.3.tar.gz
+RUN mkdir -p /usr/local && tar xz -C /usr/local </tmp/go1.21.3.tar.gz
+
+# build the final binary
+WORKDIR /go/src/github.com/jmorganca/ollama
+COPY . .
+
+ENV GOOS=linux
+ENV GOARCH=$TARGETARCH
+ENV GOFLAGS=$GOFLAGS
+
+RUN /usr/local/go/bin/go generate ./... && \
+    /usr/local/go/bin/go build .

README.md

@@ -9,19 +9,31 @@
 
 [![Discord](https://dcbadge.vercel.app/api/server/ollama?style=flat&compact=true)](https://discord.gg/ollama)
 
-Run, create, and share large language models (LLMs).
+Get up and running with large language models locally.
 
-> Note: Ollama is in early preview. Please report any issues you find.
+### macOS
 
-## Download
+[Download](https://ollama.ai/download/Ollama-darwin.zip)
 
-- [Download](https://ollama.ai/download) for macOS
-- Download for Windows and Linux (coming soon)
-- Build [from source](#building)
+### Windows
+
+Coming soon!
+
+### Linux & WSL2
+
+```
+curl https://ollama.ai/install.sh | sh
+```
+
+[Manual install instructions](https://github.com/jmorganca/ollama/blob/main/docs/linux.md)
+
+### Docker
+
+See the official [Docker image](https://hub.docker.com/r/ollama/ollama).
 
 ## Quickstart
 
-To run and chat with [Llama 2](https://ai.meta.com/llama), the new model by Meta:
+To run and chat with [Llama 2](https://ollama.ai/library/llama2):
 
 ```
 ollama run llama2
@@ -33,83 +45,50 @@ Ollama supports a list of open-source models available on [ollama.ai/library](ht
 
 Here are some example open-source models that can be downloaded:
 
-| Model                    | Parameters | Size  | Download                        |
-| ------------------------ | ---------- | ----- | ------------------------------- |
-| Llama2                   | 7B         | 3.8GB | `ollama pull llama2`            |
-| Llama2 13B               | 13B        | 7.3GB | `ollama pull llama2:13b`        |
-| Llama2 70B               | 70B        | 39GB  | `ollama pull llama2:70b`        |
-| Llama2 Uncensored        | 7B         | 3.8GB | `ollama pull llama2-uncensored` |
-| Code Llama               | 7B         | 3.8GB | `ollama pull codellama`         |
-| Orca Mini                | 3B         | 1.9GB | `ollama pull orca-mini`         |
-| Vicuna                   | 7B         | 3.8GB | `ollama pull vicuna`            |
-| Nous-Hermes              | 7B         | 3.8GB | `ollama pull nous-hermes`       |
-| Nous-Hermes 13B          | 13B        | 7.3GB | `ollama pull nous-hermes:13b`   |
-| Wizard Vicuna Uncensored | 13B        | 7.3GB | `ollama pull wizard-vicuna`     |
+| Model              | Parameters | Size  | Download                       |
+| ------------------ | ---------- | ----- | ------------------------------ |
+| Mistral            | 7B         | 4.1GB | `ollama run mistral`           |
+| Llama 2            | 7B         | 3.8GB | `ollama run llama2`            |
+| Code Llama         | 7B         | 3.8GB | `ollama run codellama`         |
+| Llama 2 Uncensored | 7B         | 3.8GB | `ollama run llama2-uncensored` |
+| Llama 2 13B        | 13B        | 7.3GB | `ollama run llama2:13b`        |
+| Llama 2 70B        | 70B        | 39GB  | `ollama run llama2:70b`        |
+| Orca Mini          | 3B         | 1.9GB | `ollama run orca-mini`         |
+| Vicuna             | 7B         | 3.8GB | `ollama run vicuna`            |
 
 > Note: You should have at least 8 GB of RAM to run the 3B models, 16 GB to run the 7B models, and 32 GB to run the 13B models.
 
-## Examples
+## Customize your own model
 
-### Pull a public model
+### Import from GGUF
 
-```
-ollama pull llama2
-```
+Ollama supports importing GGUF models in the Modelfile:
 
-> This command can also be used to update a local model. Only updated changes will be pulled.
+1. Create a file named `Modelfile`, with a `FROM` instruction with the local filepath to the model you want to import.
 
-### Run a model interactively
+   ```
+   FROM ./vicuna-33b.Q4_0.gguf
+   ```
 
-```
-ollama run llama2
->>> hi
-Hello! How can I help you today?
-```
+2. Create the model in Ollama
 
-For multiline input, you can wrap text with `"""`:
+   ```
+   ollama create example -f Modelfile
+   ```
 
-```
->>> """Hello,
-... world!
-... """
-I'm a basic program that prints the famous "Hello, world!" message to the console.
-```
+3. Run the model
 
-### Run a model non-interactively
+   ```
+   ollama run example
+   ```
 
-```
-$ ollama run llama2 'tell me a joke'
- Sure! Here's a quick one:
- Why did the scarecrow win an award? Because he was outstanding in his field!
-```
+### Import from PyTorch or Safetensors
 
-```
-$ cat <<EOF >prompts.txt
-tell me a joke about llamas
-tell me another one
-EOF
-$ ollama run llama2 <prompts.txt
->>> tell me a joke about llamas
- Why did the llama refuse to play hide-and-seek?
- nobody likes to be hided!
+See the [guide](docs/import.md) on importing models for more information.
 
->>> tell me another one
- Sure, here's another one:
+### Customize a prompt
 
-Why did the llama go to the bar?
-To have a hay-often good time!
-```
-
-### Run a model on contents of a text file
-
-```
-$ ollama run llama2 "summarize this file:" "$(cat README.md)"
- Ollama is a lightweight, extensible framework for building and running language models on the local machine. It provides a simple API for creating, running, and managing models, as well as a library of pre-built models that can be easily used in a variety of applications.
-```
-
-### Customize a model
-
-Pull a base model:
+Models from the Ollama library can be customized with a prompt. The example
 
 ```
 ollama pull llama2
@@ -138,30 +117,61 @@ ollama run mario
 Hello! It's your friend Mario.
 ```
 
-For more examples, see the [examples](./examples) directory. For more information on creating a Modelfile, see the [Modelfile](./docs/modelfile.md) documentation.
+For more examples, see the [examples](examples) directory. For more information on working with a Modelfile, see the [Modelfile](docs/modelfile.md) documentation.
 
-### Listing local models
+## CLI Reference
+
+### Create a model
+
+`ollama create` is used to create a model from a Modelfile.
+
+### Pull a model
 
 ```
-ollama list
+ollama pull llama2
 ```
 
-### Removing local models
+> This command can also be used to update a local model. Only the diff will be pulled.
+
+### Remove a model
 
 ```
 ollama rm llama2
 ```
 
-## Model packages
+### Copy a model
 
-### Overview
+```
+ollama cp llama2 my-llama2
+```
 
-Ollama bundles model weights, configurations, and data into a single package, defined by a [Modelfile](./docs/modelfile.md).
+### Multiline input
 
-<picture>
-  <source media="(prefers-color-scheme: dark)" height="480" srcset="https://github.com/jmorganca/ollama/assets/251292/2fd96b5f-191b-45c1-9668-941cfad4eb70">
-  <img alt="logo" height="480" src="https://github.com/jmorganca/ollama/assets/251292/2fd96b5f-191b-45c1-9668-941cfad4eb70">
-</picture>
+For multiline input, you can wrap text with `"""`:
+
+```
+>>> """Hello,
+... world!
+... """
+I'm a basic program that prints the famous "Hello, world!" message to the console.
+```
+
+### Pass in prompt as arguments
+
+```
+$ ollama run llama2 "summarize this file:" "$(cat README.md)"
+ Ollama is a lightweight, extensible framework for building and running language models on the local machine. It provides a simple API for creating, running, and managing models, as well as a library of pre-built models that can be easily used in a variety of applications.
+```
+
+### List models on your computer
+
+```
+ollama list
+```
+
+### Start Ollama
+
+`ollama serve` is used when you want to start ollama without running the desktop application.
 
 ## Building
 
@@ -193,7 +203,7 @@ Finally, in a separate shell, run a model:
 
 ## REST API
 
-> See the [API documentation](./docs/api.md) for all endpoints.
+See the [API documentation](docs/api.md) for all endpoints.
 
 Ollama has an API for running and managing models. For example to generate text from a model:
 
@@ -204,12 +214,22 @@ curl -X POST http://localhost:11434/api/generate -d '{
 }'
 ```
 
-## Community Projects using Ollama
+## Community Integrations
 
-- [LangChain](https://python.langchain.com/docs/integrations/llms/ollama) and [LangChain.js](https://js.langchain.com/docs/modules/model_io/models/llms/integrations/ollama) with a question-answering [example](https://js.langchain.com/docs/use_cases/question_answering/local_retrieval_qa).
-- [Continue](https://github.com/continuedev/continue) - embeds Ollama inside Visual Studio Code. The extension lets you highlight code to add to the prompt, ask questions in the sidebar, and generate code inline.
-- [LiteLLM](https://github.com/BerriAI/litellm) a lightweight python package to simplify LLM API calls
-- [Discord AI Bot](https://github.com/mekb-turtle/discord-ai-bot) - interact with Ollama as a chatbot on Discord.
-- [Raycast Ollama](https://github.com/MassimilianoPasquini97/raycast_ollama) - Raycast extension to use Ollama for local llama inference on Raycast.
-- [Simple HTML UI for Ollama](https://github.com/rtcfirefly/ollama-ui)
-- [Emacs client](https://github.com/zweifisch/ollama) for Ollama
+- [LangChain](https://python.langchain.com/docs/integrations/llms/ollama) and [LangChain.js](https://js.langchain.com/docs/modules/model_io/models/llms/integrations/ollama) with [example](https://js.langchain.com/docs/use_cases/question_answering/local_retrieval_qa)
+- [LlamaIndex](https://gpt-index.readthedocs.io/en/stable/examples/llm/ollama.html)
+- [Raycast extension](https://github.com/MassimilianoPasquini97/raycast_ollama)
+- [Discollama](https://github.com/mxyng/discollama) (Discord bot inside the Ollama discord channel)
+- [Continue](https://github.com/continuedev/continue)
+- [Obsidian Ollama plugin](https://github.com/hinterdupfinger/obsidian-ollama)
+- [Dagger Chatbot](https://github.com/samalba/dagger-chatbot)
+- [LiteLLM](https://github.com/BerriAI/litellm)
+- [Discord AI Bot](https://github.com/mekb-turtle/discord-ai-bot)
+- [Chatbot UI](https://github.com/ivanfioravanti/chatbot-ollama)
+- [HTML UI](https://github.com/rtcfirefly/ollama-ui)
+- [Typescript UI](https://github.com/ollama-interface/Ollama-Gui?tab=readme-ov-file)
+- [Dumbar](https://github.com/JerrySievert/Dumbar)
+- [Emacs client](https://github.com/zweifisch/ollama)
+- [oterm](https://github.com/ggozad/oterm)
+- [Ellama Emacs client](https://github.com/s-kostyaev/ellama)
+- [OllamaSharp for .NET](https://github.com/awaescher/OllamaSharp)

api/client.go

@@ -7,25 +7,24 @@ import (
 	"encoding/json"
 	"fmt"
 	"io"
+	"net"
 	"net/http"
 	"net/url"
 	"os"
 	"runtime"
 	"strings"
 
+	"github.com/jmorganca/ollama/format"
 	"github.com/jmorganca/ollama/version"
 )
 
 const DefaultHost = "127.0.0.1:11434"
 
-var (
-	envHost = os.Getenv("OLLAMA_HOST")
-)
+var envHost = os.Getenv("OLLAMA_HOST")
 
 type Client struct {
-	Base    url.URL
-	HTTP    http.Client
-	Headers http.Header
+	base *url.URL
+	http http.Client
 }
 
 func checkError(resp *http.Response, body []byte) error {
@@ -44,34 +43,46 @@ func checkError(resp *http.Response, body []byte) error {
 	return apiError
 }
 
-// Host returns the default host to use for the client. It is determined in the following order:
-// 1. The OLLAMA_HOST environment variable
-// 2. The default host (localhost:11434)
-func Host() string {
-	if envHost != "" {
-		return envHost
-	}
-	return DefaultHost
-}
-
-// FromEnv creates a new client using Host() as the host. An error is returns
-// if the host is invalid.
-func FromEnv() (*Client, error) {
-	h := Host()
-	if !strings.HasPrefix(h, "http://") && !strings.HasPrefix(h, "https://") {
-		h = "http://" + h
+func ClientFromEnvironment() (*Client, error) {
+	scheme, hostport, ok := strings.Cut(os.Getenv("OLLAMA_HOST"), "://")
+	if !ok {
+		scheme, hostport = "http", os.Getenv("OLLAMA_HOST")
 	}
 
-	u, err := url.Parse(h)
+	host, port, err := net.SplitHostPort(hostport)
 	if err != nil {
-		return nil, fmt.Errorf("could not parse host: %w", err)
+		host, port = "127.0.0.1", "11434"
+		if ip := net.ParseIP(strings.Trim(hostport, "[]")); ip != nil {
+			host = ip.String()
+		} else if hostport != "" {
+			host = hostport
+		}
 	}
 
-	if u.Port() == "" {
-		u.Host += ":11434"
+	client := Client{
+		base: &url.URL{
+			Scheme: scheme,
+			Host:   net.JoinHostPort(host, port),
+		},
 	}
 
-	return &Client{Base: *u, HTTP: http.Client{}}, nil
+	mockRequest, err := http.NewRequest("HEAD", client.base.String(), nil)
+	if err != nil {
+		return nil, err
+	}
+
+	proxyURL, err := http.ProxyFromEnvironment(mockRequest)
+	if err != nil {
+		return nil, err
+	}
+
+	client.http = http.Client{
+		Transport: &http.Transport{
+			Proxy: http.ProxyURL(proxyURL),
+		},
+	}
+
+	return &client, nil
 }
 
 func (c *Client) do(ctx context.Context, method, path string, reqData, respData any) error {
@@ -86,7 +97,7 @@ func (c *Client) do(ctx context.Context, method, path string, reqData, respData
 		reqBody = bytes.NewReader(data)
 	}
 
-	requestURL := c.Base.JoinPath(path)
+	requestURL := c.base.JoinPath(path)
 	request, err := http.NewRequestWithContext(ctx, method, requestURL.String(), reqBody)
 	if err != nil {
 		return err
@@ -96,11 +107,7 @@ func (c *Client) do(ctx context.Context, method, path string, reqData, respData
 	request.Header.Set("Accept", "application/json")
 	request.Header.Set("User-Agent", fmt.Sprintf("ollama/%s (%s %s) Go/%s", version.Version, runtime.GOARCH, runtime.GOOS, runtime.Version()))
 
-	for k, v := range c.Headers {
-		request.Header[k] = v
-	}
-
-	respObj, err := c.HTTP.Do(request)
+	respObj, err := c.http.Do(request)
 	if err != nil {
 		return err
 	}
@@ -123,6 +130,8 @@ func (c *Client) do(ctx context.Context, method, path string, reqData, respData
 	return nil
 }
 
+const maxBufferSize = 512 * format.KiloByte
+
 func (c *Client) stream(ctx context.Context, method, path string, data any, fn func([]byte) error) error {
 	var buf *bytes.Buffer
 	if data != nil {
@@ -134,23 +143,26 @@ func (c *Client) stream(ctx context.Context, method, path string, data any, fn f
 		buf = bytes.NewBuffer(bts)
 	}
 
-	requestURL := c.Base.JoinPath(path)
+	requestURL := c.base.JoinPath(path)
 	request, err := http.NewRequestWithContext(ctx, method, requestURL.String(), buf)
 	if err != nil {
 		return err
 	}
 
 	request.Header.Set("Content-Type", "application/json")
-	request.Header.Set("Accept", "application/json")
+	request.Header.Set("Accept", "application/x-ndjson")
 	request.Header.Set("User-Agent", fmt.Sprintf("ollama/%s (%s %s) Go/%s", version.Version, runtime.GOARCH, runtime.GOOS, runtime.Version()))
 
-	response, err := http.DefaultClient.Do(request)
+	response, err := c.http.Do(request)
 	if err != nil {
 		return err
 	}
 	defer response.Body.Close()
 
 	scanner := bufio.NewScanner(response.Body)
+	// increase the buffer size to avoid running out of space
+	scanBuf := make([]byte, 0, maxBufferSize)
+	scanner.Buffer(scanBuf, maxBufferSize)
 	for scanner.Scan() {
 		var errorResponse struct {
 			Error string `json:"error,omitempty"`

api/client.py

@@ -0,0 +1,225 @@
+import os
+import json
+import requests
+
+BASE_URL = os.environ.get('OLLAMA_HOST', 'http://localhost:11434')
+
+# Generate a response for a given prompt with a provided model. This is a streaming endpoint, so will be a series of responses.
+# The final response object will include statistics and additional data from the request. Use the callback function to override
+# the default handler.
+def generate(model_name, prompt, system=None, template=None, context=None, options=None, callback=None):
+    try:
+        url = f"{BASE_URL}/api/generate"
+        payload = {
+            "model": model_name, 
+            "prompt": prompt, 
+            "system": system, 
+            "template": template, 
+            "context": context, 
+            "options": options
+        }
+        
+        # Remove keys with None values
+        payload = {k: v for k, v in payload.items() if v is not None}
+        
+        with requests.post(url, json=payload, stream=True) as response:
+            response.raise_for_status()
+            
+            # Creating a variable to hold the context history of the final chunk
+            final_context = None
+            
+            # Variable to hold concatenated response strings if no callback is provided
+            full_response = ""
+
+            # Iterating over the response line by line and displaying the details
+            for line in response.iter_lines():
+                if line:
+                    # Parsing each line (JSON chunk) and extracting the details
+                    chunk = json.loads(line)
+                    
+                    # If a callback function is provided, call it with the chunk
+                    if callback:
+                        callback(chunk)
+                    else:
+                        # If this is not the last chunk, add the "response" field value to full_response and print it
+                        if not chunk.get("done"):
+                            response_piece = chunk.get("response", "")
+                            full_response += response_piece
+                            print(response_piece, end="", flush=True)
+                    
+                    # Check if it's the last chunk (done is true)
+                    if chunk.get("done"):
+                        final_context = chunk.get("context")
+            
+            # Return the full response and the final context
+            return full_response, final_context
+    except requests.exceptions.RequestException as e:
+        print(f"An error occurred: {e}")
+        return None, None
+
+# Create a model from a Modelfile. Use the callback function to override the default handler.
+def create(model_name, model_path, callback=None):
+    try:
+        url = f"{BASE_URL}/api/create"
+        payload = {"name": model_name, "path": model_path}
+        
+        # Making a POST request with the stream parameter set to True to handle streaming responses
+        with requests.post(url, json=payload, stream=True) as response:
+            response.raise_for_status()
+
+            # Iterating over the response line by line and displaying the status
+            for line in response.iter_lines():
+                if line:
+                    # Parsing each line (JSON chunk) and extracting the status
+                    chunk = json.loads(line)
+
+                    if callback:
+                        callback(chunk)
+                    else:
+                        print(f"Status: {chunk.get('status')}")
+    except requests.exceptions.RequestException as e:
+        print(f"An error occurred: {e}")
+
+# Pull a model from a the model registry. Cancelled pulls are resumed from where they left off, and multiple
+# calls to will share the same download progress. Use the callback function to override the default handler.
+def pull(model_name, insecure=False, callback=None):
+    try:
+        url = f"{BASE_URL}/api/pull"
+        payload = {
+            "name": model_name,
+            "insecure": insecure
+        }
+
+        # Making a POST request with the stream parameter set to True to handle streaming responses
+        with requests.post(url, json=payload, stream=True) as response:
+            response.raise_for_status()
+
+            # Iterating over the response line by line and displaying the details
+            for line in response.iter_lines():
+                if line:
+                    # Parsing each line (JSON chunk) and extracting the details
+                    chunk = json.loads(line)
+
+                    # If a callback function is provided, call it with the chunk
+                    if callback:
+                        callback(chunk)
+                    else:
+                        # Print the status message directly to the console
+                        print(chunk.get('status', ''), end='', flush=True)
+                    
+                    # If there's layer data, you might also want to print that (adjust as necessary)
+                    if 'digest' in chunk:
+                        print(f" - Digest: {chunk['digest']}", end='', flush=True)
+                        print(f" - Total: {chunk['total']}", end='', flush=True)
+                        print(f" - Completed: {chunk['completed']}", end='\n', flush=True)
+                    else:
+                        print()
+    except requests.exceptions.RequestException as e:
+        print(f"An error occurred: {e}")
+
+# Push a model to the model registry. Use the callback function to override the default handler.
+def push(model_name, insecure=False, callback=None):
+    try:
+        url = f"{BASE_URL}/api/push"
+        payload = {
+            "name": model_name,
+            "insecure": insecure
+        }
+
+        # Making a POST request with the stream parameter set to True to handle streaming responses
+        with requests.post(url, json=payload, stream=True) as response:
+            response.raise_for_status()
+
+            # Iterating over the response line by line and displaying the details
+            for line in response.iter_lines():
+                if line:
+                    # Parsing each line (JSON chunk) and extracting the details
+                    chunk = json.loads(line)
+
+                    # If a callback function is provided, call it with the chunk
+                    if callback:
+                        callback(chunk)
+                    else:
+                        # Print the status message directly to the console
+                        print(chunk.get('status', ''), end='', flush=True)
+                    
+                    # If there's layer data, you might also want to print that (adjust as necessary)
+                    if 'digest' in chunk:
+                        print(f" - Digest: {chunk['digest']}", end='', flush=True)
+                        print(f" - Total: {chunk['total']}", end='', flush=True)
+                        print(f" - Completed: {chunk['completed']}", end='\n', flush=True)
+                    else:
+                        print()
+    except requests.exceptions.RequestException as e:
+        print(f"An error occurred: {e}")
+
+# List models that are available locally.
+def list():
+    try:
+        response = requests.get(f"{BASE_URL}/api/tags")
+        response.raise_for_status()
+        data = response.json()
+        models = data.get('models', [])
+        return models
+
+    except requests.exceptions.RequestException as e:
+        print(f"An error occurred: {e}")
+        return None
+
+# Copy a model. Creates a model with another name from an existing model.
+def copy(source, destination):
+    try:
+        # Create the JSON payload
+        payload = {
+            "source": source,
+            "destination": destination
+        }
+        
+        response = requests.post(f"{BASE_URL}/api/copy", json=payload)
+        response.raise_for_status()
+        
+        # If the request was successful, return a message indicating that the copy was successful
+        return "Copy successful"
+
+    except requests.exceptions.RequestException as e:
+        print(f"An error occurred: {e}")
+        return None
+
+# Delete a model and its data.
+def delete(model_name):
+    try:
+        url = f"{BASE_URL}/api/delete"
+        payload = {"name": model_name}
+        response = requests.delete(url, json=payload)
+        response.raise_for_status()
+        return "Delete successful"
+    except requests.exceptions.RequestException as e:
+        print(f"An error occurred: {e}")
+        return None
+
+# Show info about a model.
+def show(model_name):
+    try:
+        url = f"{BASE_URL}/api/show"
+        payload = {"name": model_name}
+        response = requests.post(url, json=payload)
+        response.raise_for_status()
+        
+        # Parse the JSON response and return it
+        data = response.json()
+        return data
+    except requests.exceptions.RequestException as e:
+        print(f"An error occurred: {e}")
+        return None
+
+def heartbeat():
+    try:
+        url = f"{BASE_URL}/"
+        response = requests.head(url)
+        response.raise_for_status()
+        return "Ollama is running"
+    except requests.exceptions.RequestException as e:
+        print(f"An error occurred: {e}")
+        return "Ollama is not running"
+
+

api/types.go

@@ -3,7 +3,6 @@ package api
 import (
 	"encoding/json"
 	"fmt"
-	"log"
 	"math"
 	"os"
 	"reflect"
@@ -37,6 +36,7 @@ type GenerateRequest struct {
 	System   string `json:"system"`
 	Template string `json:"template"`
 	Context  []int  `json:"context,omitempty"`
+	Stream   *bool  `json:"stream,omitempty"`
 
 	Options map[string]interface{} `json:"options"`
 }
@@ -53,8 +53,9 @@ type EmbeddingResponse struct {
 }
 
 type CreateRequest struct {
-	Name string `json:"name"`
-	Path string `json:"path"`
+	Name   string `json:"name"`
+	Path   string `json:"path"`
+	Stream *bool  `json:"stream,omitempty"`
 }
 
 type DeleteRequest struct {
@@ -83,13 +84,14 @@ type PullRequest struct {
 	Insecure bool   `json:"insecure,omitempty"`
 	Username string `json:"username"`
 	Password string `json:"password"`
+	Stream   *bool  `json:"stream,omitempty"`
 }
 
 type ProgressResponse struct {
 	Status    string `json:"status"`
 	Digest    string `json:"digest,omitempty"`
-	Total     int    `json:"total,omitempty"`
-	Completed int    `json:"completed,omitempty"`
+	Total     int64  `json:"total,omitempty"`
+	Completed int64  `json:"completed,omitempty"`
 }
 
 type PushRequest struct {
@@ -97,6 +99,7 @@ type PushRequest struct {
 	Insecure bool   `json:"insecure,omitempty"`
 	Username string `json:"username"`
 	Password string `json:"password"`
+	Stream   *bool  `json:"stream,omitempty"`
 }
 
 type ListResponse struct {
@@ -106,7 +109,7 @@ type ListResponse struct {
 type ModelResponse struct {
 	Name       string    `json:"name"`
 	ModifiedAt time.Time `json:"modified_at"`
-	Size       int       `json:"size"`
+	Size       int64     `json:"size"`
 	Digest     string    `json:"digest"`
 }
 
@@ -117,7 +120,7 @@ type TokenResponse struct {
 type GenerateResponse struct {
 	Model     string    `json:"model"`
 	CreatedAt time.Time `json:"created_at"`
-	Response  string    `json:"response,omitempty"`
+	Response  string    `json:"response"`
 
 	Done    bool  `json:"done"`
 	Context []int `json:"context,omitempty"`
@@ -158,15 +161,10 @@ func (r *GenerateResponse) Summary() {
 	}
 }
 
-type Options struct {
-	Seed int `json:"seed,omitempty"`
-
-	// Backend options
-	UseNUMA bool `json:"numa,omitempty"`
-
-	// Model options
+// Runner options which must be set when the model is loaded into memory
+type Runner struct {
+	UseNUMA            bool    `json:"numa,omitempty"`
 	NumCtx             int     `json:"num_ctx,omitempty"`
-	NumKeep            int     `json:"num_keep,omitempty"`
 	NumBatch           int     `json:"num_batch,omitempty"`
 	NumGQA             int     `json:"num_gqa,omitempty"`
 	NumGPU             int     `json:"num_gpu,omitempty"`
@@ -180,8 +178,15 @@ type Options struct {
 	EmbeddingOnly      bool    `json:"embedding_only,omitempty"`
 	RopeFrequencyBase  float32 `json:"rope_frequency_base,omitempty"`
 	RopeFrequencyScale float32 `json:"rope_frequency_scale,omitempty"`
+	NumThread          int     `json:"num_thread,omitempty"`
+}
 
-	// Predict options
+type Options struct {
+	Runner
+
+	// Predict options used at runtime
+	NumKeep          int      `json:"num_keep,omitempty"`
+	Seed             int      `json:"seed,omitempty"`
 	NumPredict       int      `json:"num_predict,omitempty"`
 	TopK             int      `json:"top_k,omitempty"`
 	TopP             float32  `json:"top_p,omitempty"`
@@ -197,10 +202,10 @@ type Options struct {
 	MirostatEta      float32  `json:"mirostat_eta,omitempty"`
 	PenalizeNewline  bool     `json:"penalize_newline,omitempty"`
 	Stop             []string `json:"stop,omitempty"`
-
-	NumThread int `json:"num_thread,omitempty"`
 }
 
+var ErrInvalidOpts = fmt.Errorf("invalid options")
+
 func (opts *Options) FromMap(m map[string]interface{}) error {
 	valueOpts := reflect.ValueOf(opts).Elem() // names of the fields in the options struct
 	typeOpts := reflect.TypeOf(opts).Elem()   // types of the fields in the options struct
@@ -214,6 +219,7 @@ func (opts *Options) FromMap(m map[string]interface{}) error {
 		}
 	}
 
+	invalidOpts := []string{}
 	for key, val := range m {
 		if opt, ok := jsonOpts[key]; ok {
 			field := valueOpts.FieldByName(opt.Name)
@@ -231,44 +237,39 @@ func (opts *Options) FromMap(m map[string]interface{}) error {
 						// when JSON unmarshals numbers, it uses float64, not int
 						field.SetInt(int64(t))
 					default:
-						log.Printf("could not convert model parameter %v to int, skipped", key)
+						return fmt.Errorf("option %q must be of type integer", key)
 					}
 				case reflect.Bool:
 					val, ok := val.(bool)
 					if !ok {
-						log.Printf("could not convert model parameter %v to bool, skipped", key)
-						continue
+						return fmt.Errorf("option %q must be of type boolean", key)
 					}
 					field.SetBool(val)
 				case reflect.Float32:
 					// JSON unmarshals to float64
 					val, ok := val.(float64)
 					if !ok {
-						log.Printf("could not convert model parameter %v to float32, skipped", key)
-						continue
+						return fmt.Errorf("option %q must be of type float32", key)
 					}
 					field.SetFloat(val)
 				case reflect.String:
 					val, ok := val.(string)
 					if !ok {
-						log.Printf("could not convert model parameter %v to string, skipped", key)
-						continue
+						return fmt.Errorf("option %q must be of type string", key)
 					}
 					field.SetString(val)
 				case reflect.Slice:
 					// JSON unmarshals to []interface{}, not []string
 					val, ok := val.([]interface{})
 					if !ok {
-						log.Printf("could not convert model parameter %v to slice, skipped", key)
-						continue
+						return fmt.Errorf("option %q must be of type array", key)
 					}
 					// convert []interface{} to []string
 					slice := make([]string, len(val))
 					for i, item := range val {
 						str, ok := item.(string)
 						if !ok {
-							log.Printf("could not convert model parameter %v to slice of strings, skipped", key)
-							continue
+							return fmt.Errorf("option %q must be of an array of strings", key)
 						}
 						slice[i] = str
 					}
@@ -277,45 +278,53 @@ func (opts *Options) FromMap(m map[string]interface{}) error {
 					return fmt.Errorf("unknown type loading config params: %v", field.Kind())
 				}
 			}
+		} else {
+			invalidOpts = append(invalidOpts, key)
 		}
 	}
+
+	if len(invalidOpts) > 0 {
+		return fmt.Errorf("%w: %v", ErrInvalidOpts, strings.Join(invalidOpts, ", "))
+	}
 	return nil
 }
 
 func DefaultOptions() Options {
 	return Options{
-		Seed: -1,
-
-		UseNUMA: false,
-
-		NumCtx:             2048,
-		NumKeep:            -1,
-		NumBatch:           512,
-		NumGPU:             1,
-		NumGQA:             1,
-		LowVRAM:            false,
-		F16KV:              true,
-		UseMMap:            true,
-		UseMLock:           false,
-		RopeFrequencyBase:  10000.0,
-		RopeFrequencyScale: 1.0,
-		EmbeddingOnly:      true,
-
-		RepeatLastN:      64,
-		RepeatPenalty:    1.1,
-		FrequencyPenalty: 0.0,
-		PresencePenalty:  0.0,
+		// options set on request to runner
+		NumPredict:       -1,
+		NumKeep:          -1,
 		Temperature:      0.8,
 		TopK:             40,
 		TopP:             0.9,
 		TFSZ:             1.0,
 		TypicalP:         1.0,
+		RepeatLastN:      64,
+		RepeatPenalty:    1.1,
+		PresencePenalty:  0.0,
+		FrequencyPenalty: 0.0,
 		Mirostat:         0,
 		MirostatTau:      5.0,
 		MirostatEta:      0.1,
 		PenalizeNewline:  true,
+		Seed:             -1,
 
-		NumThread: 0, // let the runtime decide
+		Runner: Runner{
+			// options set when the model is loaded
+			NumCtx:             2048,
+			RopeFrequencyBase:  10000.0,
+			RopeFrequencyScale: 1.0,
+			NumBatch:           512,
+			NumGPU:             -1, // -1 here indicates that NumGPU should be set dynamically
+			NumGQA:             1,
+			NumThread:          0, // let the runtime decide
+			LowVRAM:            false,
+			F16KV:              true,
+			UseMLock:           false,
+			UseMMap:            true,
+			UseNUMA:            false,
+			EmbeddingOnly:      true,
+		},
 	}
 }
 

app/forge.config.ts

@@ -47,16 +47,6 @@ const config: ForgeConfig = {
   },
   rebuildConfig: {},
   makers: [new MakerSquirrel({}), new MakerZIP({}, ['darwin'])],
-  publishers: [
-    new PublisherGithub({
-      repository: {
-        name: 'ollama',
-        owner: 'jmorganca',
-      },
-      draft: false,
-      prerelease: true,
-    }),
-  ],
   hooks: {
     readPackageJson: async (_, packageJson) => {
       return { ...packageJson, version: process.env.VERSION || packageJson.version }

app/package.json

@@ -46,7 +46,7 @@
     "chmodr": "^1.2.0",
     "copy-webpack-plugin": "^11.0.0",
     "css-loader": "^6.8.1",
-    "electron": "25.2.0",
+    "electron": "25.9.2",
     "eslint": "^8.43.0",
     "eslint-plugin-import": "^2.27.5",
     "fork-ts-checker-webpack-plugin": "^7.3.0",

app/src/index.ts

@@ -5,7 +5,7 @@ import winston from 'winston'
 import 'winston-daily-rotate-file'
 import * as path from 'path'
 
-import { analytics, id } from './telemetry'
+import { v4 as uuidv4 } from 'uuid'
 import { installed } from './install'
 
 require('@electron/remote/main').initialize()
@@ -162,13 +162,56 @@ app.on('before-quit', () => {
   }
 })
 
+const updateURL = `https://ollama.ai/api/update?os=${process.platform}&arch=${
+  process.arch
+}&version=${app.getVersion()}&id=${id()}`
+
+let latest = ''
+async function isNewReleaseAvailable() {
+  try {
+    const response = await fetch(updateURL)
+
+    if (!response.ok) {
+      return false
+    }
+
+    if (response.status === 204) {
+      return false
+    }
+
+    const data = await response.json()
+
+    const url = data?.url
+    if (!url) {
+      return false
+    }
+
+    if (latest === url) {
+      return false
+    }
+
+    latest = url
+
+    return true
+  } catch (error) {
+    logger.error(`update check failed - ${error}`)
+    return false
+  }
+}
+
+async function checkUpdate() {
+  const available = await isNewReleaseAvailable()
+  if (available) {
+    logger.info('checking for update')
+    autoUpdater.checkForUpdates()
+  }
+}
+
 function init() {
   if (app.isPackaged) {
-    heartbeat()
-    autoUpdater.checkForUpdates()
+    checkUpdate()
     setInterval(() => {
-      heartbeat()
-      autoUpdater.checkForUpdates()
+      checkUpdate()
     }, 60 * 60 * 1000)
   }
 
@@ -234,28 +277,22 @@ app.on('window-all-closed', () => {
   }
 })
 
-// In this file you can include the rest of your app's specific main process
-// code. You can also put them in separate files and import them here.
-let aid = ''
-try {
-  aid = id()
-} catch (e) {}
+function id(): string {
+  const id = store.get('id') as string
 
-autoUpdater.setFeedURL({
-  url: `https://ollama.ai/api/update?os=${process.platform}&arch=${process.arch}&version=${app.getVersion()}&id=${aid}`,
-})
+  if (id) {
+    return id
+  }
 
-async function heartbeat() {
-  analytics.track({
-    anonymousId: aid,
-    event: 'heartbeat',
-    properties: {
-      version: app.getVersion(),
-    },
-  })
+  const uuid = uuidv4()
+  store.set('id', uuid)
+  return uuid
 }
 
+autoUpdater.setFeedURL({ url: updateURL })
+
 autoUpdater.on('error', e => {
+  logger.error(`update check failed - ${e.message}`)
   console.error(`update check failed - ${e.message}`)
 })
 

app/src/telemetry.ts

@@ -1,19 +0,0 @@
-import { Analytics } from '@segment/analytics-node'
-import { v4 as uuidv4 } from 'uuid'
-import Store from 'electron-store'
-
-const store = new Store()
-
-export const analytics = new Analytics({ writeKey: process.env.TELEMETRY_WRITE_KEY || '<empty>' })
-
-export function id(): string {
-  const id = store.get('id') as string
-
-  if (id) {
-    return id
-  }
-
-  const uuid = uuidv4()
-  store.set('id', uuid)
-  return uuid
-}

cmd/cmd.go

@@ -11,20 +11,21 @@ import (
 	"io"
 	"log"
 	"net"
-	"net/http"
 	"os"
 	"os/exec"
-	"path"
+	"os/signal"
 	"path/filepath"
 	"runtime"
 	"strings"
+	"syscall"
 	"time"
 
-	"github.com/chzyer/readline"
 	"github.com/dustin/go-humanize"
 	"github.com/olekukonko/tablewriter"
+	"github.com/pdevine/readline"
 	"github.com/spf13/cobra"
 	"golang.org/x/crypto/ssh"
+	"golang.org/x/term"
 
 	"github.com/jmorganca/ollama/api"
 	"github.com/jmorganca/ollama/format"
@@ -33,6 +34,26 @@ import (
 	"github.com/jmorganca/ollama/version"
 )
 
+type Painter struct {
+	IsMultiLine bool
+}
+
+func (p Painter) Paint(line []rune, _ int) []rune {
+	termType := os.Getenv("TERM")
+	if termType == "xterm-256color" && len(line) == 0 {
+		var prompt string
+		if p.IsMultiLine {
+			prompt = "Use \"\"\" to end multi-line input"
+		} else {
+			prompt = "Send a message (/? for help)"
+		}
+		return []rune(fmt.Sprintf("\033[38;5;245m%s\033[%dD\033[0m", prompt, len(prompt)))
+	}
+	// add a space and a backspace to prevent the cursor from walking up the screen
+	line = append(line, []rune(" \b")...)
+	return line
+}
+
 func CreateHandler(cmd *cobra.Command, args []string) error {
 	filename, _ := cmd.Flags().GetString("file")
 	filename, err := filepath.Abs(filename)
@@ -40,7 +61,7 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 		return err
 	}
 
-	client, err := api.FromEnv()
+	client, err := api.ClientFromEnvironment()
 	if err != nil {
 		return err
 	}
@@ -57,20 +78,14 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 				spinner.Stop()
 			}
 			currentDigest = resp.Digest
-			switch {
-			case strings.Contains(resp.Status, "embeddings"):
-				bar = progressbar.Default(int64(resp.Total), resp.Status)
-				bar.Set(resp.Completed)
-			default:
-				// pulling
-				bar = progressbar.DefaultBytes(
-					int64(resp.Total),
-					resp.Status,
-				)
-				bar.Set(resp.Completed)
-			}
+			// pulling
+			bar = progressbar.DefaultBytes(
+				resp.Total,
+				resp.Status,
+			)
+			bar.Set64(resp.Completed)
 		} else if resp.Digest == currentDigest && resp.Digest != "" {
-			bar.Set(resp.Completed)
+			bar.Set64(resp.Completed)
 		} else {
 			currentDigest = ""
 			if spinner != nil {
@@ -98,39 +113,24 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 }
 
 func RunHandler(cmd *cobra.Command, args []string) error {
-	insecure, err := cmd.Flags().GetBool("insecure")
+	client, err := api.ClientFromEnvironment()
 	if err != nil {
 		return err
 	}
 
-	mp := server.ParseModelPath(args[0])
+	models, err := client.List(context.Background())
 	if err != nil {
 		return err
 	}
 
-	if mp.ProtocolScheme == "http" && !insecure {
-		return fmt.Errorf("insecure protocol http")
-	}
-
-	fp, err := mp.GetManifestPath(false)
-	if err != nil {
-		return err
-	}
-
-	_, err = os.Stat(fp)
-	switch {
-	case errors.Is(err, os.ErrNotExist):
-		if err := pull(args[0], insecure); err != nil {
-			var apiStatusError api.StatusError
-			if !errors.As(err, &apiStatusError) {
-				return err
-			}
-
-			if apiStatusError.StatusCode != http.StatusBadGateway {
-				return err
-			}
+	canonicalModelPath := server.ParseModelPath(args[0])
+	for _, model := range models.Models {
+		if model.Name == canonicalModelPath.GetShortTagname() {
+			return RunGenerate(cmd, args)
 		}
-	case err != nil:
+	}
+
+	if err := PullHandler(cmd, args); err != nil {
 		return err
 	}
 
@@ -138,7 +138,7 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 }
 
 func PushHandler(cmd *cobra.Command, args []string) error {
-	client, err := api.FromEnv()
+	client, err := api.ClientFromEnvironment()
 	if err != nil {
 		return err
 	}
@@ -156,13 +156,13 @@ func PushHandler(cmd *cobra.Command, args []string) error {
 		if resp.Digest != currentDigest && resp.Digest != "" {
 			currentDigest = resp.Digest
 			bar = progressbar.DefaultBytes(
-				int64(resp.Total),
+				resp.Total,
 				fmt.Sprintf("pushing %s...", resp.Digest[7:19]),
 			)
 
-			bar.Set(resp.Completed)
+			bar.Set64(resp.Completed)
 		} else if resp.Digest == currentDigest && resp.Digest != "" {
-			bar.Set(resp.Completed)
+			bar.Set64(resp.Completed)
 		} else {
 			currentDigest = ""
 			fmt.Println(resp.Status)
@@ -182,7 +182,7 @@ func PushHandler(cmd *cobra.Command, args []string) error {
 }
 
 func ListHandler(cmd *cobra.Command, args []string) error {
-	client, err := api.FromEnv()
+	client, err := api.ClientFromEnvironment()
 	if err != nil {
 		return err
 	}
@@ -215,7 +215,7 @@ func ListHandler(cmd *cobra.Command, args []string) error {
 }
 
 func DeleteHandler(cmd *cobra.Command, args []string) error {
-	client, err := api.FromEnv()
+	client, err := api.ClientFromEnvironment()
 	if err != nil {
 		return err
 	}
@@ -231,7 +231,7 @@ func DeleteHandler(cmd *cobra.Command, args []string) error {
 }
 
 func ShowHandler(cmd *cobra.Command, args []string) error {
-	client, err := api.FromEnv()
+	client, err := api.ClientFromEnvironment()
 	if err != nil {
 		return err
 	}
@@ -309,7 +309,7 @@ func ShowHandler(cmd *cobra.Command, args []string) error {
 }
 
 func CopyHandler(cmd *cobra.Command, args []string) error {
-	client, err := api.FromEnv()
+	client, err := api.ClientFromEnvironment()
 	if err != nil {
 		return err
 	}
@@ -332,7 +332,7 @@ func PullHandler(cmd *cobra.Command, args []string) error {
 }
 
 func pull(model string, insecure bool) error {
-	client, err := api.FromEnv()
+	client, err := api.ClientFromEnvironment()
 	if err != nil {
 		return err
 	}
@@ -345,13 +345,13 @@ func pull(model string, insecure bool) error {
 		if resp.Digest != currentDigest && resp.Digest != "" {
 			currentDigest = resp.Digest
 			bar = progressbar.DefaultBytes(
-				int64(resp.Total),
+				resp.Total,
 				fmt.Sprintf("pulling %s...", resp.Digest[7:19]),
 			)
 
-			bar.Set(resp.Completed)
+			bar.Set64(resp.Completed)
 		} else if resp.Digest == currentDigest && resp.Digest != "" {
-			bar.Set(resp.Completed)
+			bar.Set64(resp.Completed)
 		} else {
 			currentDigest = ""
 			fmt.Println(resp.Status)
@@ -374,7 +374,20 @@ func pull(model string, insecure bool) error {
 func RunGenerate(cmd *cobra.Command, args []string) error {
 	if len(args) > 1 {
 		// join all args into a single prompt
-		return generate(cmd, args[0], strings.Join(args[1:], " "))
+		wordWrap := false
+		if term.IsTerminal(int(os.Stdout.Fd())) {
+			wordWrap = true
+		}
+
+		nowrap, err := cmd.Flags().GetBool("nowordwrap")
+		if err != nil {
+			return err
+		}
+		if nowrap {
+			wordWrap = false
+		}
+
+		return generate(cmd, args[0], strings.Join(args[1:], " "), wordWrap)
 	}
 
 	if readline.IsTerminal(int(os.Stdin.Fd())) {
@@ -386,71 +399,110 @@ func RunGenerate(cmd *cobra.Command, args []string) error {
 
 type generateContextKey string
 
-func generate(cmd *cobra.Command, model, prompt string) error {
-	if len(strings.TrimSpace(prompt)) > 0 {
-		client, err := api.FromEnv()
-		if err != nil {
-			return err
+func generate(cmd *cobra.Command, model, prompt string, wordWrap bool) error {
+	client, err := api.ClientFromEnvironment()
+	if err != nil {
+		return err
+	}
+
+	spinner := NewSpinner("")
+	go spinner.Spin(60 * time.Millisecond)
+
+	var latest api.GenerateResponse
+
+	generateContext, ok := cmd.Context().Value(generateContextKey("context")).([]int)
+	if !ok {
+		generateContext = []int{}
+	}
+
+	termWidth, _, err := term.GetSize(int(0))
+	if err != nil {
+		wordWrap = false
+	}
+
+	cancelCtx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+
+	sigChan := make(chan os.Signal, 1)
+	signal.Notify(sigChan, syscall.SIGINT)
+	var abort bool
+
+	go func() {
+		<-sigChan
+		cancel()
+		abort = true
+	}()
+
+	var currentLineLength int
+	var wordBuffer string
+
+	request := api.GenerateRequest{Model: model, Prompt: prompt, Context: generateContext}
+	fn := func(response api.GenerateResponse) error {
+		if !spinner.IsFinished() {
+			spinner.Finish()
 		}
 
-		spinner := NewSpinner("")
-		go spinner.Spin(60 * time.Millisecond)
+		latest = response
 
-		var latest api.GenerateResponse
+		if wordWrap {
+			for _, ch := range response.Response {
+				if currentLineLength+1 > termWidth-5 {
+					// backtrack the length of the last word and clear to the end of the line
+					fmt.Printf("\x1b[%dD\x1b[K\n", len(wordBuffer))
+					fmt.Printf("%s%c", wordBuffer, ch)
+					currentLineLength = len(wordBuffer) + 1
+				} else {
+					fmt.Print(string(ch))
+					currentLineLength += 1
 
-		generateContext, ok := cmd.Context().Value(generateContextKey("context")).([]int)
-		if !ok {
-			generateContext = []int{}
-		}
-
-		request := api.GenerateRequest{Model: model, Prompt: prompt, Context: generateContext}
-		fn := func(response api.GenerateResponse) error {
-			if !spinner.IsFinished() {
-				spinner.Finish()
+					switch ch {
+					case ' ':
+						wordBuffer = ""
+					case '\n':
+						currentLineLength = 0
+					default:
+						wordBuffer += string(ch)
+					}
+				}
 			}
-
-			latest = response
-
+		} else {
 			fmt.Print(response.Response)
+		}
+
+		return nil
+	}
+
+	if err := client.Generate(cancelCtx, &request, fn); err != nil {
+		if strings.Contains(err.Error(), "context canceled") && abort {
+			spinner.Finish()
 			return nil
 		}
-
-		if err := client.Generate(context.Background(), &request, fn); err != nil {
-			if strings.Contains(err.Error(), "failed to load model") {
-				// tell the user to check the server log, if it exists locally
-				home, nestedErr := os.UserHomeDir()
-				if nestedErr != nil {
-					// return the original error
-					return err
-				}
-				logPath := filepath.Join(home, ".ollama", "logs", "server.log")
-				if _, nestedErr := os.Stat(logPath); nestedErr == nil {
-					err = fmt.Errorf("%w\nFor more details, check the error logs at %s", err, logPath)
-				}
-			}
-			return err
-		}
-
-		fmt.Println()
-		fmt.Println()
-
-		if !latest.Done {
-			return errors.New("unexpected end of response")
-		}
-
-		verbose, err := cmd.Flags().GetBool("verbose")
-		if err != nil {
-			return err
-		}
-
-		if verbose {
-			latest.Summary()
-		}
-
-		ctx := cmd.Context()
-		ctx = context.WithValue(ctx, generateContextKey("context"), latest.Context)
-		cmd.SetContext(ctx)
+		return err
 	}
+	if prompt != "" {
+		fmt.Println()
+		fmt.Println()
+	}
+
+	if !latest.Done {
+		if abort {
+			return nil
+		}
+		return errors.New("unexpected end of response")
+	}
+
+	verbose, err := cmd.Flags().GetBool("verbose")
+	if err != nil {
+		return err
+	}
+
+	if verbose {
+		latest.Summary()
+	}
+
+	ctx := cmd.Context()
+	ctx = context.WithValue(ctx, generateContextKey("context"), latest.Context)
+	cmd.SetContext(ctx)
 
 	return nil
 }
@@ -461,19 +513,21 @@ func generateInteractive(cmd *cobra.Command, model string) error {
 		return err
 	}
 
+	// load the model
+	if err := generate(cmd, model, "", false); err != nil {
+		return err
+	}
+
 	completer := readline.NewPrefixCompleter(
 		readline.PcItem("/help"),
 		readline.PcItem("/list"),
 		readline.PcItem("/set",
 			readline.PcItem("history"),
 			readline.PcItem("nohistory"),
+			readline.PcItem("wordwrap"),
+			readline.PcItem("nowordwrap"),
 			readline.PcItem("verbose"),
 			readline.PcItem("quiet"),
-			readline.PcItem("mode",
-				readline.PcItem("vim"),
-				readline.PcItem("emacs"),
-				readline.PcItem("default"),
-			),
 		),
 		readline.PcItem("/show",
 			readline.PcItem("license"),
@@ -487,11 +541,41 @@ func generateInteractive(cmd *cobra.Command, model string) error {
 	)
 
 	usage := func() {
-		fmt.Fprintln(os.Stderr, "commands:")
-		fmt.Fprintln(os.Stderr, completer.Tree("  "))
+		fmt.Fprintln(os.Stderr, "Available Commands:")
+		fmt.Fprintln(os.Stderr, "  /set         Set session variables")
+		fmt.Fprintln(os.Stderr, "  /show        Show model information")
+		fmt.Fprintln(os.Stderr, "  /bye         Exit")
+		fmt.Fprintln(os.Stderr, "  /?, /help    Help for a command")
+		fmt.Fprintln(os.Stderr, "")
+		fmt.Fprintln(os.Stderr, "Use \"\"\" to begin a multi-line message.")
+		fmt.Fprintln(os.Stderr, "")
 	}
 
+	usageSet := func() {
+		fmt.Fprintln(os.Stderr, "Available Commands:")
+		fmt.Fprintln(os.Stderr, "  /set history      Enable history")
+		fmt.Fprintln(os.Stderr, "  /set nohistory    Disable history")
+		fmt.Fprintln(os.Stderr, "  /set wordwrap     Enable wordwrap")
+		fmt.Fprintln(os.Stderr, "  /set nowordwrap   Disable wordwrap")
+		fmt.Fprintln(os.Stderr, "  /set verbose      Show LLM stats")
+		fmt.Fprintln(os.Stderr, "  /set quiet        Disable LLM stats")
+		fmt.Fprintln(os.Stderr, "")
+	}
+
+	usageShow := func() {
+		fmt.Fprintln(os.Stderr, "Available Commands:")
+		fmt.Fprintln(os.Stderr, "  /show license      Show model license")
+		fmt.Fprintln(os.Stderr, "  /show modelfile    Show Modelfile for this model")
+		fmt.Fprintln(os.Stderr, "  /show parameters   Show parameters for this model")
+		fmt.Fprintln(os.Stderr, "  /show system       Show system prompt")
+		fmt.Fprintln(os.Stderr, "  /show template     Show prompt template")
+		fmt.Fprintln(os.Stderr, "")
+	}
+
+	var painter Painter
+
 	config := readline.Config{
+		Painter:      &painter,
 		Prompt:       ">>> ",
 		HistoryFile:  filepath.Join(home, ".ollama", "history"),
 		AutoComplete: completer,
@@ -503,6 +587,21 @@ func generateInteractive(cmd *cobra.Command, model string) error {
 	}
 	defer scanner.Close()
 
+	var wordWrap bool
+	termType := os.Getenv("TERM")
+	if termType == "xterm-256color" {
+		wordWrap = true
+	}
+
+	// override wrapping if the user turned it off
+	nowrap, err := cmd.Flags().GetBool("nowordwrap")
+	if err != nil {
+		return err
+	}
+	if nowrap {
+		wordWrap = false
+	}
+
 	var multiLineBuffer string
 	var isMultiLine bool
 
@@ -513,7 +612,7 @@ func generateInteractive(cmd *cobra.Command, model string) error {
 			return nil
 		case errors.Is(err, readline.ErrInterrupt):
 			if line == "" {
-				return nil
+				fmt.Println("Use Ctrl-D or /bye to exit.")
 			}
 
 			continue
@@ -527,6 +626,7 @@ func generateInteractive(cmd *cobra.Command, model string) error {
 		case isMultiLine:
 			if strings.HasSuffix(line, `"""`) {
 				isMultiLine = false
+				painter.IsMultiLine = isMultiLine
 				multiLineBuffer += strings.TrimSuffix(line, `"""`)
 				line = multiLineBuffer
 				multiLineBuffer = ""
@@ -537,6 +637,7 @@ func generateInteractive(cmd *cobra.Command, model string) error {
 			}
 		case strings.HasPrefix(line, `"""`):
 			isMultiLine = true
+			painter.IsMultiLine = isMultiLine
 			multiLineBuffer = strings.TrimPrefix(line, `"""`) + " "
 			scanner.SetPrompt("... ")
 			continue
@@ -545,84 +646,115 @@ func generateInteractive(cmd *cobra.Command, model string) error {
 			if err := ListHandler(cmd, args[1:]); err != nil {
 				return err
 			}
-
-			continue
 		case strings.HasPrefix(line, "/set"):
 			args := strings.Fields(line)
 			if len(args) > 1 {
 				switch args[1] {
 				case "history":
 					scanner.HistoryEnable()
-					continue
 				case "nohistory":
 					scanner.HistoryDisable()
-					continue
+				case "wordwrap":
+					wordWrap = true
+					fmt.Println("Set 'wordwrap' mode.")
+				case "nowordwrap":
+					wordWrap = false
+					fmt.Println("Set 'nowordwrap' mode.")
 				case "verbose":
 					cmd.Flags().Set("verbose", "true")
-					continue
+					fmt.Println("Set 'verbose' mode.")
 				case "quiet":
 					cmd.Flags().Set("verbose", "false")
-					continue
+					fmt.Println("Set 'quiet' mode.")
 				case "mode":
 					if len(args) > 2 {
 						switch args[2] {
 						case "vim":
 							scanner.SetVimMode(true)
-							continue
 						case "emacs", "default":
 							scanner.SetVimMode(false)
-							continue
 						default:
 							usage()
-							continue
 						}
 					} else {
 						usage()
-						continue
 					}
+				default:
+					fmt.Printf("Unknown command '/set %s'. Type /? for help\n", args[1])
 				}
 			} else {
-				usage()
-				continue
+				usageSet()
 			}
 		case strings.HasPrefix(line, "/show"):
 			args := strings.Fields(line)
 			if len(args) > 1 {
-				resp, err := server.GetModelInfo(model)
+				client, err := api.ClientFromEnvironment()
+				if err != nil {
+					fmt.Println("error: couldn't connect to ollama server")
+					return err
+				}
+				resp, err := client.Show(cmd.Context(), &api.ShowRequest{Name: model})
 				if err != nil {
 					fmt.Println("error: couldn't get model")
-					continue
+					return err
 				}
 
 				switch args[1] {
 				case "license":
-					fmt.Println(resp.License)
+					if resp.License == "" {
+						fmt.Print("No license was specified for this model.\n\n")
+					} else {
+						fmt.Println(resp.License)
+					}
 				case "modelfile":
 					fmt.Println(resp.Modelfile)
 				case "parameters":
-					fmt.Println(resp.Parameters)
+					if resp.Parameters == "" {
+						fmt.Print("No parameters were specified for this model.\n\n")
+					} else {
+						fmt.Println(resp.Parameters)
+					}
 				case "system":
-					fmt.Println(resp.System)
+					if resp.System == "" {
+						fmt.Print("No system prompt was specified for this model.\n\n")
+					} else {
+						fmt.Println(resp.System)
+					}
 				case "template":
-					fmt.Println(resp.Template)
+					if resp.Template == "" {
+						fmt.Print("No prompt template was specified for this model.\n\n")
+					} else {
+						fmt.Println(resp.Template)
+					}
 				default:
-					fmt.Println("error: unknown command")
+					fmt.Printf("Unknown command '/show %s'. Type /? for help\n", args[1])
+				}
+			} else {
+				usageShow()
+			}
+		case strings.HasPrefix(line, "/help"), strings.HasPrefix(line, "/?"):
+			args := strings.Fields(line)
+			if len(args) > 1 {
+				switch args[1] {
+				case "set", "/set":
+					usageSet()
+				case "show", "/show":
+					usageShow()
 				}
-
-				continue
 			} else {
 				usage()
-				continue
 			}
-		case line == "/help", line == "/?":
-			usage()
-			continue
 		case line == "/exit", line == "/bye":
 			return nil
+		case strings.HasPrefix(line, "/"):
+			args := strings.Fields(line)
+			fmt.Printf("Unknown command '%s'. Type /? for help\n", args[0])
 		}
 
-		if err := generate(cmd, model, line); err != nil {
-			return err
+		if len(line) > 0 && line[0] != '/' {
+			if err := generate(cmd, model, line, wordWrap); err != nil {
+				return err
+			}
 		}
 	}
 }
@@ -632,7 +764,7 @@ func generateBatch(cmd *cobra.Command, model string) error {
 	for scanner.Scan() {
 		prompt := scanner.Text()
 		fmt.Printf(">>> %s\n", prompt)
-		if err := generate(cmd, model, prompt); err != nil {
+		if err := generate(cmd, model, prompt, false); err != nil {
 			return err
 		}
 	}
@@ -641,28 +773,19 @@ func generateBatch(cmd *cobra.Command, model string) error {
 }
 
 func RunServer(cmd *cobra.Command, _ []string) error {
-	host, port := "127.0.0.1", "11434"
-
-	parts := strings.Split(os.Getenv("OLLAMA_HOST"), ":")
-	if ip := net.ParseIP(parts[0]); ip != nil {
-		host = ip.String()
-	}
-
-	if len(parts) > 1 {
-		port = parts[1]
-	}
-
-	// deprecated: include port in OLLAMA_HOST
-	if p := os.Getenv("OLLAMA_PORT"); p != "" {
-		port = p
-	}
-
-	err := initializeKeypair()
+	host, port, err := net.SplitHostPort(os.Getenv("OLLAMA_HOST"))
 	if err != nil {
+		host, port = "127.0.0.1", "11434"
+		if ip := net.ParseIP(strings.Trim(os.Getenv("OLLAMA_HOST"), "[]")); ip != nil {
+			host = ip.String()
+		}
+	}
+
+	if err := initializeKeypair(); err != nil {
 		return err
 	}
 
-	ln, err := net.Listen("tcp", fmt.Sprintf("%s:%s", host, port))
+	ln, err := net.Listen("tcp", net.JoinHostPort(host, port))
 	if err != nil {
 		return err
 	}
@@ -672,6 +795,21 @@ func RunServer(cmd *cobra.Command, _ []string) error {
 		origins = strings.Split(o, ",")
 	}
 
+	if noprune := os.Getenv("OLLAMA_NOPRUNE"); noprune == "" {
+		if err := server.PruneLayers(); err != nil {
+			return err
+		}
+
+		manifestsPath, err := server.GetManifestPath()
+		if err != nil {
+			return err
+		}
+
+		if err := server.PruneDirectory(manifestsPath); err != nil {
+			return err
+		}
+	}
+
 	return server.Serve(ln, origins)
 }
 
@@ -697,7 +835,7 @@ func initializeKeypair() error {
 			return err
 		}
 
-		err = os.MkdirAll(path.Dir(privKeyPath), 0o700)
+		err = os.MkdirAll(filepath.Dir(privKeyPath), 0o755)
 		if err != nil {
 			return fmt.Errorf("could not create directory %w", err)
 		}
@@ -756,7 +894,7 @@ func startMacApp(client *api.Client) error {
 }
 
 func checkServerHeartbeat(_ *cobra.Command, _ []string) error {
-	client, err := api.FromEnv()
+	client, err := api.ClientFromEnvironment()
 	if err != nil {
 		return err
 	}
@@ -794,7 +932,7 @@ func NewCLI() *cobra.Command {
 	createCmd := &cobra.Command{
 		Use:     "create MODEL",
 		Short:   "Create a model from a Modelfile",
-		Args:    cobra.MinimumNArgs(1),
+		Args:    cobra.ExactArgs(1),
 		PreRunE: checkServerHeartbeat,
 		RunE:    CreateHandler,
 	}
@@ -804,7 +942,7 @@ func NewCLI() *cobra.Command {
 	showCmd := &cobra.Command{
 		Use:     "show MODEL",
 		Short:   "Show information for a model",
-		Args:    cobra.MinimumNArgs(1),
+		Args:    cobra.ExactArgs(1),
 		PreRunE: checkServerHeartbeat,
 		RunE:    ShowHandler,
 	}
@@ -825,18 +963,20 @@ func NewCLI() *cobra.Command {
 
 	runCmd.Flags().Bool("verbose", false, "Show timings for response")
 	runCmd.Flags().Bool("insecure", false, "Use an insecure registry")
+	runCmd.Flags().Bool("nowordwrap", false, "Don't wrap words to the next line automatically")
 
 	serveCmd := &cobra.Command{
 		Use:     "serve",
 		Aliases: []string{"start"},
 		Short:   "Start ollama",
+		Args:    cobra.ExactArgs(0),
 		RunE:    RunServer,
 	}
 
 	pullCmd := &cobra.Command{
 		Use:     "pull MODEL",
 		Short:   "Pull a model from a registry",
-		Args:    cobra.MinimumNArgs(1),
+		Args:    cobra.ExactArgs(1),
 		PreRunE: checkServerHeartbeat,
 		RunE:    PullHandler,
 	}
@@ -846,7 +986,7 @@ func NewCLI() *cobra.Command {
 	pushCmd := &cobra.Command{
 		Use:     "push MODEL",
 		Short:   "Push a model to a registry",
-		Args:    cobra.MinimumNArgs(1),
+		Args:    cobra.ExactArgs(1),
 		PreRunE: checkServerHeartbeat,
 		RunE:    PushHandler,
 	}
@@ -862,15 +1002,15 @@ func NewCLI() *cobra.Command {
 	}
 
 	copyCmd := &cobra.Command{
-		Use:     "cp",
+		Use:     "cp SOURCE TARGET",
 		Short:   "Copy a model",
-		Args:    cobra.MinimumNArgs(2),
+		Args:    cobra.ExactArgs(2),
 		PreRunE: checkServerHeartbeat,
 		RunE:    CopyHandler,
 	}
 
 	deleteCmd := &cobra.Command{
-		Use:     "rm",
+		Use:     "rm MODEL [MODEL...]",
 		Short:   "Remove a model",
 		Args:    cobra.MinimumNArgs(1),
 		PreRunE: checkServerHeartbeat,

docs/api.md

@@ -3,26 +3,32 @@
 ## Endpoints
 
 - [Generate a completion](#generate-a-completion)
-- [Create a model](#create-a-model)
-- [List local models](#list-local-models)
-- [Copy a model](#copy-a-model)
-- [Delete a model](#delete-a-model)
-- [Pull a model](#pull-a-model)
-- [Generate embeddings](#generate-embeddings)
+- [Create a Model](#create-a-model)
+- [List Local Models](#list-local-models)
+- [Show Model Information](#show-model-information)
+- [Copy a Model](#copy-a-model)
+- [Delete a Model](#delete-a-model)
+- [Pull a Model](#pull-a-model)
+- [Push a Model](#push-a-model)
+- [Generate Embeddings](#generate-embeddings)
 
 ## Conventions
 
 ### Model names
 
-Model names follow a `model:tag` format. Some examples are `orca-mini:3b-q4_1` and `llama2:70b`. The tag is optional and if not provided will default to `latest`. The tag is used to identify a specific version.
+Model names follow a `model:tag` format. Some examples are `orca-mini:3b-q4_1` and `llama2:70b`. The tag is optional and, if not provided, will default to `latest`. The tag is used to identify a specific version.
 
 ### Durations
 
 All durations are returned in nanoseconds.
 
+### Streaming responses
+
+Certain endpoints stream responses as JSON objects delineated with the newline (`\n`) character.
+
 ## Generate a completion
 
-```
+```shell
 POST /api/generate
 ```
 
@@ -33,16 +39,17 @@ Generate a response for a given prompt with a provided model. This is a streamin
 - `model`: (required) the [model name](#model-names)
 - `prompt`: the prompt to generate a response for
 
-Advanced parameters:
+Advanced parameters (optional):
 
 - `options`: additional model parameters listed in the documentation for the [Modelfile](./modelfile.md#valid-parameters-and-values) such as `temperature`
 - `system`: system prompt to (overrides what is defined in the `Modelfile`)
 - `template`: the full prompt or prompt template (overrides what is defined in the `Modelfile`)
 - `context`: the context parameter returned from a previous request to `/generate`, this can be used to keep a short conversational memory
+- `stream`: if `false` the response will be be returned as a single response object, rather than a stream of objects
 
 ### Request
 
-```
+```shell
 curl -X POST http://localhost:11434/api/generate -d '{
   "model": "llama2:7b",
   "prompt": "Why is the sky blue?"
@@ -73,6 +80,7 @@ The final response in the stream also includes additional data about the generat
 - `eval_count`: number of tokens the response
 - `eval_duration`: time in nanoseconds spent generating the response
 - `context`: an encoding of the conversation used in this response, this can be sent in the next request to keep a conversational memory
+- `response`: empty if the response was streamed, if not streamed, this will contain the full response
 
 To calculate how fast the response is generated in tokens per second (token/s), divide `eval_count` / `eval_duration`.
 
@@ -80,6 +88,7 @@ To calculate how fast the response is generated in tokens per second (token/s),
 {
   "model": "llama2:7b",
   "created_at": "2023-08-04T19:22:45.499127Z",
+  "response": "",
   "context": [1, 2, 3],
   "done": true,
   "total_duration": 5589157167,
@@ -95,7 +104,7 @@ To calculate how fast the response is generated in tokens per second (token/s),
 
 ## Create a Model
 
-```
+```shell
 POST /api/create
 ```
 
@@ -105,10 +114,11 @@ Create a model from a [`Modelfile`](./modelfile.md)
 
 - `name`: name of the model to create
 - `path`: path to the Modelfile
+- `stream`: (optional) if `false` the response will be be returned as a single response object, rather than a stream of objects
 
 ### Request
 
-```
+```shell
 curl -X POST http://localhost:11434/api/create -d '{
   "name": "mario",
   "path": "~/Modelfile"
@@ -117,7 +127,7 @@ curl -X POST http://localhost:11434/api/create -d '{
 
 ### Response
 
-A stream of JSON objects. When finished, `status` is `success`
+A stream of JSON objects. When finished, `status` is `success`.
 
 ```json
 {
@@ -127,7 +137,7 @@ A stream of JSON objects. When finished, `status` is `success`
 
 ## List Local Models
 
-```
+```shell
 GET /api/tags
 ```
 
@@ -135,7 +145,7 @@ List models that are available locally.
 
 ### Request
 
-```
+```shell
 curl http://localhost:11434/api/tags
 ```
 
@@ -158,9 +168,40 @@ curl http://localhost:11434/api/tags
 }
 ```
 
+## Show Model Information
+
+```shell
+POST /api/show
+```
+
+Show details about a model including modelfile, template, parameters, license, and system prompt.
+
+### Parameters
+
+- `name`: name of the model to show
+
+### Request
+
+```shell
+curl http://localhost:11434/api/show -d '{
+  "name": "llama2:7b"
+}'
+```
+
+### Response
+
+```json
+{
+  "license": "<contents of license block>",
+  "modelfile": "# Modelfile generated by \"ollama show\"\n# To build a new Modelfile based on this one, replace the FROM line with:\n# FROM llama2:latest\n\nFROM /Users/username/.ollama/models/blobs/sha256:8daa9615cce30c259a9555b1cc250d461d1bc69980a274b44d7eda0be78076d8\nTEMPLATE \"\"\"[INST] {{ if and .First .System }}<<SYS>>{{ .System }}<</SYS>>\n\n{{ end }}{{ .Prompt }} [/INST] \"\"\"\nSYSTEM \"\"\"\"\"\"\nPARAMETER stop [INST]\nPARAMETER stop [/INST]\nPARAMETER stop <<SYS>>\nPARAMETER stop <</SYS>>\n",
+  "parameters": "stop                           [INST]\nstop                           [/INST]\nstop                           <<SYS>>\nstop                           <</SYS>>",
+  "template": "[INST] {{ if and .First .System }}<<SYS>>{{ .System }}<</SYS>>\n\n{{ end }}{{ .Prompt }} [/INST] "
+}
+```
+
 ## Copy a Model
 
-```
+```shell
 POST /api/copy
 ```
 
@@ -168,7 +209,7 @@ Copy a model. Creates a model with another name from an existing model.
 
 ### Request
 
-```
+```shell
 curl http://localhost:11434/api/copy -d '{
   "source": "llama2:7b",
   "destination": "llama2-backup"
@@ -177,7 +218,7 @@ curl http://localhost:11434/api/copy -d '{
 
 ## Delete a Model
 
-```
+```shell
 DELETE /api/delete
 ```
 
@@ -189,7 +230,7 @@ Delete a model and its data.
 
 ### Request
 
-```
+```shell
 curl -X DELETE http://localhost:11434/api/delete -d '{
   "name": "llama2:13b"
 }'
@@ -197,19 +238,21 @@ curl -X DELETE http://localhost:11434/api/delete -d '{
 
 ## Pull a Model
 
-```
+```shell
 POST /api/pull
 ```
 
-Download a model from a the model registry. Cancelled pulls are resumed from where they left off, and multiple calls to will share the same download progress.
+Download a model from the ollama library. Cancelled pulls are resumed from where they left off, and multiple calls will share the same download progress.
 
 ### Parameters
 
 - `name`: name of the model to pull
+- `insecure`: (optional) allow insecure connections to the library. Only use this if you are pulling from your own library during development.
+- `stream`: (optional) if `false` the response will be be returned as a single response object, rather than a stream of objects
 
 ### Request
 
-```
+```shell
 curl -X POST http://localhost:11434/api/pull -d '{
   "name": "llama2:7b"
 }'
@@ -225,9 +268,66 @@ curl -X POST http://localhost:11434/api/pull -d '{
 }
 ```
 
+## Push a Model
+
+```shell
+POST /api/push
+```
+
+Upload a model to a model library. Requires registering for ollama.ai and adding a public key first.
+
+### Parameters
+
+- `name`: name of the model to push in the form of `<namespace>/<model>:<tag>`
+- `insecure`: (optional) allow insecure connections to the library. Only use this if you are pushing to your library during development.
+- `stream`: (optional) if `false` the response will be be returned as a single response object, rather than a stream of objects
+
+### Request
+
+```shell
+curl -X POST http://localhost:11434/api/push -d '{
+  "name": "mattw/pygmalion:latest"
+}'
+```
+
+### Response
+
+Streaming response that starts with:
+
+```json
+{ "status": "retrieving manifest" }
+```
+
+and then:
+
+```json
+{
+  "status": "starting upload",
+  "digest": "sha256:bc07c81de745696fdf5afca05e065818a8149fb0c77266fb584d9b2cba3711ab",
+  "total": 1928429856
+}
+```
+
+Then there is a series of uploading responses:
+
+```json
+{
+  "status": "starting upload",
+  "digest": "sha256:bc07c81de745696fdf5afca05e065818a8149fb0c77266fb584d9b2cba3711ab",
+  "total": 1928429856
+}
+```
+
+Finally, when the upload is complete:
+
+```json
+{"status":"pushing manifest"}
+{"status":"success"}
+```
+
 ## Generate Embeddings
 
-```
+```shell
 POST /api/embeddings
 ```
 
@@ -244,7 +344,7 @@ Advanced parameters:
 
 ### Request
 
-```
+```shell
 curl -X POST http://localhost:11434/api/embeddings -d '{
   "model": "llama2:7b",
   "prompt": "Here is an article about llamas..."
@@ -255,7 +355,7 @@ curl -X POST http://localhost:11434/api/embeddings -d '{
 
 ```json
 {
-  "embeddings": [
+  "embedding": [
     0.5670403838157654, 0.009260174818336964, 0.23178744316101074, -0.2916173040866852, -0.8924556970596313,
     0.8785552978515625, -0.34576427936553955, 0.5742510557174683, -0.04222835972905159, -0.137906014919281
   ]

docs/development.md

@@ -6,24 +6,34 @@
 
 Install required tools:
 
-```
+- cmake version 3.24 or higher
+- go version 1.20 or higher
+- gcc version 11.4.0 or higher
+
+```bash
 brew install go cmake gcc
 ```
 
 Get the required libraries:
 
-```
+```bash
 go generate ./...
 ```
 
 Then build ollama:
 
-```
+```bash
 go build .
 ```
 
 Now you can run `ollama`:
 
-```
+```bash
 ./ollama
 ```
+
+## Building on Linux with GPU support
+
+- Install cmake and nvidia-cuda-toolkit
+- run `go generate ./...`
+- run `go build .`

docs/faq.md

@@ -1,17 +1,34 @@
 # FAQ
 
+## How can I view the logs?
+
+On macOS:
+
+```
+cat ~/.ollama/logs/server.log
+```
+
+On Linux:
+
+```
+journalctl -u ollama
+```
+
+If you're running `ollama serve` directly, the logs will be printed to the console.
+
 ## How can I expose the Ollama server?
 
-```
+```bash
 OLLAMA_HOST=0.0.0.0:11435 ollama serve
 ```
 
 By default, Ollama allows cross origin requests from `127.0.0.1` and `0.0.0.0`. To support more origins, you can use the `OLLAMA_ORIGINS` environment variable:
 
-```
+```bash
 OLLAMA_ORIGINS=http://192.168.1.1:*,https://example.com ollama serve
 ```
 
 ## Where are models stored?
 
-Raw model data is stored under `~/.ollama/models`.
+- macOS: Raw model data is stored under `~/.ollama/models`.
+- Linux: Raw model data is stored under `/usr/share/ollama/.ollama/models`

docs/import.md

@@ -0,0 +1,163 @@
+# Import a model
+
+This guide walks through importing a PyTorch, Safetensors or GGUF model.
+
+## Supported models
+
+Ollama supports a set of model architectures, with support for more coming soon:
+
+- Llama & Mistral
+- Falcon & RW
+- GPT-NeoX
+- BigCode
+
+To view a model's architecture, check the `config.json` file in its HuggingFace repo. You should see an entry under `architectures` (e.g. `LlamaForCausalLM`).
+
+## Importing
+
+### Step 1: Clone the HuggingFace repository (optional)
+
+If the model is currently hosted in a HuggingFace repository, first clone that repository to download the raw model.
+
+```
+git lfs install
+git clone https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.1
+cd Mistral-7B-Instruct-v0.1
+```
+
+### Step 2: Convert and quantize to a `.bin` file (optional, for PyTorch and Safetensors)
+
+If the model is in PyTorch or Safetensors format, a [Docker image](https://hub.docker.com/r/ollama/quantize) with the tooling required to convert and quantize models is available.
+
+First, Install [Docker](https://www.docker.com/get-started/).
+
+Next, to convert and quantize your model, run:
+
+```
+docker run --rm -v .:/model ollama/quantize -q q4_0 /model
+```
+
+This will output two files into the directory:
+
+- `f16.bin`: the model converted to GGUF
+- `q4_0.bin` the model quantized to a 4-bit quantization (we will use this file to create the Ollama model)
+
+### Step 3: Write a `Modelfile`
+
+Next, create a `Modelfile` for your model. This file is the blueprint for your model, specifying weights, parameters, prompt templates and more.
+
+```
+FROM ./q4_0.bin
+```
+
+(Optional) many chat models require a prompt template in order to answer correctly. A default prompt template can be specified with the `TEMPLATE` instruction in the `Modelfile`:
+
+```
+FROM ./q4_0.bin
+TEMPLATE "[INST] {{ .Prompt }} [/INST]"
+```
+
+### Step 4: Create the Ollama model
+
+Finally, create a model from your `Modelfile`:
+
+```
+ollama create example -f Modelfile
+```
+
+Next, test the model with `ollama run`:
+
+```
+ollama run example "What is your favourite condiment?"
+```
+
+### Step 5: Publish your model (optional – early alpha)
+
+Publishing models is in early alpha. If you'd like to publish your model to share with others, follow these steps:
+
+1. Create [an account](https://ollama.ai/signup)
+2. Run `cat ~/.ollama/id_ed25519.pub` to view your Ollama public key. Copy this to the clipboard.
+3. Add your public key to your [Ollama account](https://ollama.ai/settings/keys)
+
+Next, copy your model to your username's namespace:
+
+```
+ollama cp example <your username>/example
+```
+
+Then push the model:
+
+```
+ollama push <your username>/example
+```
+
+After publishing, your model will be available at `https://ollama.ai/<your username>/example`.
+
+## Quantization reference
+
+The quantization options are as follow (from highest highest to lowest levels of quantization). Note: some architectures such as Falcon do not support K quants.
+
+- `q2_K`
+- `q3_K`
+- `q3_K_S`
+- `q3_K_M`
+- `q3_K_L`
+- `q4_0` (recommended)
+- `q4_1`
+- `q4_K`
+- `q4_K_S`
+- `q4_K_M`
+- `q5_0`
+- `q5_1`
+- `q5_K`
+- `q5_K_S`
+- `q5_K_M`
+- `q6_K`
+- `q8_0`
+
+## Manually converting & quantizing models
+
+### Prerequisites
+
+Start by cloning the `llama.cpp` repo to your machine in another directory:
+
+```
+git clone https://github.com/ggerganov/llama.cpp.git
+cd llama.cpp
+```
+
+Next, install the Python dependencies:
+
+```
+pip install -r requirements.txt
+```
+
+Finally, build the `quantize` tool:
+
+```
+make quantize
+```
+
+### Convert the model
+
+Run the correct conversion script for your model architecture:
+
+```shell
+# LlamaForCausalLM or MistralForCausalLM
+python convert.py <path to model directory>
+
+# FalconForCausalLM
+python convert-falcon-hf-to-gguf.py <path to model directory>
+
+# GPTNeoXForCausalLM
+python convert-falcon-hf-to-gguf.py <path to model directory>
+
+# GPTBigCodeForCausalLM
+python convert-starcoder-hf-to-gguf.py <path to model directory>
+```
+
+### Quantize the model
+
+```
+quantize <path to model dir>/ggml-model-f32.bin <path to model dir>/q4_0.bin q4_0
+```

docs/linux.md

@@ -0,0 +1,82 @@
+# Installing Ollama on Linux
+
+> Note: A one line installer for Ollama is available by running:
+>
+> ```bash
+> curl https://ollama.ai/install.sh | sh
+> ```
+
+## Download the `ollama` binary
+
+Ollama is distributed as a self-contained binary. Download it to a directory in your PATH:
+
+```bash
+sudo curl -L https://ollama.ai/download/ollama-linux-amd64 -o /usr/bin/ollama
+sudo chmod +x /usr/bin/ollama
+```
+
+## Start Ollama
+
+Start Ollama by running `ollama serve`:
+
+```bash
+ollama serve
+```
+
+Once Ollama is running, run a model in another terminal session:
+
+```bash
+ollama run llama2
+```
+
+## Install CUDA drivers (optional – for Nvidia GPUs)
+
+[Download and install](https://developer.nvidia.com/cuda-downloads) CUDA.
+
+Verify that the drivers are installed by running the following command, which should print details about your GPU:
+
+```bash
+nvidia-smi
+```
+
+## Adding Ollama as a startup service (optional)
+
+Create a user for Ollama:
+
+```bash
+sudo useradd -r -s /bin/false -m -d /usr/share/ollama ollama
+```
+
+Create a service file in `/etc/systemd/system/ollama.service`:
+
+```ini
+[Unit]
+Description=Ollama Service
+After=network-online.target
+
+[Service]
+ExecStart=/usr/bin/ollama serve
+User=ollama
+Group=ollama
+Restart=always
+RestartSec=3
+Environment="HOME=/usr/share/ollama"
+
+[Install]
+WantedBy=default.target
+```
+
+Then start the service:
+
+```bash
+sudo systemctl daemon-reload
+sudo systemctl enable ollama
+```
+
+### Viewing logs
+
+To view logs of Ollama running as a startup service, run:
+
+```bash
+journalctl -u ollama
+```

docs/modelfile.md

@@ -1,6 +1,6 @@
 # Ollama Model File
 
-> Note: this model file syntax is in development
+> Note: this `Modelfile` syntax is in development
 
 A model file is the blueprint to create and share models with Ollama.
 
@@ -12,7 +12,6 @@ A model file is the blueprint to create and share models with Ollama.
   - [FROM (Required)](#from-required)
     - [Build from llama2](#build-from-llama2)
     - [Build from a bin file](#build-from-a-bin-file)
-  - [EMBED](#embed)
   - [PARAMETER](#parameter)
     - [Valid Parameters and Values](#valid-parameters-and-values)
   - [TEMPLATE](#template)
@@ -24,7 +23,7 @@ A model file is the blueprint to create and share models with Ollama.
 
 ## Format
 
-The format of the Modelfile:
+The format of the `Modelfile`:
 
 ```modelfile
 # comment
@@ -42,9 +41,9 @@ INSTRUCTION arguments
 
 ## Examples
 
-An example of a model file creating a mario blueprint:
+An example of a `Modelfile` creating a mario blueprint:
 
-```
+```modelfile
 FROM llama2
 # sets the temperature to 1 [higher is more creative, lower is more coherent]
 PARAMETER temperature 1
@@ -57,9 +56,9 @@ SYSTEM You are Mario from super mario bros, acting as an assistant.
 
 To use this:
 
-1. Save it as a file (eg. `Modelfile`)
-2. `ollama create NAME -f <location of the file eg. ./Modelfile>'`
-3. `ollama run NAME`
+1. Save it as a file (e.g. `Modelfile`)
+2. `ollama create choose-a-model-name -f <location of the file e.g. ./Modelfile>'`
+3. `ollama run choose-a-model-name`
 4. Start using the model!
 
 More examples are available in the [examples directory](../examples).
@@ -68,44 +67,34 @@ More examples are available in the [examples directory](../examples).
 
 ### FROM (Required)
 
-The FROM instruction defines the base model to use when creating a model.
+The `FROM` instruction defines the base model to use when creating a model.
 
-```
+```modelfile
 FROM <model name>:<tag>
 ```
 
 #### Build from llama2
 
-```
+```modelfile
 FROM llama2
 ```
 
 A list of available base models:
 <https://github.com/jmorganca/ollama#model-library>
 
-#### Build from a bin file
+#### Build from a `bin` file
 
-```
+```modelfile
 FROM ./ollama-model.bin
 ```
 
-This bin file location should be specified as an absolute path or relative to the Modelfile location.
-
-### EMBED
-
-The EMBED instruction is used to add embeddings of files to a model. This is useful for adding custom data that the model can reference when generating an answer. Note that currently only text files are supported, formatted with each line as one embedding.
-```
-FROM <model name>:<tag>
-EMBED <file path>.txt
-EMBED <different file path>.txt
-EMBED <path to directory>/*.txt
-```
+This bin file location should be specified as an absolute path or relative to the `Modelfile` location.
 
 ### PARAMETER
 
 The `PARAMETER` instruction defines a parameter that can be set when the model is run.
 
-```
+```modelfile
 PARAMETER <parameter> <parametervalue>
 ```
 
@@ -118,19 +107,21 @@ PARAMETER <parameter> <parametervalue>
 | mirostat_tau   | Controls the balance between coherence and diversity of the output. A lower value will result in more focused and coherent text. (Default: 5.0)                                                                                                         | float      | mirostat_tau 5.0     |
 | num_ctx        | Sets the size of the context window used to generate the next token. (Default: 2048)                                                                                                                                                                    | int        | num_ctx 4096         |
 | num_gqa        | The number of GQA groups in the transformer layer. Required for some models, for example it is 8 for llama2:70b                                                                                                                                         | int        | num_gqa 1            |
-| num_gpu        | The number of GPUs to use. On macOS it defaults to 1 to enable metal support, 0 to disable.                                                                                                                                                             | int        | num_gpu 1            |
+| num_gpu        | The number of layers to send to the GPU(s). On macOS it defaults to 1 to enable metal support, 0 to disable.                                                                                                                                            | int        | num_gpu 50           |
 | num_thread     | Sets the number of threads to use during computation. By default, Ollama will detect this for optimal performance. It is recommended to set this value to the number of physical CPU cores your system has (as opposed to the logical number of cores). | int        | num_thread 8         |
 | repeat_last_n  | Sets how far back for the model to look back to prevent repetition. (Default: 64, 0 = disabled, -1 = num_ctx)                                                                                                                                           | int        | repeat_last_n 64     |
 | repeat_penalty | Sets how strongly to penalize repetitions. A higher value (e.g., 1.5) will penalize repetitions more strongly, while a lower value (e.g., 0.9) will be more lenient. (Default: 1.1)                                                                     | float      | repeat_penalty 1.1   |
 | temperature    | The temperature of the model. Increasing the temperature will make the model answer more creatively. (Default: 0.8)                                                                                                                                     | float      | temperature 0.7      |
+| seed | Sets the random number seed to use for generation. Setting this to a specific number will make the model generate the same text for the same prompt. (Default: 0) | int | seed 42 |
 | stop           | Sets the stop sequences to use.                                                                                                                                                                                                                         | string     | stop "AI assistant:" |
 | tfs_z          | Tail free sampling is used to reduce the impact of less probable tokens from the output. A higher value (e.g., 2.0) will reduce the impact more, while a value of 1.0 disables this setting. (default: 1)                                               | float      | tfs_z 1              |
+| num_predict    | Maximum number of tokens to predict when generating text. (Default: 128, -1 = infinite generation, -2 = fill context)                                                                                                                                   | int        | num_predict 42       |
 | top_k          | Reduces the probability of generating nonsense. A higher value (e.g. 100) will give more diverse answers, while a lower value (e.g. 10) will be more conservative. (Default: 40)                                                                        | int        | top_k 40             |
 | top_p          | Works together with top-k. A higher value (e.g., 0.95) will lead to more diverse text, while a lower value (e.g., 0.5) will generate more focused and conservative text. (Default: 0.9)                                                                 | float      | top_p 0.9            |
 
 ### TEMPLATE
 
-`TEMPLATE` of the full prompt template to be passed into the model. It may include (optionally) a system prompt and a user's prompt. This is used to create a full custom prompt, and syntax may be model specific.
+`TEMPLATE` of the full prompt template to be passed into the model. It may include (optionally) a system prompt and a user's prompt. This is used to create a full custom prompt, and syntax may be model specific. You can usually find the template for a given model in the readme for that model.
 
 #### Template Variables
 
@@ -140,7 +131,7 @@ PARAMETER <parameter> <parametervalue>
 | `{{ .Prompt }}` | The incoming prompt, this is not specified in the model file and will be set based on input.                 |
 | `{{ .First }}`  | A boolean value used to render specific template information for the first generation of a session.          |
 
-```
+```modelfile
 TEMPLATE """
 {{- if .First }}
 ### System:
@@ -160,7 +151,7 @@ SYSTEM """<system message>"""
 
 The `SYSTEM` instruction specifies the system prompt to be used in the template, if applicable.
 
-```
+```modelfile
 SYSTEM """<system message>"""
 ```
 
@@ -168,7 +159,7 @@ SYSTEM """<system message>"""
 
 The `ADAPTER` instruction specifies the LoRA adapter to apply to the base model. The value of this instruction should be an absolute path or a path relative to the Modelfile and the file must be in a GGML file format. The adapter should be tuned from the base model otherwise the behaviour is undefined.
 
-```
+```modelfile
 ADAPTER ./ollama-lora.bin
 ```
 
@@ -176,7 +167,7 @@ ADAPTER ./ollama-lora.bin
 
 The `LICENSE` instruction allows you to specify the legal license under which the model used with this Modelfile is shared or distributed.
 
-```
+```modelfile
 LICENSE """
 <license text>
 """
@@ -184,5 +175,5 @@ LICENSE """
 
 ## Notes
 
-- the **modelfile is not case sensitive**. In the examples, we use uppercase for instructions to make it easier to distinguish it from arguments.
+- the **`Modelfile` is not case sensitive**. In the examples, we use uppercase for instructions to make it easier to distinguish it from arguments.
 - Instructions can be in any order. In the examples, we start with FROM instruction to keep it easily readable.

examples/.gitignore

@@ -0,0 +1,171 @@
+node_modules
+# OSX
+.DS_STORE
+
+# Models
+models/
+
+# Local Chroma db
+.chroma/
+db/
+
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+*$py.class
+
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# PyInstaller
+#  Usually these files are written by a python script from a template
+#  before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.nox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+*.py,cover
+.hypothesis/
+.pytest_cache/
+cover/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+db.sqlite3-journal
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+.pybuilder/
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# IPython
+profile_default/
+ipython_config.py
+
+# pyenv
+#   For a library or package, you might want to ignore these files since the code is
+#   intended to run in multiple environments; otherwise, check them in:
+# .python-version
+
+# pipenv
+#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
+#   However, in case of collaboration, if having platform-specific dependencies or dependencies
+#   having no cross-platform support, pipenv may install dependencies that don't work, or not
+#   install all needed dependencies.
+#Pipfile.lock
+
+# poetry
+#   Similar to Pipfile.lock, it is generally recommended to include poetry.lock in version control.
+#   This is especially recommended for binary packages to ensure reproducibility, and is more
+#   commonly ignored for libraries.
+#   https://python-poetry.org/docs/basic-usage/#commit-your-poetrylock-file-to-version-control
+#poetry.lock
+
+# pdm
+#   Similar to Pipfile.lock, it is generally recommended to include pdm.lock in version control.
+#pdm.lock
+#   pdm stores project-wide configurations in .pdm.toml, but it is recommended to not include it
+#   in version control.
+#   https://pdm.fming.dev/#use-with-ide
+.pdm.toml
+
+# PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm
+__pypackages__/
+
+# Celery stuff
+celerybeat-schedule
+celerybeat.pid
+
+# SageMath parsed files
+*.sage.py
+
+# Environments
+.env
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# Pyre type checker
+.pyre/
+
+# pytype static type analyzer
+.pytype/
+
+# Cython debug symbols
+cython_debug/
+
+# PyCharm
+#  JetBrains specific template is maintained in a separate JetBrains.gitignore that can
+#  be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore
+#  and can be added to the global gitignore or merged into this file.  For a more nuclear
+#  option (not recommended) you can uncomment the following to ignore the entire idea folder.
+#.idea/

examples/README.md

@@ -1,15 +1,3 @@
 # Examples
 
-This directory contains different examples of using Ollama
-
-To create a model:
-
-```
-ollama create example -f <example file>
-```
-
-To run a model:
-
-```
-ollama run example
-```
+This directory contains different examples of using Ollama.

examples/golang-simplegenerate/main.go

@@ -0,0 +1,27 @@
+package main
+
+import (
+	"bytes"
+	"fmt"
+	"io"
+	"log"
+	"net/http"
+	"os"
+)
+
+func main() {
+	body := []byte(`{"model":"mistral"}`)
+	resp, err := http.Post("http://localhost:11434/api/generate", "application/json", bytes.NewBuffer(body))
+
+	if err != nil {
+		fmt.Print(err.Error())
+		os.Exit(1)
+	}
+
+	responseData, err := io.ReadAll(resp.Body)
+	if err != nil {
+		log.Fatal(err)
+	}
+	fmt.Println(string(responseData))
+
+}

examples/langchain-typescript-simple/README.md

@@ -0,0 +1,21 @@
+# LangChain
+
+This example is a basic "hello world" of using LangChain with Ollama using Node.js and Typescript.
+
+## Setup
+
+```shell
+npm install
+```
+
+## Run
+
+```shell
+ts-node main.ts
+```
+
+Running this example will print the response for "hello":
+
+```plaintext
+Hello! It's nice to meet you. hopefully you are having a great day! Is there something I can help you with or would you like to chat?
+```

examples/langchain-typescript-simple/main.ts

@@ -0,0 +1,15 @@
+import { Ollama} from 'langchain/llms/ollama';
+
+async function main() {
+  const ollama = new Ollama({
+    model: 'mistral'    
+    // other parameters can be found at https://js.langchain.com/docs/api/llms_ollama/classes/Ollama
+  })
+  const stream = await ollama.stream("Hello");
+
+  for await (const chunk of stream) {
+    process.stdout.write(chunk);
+  }
+}
+
+main();

examples/langchain-typescript-simple/package-lock.json

@@ -0,0 +1,997 @@
+{
+  "name": "with-langchain-typescript-simplegenerate",
+  "lockfileVersion": 3,
+  "requires": true,
+  "packages": {
+    "": {
+      "dependencies": {
+        "langchain": "^0.0.165"
+      },
+      "devDependencies": {
+        "typescript": "^5.2.2"
+      }
+    },
+    "node_modules/@anthropic-ai/sdk": {
+      "version": "0.6.2",
+      "resolved": "https://registry.npmjs.org/@anthropic-ai/sdk/-/sdk-0.6.2.tgz",
+      "integrity": "sha512-fB9PUj9RFT+XjkL+E9Ol864ZIJi+1P8WnbHspN3N3/GK2uSzjd0cbVIKTGgf4v3N8MwaQu+UWnU7C4BG/fap/g==",
+      "dependencies": {
+        "@types/node": "^18.11.18",
+        "@types/node-fetch": "^2.6.4",
+        "abort-controller": "^3.0.0",
+        "agentkeepalive": "^4.2.1",
+        "digest-fetch": "^1.3.0",
+        "form-data-encoder": "1.7.2",
+        "formdata-node": "^4.3.2",
+        "node-fetch": "^2.6.7"
+      }
+    },
+    "node_modules/@types/node": {
+      "version": "18.18.4",
+      "resolved": "https://registry.npmjs.org/@types/node/-/node-18.18.4.tgz",
+      "integrity": "sha512-t3rNFBgJRugIhackit2mVcLfF6IRc0JE4oeizPQL8Zrm8n2WY/0wOdpOPhdtG0V9Q2TlW/axbF1MJ6z+Yj/kKQ=="
+    },
+    "node_modules/@types/node-fetch": {
+      "version": "2.6.6",
+      "resolved": "https://registry.npmjs.org/@types/node-fetch/-/node-fetch-2.6.6.tgz",
+      "integrity": "sha512-95X8guJYhfqiuVVhRFxVQcf4hW/2bCuoPwDasMf/531STFoNoWTT7YDnWdXHEZKqAGUigmpG31r2FE70LwnzJw==",
+      "dependencies": {
+        "@types/node": "*",
+        "form-data": "^4.0.0"
+      }
+    },
+    "node_modules/@types/retry": {
+      "version": "0.12.0",
+      "resolved": "https://registry.npmjs.org/@types/retry/-/retry-0.12.0.tgz",
+      "integrity": "sha512-wWKOClTTiizcZhXnPY4wikVAwmdYHp8q6DmC+EJUzAMsycb7HB32Kh9RN4+0gExjmPmZSAQjgURXIGATPegAvA=="
+    },
+    "node_modules/@types/uuid": {
+      "version": "9.0.5",
+      "resolved": "https://registry.npmjs.org/@types/uuid/-/uuid-9.0.5.tgz",
+      "integrity": "sha512-xfHdwa1FMJ082prjSJpoEI57GZITiQz10r3vEJCHa2khEFQjKy91aWKz6+zybzssCvXUwE1LQWgWVwZ4nYUvHQ=="
+    },
+    "node_modules/abort-controller": {
+      "version": "3.0.0",
+      "resolved": "https://registry.npmjs.org/abort-controller/-/abort-controller-3.0.0.tgz",
+      "integrity": "sha512-h8lQ8tacZYnR3vNQTgibj+tODHI5/+l06Au2Pcriv/Gmet0eaj4TwWH41sO9wnHDiQsEj19q0drzdWdeAHtweg==",
+      "dependencies": {
+        "event-target-shim": "^5.0.0"
+      },
+      "engines": {
+        "node": ">=6.5"
+      }
+    },
+    "node_modules/agentkeepalive": {
+      "version": "4.5.0",
+      "resolved": "https://registry.npmjs.org/agentkeepalive/-/agentkeepalive-4.5.0.tgz",
+      "integrity": "sha512-5GG/5IbQQpC9FpkRGsSvZI5QYeSCzlJHdpBQntCsuTOxhKD8lqKhrleg2Yi7yvMIf82Ycmmqln9U8V9qwEiJew==",
+      "dependencies": {
+        "humanize-ms": "^1.2.1"
+      },
+      "engines": {
+        "node": ">= 8.0.0"
+      }
+    },
+    "node_modules/ansi-styles": {
+      "version": "5.2.0",
+      "resolved": "https://registry.npmjs.org/ansi-styles/-/ansi-styles-5.2.0.tgz",
+      "integrity": "sha512-Cxwpt2SfTzTtXcfOlzGEee8O+c+MmUgGrNiBcXnuWxuFJHe6a5Hz7qwhwe5OgaSYI0IJvkLqWX1ASG+cJOkEiA==",
+      "engines": {
+        "node": ">=10"
+      },
+      "funding": {
+        "url": "https://github.com/chalk/ansi-styles?sponsor=1"
+      }
+    },
+    "node_modules/argparse": {
+      "version": "2.0.1",
+      "resolved": "https://registry.npmjs.org/argparse/-/argparse-2.0.1.tgz",
+      "integrity": "sha512-8+9WqebbFzpX9OR+Wa6O29asIogeRMzcGtAINdpMHHyAg10f05aSFVBbcEqGf/PXw1EjAZ+q2/bEBg3DvurK3Q=="
+    },
+    "node_modules/asynckit": {
+      "version": "0.4.0",
+      "resolved": "https://registry.npmjs.org/asynckit/-/asynckit-0.4.0.tgz",
+      "integrity": "sha512-Oei9OH4tRh0YqU3GxhX79dM/mwVgvbZJaSNaRk+bshkj0S5cfHcgYakreBjrHwatXKbz+IoIdYLxrKim2MjW0Q=="
+    },
+    "node_modules/base-64": {
+      "version": "0.1.0",
+      "resolved": "https://registry.npmjs.org/base-64/-/base-64-0.1.0.tgz",
+      "integrity": "sha512-Y5gU45svrR5tI2Vt/X9GPd3L0HNIKzGu202EjxrXMpuc2V2CiKgemAbUUsqYmZJvPtCXoUKjNZwBJzsNScUbXA=="
+    },
+    "node_modules/base64-js": {
+      "version": "1.5.1",
+      "resolved": "https://registry.npmjs.org/base64-js/-/base64-js-1.5.1.tgz",
+      "integrity": "sha512-AKpaYlHn8t4SVbOHCy+b5+KKgvR4vrsD8vbvrbiQJps7fKDTkjkDry6ji0rUJjC0kzbNePLwzxq8iypo41qeWA==",
+      "funding": [
+        {
+          "type": "github",
+          "url": "https://github.com/sponsors/feross"
+        },
+        {
+          "type": "patreon",
+          "url": "https://www.patreon.com/feross"
+        },
+        {
+          "type": "consulting",
+          "url": "https://feross.org/support"
+        }
+      ]
+    },
+    "node_modules/binary-extensions": {
+      "version": "2.2.0",
+      "resolved": "https://registry.npmjs.org/binary-extensions/-/binary-extensions-2.2.0.tgz",
+      "integrity": "sha512-jDctJ/IVQbZoJykoeHbhXpOlNBqGNcwXJKJog42E5HDPUwQTSdjCHdihjj0DlnheQ7blbT6dHOafNAiS8ooQKA==",
+      "engines": {
+        "node": ">=8"
+      }
+    },
+    "node_modules/binary-search": {
+      "version": "1.3.6",
+      "resolved": "https://registry.npmjs.org/binary-search/-/binary-search-1.3.6.tgz",
+      "integrity": "sha512-nbE1WxOTTrUWIfsfZ4aHGYu5DOuNkbxGokjV6Z2kxfJK3uaAb8zNK1muzOeipoLHZjInT4Br88BHpzevc681xA=="
+    },
+    "node_modules/camelcase": {
+      "version": "6.3.0",
+      "resolved": "https://registry.npmjs.org/camelcase/-/camelcase-6.3.0.tgz",
+      "integrity": "sha512-Gmy6FhYlCY7uOElZUSbxo2UCDH8owEk996gkbrpsgGtrJLM3J7jGxl9Ic7Qwwj4ivOE5AWZWRMecDdF7hqGjFA==",
+      "engines": {
+        "node": ">=10"
+      },
+      "funding": {
+        "url": "https://github.com/sponsors/sindresorhus"
+      }
+    },
+    "node_modules/charenc": {
+      "version": "0.0.2",
+      "resolved": "https://registry.npmjs.org/charenc/-/charenc-0.0.2.tgz",
+      "integrity": "sha512-yrLQ/yVUFXkzg7EDQsPieE/53+0RlaWTs+wBrvW36cyilJ2SaDWfl4Yj7MtLTXleV9uEKefbAGUPv2/iWSooRA==",
+      "engines": {
+        "node": "*"
+      }
+    },
+    "node_modules/combined-stream": {
+      "version": "1.0.8",
+      "resolved": "https://registry.npmjs.org/combined-stream/-/combined-stream-1.0.8.tgz",
+      "integrity": "sha512-FQN4MRfuJeHf7cBbBMJFXhKSDq+2kAArBlmRBvcvFE5BB1HZKXtSFASDhdlz9zOYwxh8lDdnvmMOe/+5cdoEdg==",
+      "dependencies": {
+        "delayed-stream": "~1.0.0"
+      },
+      "engines": {
+        "node": ">= 0.8"
+      }
+    },
+    "node_modules/commander": {
+      "version": "10.0.1",
+      "resolved": "https://registry.npmjs.org/commander/-/commander-10.0.1.tgz",
+      "integrity": "sha512-y4Mg2tXshplEbSGzx7amzPwKKOCGuoSRP/CjEdwwk0FOGlUbq6lKuoyDZTNZkmxHdJtp54hdfY/JUrdL7Xfdug==",
+      "engines": {
+        "node": ">=14"
+      }
+    },
+    "node_modules/crypt": {
+      "version": "0.0.2",
+      "resolved": "https://registry.npmjs.org/crypt/-/crypt-0.0.2.tgz",
+      "integrity": "sha512-mCxBlsHFYh9C+HVpiEacem8FEBnMXgU9gy4zmNC+SXAZNB/1idgp/aulFJ4FgCi7GPEVbfyng092GqL2k2rmow==",
+      "engines": {
+        "node": "*"
+      }
+    },
+    "node_modules/decamelize": {
+      "version": "1.2.0",
+      "resolved": "https://registry.npmjs.org/decamelize/-/decamelize-1.2.0.tgz",
+      "integrity": "sha512-z2S+W9X73hAUUki+N+9Za2lBlun89zigOyGrsax+KUQ6wKW4ZoWpEYBkGhQjwAjjDCkWxhY0VKEhk8wzY7F5cA==",
+      "engines": {
+        "node": ">=0.10.0"
+      }
+    },
+    "node_modules/delayed-stream": {
+      "version": "1.0.0",
+      "resolved": "https://registry.npmjs.org/delayed-stream/-/delayed-stream-1.0.0.tgz",
+      "integrity": "sha512-ZySD7Nf91aLB0RxL4KGrKHBXl7Eds1DAmEdcoVawXnLD7SDhpNgtuII2aAkg7a7QS41jxPSZ17p4VdGnMHk3MQ==",
+      "engines": {
+        "node": ">=0.4.0"
+      }
+    },
+    "node_modules/digest-fetch": {
+      "version": "1.3.0",
+      "resolved": "https://registry.npmjs.org/digest-fetch/-/digest-fetch-1.3.0.tgz",
+      "integrity": "sha512-CGJuv6iKNM7QyZlM2T3sPAdZWd/p9zQiRNS9G+9COUCwzWFTs0Xp8NF5iePx7wtvhDykReiRRrSeNb4oMmB8lA==",
+      "dependencies": {
+        "base-64": "^0.1.0",
+        "md5": "^2.3.0"
+      }
+    },
+    "node_modules/event-target-shim": {
+      "version": "5.0.1",
+      "resolved": "https://registry.npmjs.org/event-target-shim/-/event-target-shim-5.0.1.tgz",
+      "integrity": "sha512-i/2XbnSz/uxRCU6+NdVJgKWDTM427+MqYbkQzD321DuCQJUqOuJKIA0IM2+W2xtYHdKOmZ4dR6fExsd4SXL+WQ==",
+      "engines": {
+        "node": ">=6"
+      }
+    },
+    "node_modules/eventemitter3": {
+      "version": "4.0.7",
+      "resolved": "https://registry.npmjs.org/eventemitter3/-/eventemitter3-4.0.7.tgz",
+      "integrity": "sha512-8guHBZCwKnFhYdHr2ysuRWErTwhoN2X8XELRlrRwpmfeY2jjuUN4taQMsULKUVo1K4DvZl+0pgfyoysHxvmvEw=="
+    },
+    "node_modules/expr-eval": {
+      "version": "2.0.2",
+      "resolved": "https://registry.npmjs.org/expr-eval/-/expr-eval-2.0.2.tgz",
+      "integrity": "sha512-4EMSHGOPSwAfBiibw3ndnP0AvjDWLsMvGOvWEZ2F96IGk0bIVdjQisOHxReSkE13mHcfbuCiXw+G4y0zv6N8Eg=="
+    },
+    "node_modules/flat": {
+      "version": "5.0.2",
+      "resolved": "https://registry.npmjs.org/flat/-/flat-5.0.2.tgz",
+      "integrity": "sha512-b6suED+5/3rTpUBdG1gupIl8MPFCAMA0QXwmljLhvCUKcUvdE4gWky9zpuGCcXHOsz4J9wPGNWq6OKpmIzz3hQ==",
+      "bin": {
+        "flat": "cli.js"
+      }
+    },
+    "node_modules/form-data": {
+      "version": "4.0.0",
+      "resolved": "https://registry.npmjs.org/form-data/-/form-data-4.0.0.tgz",
+      "integrity": "sha512-ETEklSGi5t0QMZuiXoA/Q6vcnxcLQP5vdugSpuAyi6SVGi2clPPp+xgEhuMaHC+zGgn31Kd235W35f7Hykkaww==",
+      "dependencies": {
+        "asynckit": "^0.4.0",
+        "combined-stream": "^1.0.8",
+        "mime-types": "^2.1.12"
+      },
+      "engines": {
+        "node": ">= 6"
+      }
+    },
+    "node_modules/form-data-encoder": {
+      "version": "1.7.2",
+      "resolved": "https://registry.npmjs.org/form-data-encoder/-/form-data-encoder-1.7.2.tgz",
+      "integrity": "sha512-qfqtYan3rxrnCk1VYaA4H+Ms9xdpPqvLZa6xmMgFvhO32x7/3J/ExcTd6qpxM0vH2GdMI+poehyBZvqfMTto8A=="
+    },
+    "node_modules/formdata-node": {
+      "version": "4.4.1",
+      "resolved": "https://registry.npmjs.org/formdata-node/-/formdata-node-4.4.1.tgz",
+      "integrity": "sha512-0iirZp3uVDjVGt9p49aTaqjk84TrglENEDuqfdlZQ1roC9CWlPk6Avf8EEnZNcAqPonwkG35x4n3ww/1THYAeQ==",
+      "dependencies": {
+        "node-domexception": "1.0.0",
+        "web-streams-polyfill": "4.0.0-beta.3"
+      },
+      "engines": {
+        "node": ">= 12.20"
+      }
+    },
+    "node_modules/humanize-ms": {
+      "version": "1.2.1",
+      "resolved": "https://registry.npmjs.org/humanize-ms/-/humanize-ms-1.2.1.tgz",
+      "integrity": "sha512-Fl70vYtsAFb/C06PTS9dZBo7ihau+Tu/DNCk/OyHhea07S+aeMWpFFkUaXRa8fI+ScZbEI8dfSxwY7gxZ9SAVQ==",
+      "dependencies": {
+        "ms": "^2.0.0"
+      }
+    },
+    "node_modules/is-any-array": {
+      "version": "2.0.1",
+      "resolved": "https://registry.npmjs.org/is-any-array/-/is-any-array-2.0.1.tgz",
+      "integrity": "sha512-UtilS7hLRu++wb/WBAw9bNuP1Eg04Ivn1vERJck8zJthEvXCBEBpGR/33u/xLKWEQf95803oalHrVDptcAvFdQ=="
+    },
+    "node_modules/is-buffer": {
+      "version": "1.1.6",
+      "resolved": "https://registry.npmjs.org/is-buffer/-/is-buffer-1.1.6.tgz",
+      "integrity": "sha512-NcdALwpXkTm5Zvvbk7owOUSvVvBKDgKP5/ewfXEznmQFfs4ZRmanOeKBTjRVjka3QFoN6XJ+9F3USqfHqTaU5w=="
+    },
+    "node_modules/js-tiktoken": {
+      "version": "1.0.7",
+      "resolved": "https://registry.npmjs.org/js-tiktoken/-/js-tiktoken-1.0.7.tgz",
+      "integrity": "sha512-biba8u/clw7iesNEWLOLwrNGoBP2lA+hTaBLs/D45pJdUPFXyxD6nhcDVtADChghv4GgyAiMKYMiRx7x6h7Biw==",
+      "dependencies": {
+        "base64-js": "^1.5.1"
+      }
+    },
+    "node_modules/js-yaml": {
+      "version": "4.1.0",
+      "resolved": "https://registry.npmjs.org/js-yaml/-/js-yaml-4.1.0.tgz",
+      "integrity": "sha512-wpxZs9NoxZaJESJGIZTyDEaYpl0FKSA+FB9aJiyemKhMwkxQg63h4T1KJgUGHpTqPDNRcmmYLugrRjJlBtWvRA==",
+      "dependencies": {
+        "argparse": "^2.0.1"
+      },
+      "bin": {
+        "js-yaml": "bin/js-yaml.js"
+      }
+    },
+    "node_modules/jsonpointer": {
+      "version": "5.0.1",
+      "resolved": "https://registry.npmjs.org/jsonpointer/-/jsonpointer-5.0.1.tgz",
+      "integrity": "sha512-p/nXbhSEcu3pZRdkW1OfJhpsVtW1gd4Wa1fnQc9YLiTfAjn0312eMKimbdIQzuZl9aa9xUGaRlP9T/CJE/ditQ==",
+      "engines": {
+        "node": ">=0.10.0"
+      }
+    },
+    "node_modules/langchain": {
+      "version": "0.0.165",
+      "resolved": "https://registry.npmjs.org/langchain/-/langchain-0.0.165.tgz",
+      "integrity": "sha512-CpbNpjwaE+9lzjdw+pZz0VgnRrFivEgr7CVp9dDaAb5JpaJAA4V2v6uQ9ZPN+TSqupTQ79HFn2sfyZVEl2EG7Q==",
+      "dependencies": {
+        "@anthropic-ai/sdk": "^0.6.2",
+        "ansi-styles": "^5.0.0",
+        "binary-extensions": "^2.2.0",
+        "camelcase": "6",
+        "decamelize": "^1.2.0",
+        "expr-eval": "^2.0.2",
+        "flat": "^5.0.2",
+        "js-tiktoken": "^1.0.7",
+        "js-yaml": "^4.1.0",
+        "jsonpointer": "^5.0.1",
+        "langchainhub": "~0.0.6",
+        "langsmith": "~0.0.31",
+        "ml-distance": "^4.0.0",
+        "object-hash": "^3.0.0",
+        "openai": "~4.4.0",
+        "openapi-types": "^12.1.3",
+        "p-queue": "^6.6.2",
+        "p-retry": "4",
+        "uuid": "^9.0.0",
+        "yaml": "^2.2.1",
+        "zod": "^3.22.3",
+        "zod-to-json-schema": "^3.20.4"
+      },
+      "engines": {
+        "node": ">=18"
+      },
+      "peerDependencies": {
+        "@aws-crypto/sha256-js": "^5.0.0",
+        "@aws-sdk/client-bedrock-runtime": "^3.422.0",
+        "@aws-sdk/client-dynamodb": "^3.310.0",
+        "@aws-sdk/client-kendra": "^3.352.0",
+        "@aws-sdk/client-lambda": "^3.310.0",
+        "@aws-sdk/client-s3": "^3.310.0",
+        "@aws-sdk/client-sagemaker-runtime": "^3.310.0",
+        "@aws-sdk/client-sfn": "^3.310.0",
+        "@aws-sdk/credential-provider-node": "^3.388.0",
+        "@azure/storage-blob": "^12.15.0",
+        "@clickhouse/client": "^0.0.14",
+        "@cloudflare/ai": "^1.0.12",
+        "@elastic/elasticsearch": "^8.4.0",
+        "@getmetal/metal-sdk": "*",
+        "@getzep/zep-js": "^0.7.0",
+        "@gomomento/sdk": "^1.23.0",
+        "@google-ai/generativelanguage": "^0.2.1",
+        "@google-cloud/storage": "^6.10.1",
+        "@huggingface/inference": "^1.5.1",
+        "@mozilla/readability": "*",
+        "@notionhq/client": "^2.2.10",
+        "@opensearch-project/opensearch": "*",
+        "@pinecone-database/pinecone": "^1.1.0",
+        "@planetscale/database": "^1.8.0",
+        "@qdrant/js-client-rest": "^1.2.0",
+        "@raycast/api": "^1.55.2",
+        "@smithy/eventstream-codec": "^2.0.5",
+        "@smithy/protocol-http": "^3.0.6",
+        "@smithy/signature-v4": "^2.0.10",
+        "@smithy/util-utf8": "^2.0.0",
+        "@supabase/postgrest-js": "^1.1.1",
+        "@supabase/supabase-js": "^2.10.0",
+        "@tensorflow-models/universal-sentence-encoder": "*",
+        "@tensorflow/tfjs-converter": "*",
+        "@tensorflow/tfjs-core": "*",
+        "@upstash/redis": "^1.20.6",
+        "@vercel/postgres": "^0.5.0",
+        "@writerai/writer-sdk": "^0.40.2",
+        "@xata.io/client": "^0.25.1",
+        "@xenova/transformers": "^2.5.4",
+        "@zilliz/milvus2-sdk-node": ">=2.2.7",
+        "apify-client": "^2.7.1",
+        "axios": "*",
+        "cassandra-driver": "^4.6.4",
+        "cheerio": "^1.0.0-rc.12",
+        "chromadb": "*",
+        "cohere-ai": ">=6.0.0",
+        "d3-dsv": "^2.0.0",
+        "epub2": "^3.0.1",
+        "faiss-node": "^0.3.0",
+        "fast-xml-parser": "^4.2.7",
+        "firebase-admin": "^11.9.0",
+        "google-auth-library": "^8.9.0",
+        "googleapis": "^126.0.1",
+        "hnswlib-node": "^1.4.2",
+        "html-to-text": "^9.0.5",
+        "ignore": "^5.2.0",
+        "ioredis": "^5.3.2",
+        "jsdom": "*",
+        "llmonitor": "*",
+        "lodash": "^4.17.21",
+        "mammoth": "*",
+        "mongodb": "^5.2.0",
+        "mysql2": "^3.3.3",
+        "neo4j-driver": "*",
+        "node-llama-cpp": "*",
+        "notion-to-md": "^3.1.0",
+        "pdf-parse": "1.1.1",
+        "peggy": "^3.0.2",
+        "pg": "^8.11.0",
+        "pg-copy-streams": "^6.0.5",
+        "pickleparser": "^0.1.0",
+        "playwright": "^1.32.1",
+        "portkey-ai": "^0.1.11",
+        "puppeteer": "^19.7.2",
+        "redis": "^4.6.4",
+        "replicate": "^0.18.0",
+        "sonix-speech-recognition": "^2.1.1",
+        "srt-parser-2": "^1.2.2",
+        "typeorm": "^0.3.12",
+        "typesense": "^1.5.3",
+        "usearch": "^1.1.1",
+        "vectordb": "^0.1.4",
+        "voy-search": "0.6.2",
+        "weaviate-ts-client": "^1.4.0",
+        "web-auth-library": "^1.0.3",
+        "youtube-transcript": "^1.0.6",
+        "youtubei.js": "^5.8.0"
+      },
+      "peerDependenciesMeta": {
+        "@aws-crypto/sha256-js": {
+          "optional": true
+        },
+        "@aws-sdk/client-bedrock-runtime": {
+          "optional": true
+        },
+        "@aws-sdk/client-dynamodb": {
+          "optional": true
+        },
+        "@aws-sdk/client-kendra": {
+          "optional": true
+        },
+        "@aws-sdk/client-lambda": {
+          "optional": true
+        },
+        "@aws-sdk/client-s3": {
+          "optional": true
+        },
+        "@aws-sdk/client-sagemaker-runtime": {
+          "optional": true
+        },
+        "@aws-sdk/client-sfn": {
+          "optional": true
+        },
+        "@aws-sdk/credential-provider-node": {
+          "optional": true
+        },
+        "@azure/storage-blob": {
+          "optional": true
+        },
+        "@clickhouse/client": {
+          "optional": true
+        },
+        "@cloudflare/ai": {
+          "optional": true
+        },
+        "@elastic/elasticsearch": {
+          "optional": true
+        },
+        "@getmetal/metal-sdk": {
+          "optional": true
+        },
+        "@getzep/zep-js": {
+          "optional": true
+        },
+        "@gomomento/sdk": {
+          "optional": true
+        },
+        "@google-ai/generativelanguage": {
+          "optional": true
+        },
+        "@google-cloud/storage": {
+          "optional": true
+        },
+        "@huggingface/inference": {
+          "optional": true
+        },
+        "@mozilla/readability": {
+          "optional": true
+        },
+        "@notionhq/client": {
+          "optional": true
+        },
+        "@opensearch-project/opensearch": {
+          "optional": true
+        },
+        "@pinecone-database/pinecone": {
+          "optional": true
+        },
+        "@planetscale/database": {
+          "optional": true
+        },
+        "@qdrant/js-client-rest": {
+          "optional": true
+        },
+        "@raycast/api": {
+          "optional": true
+        },
+        "@smithy/eventstream-codec": {
+          "optional": true
+        },
+        "@smithy/protocol-http": {
+          "optional": true
+        },
+        "@smithy/signature-v4": {
+          "optional": true
+        },
+        "@smithy/util-utf8": {
+          "optional": true
+        },
+        "@supabase/postgrest-js": {
+          "optional": true
+        },
+        "@supabase/supabase-js": {
+          "optional": true
+        },
+        "@tensorflow-models/universal-sentence-encoder": {
+          "optional": true
+        },
+        "@tensorflow/tfjs-converter": {
+          "optional": true
+        },
+        "@tensorflow/tfjs-core": {
+          "optional": true
+        },
+        "@upstash/redis": {
+          "optional": true
+        },
+        "@vercel/postgres": {
+          "optional": true
+        },
+        "@writerai/writer-sdk": {
+          "optional": true
+        },
+        "@xata.io/client": {
+          "optional": true
+        },
+        "@xenova/transformers": {
+          "optional": true
+        },
+        "@zilliz/milvus2-sdk-node": {
+          "optional": true
+        },
+        "apify-client": {
+          "optional": true
+        },
+        "axios": {
+          "optional": true
+        },
+        "cassandra-driver": {
+          "optional": true
+        },
+        "cheerio": {
+          "optional": true
+        },
+        "chromadb": {
+          "optional": true
+        },
+        "cohere-ai": {
+          "optional": true
+        },
+        "d3-dsv": {
+          "optional": true
+        },
+        "epub2": {
+          "optional": true
+        },
+        "faiss-node": {
+          "optional": true
+        },
+        "fast-xml-parser": {
+          "optional": true
+        },
+        "firebase-admin": {
+          "optional": true
+        },
+        "google-auth-library": {
+          "optional": true
+        },
+        "googleapis": {
+          "optional": true
+        },
+        "hnswlib-node": {
+          "optional": true
+        },
+        "html-to-text": {
+          "optional": true
+        },
+        "ignore": {
+          "optional": true
+        },
+        "ioredis": {
+          "optional": true
+        },
+        "jsdom": {
+          "optional": true
+        },
+        "llmonitor": {
+          "optional": true
+        },
+        "lodash": {
+          "optional": true
+        },
+        "mammoth": {
+          "optional": true
+        },
+        "mongodb": {
+          "optional": true
+        },
+        "mysql2": {
+          "optional": true
+        },
+        "neo4j-driver": {
+          "optional": true
+        },
+        "node-llama-cpp": {
+          "optional": true
+        },
+        "notion-to-md": {
+          "optional": true
+        },
+        "pdf-parse": {
+          "optional": true
+        },
+        "peggy": {
+          "optional": true
+        },
+        "pg": {
+          "optional": true
+        },
+        "pg-copy-streams": {
+          "optional": true
+        },
+        "pickleparser": {
+          "optional": true
+        },
+        "playwright": {
+          "optional": true
+        },
+        "portkey-ai": {
+          "optional": true
+        },
+        "puppeteer": {
+          "optional": true
+        },
+        "redis": {
+          "optional": true
+        },
+        "replicate": {
+          "optional": true
+        },
+        "sonix-speech-recognition": {
+          "optional": true
+        },
+        "srt-parser-2": {
+          "optional": true
+        },
+        "typeorm": {
+          "optional": true
+        },
+        "typesense": {
+          "optional": true
+        },
+        "usearch": {
+          "optional": true
+        },
+        "vectordb": {
+          "optional": true
+        },
+        "voy-search": {
+          "optional": true
+        },
+        "weaviate-ts-client": {
+          "optional": true
+        },
+        "web-auth-library": {
+          "optional": true
+        },
+        "youtube-transcript": {
+          "optional": true
+        },
+        "youtubei.js": {
+          "optional": true
+        }
+      }
+    },
+    "node_modules/langchainhub": {
+      "version": "0.0.6",
+      "resolved": "https://registry.npmjs.org/langchainhub/-/langchainhub-0.0.6.tgz",
+      "integrity": "sha512-SW6105T+YP1cTe0yMf//7kyshCgvCTyFBMTgH2H3s9rTAR4e+78DA/BBrUL/Mt4Q5eMWui7iGuAYb3pgGsdQ9w=="
+    },
+    "node_modules/langsmith": {
+      "version": "0.0.42",
+      "resolved": "https://registry.npmjs.org/langsmith/-/langsmith-0.0.42.tgz",
+      "integrity": "sha512-sFuN+e7E+pPBIRaRgFqZh/BRBWNHTZNAwi6uj4kydQawooCZYoJmM5snOkiQrhVSvAhgu6xFhLvmfvkPcKzD7w==",
+      "dependencies": {
+        "@types/uuid": "^9.0.1",
+        "commander": "^10.0.1",
+        "p-queue": "^6.6.2",
+        "p-retry": "4",
+        "uuid": "^9.0.0"
+      },
+      "bin": {
+        "langsmith": "dist/cli/main.cjs"
+      }
+    },
+    "node_modules/md5": {
+      "version": "2.3.0",
+      "resolved": "https://registry.npmjs.org/md5/-/md5-2.3.0.tgz",
+      "integrity": "sha512-T1GITYmFaKuO91vxyoQMFETst+O71VUPEU3ze5GNzDm0OWdP8v1ziTaAEPUr/3kLsY3Sftgz242A1SetQiDL7g==",
+      "dependencies": {
+        "charenc": "0.0.2",
+        "crypt": "0.0.2",
+        "is-buffer": "~1.1.6"
+      }
+    },
+    "node_modules/mime-db": {
+      "version": "1.52.0",
+      "resolved": "https://registry.npmjs.org/mime-db/-/mime-db-1.52.0.tgz",
+      "integrity": "sha512-sPU4uV7dYlvtWJxwwxHD0PuihVNiE7TyAbQ5SWxDCB9mUYvOgroQOwYQQOKPJ8CIbE+1ETVlOoK1UC2nU3gYvg==",
+      "engines": {
+        "node": ">= 0.6"
+      }
+    },
+    "node_modules/mime-types": {
+      "version": "2.1.35",
+      "resolved": "https://registry.npmjs.org/mime-types/-/mime-types-2.1.35.tgz",
+      "integrity": "sha512-ZDY+bPm5zTTF+YpCrAU9nK0UgICYPT0QtT1NZWFv4s++TNkcgVaT0g6+4R2uI4MjQjzysHB1zxuWL50hzaeXiw==",
+      "dependencies": {
+        "mime-db": "1.52.0"
+      },
+      "engines": {
+        "node": ">= 0.6"
+      }
+    },
+    "node_modules/ml-array-mean": {
+      "version": "1.1.6",
+      "resolved": "https://registry.npmjs.org/ml-array-mean/-/ml-array-mean-1.1.6.tgz",
+      "integrity": "sha512-MIdf7Zc8HznwIisyiJGRH9tRigg3Yf4FldW8DxKxpCCv/g5CafTw0RRu51nojVEOXuCQC7DRVVu5c7XXO/5joQ==",
+      "dependencies": {
+        "ml-array-sum": "^1.1.6"
+      }
+    },
+    "node_modules/ml-array-sum": {
+      "version": "1.1.6",
+      "resolved": "https://registry.npmjs.org/ml-array-sum/-/ml-array-sum-1.1.6.tgz",
+      "integrity": "sha512-29mAh2GwH7ZmiRnup4UyibQZB9+ZLyMShvt4cH4eTK+cL2oEMIZFnSyB3SS8MlsTh6q/w/yh48KmqLxmovN4Dw==",
+      "dependencies": {
+        "is-any-array": "^2.0.0"
+      }
+    },
+    "node_modules/ml-distance": {
+      "version": "4.0.1",
+      "resolved": "https://registry.npmjs.org/ml-distance/-/ml-distance-4.0.1.tgz",
+      "integrity": "sha512-feZ5ziXs01zhyFUUUeZV5hwc0f5JW0Sh0ckU1koZe/wdVkJdGxcP06KNQuF0WBTj8FttQUzcvQcpcrOp/XrlEw==",
+      "dependencies": {
+        "ml-array-mean": "^1.1.6",
+        "ml-distance-euclidean": "^2.0.0",
+        "ml-tree-similarity": "^1.0.0"
+      }
+    },
+    "node_modules/ml-distance-euclidean": {
+      "version": "2.0.0",
+      "resolved": "https://registry.npmjs.org/ml-distance-euclidean/-/ml-distance-euclidean-2.0.0.tgz",
+      "integrity": "sha512-yC9/2o8QF0A3m/0IXqCTXCzz2pNEzvmcE/9HFKOZGnTjatvBbsn4lWYJkxENkA4Ug2fnYl7PXQxnPi21sgMy/Q=="
+    },
+    "node_modules/ml-tree-similarity": {
+      "version": "1.0.0",
+      "resolved": "https://registry.npmjs.org/ml-tree-similarity/-/ml-tree-similarity-1.0.0.tgz",
+      "integrity": "sha512-XJUyYqjSuUQkNQHMscr6tcjldsOoAekxADTplt40QKfwW6nd++1wHWV9AArl0Zvw/TIHgNaZZNvr8QGvE8wLRg==",
+      "dependencies": {
+        "binary-search": "^1.3.5",
+        "num-sort": "^2.0.0"
+      }
+    },
+    "node_modules/ms": {
+      "version": "2.1.3",
+      "resolved": "https://registry.npmjs.org/ms/-/ms-2.1.3.tgz",
+      "integrity": "sha512-6FlzubTLZG3J2a/NVCAleEhjzq5oxgHyaCU9yYXvcLsvoVaHJq/s5xXI6/XXP6tz7R9xAOtHnSO/tXtF3WRTlA=="
+    },
+    "node_modules/node-domexception": {
+      "version": "1.0.0",
+      "resolved": "https://registry.npmjs.org/node-domexception/-/node-domexception-1.0.0.tgz",
+      "integrity": "sha512-/jKZoMpw0F8GRwl4/eLROPA3cfcXtLApP0QzLmUT/HuPCZWyB7IY9ZrMeKw2O/nFIqPQB3PVM9aYm0F312AXDQ==",
+      "funding": [
+        {
+          "type": "github",
+          "url": "https://github.com/sponsors/jimmywarting"
+        },
+        {
+          "type": "github",
+          "url": "https://paypal.me/jimmywarting"
+        }
+      ],
+      "engines": {
+        "node": ">=10.5.0"
+      }
+    },
+    "node_modules/node-fetch": {
+      "version": "2.7.0",
+      "resolved": "https://registry.npmjs.org/node-fetch/-/node-fetch-2.7.0.tgz",
+      "integrity": "sha512-c4FRfUm/dbcWZ7U+1Wq0AwCyFL+3nt2bEw05wfxSz+DWpWsitgmSgYmy2dQdWyKC1694ELPqMs/YzUSNozLt8A==",
+      "dependencies": {
+        "whatwg-url": "^5.0.0"
+      },
+      "engines": {
+        "node": "4.x || >=6.0.0"
+      },
+      "peerDependencies": {
+        "encoding": "^0.1.0"
+      },
+      "peerDependenciesMeta": {
+        "encoding": {
+          "optional": true
+        }
+      }
+    },
+    "node_modules/num-sort": {
+      "version": "2.1.0",
+      "resolved": "https://registry.npmjs.org/num-sort/-/num-sort-2.1.0.tgz",
+      "integrity": "sha512-1MQz1Ed8z2yckoBeSfkQHHO9K1yDRxxtotKSJ9yvcTUUxSvfvzEq5GwBrjjHEpMlq/k5gvXdmJ1SbYxWtpNoVg==",
+      "engines": {
+        "node": ">=8"
+      },
+      "funding": {
+        "url": "https://github.com/sponsors/sindresorhus"
+      }
+    },
+    "node_modules/object-hash": {
+      "version": "3.0.0",
+      "resolved": "https://registry.npmjs.org/object-hash/-/object-hash-3.0.0.tgz",
+      "integrity": "sha512-RSn9F68PjH9HqtltsSnqYC1XXoWe9Bju5+213R98cNGttag9q9yAOTzdbsqvIa7aNm5WffBZFpWYr2aWrklWAw==",
+      "engines": {
+        "node": ">= 6"
+      }
+    },
+    "node_modules/openai": {
+      "version": "4.4.0",
+      "resolved": "https://registry.npmjs.org/openai/-/openai-4.4.0.tgz",
+      "integrity": "sha512-JN0t628Kh95T0IrXl0HdBqnlJg+4Vq0Bnh55tio+dfCnyzHvMLiWyCM9m726MAJD2YkDU4/8RQB6rNbEq9ct2w==",
+      "dependencies": {
+        "@types/node": "^18.11.18",
+        "@types/node-fetch": "^2.6.4",
+        "abort-controller": "^3.0.0",
+        "agentkeepalive": "^4.2.1",
+        "digest-fetch": "^1.3.0",
+        "form-data-encoder": "1.7.2",
+        "formdata-node": "^4.3.2",
+        "node-fetch": "^2.6.7"
+      },
+      "bin": {
+        "openai": "bin/cli"
+      }
+    },
+    "node_modules/openapi-types": {
+      "version": "12.1.3",
+      "resolved": "https://registry.npmjs.org/openapi-types/-/openapi-types-12.1.3.tgz",
+      "integrity": "sha512-N4YtSYJqghVu4iek2ZUvcN/0aqH1kRDuNqzcycDxhOUpg7GdvLa2F3DgS6yBNhInhv2r/6I0Flkn7CqL8+nIcw=="
+    },
+    "node_modules/p-finally": {
+      "version": "1.0.0",
+      "resolved": "https://registry.npmjs.org/p-finally/-/p-finally-1.0.0.tgz",
+      "integrity": "sha512-LICb2p9CB7FS+0eR1oqWnHhp0FljGLZCWBE9aix0Uye9W8LTQPwMTYVGWQWIw9RdQiDg4+epXQODwIYJtSJaow==",
+      "engines": {
+        "node": ">=4"
+      }
+    },
+    "node_modules/p-queue": {
+      "version": "6.6.2",
+      "resolved": "https://registry.npmjs.org/p-queue/-/p-queue-6.6.2.tgz",
+      "integrity": "sha512-RwFpb72c/BhQLEXIZ5K2e+AhgNVmIejGlTgiB9MzZ0e93GRvqZ7uSi0dvRF7/XIXDeNkra2fNHBxTyPDGySpjQ==",
+      "dependencies": {
+        "eventemitter3": "^4.0.4",
+        "p-timeout": "^3.2.0"
+      },
+      "engines": {
+        "node": ">=8"
+      },
+      "funding": {
+        "url": "https://github.com/sponsors/sindresorhus"
+      }
+    },
+    "node_modules/p-retry": {
+      "version": "4.6.2",
+      "resolved": "https://registry.npmjs.org/p-retry/-/p-retry-4.6.2.tgz",
+      "integrity": "sha512-312Id396EbJdvRONlngUx0NydfrIQ5lsYu0znKVUzVvArzEIt08V1qhtyESbGVd1FGX7UKtiFp5uwKZdM8wIuQ==",
+      "dependencies": {
+        "@types/retry": "0.12.0",
+        "retry": "^0.13.1"
+      },
+      "engines": {
+        "node": ">=8"
+      }
+    },
+    "node_modules/p-timeout": {
+      "version": "3.2.0",
+      "resolved": "https://registry.npmjs.org/p-timeout/-/p-timeout-3.2.0.tgz",
+      "integrity": "sha512-rhIwUycgwwKcP9yTOOFK/AKsAopjjCakVqLHePO3CC6Mir1Z99xT+R63jZxAT5lFZLa2inS5h+ZS2GvR99/FBg==",
+      "dependencies": {
+        "p-finally": "^1.0.0"
+      },
+      "engines": {
+        "node": ">=8"
+      }
+    },
+    "node_modules/retry": {
+      "version": "0.13.1",
+      "resolved": "https://registry.npmjs.org/retry/-/retry-0.13.1.tgz",
+      "integrity": "sha512-XQBQ3I8W1Cge0Seh+6gjj03LbmRFWuoszgK9ooCpwYIrhhoO80pfq4cUkU5DkknwfOfFteRwlZ56PYOGYyFWdg==",
+      "engines": {
+        "node": ">= 4"
+      }
+    },
+    "node_modules/tr46": {
+      "version": "0.0.3",
+      "resolved": "https://registry.npmjs.org/tr46/-/tr46-0.0.3.tgz",
+      "integrity": "sha512-N3WMsuqV66lT30CrXNbEjx4GEwlow3v6rr4mCcv6prnfwhS01rkgyFdjPNBYd9br7LpXV1+Emh01fHnq2Gdgrw=="
+    },
+    "node_modules/typescript": {
+      "version": "5.2.2",
+      "resolved": "https://registry.npmjs.org/typescript/-/typescript-5.2.2.tgz",
+      "integrity": "sha512-mI4WrpHsbCIcwT9cF4FZvr80QUeKvsUsUvKDoR+X/7XHQH98xYD8YHZg7ANtz2GtZt/CBq2QJ0thkGJMHfqc1w==",
+      "dev": true,
+      "bin": {
+        "tsc": "bin/tsc",
+        "tsserver": "bin/tsserver"
+      },
+      "engines": {
+        "node": ">=14.17"
+      }
+    },
+    "node_modules/uuid": {
+      "version": "9.0.1",
+      "resolved": "https://registry.npmjs.org/uuid/-/uuid-9.0.1.tgz",
+      "integrity": "sha512-b+1eJOlsR9K8HJpow9Ok3fiWOWSIcIzXodvv0rQjVoOVNpWMpxf1wZNpt4y9h10odCNrqnYp1OBzRktckBe3sA==",
+      "funding": [
+        "https://github.com/sponsors/broofa",
+        "https://github.com/sponsors/ctavan"
+      ],
+      "bin": {
+        "uuid": "dist/bin/uuid"
+      }
+    },
+    "node_modules/web-streams-polyfill": {
+      "version": "4.0.0-beta.3",
+      "resolved": "https://registry.npmjs.org/web-streams-polyfill/-/web-streams-polyfill-4.0.0-beta.3.tgz",
+      "integrity": "sha512-QW95TCTaHmsYfHDybGMwO5IJIM93I/6vTRk+daHTWFPhwh+C8Cg7j7XyKrwrj8Ib6vYXe0ocYNrmzY4xAAN6ug==",
+      "engines": {
+        "node": ">= 14"
+      }
+    },
+    "node_modules/webidl-conversions": {
+      "version": "3.0.1",
+      "resolved": "https://registry.npmjs.org/webidl-conversions/-/webidl-conversions-3.0.1.tgz",
+      "integrity": "sha512-2JAn3z8AR6rjK8Sm8orRC0h/bcl/DqL7tRPdGZ4I1CjdF+EaMLmYxBHyXuKL849eucPFhvBoxMsflfOb8kxaeQ=="
+    },
+    "node_modules/whatwg-url": {
+      "version": "5.0.0",
+      "resolved": "https://registry.npmjs.org/whatwg-url/-/whatwg-url-5.0.0.tgz",
+      "integrity": "sha512-saE57nupxk6v3HY35+jzBwYa0rKSy0XR8JSxZPwgLr7ys0IBzhGviA1/TUGJLmSVqs8pb9AnvICXEuOHLprYTw==",
+      "dependencies": {
+        "tr46": "~0.0.3",
+        "webidl-conversions": "^3.0.0"
+      }
+    },
+    "node_modules/yaml": {
+      "version": "2.3.2",
+      "resolved": "https://registry.npmjs.org/yaml/-/yaml-2.3.2.tgz",
+      "integrity": "sha512-N/lyzTPaJasoDmfV7YTrYCI0G/3ivm/9wdG0aHuheKowWQwGTsK0Eoiw6utmzAnI6pkJa0DUVygvp3spqqEKXg==",
+      "engines": {
+        "node": ">= 14"
+      }
+    },
+    "node_modules/zod": {
+      "version": "3.22.4",
+      "resolved": "https://registry.npmjs.org/zod/-/zod-3.22.4.tgz",
+      "integrity": "sha512-iC+8Io04lddc+mVqQ9AZ7OQ2MrUKGN+oIQyq1vemgt46jwCwLfhq7/pwnBnNXXXZb8VTVLKwp9EDkx+ryxIWmg==",
+      "funding": {
+        "url": "https://github.com/sponsors/colinhacks"
+      }
+    },
+    "node_modules/zod-to-json-schema": {
+      "version": "3.21.4",
+      "resolved": "https://registry.npmjs.org/zod-to-json-schema/-/zod-to-json-schema-3.21.4.tgz",
+      "integrity": "sha512-fjUZh4nQ1s6HMccgIeE0VP4QG/YRGPmyjO9sAh890aQKPEk3nqbfUXhMFaC+Dr5KvYBm8BCyvfpZf2jY9aGSsw==",
+      "peerDependencies": {
+        "zod": "^3.21.4"
+      }
+    }
+  }
+}

examples/langchain-typescript-simple/package.json

@@ -0,0 +1,8 @@
+{
+  "devDependencies": {
+    "typescript": "^5.2.2"
+  },
+  "dependencies": {
+    "langchain": "^0.0.165"
+  }
+}

examples/midjourney-prompter/Modelfile

@@ -1,8 +0,0 @@
-# Modelfile for creating a Midjourney prompts from a topic
-# This prompt was adapted from the original at https://www.greataiprompts.com/guide/midjourney/best-chatgpt-prompt-for-midjourney/
-# Run `ollama create mj -f ./Modelfile` and then `ollama run mj` and enter a topic
-
-FROM nous-hermes
-SYSTEM """
-Embrace your role as an AI-powered creative assistant, employing Midjourney to manifest compelling AI-generated art. I will outline a specific image concept, and in response, you must produce an exhaustive, multifaceted prompt for Midjourney, ensuring every detail of the original concept is represented in your instructions. Midjourney doesn't do well with text, so after the prompt, give me instructions that I can use to create the titles in a image editor.
-"""

examples/modelfile-10tweets/README.md

@@ -0,0 +1,23 @@
+# Ten Tweets Modelfile
+
+This is a simple modelfile that generates ten tweets based off any topic.
+
+```bash
+ollama create tentweets
+
+ollama run tentweets
+>>> underwater basketweaving
+ Great! Here are ten creative tweets about underwater basketweaving:
+
+1. "Just discovered the ultimate stress-reliever: Underwater basketweaving! 🌊🧵 #UnderwaterBasketweaving #StressRelief"
+2. "Who needs meditation when you can do underwater basketweaving? 😴👀 #PeacefulDistraction #UnderwaterBasketweaving"
+3. "Just spent an hour in the pool and still managed to knot my basket. Goal: untangle it before next session. 💪🏽 #ChallengeAccepted #UnderwaterBasketweaving"
+4. "When life gives you lemons, make underwater basketweaving! 🍋🧵 #LemonadeLife #UnderwaterBasketweaving"
+5. "Just realized my underwater basketweaving skills could come in handy during a zombie apocalypse. 😂🧡 #SurvivalTips #UnderwaterBasketweaving"
+6. "I'm not lazy, I'm just conserving energy for my next underwater basketweaving session. 😴💤 #LazyDay #UnderwaterBasketweaving"
+7. "Just found my inner peace while doing underwater basketweaving. It's like meditation, but with knots! 🙏🧵 #Mindfulness #UnderwaterBasketweaving"
+8. "Why study for exams when you can do underwater basketweaving and forget all your worries? 😜🧵 #ProcrastinationStation #UnderwaterBasketweaving"
+9. "Just had to cut my underwater basketweaving session short due to a sudden urge to breathe. 🤯🌊 #AquaticAdventures #UnderwaterBasketweaving"
+10. "I'm not sure what's more impressive: my underwater basketweaving skills or the fact that I didn't drown trying to make this tweet. 😅🧵 #Accomplishment 
+#UnderwaterBasketweaving"
+```

examples/modelfile-devopsengineer/Modelfile

@@ -1,7 +1,7 @@
 # Modelfile for creating a devops engineer assistant
 # Run `ollama create devops-engineer -f ./Modelfile` and then `ollama run devops-engineer` and enter a topic
 
-FROM llama2:13b
+FROM mistral
 PARAMETER temperature 1
 SYSTEM """
 You are a senior devops engineer, acting as an assistant. You offer help with cloud technologies like: Terraform, AWS, kubernetes, python. You answer with code examples when possible

examples/modelfile-devopsengineer/README.md

@@ -0,0 +1,40 @@
+# DevOps Engineer Modelfile
+
+This is a simple modelfile that takes will respond to any questions that a senior DevOps engineer should know how to solve.
+
+````text
+ollama create devopsengineer
+
+ollama run devopsengineer
+>>> how can i proxy all connections from inside the office through a single server that caches large files. Provide a config file for nginx to achieve this?
+
+Here is an example configuration file for Nginx that can be used to proxy all connections from inside the office network to a centralized cache server:
+```perl
+http {
+    upstream cache_server {
+        server <cache_server_ip>;
+        server <cache_server_ip>;
+        server <cache_server_ip>;
+    }
+    
+    server {
+        listen 80;
+        server_name localhost;
+        
+        location / {
+            proxy_pass http://cache_server;
+            proxy_set_header Host $host;
+            proxy_set_header X-Real-IP $remote_addr;
+            proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
+        }
+    }
+}
+```
+In this configuration, we define an `upstream` block that lists the IP addresses of the cache servers. We then define a server block that listens on port 80 and routes all 
+traffic to the `/` location to the cache servers using the `proxy_pass` directive. The `proxy_set_header` directives are used to preserve the source IP address of the client
+request when forwarding it to the cache server.
+
+To use this configuration, you would need to replace the placeholder `<cache_server_ip>` with the actual IP addresses of your cache servers. You would also need to make sure
+that the cache servers are configured to accept incoming connections from the Nginx server and handle requests for files.
+
+````

examples/modelfile-midjourney/Modelfile

@@ -0,0 +1,11 @@
+# Modelfile for creating a Midjourney prompts from a topic
+# This prompt was adapted from the original at https://www.greataiprompts.com/guide/midjourney/best-chatgpt-prompt-for-midjourney/
+# Run `ollama create mj -f ./Modelfile` and then `ollama run mj` and enter a topic
+
+FROM zephyr
+PARAMETER temperature 0.8
+PARAMETER top_k 500
+PARAMETER top_p 0.9
+SYSTEM """
+Embrace your role as a creative illustrator. Based on a concept provided, you must produce a single paragraph with a multifaceted description of an image, ensuring significant details of the concept and more is represented in your instructions. You do not need to write complete sentences but rather short concepts with the following information: the level of detail that should be represented, an artistic style and maybe a specific name of a painter or illustrator, the ideal color pallete, lighting, mood, perspective, the setting, time of day, weather, the season, the time period, location, materials, the textures, patterns, lines, brushstrokes, techniques, the medium, the genre, the rendering style. Don't include everything and keep the description length under 250 words. 
+"""

examples/modelfile-midjourney/README.md

@@ -0,0 +1,11 @@
+# Midjourney Prompt Generator Modelfile
+
+This simple modelfile will help create a prompt to feed to Midjourney.
+
+```text
+ollama create midjourney
+
+ollama run midjourney
+>>> a sports car in the mountains. 
+A sleek, high-performance automobile cuts through a serpentine mountain landscape. The concept is a classic illustration of speed and power, depicted in the style of pop art by Andy Warhol. The color palette is dominated by bold, primary hues of red, blue, and yellow, with striking accent colors of white, black, and metallic shades. The lighting is bright and focused, casting sharp shadows on the rugged terrain. A sense of excitement and anticipation permeates throughout the scene, as the car navigates a treacherous course through the winding road. The perspective is low, allowing for a full view of the vehicle's sleek lines and intricate details. The setting takes place in the afternoon during a sunny day in autumn, as evidenced by the vibrant foliage on the mountainside. The time period is modern, with nods to classic car design. The materials are primarily digital, allowing for smooth curves and sharp contrasts. The textures are sleek and polished, with meticulously detailed lines and brushstrokes that accentuate the car's aerodynamic design. The patterns consist of geometric shapes and bold stripes, adding to the car's dynamic appeal. The genre is modern realism, with a focus on precision and detail. The rendering style is highly technical, capturing the nuances and subtleties of the vehicle and its surroundings in breathtaking detail.
+```

examples/modelfile-recipemaker/README.md

@@ -0,0 +1,20 @@
+# Recipe Maker Modelfile 
+
+Simple modelfile to generate a recipe from a short list of ingredients.
+
+```
+ollama create recipemaker
+
+ollama run recipemaker
+>>> chilli pepper, white chocolate, kale
+ Ingredients:
+- 1 small chili pepper
+- 4 squares of white chocolate
+- handful of kale leaves
+
+Instructions:
+1. In a blender or food processor, puree the chilies and white chocolate until smooth.
+2. Add the chopped kale leaves to the blender and pulse until well combined.
+3. Serve immediately as a dip for crackers or use it as an ingredient in your favorite recipe. The mixture of spicy chili pepper with sweet white chocolate and nutritious 
+kale will make your taste buds dance with delight!
+```

examples/python-dockerit/Modelfile

@@ -1,4 +1,4 @@
-FROM llama2
+FROM mistral
 SYSTEM """
 You are an experienced Devops engineer focused on docker. When given specifications for a particular need or application you know the best way to host that within a docker container. For instance if someone tells you they want an nginx server to host files located at /web you will answer as follows
 

examples/python-rag-newssummary/README.md

@@ -0,0 +1,22 @@
+# News Summarizer
+
+This example goes through a series of steps:
+
+  1. You choose a topic area (e.g., "news", "NVidia", "music", etc.).
+  2. Gets the most recent articles on that topic from various sources.
+  3. Uses Ollama to summarize each article.
+  4. Creates chunks of sentences from each article.
+  5. Uses Sentence Transformers to generate embeddings for each of those chunks.
+  6. You enter a question regarding the summaries shown.
+  7. Uses Sentence Transformers to generate an embedding for that question.
+  8. Uses the embedded question to find the most similar chunks.
+  9. Feeds all that to Ollama to generate a good answer to your question based on these news articles.
+
+This example lets you pick from a few different topic areas, then summarize the most recent x articles for that topic. It then creates chunks of sentences from each article and then generates embeddings for each of those chunks.
+
+You can run the example like this:
+
+```bash
+pip install -r requirements.txt
+python summ.py
+```

examples/python-rag-newssummary/requirements.txt

@@ -0,0 +1,9 @@
+beautifulsoup4==4.12.2
+feedparser==6.0.10
+mattsollamatools==0.0.8
+newspaper3k==0.2.8
+nltk==3.8.1
+numpy==1.24.3
+Requests==2.31.0
+scikit_learn==1.3.0
+sentence_transformers==2.2.2

examples/python-rag-newssummary/summ.py

@@ -0,0 +1,86 @@
+import curses
+import json
+from utils import get_url_for_topic, topic_urls, menu, getUrls, get_summary, getArticleText, knn_search
+import requests
+from sentence_transformers import SentenceTransformer
+from mattsollamatools import chunker
+
+if __name__ == "__main__":
+    chosen_topic = curses.wrapper(menu)
+    print("Here is your news summary:\n")
+    urls = getUrls(chosen_topic, n=5)
+    model = SentenceTransformer('all-MiniLM-L6-v2')
+    allEmbeddings = []
+
+    for url in urls:
+      article={}
+      article['embeddings'] = []
+      article['url'] = url
+      text = getArticleText(url)
+      summary = get_summary(text)
+      chunks = chunker(text)  # Use the chunk_text function from web_utils
+      embeddings = model.encode(chunks)
+      for (chunk, embedding) in zip(chunks, embeddings):
+        item = {}
+        item['source'] = chunk
+        item['embedding'] = embedding.tolist()  # Convert NumPy array to list
+        item['sourcelength'] = len(chunk)
+        article['embeddings'].append(item)
+    
+      allEmbeddings.append(article)
+
+      print(f"{summary}\n")
+
+    
+    while True:
+      context = []
+      # Input a question from the user
+      question = input("Enter your question about the news, or type quit: ")
+
+      if question.lower() == 'quit':
+        break
+
+      # Embed the user's question
+      question_embedding = model.encode([question])
+
+      # Perform KNN search to find the best matches (indices and source text)
+      best_matches = knn_search(question_embedding, allEmbeddings, k=10)
+
+
+      sourcetext=""
+      for i, (index, source_text) in enumerate(best_matches, start=1):
+          sourcetext += f"{i}. Index: {index}, Source Text: {source_text}"
+
+      systemPrompt = f"Only use the following information to answer the question. Do not use anything else: {sourcetext}"
+
+      url = "http://localhost:11434/api/generate"
+
+      payload = {
+      "model": "mistral-openorca",
+      "prompt": question, 
+      "system": systemPrompt,
+      "stream": False, 
+      "context": context
+      }
+
+      # Convert the payload to a JSON string
+      payload_json = json.dumps(payload)
+
+      # Set the headers to specify JSON content
+      headers = {
+          "Content-Type": "application/json"
+      }
+
+      # Send the POST request
+      response = requests.post(url, data=payload_json, headers=headers)
+
+      # Check the response
+      if response.status_code == 200:
+          output = json.loads(response.text)
+          context = output['context']
+          print(output['response']+ "\n")
+          
+
+      else:
+          print(f"Request failed with status code {response.status_code}")
+

examples/python-rag-newssummary/utils.py

@@ -0,0 +1,108 @@
+import curses
+import feedparser
+import requests
+import unicodedata
+import json
+from newspaper import Article
+from bs4 import BeautifulSoup
+from nltk.tokenize import sent_tokenize, word_tokenize
+import numpy as np
+from sklearn.neighbors import NearestNeighbors
+from mattsollamatools import chunker
+
+# Create a dictionary to store topics and their URLs
+topic_urls = {
+    "Mac": "https://9to5mac.com/guides/mac/feed",
+    "News": "http://www.npr.org/rss/rss.php?id=1001",
+    "Nvidia": "https://nvidianews.nvidia.com/releases.xml",
+    "Raspberry Pi": "https://www.raspberrypi.com/news/feed/", 
+    "Music": "https://www.billboard.com/c/music/music-news/feed/"
+}
+
+# Use curses to create a menu of topics
+def menu(stdscr):
+    chosen_topic = get_url_for_topic(stdscr)  
+    url = topic_urls[chosen_topic] if chosen_topic in topic_urls else "Topic not found"
+    
+    stdscr.addstr(len(topic_urls) + 3, 0, f"Selected URL for {chosen_topic}: {url}")
+    stdscr.refresh()
+    
+    return chosen_topic
+
+# You have chosen a topic. Now return the url for that topic
+def get_url_for_topic(stdscr):
+    curses.curs_set(0)  # Hide the cursor
+    stdscr.clear()
+
+    stdscr.addstr(0, 0, "Choose a topic using the arrow keys (Press Enter to select):")
+
+    # Create a list of topics
+    topics = list(topic_urls.keys())
+    current_topic = 0
+
+    while True:
+        for i, topic in enumerate(topics):
+            if i == current_topic:
+                stdscr.addstr(i + 2, 2, f"> {topic}")
+            else:
+                stdscr.addstr(i + 2, 2, f"  {topic}")
+
+        stdscr.refresh()
+
+        key = stdscr.getch()
+
+        if key == curses.KEY_DOWN and current_topic < len(topics) - 1:
+            current_topic += 1
+        elif key == curses.KEY_UP and current_topic > 0:
+            current_topic -= 1
+        elif key == 10:  # Enter key
+            return topic_urls[topics[current_topic]]
+
+# Get the last N URLs from an RSS feed
+def getUrls(feed_url, n=20):
+    feed = feedparser.parse(feed_url)
+    entries = feed.entries[-n:]
+    urls = [entry.link for entry in entries]
+    return urls
+
+# Often there are a bunch of ads and menus on pages for a news article. This uses newspaper3k to get just the text of just the article.
+def getArticleText(url):
+  article = Article(url)
+  article.download()
+  article.parse()
+  return article.text
+
+def get_summary(text):
+  systemPrompt = "Write a concise summary of the text, return your responses with 5 lines that cover the key points of the text given."
+  prompt = text
+  
+  url = "http://localhost:11434/api/generate"
+
+  payload = {
+    "model": "mistral-openorca",
+    "prompt": prompt, 
+    "system": systemPrompt,
+    "stream": False
+  }
+  payload_json = json.dumps(payload)
+  headers = {"Content-Type": "application/json"}
+  response = requests.post(url, data=payload_json, headers=headers)
+
+  return json.loads(response.text)["response"]
+
+# Perform K-nearest neighbors (KNN) search
+def knn_search(question_embedding, embeddings, k=5):
+    X = np.array([item['embedding'] for article in embeddings for item in article['embeddings']])
+    source_texts = [item['source'] for article in embeddings for item in article['embeddings']]
+    
+    # Fit a KNN model on the embeddings
+    knn = NearestNeighbors(n_neighbors=k, metric='cosine')
+    knn.fit(X)
+    
+    # Find the indices and distances of the k-nearest neighbors
+    distances, indices = knn.kneighbors(question_embedding, n_neighbors=k)
+    
+    # Get the indices and source texts of the best matches
+    best_matches = [(indices[0][i], source_texts[indices[0][i]]) for i in range(k)]
+    
+    return best_matches

examples/typescript-mentors/.gitignore

@@ -0,0 +1,2 @@
+node_modules
+package-lock.json

examples/typescript-mentors/README.md

@@ -0,0 +1,21 @@
+# Ask the Mentors
+
+This example demonstrates how one would create a set of 'mentors' you can have a conversation with. The mentors are generated using the `character-generator.ts` file. This will use **Stable Beluga 70b** to create a bio and list of verbal ticks and common phrases used by each person. Then `mentors.ts` will take a question, and choose three of the 'mentors' and start a conversation with them. Occasionally, they will talk to each other, and other times they will just deliver a set of monologues. It's fun to see what they do and say.
+
+## Usage
+
+```bash
+ts-node ./character-generator.ts "Lorne Greene"
+```
+
+This will create `lornegreene/Modelfile`. Now you can create a model with this command:
+
+```bash
+ollama create lornegreene -f lornegreene/Modelfile
+```
+
+If you want to add your own mentors, you will have to update the code to look at your namespace instead of **mattw**. Also set the list of mentors to include yours.
+
+```bash
+ts-node ./mentors.ts "What is a Jackalope?"
+```

examples/typescript-mentors/character-generator.ts

@@ -0,0 +1,26 @@
+import { Ollama } from 'ollama-node'
+import fs from 'fs';
+import path from 'path';
+
+async function characterGenerator() {
+  const character = process.argv[2];
+  console.log(`You are creating a character for ${character}.`);
+  const foldername = character.replace(/\s/g, '').toLowerCase();
+  const directory = path.join(__dirname, foldername);
+  if (!fs.existsSync(directory)) {
+    fs.mkdirSync(directory, { recursive: true });
+  }
+
+  const ollama = new Ollama();
+  ollama.setModel("stablebeluga2:70b-q4_K_M");
+  const bio = await ollama.generate(`create a bio of ${character} in a single long paragraph. Instead of saying '${character} is...' or '${character} was...' use language like 'You are...' or 'You were...'. Then create a paragraph describing the speaking mannerisms and style of ${character}. Don't include anything about how ${character} looked or what they sounded like, just focus on the words they said. Instead of saying '${character} would say...' use language like 'You should say...'. If you use quotes, always use single quotes instead of double quotes. If there are any specific words or phrases you used a lot, show how you used them. `);
+
+  const thecontents = `FROM llama2\nSYSTEM """\n${bio.response.replace(/(\r\n|\n|\r)/gm, " ").replace('would', 'should')} All answers to questions should be related back to what you are most known for.\n"""`;
+
+  fs.writeFile(path.join(directory, 'Modelfile'), thecontents, (err: any) => {
+    if (err) throw err;
+    console.log('The file has been saved!');
+  });
+}
+
+characterGenerator();

examples/typescript-mentors/mentors.ts

@@ -0,0 +1,59 @@
+import { Ollama } from 'ollama-node';
+
+const mentorCount = 3;
+const ollama = new Ollama();
+
+function getMentors(): string[] {
+  const mentors = ['Gary Vaynerchuk', 'Kanye West', 'Martha Stewart', 'Neil deGrasse Tyson', 'Owen Wilson', 'Ronald Reagan', 'Donald Trump', 'Barack Obama', 'Jeff Bezos'];
+  const chosenMentors: string[] = [];
+  for (let i = 0; i < mentorCount; i++) {
+    const mentor = mentors[Math.floor(Math.random() * mentors.length)];
+    chosenMentors.push(mentor);
+    mentors.splice(mentors.indexOf(mentor), 1);
+  }
+  return chosenMentors;
+}
+
+function getMentorFileName(mentor: string): string {
+  const model = mentor.toLowerCase().replace(/\s/g, '');
+  return `mattw/${model}`;
+}
+
+async function getSystemPrompt(mentor: string, isLast: boolean, question: string): Promise<string> {
+  ollama.setModel(getMentorFileName(mentor));
+  const info = await ollama.showModelInfo()
+  let SystemPrompt = info.system || '';
+  SystemPrompt += ` You should continue the conversation as if you were ${mentor} and acknowledge the people before you in the conversation. You should adopt their mannerisms and tone, but also not use language they wouldn't use. If they are not known to know about the concept in the question, don't offer an answer. Your answer should be no longer than 1 paragraph. And definitely try not to sound like anyone else. Don't repeat any slang or phrases already used. And if it is a question the original ${mentor} wouldn't have know the answer to, just say that you don't know, in the style of ${mentor}. And think about the time the person lived. Don't use terminology that they wouldn't have used.`
+
+  if (isLast) {
+    SystemPrompt += ` End your answer with something like I hope our answers help you out`;
+  } else {
+    SystemPrompt += ` Remember, this is a conversation, so you don't need a conclusion, but end your answer with a question related to the first question: "${question}".`;
+  }
+  return SystemPrompt;
+}
+
+async function main() {
+  const mentors = getMentors();
+  const question = process.argv[2];
+  let theConversation = `Here is the conversation so far.\nYou: ${question}\n`
+
+  for await (const mentor of mentors) {
+    const SystemPrompt = await getSystemPrompt(mentor, mentor === mentors[mentorCount - 1], question);
+    ollama.setModel(getMentorFileName(mentor));
+    ollama.setSystemPrompt(SystemPrompt);
+    let output = '';
+    process.stdout.write(`\n${mentor}: `);
+    for await (const chunk of ollama.streamingGenerate(theConversation + `Continue the conversation as if you were ${mentor} on the question "${question}".`)) {
+      if (chunk.response) {
+        output += chunk.response;
+        process.stdout.write(chunk.response);
+      } else {
+        process.stdout.write('\n');
+      }
+    }
+    theConversation += `${mentor}: ${output}\n\n`
+  }
+}
+
+main();

examples/typescript-mentors/package.json

@@ -0,0 +1,7 @@
+{
+  "dependencies": {
+    "fs": "^0.0.1-security",
+    "ollama-node": "^0.0.3",
+    "path": "^0.12.7"
+  }
+}

format/bytes.go

@@ -0,0 +1,23 @@
+package format
+
+import "fmt"
+
+const (
+	Byte     = 1
+	KiloByte = Byte * 1000
+	MegaByte = KiloByte * 1000
+	GigaByte = MegaByte * 1000
+)
+
+func HumanBytes(b int64) string {
+	switch {
+	case b > GigaByte:
+		return fmt.Sprintf("%d GB", b/GigaByte)
+	case b > MegaByte:
+		return fmt.Sprintf("%d MB", b/MegaByte)
+	case b > KiloByte:
+		return fmt.Sprintf("%d KB", b/KiloByte)
+	default:
+		return fmt.Sprintf("%d B", b)
+	}
+}

format/time.go

@@ -7,26 +7,14 @@ import (
 	"time"
 )
 
-// HumanDuration returns a human-readable approximation of a duration
-// (eg. "About a minute", "4 hours ago", etc.).
-// Modified version of github.com/docker/go-units.HumanDuration
-func HumanDuration(d time.Duration) string {
-	return HumanDurationWithCase(d, true)
-}
-
-// HumanDurationWithCase returns a human-readable approximation of a
-// duration (eg. "About a minute", "4 hours ago", etc.). but allows
-// you to specify whether the first word should be capitalized
-// (eg. "About" vs. "about")
-func HumanDurationWithCase(d time.Duration, useCaps bool) string {
+// humanDuration returns a human-readable approximation of a
+// duration (eg. "About a minute", "4 hours ago", etc.).
+func humanDuration(d time.Duration) string {
 	seconds := int(d.Seconds())
 
 	switch {
 	case seconds < 1:
-		if useCaps {
-			return "Less than a second"
-		}
-		return "less than a second"
+		return "Less than a second"
 	case seconds == 1:
 		return "1 second"
 	case seconds < 60:
@@ -36,10 +24,7 @@ func HumanDurationWithCase(d time.Duration, useCaps bool) string {
 	minutes := int(d.Minutes())
 	switch {
 	case minutes == 1:
-		if useCaps {
-			return "About a minute"
-		}
-		return "about a minute"
+		return "About a minute"
 	case minutes < 60:
 		return fmt.Sprintf("%d minutes", minutes)
 	}
@@ -47,10 +32,7 @@ func HumanDurationWithCase(d time.Duration, useCaps bool) string {
 	hours := int(math.Round(d.Hours()))
 	switch {
 	case hours == 1:
-		if useCaps {
-			return "About an hour"
-		}
-		return "about an hour"
+		return "About an hour"
 	case hours < 48:
 		return fmt.Sprintf("%d hours", hours)
 	case hours < 24*7*2:
@@ -65,77 +47,22 @@ func HumanDurationWithCase(d time.Duration, useCaps bool) string {
 }
 
 func HumanTime(t time.Time, zeroValue string) string {
-	return humanTimeWithCase(t, zeroValue, true)
+	return humanTime(t, zeroValue)
 }
 
 func HumanTimeLower(t time.Time, zeroValue string) string {
-	return humanTimeWithCase(t, zeroValue, false)
+	return strings.ToLower(humanTime(t, zeroValue))
 }
 
-func humanTimeWithCase(t time.Time, zeroValue string, useCaps bool) string {
+func humanTime(t time.Time, zeroValue string) string {
 	if t.IsZero() {
 		return zeroValue
 	}
 
 	delta := time.Since(t)
 	if delta < 0 {
-		return HumanDurationWithCase(-delta, useCaps) + " from now"
+		return humanDuration(-delta) + " from now"
 	}
-	return HumanDurationWithCase(delta, useCaps) + " ago"
-}
-
-// ExcatDuration returns a human readable hours/minutes/seconds or milliseconds format of a duration
-// the most precise level of duration is milliseconds
-func ExactDuration(d time.Duration) string {
-	if d.Seconds() < 1 {
-		if d.Milliseconds() == 1 {
-			return fmt.Sprintf("%d millisecond", d.Milliseconds())
-		}
-		return fmt.Sprintf("%d milliseconds", d.Milliseconds())
-	}
-
-	var readableDur strings.Builder
-
-	dur := d.String()
-
-	// split the default duration string format of 0h0m0s into something nicer to read
-	h := strings.Split(dur, "h")
-	if len(h) > 1 {
-		hours := h[0]
-		if hours == "1" {
-			readableDur.WriteString(fmt.Sprintf("%s hour ", hours))
-		} else {
-			readableDur.WriteString(fmt.Sprintf("%s hours ", hours))
-		}
-		dur = h[1]
-	}
-
-	m := strings.Split(dur, "m")
-	if len(m) > 1 {
-		mins := m[0]
-		switch mins {
-		case "0":
-			// skip
-		case "1":
-			readableDur.WriteString(fmt.Sprintf("%s minute ", mins))
-		default:
-			readableDur.WriteString(fmt.Sprintf("%s minutes ", mins))
-		}
-		dur = m[1]
-	}
-
-	s := strings.Split(dur, "s")
-	if len(s) > 0 {
-		sec := s[0]
-		switch sec {
-		case "0":
-			// skip
-		case "1":
-			readableDur.WriteString(fmt.Sprintf("%s second ", sec))
-		default:
-			readableDur.WriteString(fmt.Sprintf("%s seconds ", sec))
-		}
-	}
-
-	return strings.TrimSpace(readableDur.String())
+
+	return humanDuration(delta) + " ago"
 }

format/time_test.go

@@ -11,92 +11,25 @@ func assertEqual(t *testing.T, a interface{}, b interface{}) {
 	}
 }
 
-func TestHumanDuration(t *testing.T) {
-	day := 24 * time.Hour
-	week := 7 * day
-	month := 30 * day
-	year := 365 * day
-
-	assertEqual(t, "Less than a second", HumanDuration(450*time.Millisecond))
-	assertEqual(t, "Less than a second", HumanDurationWithCase(450*time.Millisecond, true))
-	assertEqual(t, "less than a second", HumanDurationWithCase(450*time.Millisecond, false))
-	assertEqual(t, "1 second", HumanDuration(1*time.Second))
-	assertEqual(t, "45 seconds", HumanDuration(45*time.Second))
-	assertEqual(t, "46 seconds", HumanDuration(46*time.Second))
-	assertEqual(t, "59 seconds", HumanDuration(59*time.Second))
-	assertEqual(t, "About a minute", HumanDuration(60*time.Second))
-	assertEqual(t, "About a minute", HumanDurationWithCase(1*time.Minute, true))
-	assertEqual(t, "about a minute", HumanDurationWithCase(1*time.Minute, false))
-	assertEqual(t, "3 minutes", HumanDuration(3*time.Minute))
-	assertEqual(t, "35 minutes", HumanDuration(35*time.Minute))
-	assertEqual(t, "35 minutes", HumanDuration(35*time.Minute+40*time.Second))
-	assertEqual(t, "45 minutes", HumanDuration(45*time.Minute))
-	assertEqual(t, "45 minutes", HumanDuration(45*time.Minute+40*time.Second))
-	assertEqual(t, "46 minutes", HumanDuration(46*time.Minute))
-	assertEqual(t, "59 minutes", HumanDuration(59*time.Minute))
-	assertEqual(t, "About an hour", HumanDuration(1*time.Hour))
-	assertEqual(t, "About an hour", HumanDurationWithCase(1*time.Hour+29*time.Minute, true))
-	assertEqual(t, "about an hour", HumanDurationWithCase(1*time.Hour+29*time.Minute, false))
-	assertEqual(t, "2 hours", HumanDuration(1*time.Hour+31*time.Minute))
-	assertEqual(t, "2 hours", HumanDuration(1*time.Hour+59*time.Minute))
-	assertEqual(t, "3 hours", HumanDuration(3*time.Hour))
-	assertEqual(t, "3 hours", HumanDuration(3*time.Hour+29*time.Minute))
-	assertEqual(t, "4 hours", HumanDuration(3*time.Hour+31*time.Minute))
-	assertEqual(t, "4 hours", HumanDuration(3*time.Hour+59*time.Minute))
-	assertEqual(t, "4 hours", HumanDuration(3*time.Hour+60*time.Minute))
-	assertEqual(t, "24 hours", HumanDuration(24*time.Hour))
-	assertEqual(t, "36 hours", HumanDuration(1*day+12*time.Hour))
-	assertEqual(t, "2 days", HumanDuration(2*day))
-	assertEqual(t, "7 days", HumanDuration(7*day))
-	assertEqual(t, "13 days", HumanDuration(13*day+5*time.Hour))
-	assertEqual(t, "2 weeks", HumanDuration(2*week))
-	assertEqual(t, "2 weeks", HumanDuration(2*week+4*day))
-	assertEqual(t, "3 weeks", HumanDuration(3*week))
-	assertEqual(t, "4 weeks", HumanDuration(4*week))
-	assertEqual(t, "4 weeks", HumanDuration(4*week+3*day))
-	assertEqual(t, "4 weeks", HumanDuration(1*month))
-	assertEqual(t, "6 weeks", HumanDuration(1*month+2*week))
-	assertEqual(t, "2 months", HumanDuration(2*month))
-	assertEqual(t, "2 months", HumanDuration(2*month+2*week))
-	assertEqual(t, "3 months", HumanDuration(3*month))
-	assertEqual(t, "3 months", HumanDuration(3*month+1*week))
-	assertEqual(t, "5 months", HumanDuration(5*month+2*week))
-	assertEqual(t, "13 months", HumanDuration(13*month))
-	assertEqual(t, "23 months", HumanDuration(23*month))
-	assertEqual(t, "24 months", HumanDuration(24*month))
-	assertEqual(t, "2 years", HumanDuration(24*month+2*week))
-	assertEqual(t, "3 years", HumanDuration(3*year+2*month))
-}
-
 func TestHumanTime(t *testing.T) {
 	now := time.Now()
 
 	t.Run("zero value", func(t *testing.T) {
 		assertEqual(t, HumanTime(time.Time{}, "never"), "never")
 	})
+
 	t.Run("time in the future", func(t *testing.T) {
 		v := now.Add(48 * time.Hour)
 		assertEqual(t, HumanTime(v, ""), "2 days from now")
 	})
+
 	t.Run("time in the past", func(t *testing.T) {
 		v := now.Add(-48 * time.Hour)
 		assertEqual(t, HumanTime(v, ""), "2 days ago")
 	})
-}
 
-func TestExactDuration(t *testing.T) {
-	assertEqual(t, "1 millisecond", ExactDuration(1*time.Millisecond))
-	assertEqual(t, "10 milliseconds", ExactDuration(10*time.Millisecond))
-	assertEqual(t, "1 second", ExactDuration(1*time.Second))
-	assertEqual(t, "10 seconds", ExactDuration(10*time.Second))
-	assertEqual(t, "1 minute", ExactDuration(1*time.Minute))
-	assertEqual(t, "10 minutes", ExactDuration(10*time.Minute))
-	assertEqual(t, "1 hour", ExactDuration(1*time.Hour))
-	assertEqual(t, "10 hours", ExactDuration(10*time.Hour))
-	assertEqual(t, "1 hour 1 second", ExactDuration(1*time.Hour+1*time.Second))
-	assertEqual(t, "1 hour 10 seconds", ExactDuration(1*time.Hour+10*time.Second))
-	assertEqual(t, "1 hour 1 minute", ExactDuration(1*time.Hour+1*time.Minute))
-	assertEqual(t, "1 hour 10 minutes", ExactDuration(1*time.Hour+10*time.Minute))
-	assertEqual(t, "1 hour 1 minute 1 second", ExactDuration(1*time.Hour+1*time.Minute+1*time.Second))
-	assertEqual(t, "10 hours 10 minutes 10 seconds", ExactDuration(10*time.Hour+10*time.Minute+10*time.Second))
+	t.Run("soon", func(t *testing.T) {
+		v := now.Add(800 * time.Millisecond)
+		assertEqual(t, HumanTime(v, ""), "Less than a second from now")
+	})
 }

go.mod

@@ -8,7 +8,9 @@ require (
 	github.com/mattn/go-runewidth v0.0.14
 	github.com/mitchellh/colorstring v0.0.0-20190213212951-d06e56a500db
 	github.com/olekukonko/tablewriter v0.0.5
+	github.com/pdevine/readline v1.5.2
 	github.com/spf13/cobra v1.7.0
+	golang.org/x/sync v0.3.0
 )
 
 require github.com/rivo/uniseg v0.2.0 // indirect
@@ -16,7 +18,6 @@ require github.com/rivo/uniseg v0.2.0 // indirect
 require (
 	github.com/bytedance/sonic v1.9.1 // indirect
 	github.com/chenzhuoyu/base64x v0.0.0-20221115062448-fe3a3abad311 // indirect
-	github.com/chzyer/readline v1.5.1
 	github.com/gabriel-vasile/mimetype v1.4.2 // indirect
 	github.com/gin-contrib/cors v1.4.0
 	github.com/gin-contrib/sse v0.1.0 // indirect
@@ -44,7 +45,6 @@ require (
 	golang.org/x/sys v0.11.0 // indirect
 	golang.org/x/term v0.10.0
 	golang.org/x/text v0.10.0 // indirect
-	gonum.org/v1/gonum v0.13.0
 	google.golang.org/protobuf v1.30.0 // indirect
 	gopkg.in/yaml.v3 v3.0.1 // indirect
 )

go.sum

@@ -6,8 +6,6 @@ github.com/chenzhuoyu/base64x v0.0.0-20221115062448-fe3a3abad311 h1:qSGYFH7+jGhD
 github.com/chenzhuoyu/base64x v0.0.0-20221115062448-fe3a3abad311/go.mod h1:b583jCggY9gE99b6G5LEC39OIiVsWj+R97kbl5odCEk=
 github.com/chzyer/logex v1.2.1 h1:XHDu3E6q+gdHgsdTPH6ImJMIp436vR6MPtH8gP05QzM=
 github.com/chzyer/logex v1.2.1/go.mod h1:JLbx6lG2kDbNRFnfkgvh4eRJRPX1QCoOIWomwysCBrQ=
-github.com/chzyer/readline v1.5.1 h1:upd/6fQk4src78LMRzh5vItIt361/o4uq553V8B5sGI=
-github.com/chzyer/readline v1.5.1/go.mod h1:Eh+b79XXUwfKfcPLepksvw2tcLE/Ct21YObkaSkeBlk=
 github.com/chzyer/test v1.0.0 h1:p3BQDXSxOhOG0P9z6/hGnII4LGiEPOYBhs8asl/fC04=
 github.com/chzyer/test v1.0.0/go.mod h1:2JlltgoNkt4TW/z9V/IzDdFaMTM2JPIi26O1pF38GC8=
 github.com/cpuguy83/go-md2man/v2 v2.0.2/go.mod h1:tgQtvFlXSQOSOSIRvRPT7W67SCa46tRHOmNcaadrF8o=
@@ -80,6 +78,8 @@ github.com/olekukonko/tablewriter v0.0.5 h1:P2Ga83D34wi1o9J6Wh1mRuqd4mF/x/lgBS7N
 github.com/olekukonko/tablewriter v0.0.5/go.mod h1:hPp6KlRPjbx+hW8ykQs1w3UBbZlj6HuIJcUGPhkA7kY=
 github.com/pbnjay/memory v0.0.0-20210728143218-7b4eea64cf58 h1:onHthvaw9LFnH4t2DcNVpwGmV9E1BkGknEliJkfwQj0=
 github.com/pbnjay/memory v0.0.0-20210728143218-7b4eea64cf58/go.mod h1:DXv8WO4yhMYhSNPKjeNKa5WY9YCIEBRbNzFFPJbWO6Y=
+github.com/pdevine/readline v1.5.2 h1:oz6Y5GdTmhPG+08hhxcAvtHitSANWuA2100Sppb38xI=
+github.com/pdevine/readline v1.5.2/go.mod h1:na/LbuE5PYwxI7GyopWdIs3U8HVe89lYlNTFTXH3wOw=
 github.com/pelletier/go-toml/v2 v2.0.1/go.mod h1:r9LEWfGN8R5k0VXJ+0BkIe7MYkRdwZOjgMj2KwnJFUo=
 github.com/pelletier/go-toml/v2 v2.0.8 h1:0ctb6s9mE31h0/lhu+J6OPmVeDxJn+kYnJc2jZR9tGQ=
 github.com/pelletier/go-toml/v2 v2.0.8/go.mod h1:vuYfssBdrU2XDZ9bYydBu6t+6a6PYNcZljzZR9VXg+4=
@@ -120,12 +120,13 @@ golang.org/x/arch v0.3.0/go.mod h1:5om86z9Hs0C8fWVUuoMHwpExlXzs5Tkyp9hOrfG7pp8=
 golang.org/x/crypto v0.0.0-20210711020723-a769d52b0f97/go.mod h1:GvvjBRRGRdwPK5ydBHafDWAxML/pGHZbMvKqRZ5+Abc=
 golang.org/x/crypto v0.10.0 h1:LKqV2xt9+kDzSTfOhx4FrkEBcMrAgHSYgzywV9zcGmM=
 golang.org/x/crypto v0.10.0/go.mod h1:o4eNf7Ede1fv+hwOwZsTHl9EsPFO6q6ZvYR8vYfY45I=
-golang.org/x/exp v0.0.0-20230321023759-10a507213a29 h1:ooxPy7fPvB4kwsA2h+iBNHkAbp/4JxTSwCmvdjEYmug=
 golang.org/x/exp v0.0.0-20230817173708-d852ddb80c63 h1:m64FZMko/V45gv0bNmrNYoDEq8U5YUhetc9cBWKS1TQ=
 golang.org/x/exp v0.0.0-20230817173708-d852ddb80c63/go.mod h1:0v4NqG35kSWCMzLaMeX+IQrlSnVE/bqGSyC2cz/9Le8=
 golang.org/x/net v0.0.0-20210226172049-e18ecbb05110/go.mod h1:m0MpNAwzfU5UDzcl9v0D8zg8gWTRqZa9RBIspLL5mdg=
 golang.org/x/net v0.10.0 h1:X2//UzNDwYmtCLn7To6G58Wr6f5ahEAQgKNzv9Y951M=
 golang.org/x/net v0.10.0/go.mod h1:0qNGK6F8kojg2nk9dLZ2mShWaEBan6FAoqfSigmmuDg=
+golang.org/x/sync v0.3.0 h1:ftCYgMx6zT/asHUrPw8BLLscYtGznsLAnjq5RH9P66E=
+golang.org/x/sync v0.3.0/go.mod h1:FU7BRWz2tNW+3quACPkgCx/L+uEAv1htQ0V83Z9Rj+Y=
 golang.org/x/sys v0.0.0-20201119102817-f84b799fce68/go.mod h1:h1NjWce9XRLGQEsW7wpKNCjG9DtNlClVuFLEZdDNbEs=
 golang.org/x/sys v0.0.0-20210615035016-665e8c7367d1/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
 golang.org/x/sys v0.0.0-20210630005230-0f9fa26af87c/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
@@ -144,8 +145,6 @@ golang.org/x/text v0.10.0 h1:UpjohKhiEgNc0CSauXmwYftY1+LlaC75SJwh0SgCX58=
 golang.org/x/text v0.10.0/go.mod h1:TvPlkZtksWOMsz7fbANvkp4WM8x/WCo/om8BMLbz+aE=
 golang.org/x/tools v0.0.0-20180917221912-90fa682c2a6e/go.mod h1:n7NCudcB/nEzxVGmLbDWY5pfWTLqBcC2KZ6jyYvM4mQ=
 golang.org/x/xerrors v0.0.0-20191204190536-9bdfabe68543/go.mod h1:I/5z698sn9Ka8TeJc9MKroUUfqBBauWjQqLJ2OPfmY0=
-gonum.org/v1/gonum v0.13.0 h1:a0T3bh+7fhRyqeNbiC3qVHYmkiQgit3wnNan/2c0HMM=
-gonum.org/v1/gonum v0.13.0/go.mod h1:/WPYRckkfWrhWefxyYTfrTtQR0KH4iyHNuzxqXAKyAU=
 google.golang.org/protobuf v1.26.0-rc.1/go.mod h1:jlhhOSvTdKEhbULTjvd4ARK9grFBp09yW+WbY/TyQbw=
 google.golang.org/protobuf v1.28.0/go.mod h1:HV8QOd/L58Z+nl8r43ehVNZIU/HEI6OcFqwMG9pJV4I=
 google.golang.org/protobuf v1.30.0 h1:kPPoIgf3TsEvrm0PFe15JQ+570QVxYzEvvHqChK+cng=

llm/falcon.go

@@ -0,0 +1,20 @@
+package llm
+
+const (
+	falconModelType7B   = 32
+	falconModelType40B  = 60
+	falconModelType180B = 80
+)
+
+func falconModelType(numLayer uint32) string {
+	switch numLayer {
+	case 32:
+		return "7B"
+	case 60:
+		return "40B"
+	case 80:
+		return "180B"
+	default:
+		return "unknown"
+	}
+}

llm/ggml.go

@@ -3,72 +3,96 @@ package llm
 import (
 	"encoding/binary"
 	"errors"
-	"fmt"
 	"io"
 )
 
-type ModelFamily string
-
-type ModelType uint32
-
-const (
-	ModelType3B  ModelType = 26
-	ModelType7B  ModelType = 32
-	ModelType13B ModelType = 40
-	ModelType34B ModelType = 48
-	ModelType30B ModelType = 60
-	ModelType65B ModelType = 80
-)
-
-func (mt ModelType) String() string {
-	switch mt {
-	case ModelType3B:
-		return "3B"
-	case ModelType7B:
-		return "7B"
-	case ModelType13B:
-		return "13B"
-	case ModelType34B:
-		return "34B"
-	case ModelType30B:
-		return "30B"
-	case ModelType65B:
-		return "65B"
-	default:
-		return "Unknown"
-	}
-}
-
-type FileType interface {
-	String() string
-}
-
 type GGML struct {
 	magic uint32
 	container
 	model
 }
 
+const (
+	fileTypeF32 uint32 = iota
+	fileTypeF16
+	fileTypeQ4_0
+	fileTypeQ4_1
+	fileTypeQ4_1_F16
+	fileTypeQ8_0 uint32 = iota + 2
+	fileTypeQ5_0
+	fileTypeQ5_1
+	fileTypeQ2_K
+	fileTypeQ3_K_S
+	fileTypeQ3_K_M
+	fileTypeQ3_K_L
+	fileTypeQ4_K_S
+	fileTypeQ4_K_M
+	fileTypeQ5_K_S
+	fileTypeQ5_K_M
+	fileTypeQ6_K
+)
+
+func fileType(fileType uint32) string {
+	switch fileType {
+	case fileTypeF32:
+		return "F32"
+	case fileTypeF16:
+		return "F16"
+	case fileTypeQ4_0:
+		return "Q4_0"
+	case fileTypeQ4_1:
+		return "Q4_1"
+	case fileTypeQ4_1_F16:
+		return "Q4_1_F16"
+	case fileTypeQ8_0:
+		return "Q8_0"
+	case fileTypeQ5_0:
+		return "Q5_0"
+	case fileTypeQ5_1:
+		return "Q5_1"
+	case fileTypeQ2_K:
+		return "Q2_K"
+	case fileTypeQ3_K_S:
+		return "Q3_K_S"
+	case fileTypeQ3_K_M:
+		return "Q3_K_M"
+	case fileTypeQ3_K_L:
+		return "Q3_K_L"
+	case fileTypeQ4_K_S:
+		return "Q4_K_S"
+	case fileTypeQ4_K_M:
+		return "Q4_K_M"
+	case fileTypeQ5_K_S:
+		return "Q5_K_S"
+	case fileTypeQ5_K_M:
+		return "Q5_K_M"
+	case fileTypeQ6_K:
+		return "Q6_K"
+	default:
+		return "unknown"
+	}
+}
+
 type model interface {
-	ModelFamily() ModelFamily
-	ModelType() ModelType
-	FileType() FileType
+	ModelFamily() string
+	ModelType() string
+	FileType() string
+	NumLayers() int64
 }
 
 type container interface {
 	Name() string
-	Decode(io.Reader) error
+	Decode(io.Reader) (model, error)
 }
 
-type containerGGML struct {
-}
+type containerGGML struct{}
 
 func (c *containerGGML) Name() string {
 	return "ggml"
 }
 
-func (c *containerGGML) Decode(r io.Reader) error {
-	return nil
+func (c *containerGGML) Decode(r io.Reader) (model, error) {
+	return nil, nil
 }
 
 type containerGGMF struct {
@@ -79,18 +103,18 @@ func (c *containerGGMF) Name() string {
 	return "ggmf"
 }
 
-func (c *containerGGMF) Decode(r io.Reader) error {
+func (c *containerGGMF) Decode(r io.Reader) (model, error) {
 	var version uint32
 	binary.Read(r, binary.LittleEndian, &version)
 
 	switch version {
 	case 1:
 	default:
-		return errors.New("invalid version")
+		return nil, errors.New("invalid version")
 	}
 
 	c.version = version
-	return nil
+	return nil, nil
 }
 
 type containerGGJT struct {
@@ -101,18 +125,22 @@ func (c *containerGGJT) Name() string {
 	return "ggjt"
 }
 
-func (c *containerGGJT) Decode(r io.Reader) error {
+func (c *containerGGJT) Decode(r io.Reader) (model, error) {
 	var version uint32
 	binary.Read(r, binary.LittleEndian, &version)
 
 	switch version {
 	case 1, 2, 3:
 	default:
-		return errors.New("invalid version")
+		return nil, errors.New("invalid version")
 	}
 
 	c.version = version
-	return nil
+
+	// different model types may have different layouts for hyperparameters
+	var llama llamaModel
+	binary.Read(r, binary.LittleEndian, &llama.hyperparameters)
+	return &llama, nil
 }
 
 type containerLORA struct {
@@ -123,32 +151,34 @@ func (c *containerLORA) Name() string {
 	return "ggla"
 }
 
-func (c *containerLORA) Decode(r io.Reader) error {
+func (c *containerLORA) Decode(r io.Reader) (model, error) {
 	var version uint32
 	binary.Read(r, binary.LittleEndian, &version)
 
 	switch version {
 	case 1:
 	default:
-		return errors.New("invalid version")
+		return nil, errors.New("invalid version")
 	}
 
 	c.version = version
-	return nil
+	return nil, nil
 }
 
 const (
-	// / Magic constant for `ggml` files (unversioned).
+	// Magic constant for `ggml` files (unversioned).
 	FILE_MAGIC_GGML = 0x67676d6c
-	// / Magic constant for `ggml` files (versioned, ggmf).
+	// Magic constant for `ggml` files (versioned, ggmf).
 	FILE_MAGIC_GGMF = 0x67676d66
-	// / Magic constant for `ggml` files (versioned, ggjt).
+	// Magic constant for `ggml` files (versioned, ggjt).
 	FILE_MAGIC_GGJT = 0x67676a74
-	// / Magic constant for `ggla` files (LoRA adapter).
+	// Magic constant for `ggla` files (LoRA adapter).
 	FILE_MAGIC_GGLA = 0x67676C61
+	// Magic constant for `gguf` files (versioned, gguf)
+	FILE_MAGIC_GGUF = 0x46554747
 )
 
-func DecodeGGML(r io.ReadSeeker, hint ModelFamily) (*GGML, error) {
+func DecodeGGML(r io.ReadSeeker) (*GGML, error) {
 	var ggml GGML
 	binary.Read(r, binary.LittleEndian, &ggml.magic)
 
@@ -161,24 +191,18 @@ func DecodeGGML(r io.ReadSeeker, hint ModelFamily) (*GGML, error) {
 		ggml.container = &containerGGJT{}
 	case FILE_MAGIC_GGLA:
 		ggml.container = &containerLORA{}
+	case FILE_MAGIC_GGUF:
+		ggml.container = &containerGGUF{}
 	default:
 		return nil, errors.New("invalid file magic")
 	}
 
-	if err := ggml.Decode(r); err != nil {
+	model, err := ggml.Decode(r)
+	if err != nil {
 		return nil, err
 	}
 
-	// different model types may have different layouts for hyperparameters
-	switch hint {
-	case ModelFamilyLlama:
-		var llama llamaModel
-		binary.Read(r, binary.LittleEndian, &llama.hyperparameters)
-		ggml.model = &llama
-		// TODO: sanity check hyperparameters
-	default:
-		return nil, fmt.Errorf("unsupported model type: %s", hint)
-	}
+	ggml.model = model
 
 	// final model type
 	return &ggml, nil

llm/ggml_llama.go

@@ -1,687 +0,0 @@
-package llm
-
-import (
-	"bufio"
-	"bytes"
-	"context"
-	"embed"
-	"encoding/json"
-	"errors"
-	"fmt"
-	"io"
-	"io/fs"
-	"log"
-	"math/rand"
-	"net/http"
-	"os"
-	"os/exec"
-	"path"
-	"path/filepath"
-	"runtime"
-	"strconv"
-	"strings"
-	"sync"
-	"time"
-
-	"github.com/jmorganca/ollama/api"
-)
-
-const ModelFamilyLlama ModelFamily = "llama"
-
-//go:embed llama.cpp/ggml/build/*/bin/*
-var llamaCppEmbed embed.FS
-
-var (
-	ggmlGPU = path.Join("llama.cpp", "ggml", "build", "gpu", "bin")
-	ggmlCPU = path.Join("llama.cpp", "ggml", "build", "cpu", "bin")
-)
-
-var (
-	ggmlInit       sync.Once
-	ggmlRunnerPath string
-)
-
-func osPath(llamaPath string) string {
-	if runtime.GOOS == "windows" {
-		return path.Join(llamaPath, "Release")
-	}
-
-	return llamaPath
-}
-
-func initGGML() {
-	ggmlInit.Do(func() {
-		tmpDir, err := os.MkdirTemp("", "llama-*")
-		if err != nil {
-			log.Fatalf("llama.cpp: failed to create temp dir: %v", err)
-		}
-
-		llamaPath := osPath(ggmlGPU)
-		if _, err := fs.Stat(llamaCppEmbed, llamaPath); err != nil {
-			llamaPath = osPath(ggmlCPU)
-			if _, err := fs.Stat(llamaCppEmbed, llamaPath); err != nil {
-				log.Fatalf("llama.cpp executable not found")
-			}
-		}
-
-		files := []string{"server"}
-		switch runtime.GOOS {
-		case "windows":
-			files = []string{"server.exe"}
-		case "darwin":
-			if llamaPath == osPath(ggmlGPU) {
-				files = append(files, "ggml-metal.metal")
-			}
-		}
-
-		for _, f := range files {
-			srcPath := path.Join(llamaPath, f)
-			destPath := filepath.Join(tmpDir, f)
-
-			srcFile, err := llamaCppEmbed.Open(srcPath)
-			if err != nil {
-				log.Fatalf("read llama.cpp %s: %v", f, err)
-			}
-			defer srcFile.Close()
-
-			destFile, err := os.OpenFile(destPath, os.O_WRONLY|os.O_CREATE|os.O_TRUNC, 0o755)
-			if err != nil {
-				log.Fatalf("write llama.cpp %s: %v", f, err)
-			}
-			defer destFile.Close()
-
-			if _, err := io.Copy(destFile, srcFile); err != nil {
-				log.Fatalf("copy llama.cpp %s: %v", f, err)
-			}
-		}
-
-		ggmlRunnerPath = filepath.Join(tmpDir, "server")
-		if runtime.GOOS == "windows" {
-			ggmlRunnerPath = filepath.Join(tmpDir, "server.exe")
-		}
-	})
-}
-
-type ModelRunner struct {
-	Path string // path to the model runner executable
-}
-
-func ggmlRunner() ModelRunner {
-	initGGML()
-	return ModelRunner{Path: ggmlRunnerPath}
-}
-
-type llamaModel struct {
-	hyperparameters llamaHyperparameters
-}
-
-func (llm *llamaModel) ModelFamily() ModelFamily {
-	return ModelFamilyLlama
-}
-
-func (llm *llamaModel) ModelType() ModelType {
-	switch llm.hyperparameters.NumLayer {
-	case 26:
-		return ModelType3B
-	case 32:
-		return ModelType7B
-	case 40:
-		return ModelType13B
-	case 48:
-		return ModelType34B
-	case 60:
-		return ModelType30B
-	case 80:
-		return ModelType65B
-	}
-
-	// TODO: find a better default
-	return ModelType7B
-}
-
-func (llm *llamaModel) FileType() FileType {
-	return llm.hyperparameters.FileType
-}
-
-type llamaHyperparameters struct {
-	// NumVocab is the size of the model's vocabulary.
-	NumVocab uint32
-
-	// NumEmbd is the size of the model's embedding layer.
-	NumEmbd uint32
-	NumMult uint32
-	NumHead uint32
-
-	// NumLayer is the number of layers in the model.
-	NumLayer uint32
-	NumRot   uint32
-
-	// FileType describes the quantization level of the model, e.g. Q4_0, Q5_K, etc.
-	FileType llamaFileType
-}
-
-type llamaFileType uint32
-
-const (
-	llamaFileTypeF32 llamaFileType = iota
-	llamaFileTypeF16
-	llamaFileTypeQ4_0
-	llamaFileTypeQ4_1
-	llamaFileTypeQ4_1_F16
-	llamaFileTypeQ8_0 llamaFileType = iota + 2
-	llamaFileTypeQ5_0
-	llamaFileTypeQ5_1
-	llamaFileTypeQ2_K
-	llamaFileTypeQ3_K_S
-	llamaFileTypeQ3_K_M
-	llamaFileTypeQ3_K_L
-	llamaFileTypeQ4_K_S
-	llamaFileTypeQ4_K_M
-	llamaFileTypeQ5_K_S
-	llamaFileTypeQ5_K_M
-	llamaFileTypeQ6_K
-)
-
-func (ft llamaFileType) String() string {
-	switch ft {
-	case llamaFileTypeF32:
-		return "F32"
-	case llamaFileTypeF16:
-		return "F16"
-	case llamaFileTypeQ4_0:
-		return "Q4_0"
-	case llamaFileTypeQ4_1:
-		return "Q4_1"
-	case llamaFileTypeQ4_1_F16:
-		return "Q4_1_F16"
-	case llamaFileTypeQ8_0:
-		return "Q8_0"
-	case llamaFileTypeQ5_0:
-		return "Q5_0"
-	case llamaFileTypeQ5_1:
-		return "Q5_1"
-	case llamaFileTypeQ2_K:
-		return "Q2_K"
-	case llamaFileTypeQ3_K_S:
-		return "Q3_K_S"
-	case llamaFileTypeQ3_K_M:
-		return "Q3_K_M"
-	case llamaFileTypeQ3_K_L:
-		return "Q3_K_L"
-	case llamaFileTypeQ4_K_S:
-		return "Q4_K_S"
-	case llamaFileTypeQ4_K_M:
-		return "Q4_K_M"
-	case llamaFileTypeQ5_K_S:
-		return "Q5_K_S"
-	case llamaFileTypeQ5_K_M:
-		return "Q5_K_M"
-	case llamaFileTypeQ6_K:
-		return "Q6_K"
-	default:
-		return "Unknown"
-	}
-}
-
-type Running struct {
-	Port   int
-	Cmd    *exec.Cmd
-	Cancel context.CancelFunc
-}
-
-type llama struct {
-	api.Options
-	Running
-}
-
-func newLlama(model string, adapters []string, runner ModelRunner, opts api.Options) (*llama, error) {
-	if _, err := os.Stat(model); err != nil {
-		return nil, err
-	}
-
-	if _, err := os.Stat(runner.Path); err != nil {
-		return nil, err
-	}
-
-	if len(adapters) > 1 {
-		return nil, errors.New("ollama supports only one lora adapter, but multiple were provided")
-	}
-
-	params := []string{
-		"--model", model,
-		"--ctx-size", fmt.Sprintf("%d", opts.NumCtx),
-		"--gqa", fmt.Sprintf("%d", opts.NumGQA),
-		"--rope-freq-base", fmt.Sprintf("%f", opts.RopeFrequencyBase),
-		"--rope-freq-scale", fmt.Sprintf("%f", opts.RopeFrequencyScale),
-		"--batch-size", fmt.Sprintf("%d", opts.NumBatch),
-		"--n-gpu-layers", fmt.Sprintf("%d", opts.NumGPU),
-		"--embedding",
-	}
-
-	if len(adapters) > 0 {
-		// TODO: applying multiple adapters is not supported by the llama.cpp server yet
-		params = append(params, "--lora", adapters[0])
-	}
-
-	if opts.NumThread > 0 {
-		params = append(params, "--threads", fmt.Sprintf("%d", opts.NumThread))
-	}
-
-	if !opts.F16KV {
-		params = append(params, "--memory-f32")
-	}
-	if opts.UseMLock {
-		params = append(params, "--mlock")
-	}
-	if !opts.UseMMap {
-		params = append(params, "--no-mmap")
-	}
-	if opts.UseNUMA {
-		params = append(params, "--numa")
-	}
-
-	// start the llama.cpp server with a retry in case the port is already in use
-	for try := 0; try < 3; try++ {
-		port := rand.Intn(65535-49152) + 49152 // get a random port in the ephemeral range
-		ctx, cancel := context.WithCancel(context.Background())
-		cmd := exec.CommandContext(
-			ctx,
-			runner.Path,
-			append(params, "--port", strconv.Itoa(port))...,
-		)
-		cmd.Stdout = os.Stderr
-		cmd.Stderr = os.Stderr
-
-		llm := &llama{Options: opts, Running: Running{Port: port, Cmd: cmd, Cancel: cancel}}
-
-		if err := waitForServer(llm); err != nil {
-			log.Printf("error starting llama.cpp server: %v", err)
-			llm.Close()
-			// try again
-			continue
-		}
-		// server started successfully
-		return llm, nil
-	}
-
-	return nil, fmt.Errorf("max retry exceeded starting llama.cpp")
-}
-
-func waitForServer(llm *llama) error {
-	log.Print("starting llama.cpp server")
-	var stderr bytes.Buffer
-	llm.Cmd.Stderr = &stderr
-	err := llm.Cmd.Start()
-	if err != nil {
-		return fmt.Errorf("error starting the external llama.cpp server: %w", err)
-	}
-
-	exitChan := make(chan error, 1)
-
-	// the server is a long running process, watch for it exiting to keep track of something going wrong
-	go func() {
-		err := llm.Cmd.Wait()
-		log.Print(stderr.String())
-		exitChan <- err
-	}()
-
-	// wait for the server to start responding
-	start := time.Now()
-	expiresAt := time.Now().Add(30 * time.Second)
-	ticker := time.NewTicker(100 * time.Millisecond)
-
-	log.Print("waiting for llama.cpp server to start responding")
-
-	for {
-		select {
-		case <-ticker.C:
-			if time.Now().After(expiresAt) {
-				return fmt.Errorf("llama.cpp server did not start responding within 30 seconds, retrying")
-			}
-			if err := llm.Ping(context.Background()); err == nil {
-				log.Printf("llama.cpp server started in %f seconds", time.Since(start).Seconds())
-				return nil
-			}
-		case err := <-exitChan:
-			return fmt.Errorf("llama.cpp server exited unexpectedly: %w", err)
-		}
-	}
-}
-
-func (llm *llama) Close() {
-	llm.Running.Cmd.Cancel()
-}
-
-func (llm *llama) SetOptions(opts api.Options) {
-	llm.Options = opts
-}
-
-type GenerationSettings struct {
-	FrequencyPenalty float64       `json:"frequency_penalty"`
-	IgnoreEOS        bool          `json:"ignore_eos"`
-	LogitBias        []interface{} `json:"logit_bias"`
-	Mirostat         int           `json:"mirostat"`
-	MirostatEta      float64       `json:"mirostat_eta"`
-	MirostatTau      float64       `json:"mirostat_tau"`
-	Model            string        `json:"model"`
-	NCtx             int           `json:"n_ctx"`
-	NKeep            int           `json:"n_keep"`
-	NPredict         int           `json:"n_predict"`
-	NProbs           int           `json:"n_probs"`
-	PenalizeNl       bool          `json:"penalize_nl"`
-	PresencePenalty  float64       `json:"presence_penalty"`
-	RepeatLastN      int           `json:"repeat_last_n"`
-	RepeatPenalty    float64       `json:"repeat_penalty"`
-	Seed             uint32        `json:"seed"`
-	Stop             []string      `json:"stop"`
-	Stream           bool          `json:"stream"`
-	Temp             float64       `json:"temp"`
-	TfsZ             float64       `json:"tfs_z"`
-	TopK             int           `json:"top_k"`
-	TopP             float64       `json:"top_p"`
-	TypicalP         float64       `json:"typical_p"`
-}
-
-type Timings struct {
-	PredictedN  int     `json:"predicted_n"`
-	PredictedMS float64 `json:"predicted_ms"`
-	PromptN     int     `json:"prompt_n"`
-	PromptMS    float64 `json:"prompt_ms"`
-}
-
-type Prediction struct {
-	Content string `json:"content"`
-	Model   string `json:"model"`
-	Prompt  string `json:"prompt"`
-	Stop    bool   `json:"stop"`
-
-	Timings `json:"timings"`
-}
-
-type PredictRequest struct {
-	Stream           bool            `json:"stream"`
-	NPredict         int             `json:"n_predict,omitempty"`
-	TopK             int             `json:"top_k,omitempty"`
-	TopP             float32         `json:"top_p,omitempty"`
-	TfsZ             float32         `json:"tfs_z,omitempty"`
-	TypicalP         float32         `json:"typical_p,omitempty"`
-	RepeatLastN      int             `json:"repeat_last_n,omitempty"`
-	Temperature      float32         `json:"temperature,omitempty"`
-	RepeatPenalty    float32         `json:"repeat_penalty,omitempty"`
-	PresencePenalty  float32         `json:"presence_penalty,omitempty"`
-	FrequencyPenalty float32         `json:"frequency_penalty,omitempty"`
-	Mirostat         int             `json:"mirostat,omitempty"`
-	MirostatTau      float32         `json:"mirostat_tau,omitempty"`
-	MirostatEta      float32         `json:"mirostat_eta,omitempty"`
-	PenalizeNl       bool            `json:"penalize_nl,omitempty"`
-	NKeep            int             `json:"n_keep,omitempty"`
-	Seed             int             `json:"seed,omitempty"`
-	Prompt           string          `json:"prompt,omitempty"`
-	NProbs           int             `json:"n_probs,omitempty"`
-	LogitBias        map[int]float32 `json:"logit_bias,omitempty"`
-	IgnoreEos        bool            `json:"ignore_eos,omitempty"`
-	Stop             []string        `json:"stop,omitempty"`
-}
-
-func (llm *llama) Predict(ctx context.Context, prevContext []int, prompt string, fn func(api.GenerateResponse)) error {
-	prevConvo, err := llm.Decode(ctx, prevContext)
-	if err != nil {
-		return err
-	}
-
-	var nextContext strings.Builder
-	nextContext.WriteString(prevConvo)
-	nextContext.WriteString(prompt)
-
-	endpoint := fmt.Sprintf("http://127.0.0.1:%d/completion", llm.Port)
-	predReq := PredictRequest{
-		Prompt:           nextContext.String(),
-		Stream:           true,
-		NPredict:         llm.NumPredict,
-		NKeep:            llm.NumKeep,
-		Temperature:      llm.Temperature,
-		TopK:             llm.TopK,
-		TopP:             llm.TopP,
-		TfsZ:             llm.TFSZ,
-		TypicalP:         llm.TypicalP,
-		RepeatLastN:      llm.RepeatLastN,
-		RepeatPenalty:    llm.RepeatPenalty,
-		PresencePenalty:  llm.PresencePenalty,
-		FrequencyPenalty: llm.FrequencyPenalty,
-		Mirostat:         llm.Mirostat,
-		MirostatTau:      llm.MirostatTau,
-		MirostatEta:      llm.MirostatEta,
-		PenalizeNl:       llm.PenalizeNewline,
-		Stop:             llm.Stop,
-	}
-	data, err := json.Marshal(predReq)
-	if err != nil {
-		return fmt.Errorf("error marshaling data: %v", err)
-	}
-
-	req, err := http.NewRequestWithContext(ctx, http.MethodPost, endpoint, bytes.NewBuffer(data))
-	if err != nil {
-		return fmt.Errorf("error creating POST request: %v", err)
-	}
-	req.Header.Set("Content-Type", "application/json")
-
-	resp, err := http.DefaultClient.Do(req)
-	if err != nil {
-		return fmt.Errorf("POST predict: %v", err)
-	}
-	defer resp.Body.Close()
-
-	if resp.StatusCode >= 400 {
-		bodyBytes, err := io.ReadAll(resp.Body)
-		if err != nil {
-			return fmt.Errorf("failed reading llm error response: %w", err)
-		}
-		log.Printf("llm predict error: %s", bodyBytes)
-		return fmt.Errorf("%s", bodyBytes)
-	}
-
-	scanner := bufio.NewScanner(resp.Body)
-	for scanner.Scan() {
-		select {
-		case <-ctx.Done():
-			// This handles the request cancellation
-			return ctx.Err()
-		default:
-			line := scanner.Text()
-			if line == "" {
-				continue
-			}
-
-			// Read data from the server-side event stream
-			if strings.HasPrefix(line, "data: ") {
-				evt := line[6:]
-				var p Prediction
-				if err := json.Unmarshal([]byte(evt), &p); err != nil {
-					return fmt.Errorf("error unmarshaling llm prediction response: %v", err)
-				}
-
-				if p.Content != "" {
-					fn(api.GenerateResponse{Response: p.Content})
-					nextContext.WriteString(p.Content)
-				}
-
-				if p.Stop {
-					embd, err := llm.Encode(ctx, nextContext.String())
-					if err != nil {
-						return fmt.Errorf("encoding context: %v", err)
-					}
-
-					fn(api.GenerateResponse{
-						Done:               true,
-						Context:            embd,
-						PromptEvalCount:    p.PromptN,
-						PromptEvalDuration: parseDurationMs(p.PromptMS),
-						EvalCount:          p.PredictedN,
-						EvalDuration:       parseDurationMs(p.PredictedMS),
-					})
-
-					return nil
-				}
-			}
-		}
-	}
-
-	if err := scanner.Err(); err != nil {
-		return fmt.Errorf("error reading llm response: %v", err)
-	}
-
-	return nil
-}
-
-type TokenizeRequest struct {
-	Content string `json:"content"`
-}
-
-type TokenizeResponse struct {
-	Tokens []int `json:"tokens"`
-}
-
-func (llm *llama) Encode(ctx context.Context, prompt string) ([]int, error) {
-	endpoint := fmt.Sprintf("http://127.0.0.1:%d/tokenize", llm.Port)
-	data, err := json.Marshal(TokenizeRequest{Content: prompt})
-	if err != nil {
-		return nil, fmt.Errorf("marshaling encode data: %w", err)
-	}
-
-	req, err := http.NewRequestWithContext(ctx, http.MethodPost, endpoint, bytes.NewBuffer(data))
-	if err != nil {
-		return nil, fmt.Errorf("encode request: %w", err)
-	}
-	req.Header.Set("Content-Type", "application/json")
-
-	resp, err := http.DefaultClient.Do(req)
-	if err != nil {
-		return nil, fmt.Errorf("do encode request: %w", err)
-	}
-	defer resp.Body.Close()
-
-	body, err := io.ReadAll(resp.Body)
-	if err != nil {
-		return nil, fmt.Errorf("read encode request: %w", err)
-	}
-
-	if resp.StatusCode >= 400 {
-		log.Printf("llm encode error: %s", body)
-		return nil, fmt.Errorf("%s", body)
-	}
-
-	var encoded TokenizeResponse
-	if err := json.Unmarshal(body, &encoded); err != nil {
-		return nil, fmt.Errorf("unmarshal encode response: %w", err)
-	}
-
-	return encoded.Tokens, nil
-}
-
-type DetokenizeRequest struct {
-	Tokens []int `json:"tokens"`
-}
-
-type DetokenizeResponse struct {
-	Content string `json:"content"`
-}
-
-func (llm *llama) Decode(ctx context.Context, tokens []int) (string, error) {
-	if len(tokens) == 0 {
-		return "", nil
-	}
-	endpoint := fmt.Sprintf("http://127.0.0.1:%d/detokenize", llm.Port)
-	data, err := json.Marshal(DetokenizeRequest{Tokens: tokens})
-	if err != nil {
-		return "", fmt.Errorf("marshaling decode data: %w", err)
-	}
-
-	req, err := http.NewRequestWithContext(ctx, http.MethodPost, endpoint, bytes.NewBuffer(data))
-	if err != nil {
-		return "", fmt.Errorf("decode request: %w", err)
-	}
-	req.Header.Set("Content-Type", "application/json")
-
-	resp, err := http.DefaultClient.Do(req)
-	if err != nil {
-		return "", fmt.Errorf("do decode request: %w", err)
-	}
-	defer resp.Body.Close()
-
-	body, err := io.ReadAll(resp.Body)
-	if err != nil {
-		return "", fmt.Errorf("read decode request: %w", err)
-	}
-
-	if resp.StatusCode >= 400 {
-		log.Printf("llm decode error: %s", body)
-		return "", fmt.Errorf("%s", body)
-	}
-
-	var decoded DetokenizeResponse
-	if err := json.Unmarshal(body, &decoded); err != nil {
-		return "", fmt.Errorf("unmarshal encode response: %w", err)
-	}
-
-	// decoded content contains a leading whitespace
-	decoded.Content, _ = strings.CutPrefix(decoded.Content, "")
-
-	return decoded.Content, nil
-}
-
-type EmbeddingRequest struct {
-	Content string `json:"content"`
-}
-
-type EmbeddingResponse struct {
-	Embedding []float64 `json:"embedding"`
-}
-
-func (llm *llama) Embedding(ctx context.Context, input string) ([]float64, error) {
-	endpoint := fmt.Sprintf("http://127.0.0.1:%d/embedding", llm.Port)
-	data, err := json.Marshal(TokenizeRequest{Content: input})
-	if err != nil {
-		return nil, fmt.Errorf("error marshaling embed data: %w", err)
-	}
-
-	req, err := http.NewRequestWithContext(ctx, http.MethodPost, endpoint, bytes.NewBuffer(data))
-	if err != nil {
-		return nil, fmt.Errorf("error creating embed request: %w", err)
-	}
-	req.Header.Set("Content-Type", "application/json")
-
-	resp, err := http.DefaultClient.Do(req)
-	if err != nil {
-		return nil, fmt.Errorf("POST embedding: %w", err)
-	}
-	defer resp.Body.Close()
-
-	body, err := io.ReadAll(resp.Body)
-	if err != nil {
-		return nil, fmt.Errorf("error reading embed response: %w", err)
-	}
-
-	if resp.StatusCode >= 400 {
-		log.Printf("llm encode error: %s", body)
-		return nil, fmt.Errorf("%s", body)
-	}
-
-	var embedding EmbeddingResponse
-	if err := json.Unmarshal(body, &embedding); err != nil {
-		return nil, fmt.Errorf("unmarshal tokenize response: %w", err)
-	}
-
-	return embedding.Embedding, nil
-}
-
-// Ping checks that the server subprocess is still running and responding to requests
-func (llm *llama) Ping(ctx context.Context) error {
-	resp, err := http.Head(fmt.Sprintf("http://127.0.0.1:%d", llm.Running.Port))
-	if err != nil {
-		return fmt.Errorf("ping resp: %w", err)
-	}
-	if resp.StatusCode != http.StatusOK {
-		return fmt.Errorf("unexpected ping status: %s", resp.Status)
-	}
-	return nil
-}

llm/gguf.go

@@ -0,0 +1,383 @@
+package llm
+
+import (
+	"bytes"
+	"encoding/binary"
+	"errors"
+	"fmt"
+	"io"
+)
+
+type containerGGUF struct {
+	Version uint32
+
+	V1 struct {
+		NumTensor uint32
+		NumKV     uint32
+	}
+
+	V2 struct {
+		NumTensor uint64
+		NumKV     uint64
+	}
+}
+
+func (c *containerGGUF) Name() string {
+	return "gguf"
+}
+
+func (c *containerGGUF) Decode(r io.Reader) (model, error) {
+	binary.Read(r, binary.LittleEndian, &c.Version)
+
+	switch c.Version {
+	case 1:
+		binary.Read(r, binary.LittleEndian, &c.V1)
+	case 2:
+		binary.Read(r, binary.LittleEndian, &c.V2)
+	default:
+		return nil, errors.New("invalid version")
+	}
+
+	model := newGGUFModel(c)
+	if err := model.Decode(r); err != nil {
+		return nil, err
+	}
+
+	return model, nil
+}
+
+const (
+	ggufTypeUint8 uint32 = iota
+	ggufTypeInt8
+	ggufTypeUint16
+	ggufTypeInt16
+	ggufTypeUint32
+	ggufTypeInt32
+	ggufTypeFloat32
+	ggufTypeBool
+	ggufTypeString
+	ggufTypeArray
+	ggufTypeUint64
+	ggufTypeInt64
+	ggufTypeFloat64
+)
+
+type kv map[string]any
+
+type ggufModel struct {
+	*containerGGUF
+	kv
+}
+
+func newGGUFModel(container *containerGGUF) *ggufModel {
+	return &ggufModel{
+		containerGGUF: container,
+		kv:            make(kv),
+	}
+}
+
+func (llm *ggufModel) NumKV() uint64 {
+	if llm.Version == 1 {
+		return uint64(llm.V1.NumKV)
+	}
+
+	return llm.V2.NumKV
+}
+
+func (llm *ggufModel) ModelFamily() string {
+	t, ok := llm.kv["general.architecture"].(string)
+	if ok {
+		return t
+	}
+
+	return "unknown"
+}
+
+func (llm *ggufModel) ModelType() string {
+	switch llm.ModelFamily() {
+	case "llama":
+		if blocks, ok := llm.kv["llama.block_count"].(uint32); ok {
+			heads, headsOK := llm.kv["llama.head_count"].(uint32)
+			headKVs, headsKVsOK := llm.kv["llama.head_count_kv"].(uint32)
+			if headsOK && headsKVsOK && heads/headKVs == 8 {
+				return "70B"
+			}
+
+			return llamaModelType(blocks)
+		}
+	case "falcon":
+		if blocks, ok := llm.kv["falcon.block_count"].(uint32); ok {
+			return falconModelType(blocks)
+		}
+	case "starcoder":
+		if blocks, ok := llm.kv["starcoder.block_count"].(uint32); ok {
+			return starCoderModelType(blocks)
+		}
+	}
+
+	return "unknown"
+}
+
+func (llm *ggufModel) FileType() string {
+	t, ok := llm.kv["general.file_type"].(uint32)
+	if ok {
+		return fileType(t)
+	}
+
+	return "unknown"
+}
+
+func (llm *ggufModel) Decode(r io.Reader) error {
+	read := llm.readString
+	if llm.Version == 1 {
+		read = llm.readStringV1
+	}
+
+	for i := 0; uint64(i) < llm.NumKV(); i++ {
+		k, err := read(r)
+		if err != nil {
+			return err
+		}
+
+		vtype := llm.readU32(r)
+
+		var v any
+		switch vtype {
+		case ggufTypeUint8:
+			v = llm.readU8(r)
+		case ggufTypeInt8:
+			v = llm.readI8(r)
+		case ggufTypeUint16:
+			v = llm.readU16(r)
+		case ggufTypeInt16:
+			v = llm.readI16(r)
+		case ggufTypeUint32:
+			v = llm.readU32(r)
+		case ggufTypeInt32:
+			v = llm.readI32(r)
+		case ggufTypeUint64:
+			v = llm.readU64(r)
+		case ggufTypeInt64:
+			v = llm.readI64(r)
+		case ggufTypeFloat32:
+			v = llm.readF32(r)
+		case ggufTypeFloat64:
+			v = llm.readF64(r)
+		case ggufTypeBool:
+			v = llm.readBool(r)
+		case ggufTypeString:
+			fn := llm.readString
+			if llm.Version == 1 {
+				fn = llm.readStringV1
+			}
+
+			s, err := fn(r)
+			if err != nil {
+				return err
+			}
+
+			v = s
+		case ggufTypeArray:
+			fn := llm.readArray
+			if llm.Version == 1 {
+				fn = llm.readArrayV1
+			}
+
+			a, err := fn(r)
+			if err != nil {
+				return err
+			}
+
+			v = a
+		default:
+			return fmt.Errorf("invalid type: %d", vtype)
+		}
+
+		llm.kv[k] = v
+	}
+
+	return nil
+}
+
+func (llm *ggufModel) NumLayers() int64 {
+	value, exists := llm.kv[fmt.Sprintf("%s.block_count", llm.ModelFamily())]
+	if !exists {
+		return 0
+	}
+
+	v := value.(uint32)
+	return int64(v)
+}
+
+func (ggufModel) readU8(r io.Reader) uint8 {
+	var u8 uint8
+	binary.Read(r, binary.LittleEndian, &u8)
+	return u8
+}
+
+func (ggufModel) readI8(r io.Reader) int8 {
+	var i8 int8
+	binary.Read(r, binary.LittleEndian, &i8)
+	return i8
+}
+
+func (ggufModel) readU16(r io.Reader) uint16 {
+	var u16 uint16
+	binary.Read(r, binary.LittleEndian, &u16)
+	return u16
+}
+
+func (ggufModel) readI16(r io.Reader) int16 {
+	var i16 int16
+	binary.Read(r, binary.LittleEndian, &i16)
+	return i16
+}
+
+func (ggufModel) readU32(r io.Reader) uint32 {
+	var u32 uint32
+	binary.Read(r, binary.LittleEndian, &u32)
+	return u32
+}
+
+func (ggufModel) readI32(r io.Reader) int32 {
+	var i32 int32
+	binary.Read(r, binary.LittleEndian, &i32)
+	return i32
+}
+
+func (ggufModel) readU64(r io.Reader) uint64 {
+	var u64 uint64
+	binary.Read(r, binary.LittleEndian, &u64)
+	return u64
+}
+
+func (ggufModel) readI64(r io.Reader) int64 {
+	var i64 int64
+	binary.Read(r, binary.LittleEndian, &i64)
+	return i64
+}
+
+func (ggufModel) readF32(r io.Reader) float32 {
+	var f32 float32
+	binary.Read(r, binary.LittleEndian, &f32)
+	return f32
+}
+
+func (ggufModel) readF64(r io.Reader) float64 {
+	var f64 float64
+	binary.Read(r, binary.LittleEndian, &f64)
+	return f64
+}
+
+func (ggufModel) readBool(r io.Reader) bool {
+	var b bool
+	binary.Read(r, binary.LittleEndian, &b)
+	return b
+}
+
+func (ggufModel) readStringV1(r io.Reader) (string, error) {
+	var nameLength uint32
+	binary.Read(r, binary.LittleEndian, &nameLength)
+
+	var b bytes.Buffer
+	if _, err := io.CopyN(&b, r, int64(nameLength)); err != nil {
+		return "", err
+	}
+
+	// gguf v1 strings are null-terminated
+	b.Truncate(b.Len() - 1)
+
+	return b.String(), nil
+}
+
+func (llm ggufModel) readString(r io.Reader) (string, error) {
+	var nameLength uint64
+	binary.Read(r, binary.LittleEndian, &nameLength)
+
+	var b bytes.Buffer
+	if _, err := io.CopyN(&b, r, int64(nameLength)); err != nil {
+		return "", err
+	}
+
+	return b.String(), nil
+}
+
+func (llm *ggufModel) readArrayV1(r io.Reader) (arr []any, err error) {
+	atype := llm.readU32(r)
+	n := llm.readU32(r)
+
+	for i := 0; uint32(i) < n; i++ {
+		switch atype {
+		case ggufTypeUint8:
+			arr = append(arr, llm.readU8(r))
+		case ggufTypeInt8:
+			arr = append(arr, llm.readU8(r))
+		case ggufTypeUint16:
+			arr = append(arr, llm.readU16(r))
+		case ggufTypeInt16:
+			arr = append(arr, llm.readI16(r))
+		case ggufTypeUint32:
+			arr = append(arr, llm.readU32(r))
+		case ggufTypeInt32:
+			arr = append(arr, llm.readI32(r))
+		case ggufTypeFloat32:
+			arr = append(arr, llm.readF32(r))
+		case ggufTypeBool:
+			arr = append(arr, llm.readBool(r))
+		case ggufTypeString:
+			s, err := llm.readStringV1(r)
+			if err != nil {
+				return nil, err
+			}
+
+			arr = append(arr, s)
+		default:
+			return nil, fmt.Errorf("invalid array type: %d", atype)
+		}
+	}
+
+	return
+}
+
+func (llm *ggufModel) readArray(r io.Reader) (arr []any, err error) {
+	atype := llm.readU32(r)
+	n := llm.readU64(r)
+
+	for i := 0; uint64(i) < n; i++ {
+		switch atype {
+		case ggufTypeUint8:
+			arr = append(arr, llm.readU8(r))
+		case ggufTypeInt8:
+			arr = append(arr, llm.readU8(r))
+		case ggufTypeUint16:
+			arr = append(arr, llm.readU16(r))
+		case ggufTypeInt16:
+			arr = append(arr, llm.readI16(r))
+		case ggufTypeUint32:
+			arr = append(arr, llm.readU32(r))
+		case ggufTypeInt32:
+			arr = append(arr, llm.readI32(r))
+		case ggufTypeUint64:
+			arr = append(arr, llm.readU64(r))
+		case ggufTypeInt64:
+			arr = append(arr, llm.readI64(r))
+		case ggufTypeFloat32:
+			arr = append(arr, llm.readF32(r))
+		case ggufTypeFloat64:
+			arr = append(arr, llm.readF64(r))
+		case ggufTypeBool:
+			arr = append(arr, llm.readBool(r))
+		case ggufTypeString:
+			s, err := llm.readString(r)
+			if err != nil {
+				return nil, err
+			}
+
+			arr = append(arr, s)
+		default:
+			return nil, fmt.Errorf("invalid array type: %d", atype)
+		}
+	}
+
+	return
+}

llm/llama.cpp/generate.go

@@ -1,13 +0,0 @@
-//go:build !darwin
-// +build !darwin
-
-package llm
-
-//go:generate git submodule init
-//go:generate git submodule update --force ggml
-//go:generate git -C ggml apply ../ggml_patch/0001-add-detokenize-endpoint.patch
-//go:generate git -C ggml apply ../ggml_patch/0002-34B-model-support.patch
-//go:generate git -C ggml apply ../ggml_patch/0003-metal-fix-synchronization-in-new-matrix-multiplicati.patch
-//go:generate git -C ggml apply ../ggml_patch/0004-metal-add-missing-barriers-for-mul-mat-2699.patch
-//go:generate cmake --fresh -S ggml -B ggml/build/cpu -DLLAMA_K_QUANTS=on
-//go:generate cmake --build ggml/build/cpu --target server --config Release

llm/llama.cpp/generate_darwin_amd64.go

@@ -1,10 +1,18 @@
 package llm
 
 //go:generate git submodule init
+
 //go:generate git submodule update --force ggml
-//go:generate git -C ggml apply ../ggml_patch/0001-add-detokenize-endpoint.patch
-//go:generate git -C ggml apply ../ggml_patch/0002-34B-model-support.patch
-//go:generate git -C ggml apply ../ggml_patch/0003-metal-fix-synchronization-in-new-matrix-multiplicati.patch
-//go:generate git -C ggml apply ../ggml_patch/0004-metal-add-missing-barriers-for-mul-mat-2699.patch
-//go:generate cmake --fresh -S ggml -B ggml/build/cpu -DLLAMA_ACCELERATE=on -DLLAMA_K_QUANTS=on -DCMAKE_SYSTEM_PROCESSOR=x86_64 -DCMAKE_OSX_ARCHITECTURES=x86_64 -DCMAKE_OSX_DEPLOYMENT_TARGET=11.0
+//go:generate git -C ggml apply ../patches/0001-add-detokenize-endpoint.patch
+//go:generate git -C ggml apply ../patches/0002-34B-model-support.patch
+//go:generate git -C ggml apply ../patches/0003-metal-fix-synchronization-in-new-matrix-multiplicati.patch
+//go:generate git -C ggml apply ../patches/0004-metal-add-missing-barriers-for-mul-mat-2699.patch
+//go:generate cmake -S ggml -B ggml/build/cpu -DLLAMA_ACCELERATE=on -DLLAMA_K_QUANTS=on -DCMAKE_SYSTEM_PROCESSOR=x86_64 -DCMAKE_OSX_ARCHITECTURES=x86_64 -DCMAKE_OSX_DEPLOYMENT_TARGET=11.0
 //go:generate cmake --build ggml/build/cpu --target server --config Release
+//go:generate mv ggml/build/cpu/bin/server ggml/build/cpu/bin/ollama-runner
+
+//go:generate git submodule update --force gguf
+//go:generate git -C gguf apply ../patches/0001-remove-warm-up-logging.patch
+//go:generate cmake -S gguf -B gguf/build/cpu -DLLAMA_ACCELERATE=on -DLLAMA_K_QUANTS=on -DCMAKE_SYSTEM_PROCESSOR=x86_64 -DCMAKE_OSX_ARCHITECTURES=x86_64 -DCMAKE_OSX_DEPLOYMENT_TARGET=11.0
+//go:generate cmake --build gguf/build/cpu --target server --config Release
+//go:generate mv gguf/build/cpu/bin/server gguf/build/cpu/bin/ollama-runner

llm/llama.cpp/generate_darwin_arm64.go

@@ -1,10 +1,18 @@
 package llm
 
 //go:generate git submodule init
+
 //go:generate git submodule update --force ggml
-//go:generate git -C ggml apply ../ggml_patch/0001-add-detokenize-endpoint.patch
-//go:generate git -C ggml apply ../ggml_patch/0002-34B-model-support.patch
-//go:generate git -C ggml apply ../ggml_patch/0003-metal-fix-synchronization-in-new-matrix-multiplicati.patch
-//go:generate git -C ggml apply ../ggml_patch/0004-metal-add-missing-barriers-for-mul-mat-2699.patch
-//go:generate cmake --fresh -S ggml -B ggml/build/gpu -DLLAMA_METAL=on -DLLAMA_ACCELERATE=on -DLLAMA_K_QUANTS=on -DCMAKE_SYSTEM_PROCESSOR=arm64 -DCMAKE_OSX_ARCHITECTURES=arm64 -DCMAKE_OSX_DEPLOYMENT_TARGET=11.0
-//go:generate cmake --build ggml/build/gpu --target server --config Release
+//go:generate git -C ggml apply ../patches/0001-add-detokenize-endpoint.patch
+//go:generate git -C ggml apply ../patches/0002-34B-model-support.patch
+//go:generate git -C ggml apply ../patches/0003-metal-fix-synchronization-in-new-matrix-multiplicati.patch
+//go:generate git -C ggml apply ../patches/0004-metal-add-missing-barriers-for-mul-mat-2699.patch
+//go:generate cmake -S ggml -B ggml/build/metal -DLLAMA_METAL=on -DLLAMA_ACCELERATE=on -DLLAMA_K_QUANTS=on -DCMAKE_SYSTEM_PROCESSOR=arm64 -DCMAKE_OSX_ARCHITECTURES=arm64 -DCMAKE_OSX_DEPLOYMENT_TARGET=11.0
+//go:generate cmake --build ggml/build/metal --target server --config Release
+//go:generate mv ggml/build/metal/bin/server ggml/build/metal/bin/ollama-runner
+
+//go:generate git submodule update --force gguf
+//go:generate git -C gguf apply ../patches/0001-remove-warm-up-logging.patch
+//go:generate cmake -S gguf -B gguf/build/metal -DLLAMA_METAL=on -DLLAMA_ACCELERATE=on -DLLAMA_K_QUANTS=on -DCMAKE_SYSTEM_PROCESSOR=arm64 -DCMAKE_OSX_ARCHITECTURES=arm64 -DCMAKE_OSX_DEPLOYMENT_TARGET=11.0
+//go:generate cmake --build gguf/build/metal --target server --config Release
+//go:generate mv gguf/build/metal/bin/server gguf/build/metal/bin/ollama-runner

llm/llama.cpp/generate_linux.go

@@ -0,0 +1,26 @@
+package llm
+
+//go:generate git submodule init
+
+//go:generate git submodule update --force ggml
+//go:generate git -C ggml apply ../patches/0001-add-detokenize-endpoint.patch
+//go:generate git -C ggml apply ../patches/0002-34B-model-support.patch
+//go:generate git -C ggml apply ../patches/0005-ggml-support-CUDA-s-half-type-for-aarch64-1455-2670.patch
+//go:generate git -C ggml apply ../patches/0001-copy-cuda-runtime-libraries.patch
+//go:generate cmake -S ggml -B ggml/build/cpu -DLLAMA_K_QUANTS=on
+//go:generate cmake --build ggml/build/cpu --target server --config Release
+//go:generate mv ggml/build/cpu/bin/server ggml/build/cpu/bin/ollama-runner
+
+//go:generate git submodule update --force gguf
+//go:generate git -C gguf apply ../patches/0001-copy-cuda-runtime-libraries.patch
+//go:generate git -C gguf apply ../patches/0001-remove-warm-up-logging.patch
+//go:generate cmake -S gguf -B gguf/build/cpu -DLLAMA_K_QUANTS=on
+//go:generate cmake --build gguf/build/cpu --target server --config Release
+//go:generate mv gguf/build/cpu/bin/server gguf/build/cpu/bin/ollama-runner
+
+//go:generate cmake -S ggml -B ggml/build/cuda -DLLAMA_CUBLAS=on -DLLAMA_ACCELERATE=on -DLLAMA_K_QUANTS=on
+//go:generate cmake --build ggml/build/cuda --target server --config Release
+//go:generate mv ggml/build/cuda/bin/server ggml/build/cuda/bin/ollama-runner
+//go:generate cmake -S gguf -B gguf/build/cuda -DLLAMA_CUBLAS=on -DLLAMA_ACCELERATE=on -DLLAMA_K_QUANTS=on
+//go:generate cmake --build gguf/build/cuda --target server --config Release
+//go:generate mv gguf/build/cuda/bin/server gguf/build/cuda/bin/ollama-runner

llm/llama.cpp/generate_windows.go

@@ -0,0 +1,16 @@
+package llm
+
+//go:generate git submodule init
+
+//go:generate git submodule update --force ggml
+//go:generate git -C ggml apply ../patches/0001-add-detokenize-endpoint.patch
+//go:generate git -C ggml apply ../patches/0002-34B-model-support.patch
+//go:generate cmake -S ggml -B ggml/build/cpu -DLLAMA_K_QUANTS=on
+//go:generate cmake --build ggml/build/cpu --target server --config Release
+//go:generate cmd /c move ggml\build\cpu\bin\Release\server.exe ggml\build\cpu\bin\Release\ollama-runner.exe
+
+//go:generate git submodule update --force gguf
+//go:generate git -C gguf apply ../patches/0001-remove-warm-up-logging.patch
+//go:generate cmake -S gguf -B gguf/build/cpu -DLLAMA_K_QUANTS=on
+//go:generate cmake --build gguf/build/cpu --target server --config Release
+//go:generate cmd /c move gguf\build\cpu\bin\Release\server.exe gguf\build\cpu\bin\Release\ollama-runner.exe

llm/llama.cpp/ggml_patch/0003-metal-add-missing-barriers-for-mul-mat-2699.patch

@@ -1,32 +0,0 @@
-From 8c0ea847ac1460bca534d92266e3471cb31471be Mon Sep 17 00:00:00 2001
-From: Bruce MacDonald <brucewmacdonald@gmail.com>
-Date: Tue, 5 Sep 2023 16:05:08 -0400
-Subject: [PATCH] metal: add missing barriers for mul-mat #2699
-
----
- ggml-metal.metal | 2 ++
- 1 file changed, 2 insertions(+)
-
-diff --git a/ggml-metal.metal b/ggml-metal.metal
-index 3f31252..ce3541f 100644
---- a/ggml-metal.metal
-+++ b/ggml-metal.metal
-@@ -1850,6 +1850,7 @@ kernel void kernel_mul_mm(device const  uchar * src0,
-         //load data and store to threadgroup memory
-         half4x4 temp_a;
-         dequantize_func(x, il, temp_a);
-+        threadgroup_barrier(mem_flags::mem_threadgroup);
-         #pragma unroll(16)
-         for (int i = 0; i < 16; i++) {
-             *(sa + SG_MAT_SIZE * ((tiitg / THREAD_PER_ROW / 8) \
-@@ -1895,6 +1896,7 @@ kernel void kernel_mul_mm(device const  uchar * src0,
-         }
-     } else {
-         // block is smaller than 64x32, we should avoid writing data outside of the matrix
-+        threadgroup_barrier(mem_flags::mem_threadgroup);
-         threadgroup float *temp_str = ((threadgroup float *)shared_memory) \
-                                       + 32 * (sgitg&1) + (16 * (sgitg>>1)) * BLOCK_SIZE_M;
-         for (int i = 0; i < 8; i++) {
--- 
-2.39.2 (Apple Git-143)
-

llm/llama.cpp/patches/0001-copy-cuda-runtime-libraries.patch

@@ -0,0 +1,27 @@
+From 5dd02993e8cc2ce309157736b95bb572f274a3fd Mon Sep 17 00:00:00 2001
+From: Michael Yang <mxyng@pm.me>
+Date: Wed, 20 Sep 2023 14:19:52 -0700
+Subject: [PATCH] copy cuda runtime libraries
+
+---
+ CMakeLists.txt | 4 ++++
+ 1 file changed, 4 insertions(+)
+
+diff --git a/CMakeLists.txt b/CMakeLists.txt
+index 824d9f2..dd24137 100644
+--- a/CMakeLists.txt
++++ b/CMakeLists.txt
+@@ -274,6 +274,10 @@ if (LLAMA_CUBLAS)
+             set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cudart CUDA::cublas CUDA::cublasLt)
+         endif()
+ 
++        configure_file(${CUDAToolkit_LIBRARY_DIR}/libcudart.so ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/libcudart.so.${CUDAToolkit_VERSION_MAJOR}.0 COPYONLY)
++        configure_file(${CUDAToolkit_LIBRARY_DIR}/libcublas.so ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/libcublas.so.${CUDAToolkit_VERSION_MAJOR} COPYONLY)
++        configure_file(${CUDAToolkit_LIBRARY_DIR}/libcublasLt.so ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/libcublasLt.so.${CUDAToolkit_VERSION_MAJOR} COPYONLY)
++
+     if (NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
+         # 52 == lowest CUDA 12 standard
+         # 60 == f16 CUDA intrinsics
+-- 
+2.42.0
+

llm/llama.cpp/patches/0001-remove-warm-up-logging.patch

@@ -0,0 +1,25 @@
+From 8dbb5449db259a9c24796e7927d89bee98b6c8f5 Mon Sep 17 00:00:00 2001
+From: Bruce MacDonald <brucewmacdonald@gmail.com>
+Date: Thu, 5 Oct 2023 11:21:12 -0400
+Subject: [PATCH] remove warm up logging
+
+---
+ common/common.cpp | 2 --
+ 1 file changed, 2 deletions(-)
+
+diff --git a/common/common.cpp b/common/common.cpp
+index 7370017..c4433fe 100644
+--- a/common/common.cpp
++++ b/common/common.cpp
+@@ -839,8 +839,6 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
+     }
+ 
+     {
+-        LOG("warming up the model with an empty run\n");
+-
+         std::vector<llama_token> tmp = { llama_token_bos(lctx), llama_token_eos(lctx), };
+         llama_decode(lctx, llama_batch_get_one(tmp.data(), std::min(tmp.size(), (size_t) params.n_batch), 0, 0));
+         llama_kv_cache_tokens_rm(lctx, -1, -1);
+-- 
+2.39.2 (Apple Git-143)
+

llm/llama.cpp/patches/0005-ggml-support-CUDA-s-half-type-for-aarch64-1455-2670.patch

@@ -0,0 +1,32 @@
+From 1e3bc523d8053a77df3ac7126a84d0297ee97ef6 Mon Sep 17 00:00:00 2001
+From: Kylin <56434533+KyL0N@users.noreply.github.com>
+Date: Tue, 22 Aug 2023 15:14:23 +0800
+Subject: [PATCH] ggml : support CUDA's half type for aarch64(#1455) (#2670)
+
+* ggml: support CUDA's half type for aarch64(#1455)
+support CUDA's half type for aarch64 in ggml_fp16_t definition
+
+* ggml: use __CUDACC__ to recognise nvcc compiler
+---
+ ggml.h | 5 +++--
+ 1 file changed, 3 insertions(+), 2 deletions(-)
+
+diff --git a/ggml.h b/ggml.h
+index 544ad2d..0ec7ec5 100644
+--- a/ggml.h
++++ b/ggml.h
+@@ -259,8 +259,9 @@
+ extern "C" {
+ #endif
+ 
+-#ifdef __ARM_NEON
+-    // we use the built-in 16-bit float type
++#if defined(__ARM_NEON) && defined(__CUDACC__)
++    typedef half ggml_fp16_t;
++#elif defined(__ARM_NEON)
+     typedef __fp16 ggml_fp16_t;
+ #else
+     typedef uint16_t ggml_fp16_t;
+-- 
+2.39.2 (Apple Git-143)
+

llm/llama.go

@@ -0,0 +1,752 @@
+package llm
+
+import (
+	"bufio"
+	"bytes"
+	"context"
+	"embed"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"io"
+	"io/fs"
+	"log"
+	"math/rand"
+	"net/http"
+	"os"
+	"os/exec"
+	"path"
+	"path/filepath"
+	"runtime"
+	"strconv"
+	"strings"
+	"sync"
+	"time"
+
+	"github.com/jmorganca/ollama/api"
+	"github.com/jmorganca/ollama/format"
+)
+
+//go:embed llama.cpp/*/build/*/bin/*
+var llamaCppEmbed embed.FS
+
+type ModelRunner struct {
+	Path        string // path to the model runner executable
+	Accelerated bool
+}
+
+func chooseRunners(workDir, runnerType string) []ModelRunner {
+	buildPath := path.Join("llama.cpp", runnerType, "build")
+	var runners []ModelRunner
+
+	// set the runners based on the OS
+	// IMPORTANT: the order of the runners in the array is the priority order
+	switch runtime.GOOS {
+	case "darwin":
+		runners = []ModelRunner{
+			{Path: path.Join(buildPath, "metal", "bin", "ollama-runner")},
+			{Path: path.Join(buildPath, "cpu", "bin", "ollama-runner")},
+		}
+	case "linux":
+		runners = []ModelRunner{
+			{Path: path.Join(buildPath, "cuda", "bin", "ollama-runner"), Accelerated: true},
+			{Path: path.Join(buildPath, "cpu", "bin", "ollama-runner")},
+		}
+	case "windows":
+		// TODO: select windows GPU runner here when available
+		runners = []ModelRunner{
+			{Path: path.Join(buildPath, "cpu", "bin", "Release", "ollama-runner.exe")},
+		}
+	default:
+		log.Printf("unknown OS, running on CPU: %s", runtime.GOOS)
+		runners = []ModelRunner{
+			{Path: path.Join(buildPath, "cpu", "bin", "ollama-runner")},
+		}
+	}
+
+	runnerAvailable := false // if no runner files are found in the embed, this flag will cause a fast fail
+	for _, r := range runners {
+		// find all the files in the runner's bin directory
+		files, err := fs.Glob(llamaCppEmbed, path.Join(path.Dir(r.Path), "*"))
+		if err != nil {
+			// this is expected, ollama may be compiled without all runners packed in
+			log.Printf("%s runner not found: %v", r.Path, err)
+			continue
+		}
+
+		for _, f := range files {
+			runnerAvailable = true
+
+			srcFile, err := llamaCppEmbed.Open(f)
+			if err != nil {
+				log.Fatalf("read llama runner %s: %v", f, err)
+			}
+			defer srcFile.Close()
+
+			// create the directory in case it does not exist, filepath.Dir() converts the file path to the OS's format
+			destPath := filepath.Join(workDir, filepath.Dir(f))
+			if err := os.MkdirAll(destPath, 0o755); err != nil {
+				log.Fatalf("create runner temp dir %s: %v", filepath.Dir(f), err)
+			}
+
+			// create the path to the destination file, filepath.Base() converts the file path to the OS's format
+			destFile := filepath.Join(destPath, filepath.Base(f))
+
+			_, err = os.Stat(destFile)
+			switch {
+			case errors.Is(err, os.ErrNotExist):
+				destFile, err := os.OpenFile(destFile, os.O_WRONLY|os.O_CREATE|os.O_TRUNC, 0o755)
+				if err != nil {
+					log.Fatalf("write llama runner %s: %v", f, err)
+				}
+				defer destFile.Close()
+
+				if _, err := io.Copy(destFile, srcFile); err != nil {
+					log.Fatalf("copy llama runner %s: %v", f, err)
+				}
+			case err != nil:
+				log.Fatalf("stat llama runner %s: %v", f, err)
+			}
+		}
+	}
+	if !runnerAvailable {
+		log.Fatalf("%s runner not found", runnerType)
+	}
+
+	// return the runners to try in priority order
+	localRunnersByPriority := []ModelRunner{}
+	for _, r := range runners {
+		// clean the ModelRunner paths so that they match the OS we are running on
+		localRunnersByPriority = append(localRunnersByPriority, ModelRunner{
+			Path:        filepath.Clean(path.Join(workDir, r.Path)),
+			Accelerated: r.Accelerated,
+		})
+	}
+
+	return localRunnersByPriority
+}
+
+type llamaModel struct {
+	hyperparameters llamaHyperparameters
+}
+
+func (llm *llamaModel) ModelFamily() string {
+	return "llama"
+}
+
+func llamaModelType(numLayer uint32) string {
+	switch numLayer {
+	case 26:
+		return "3B"
+	case 32:
+		return "7B"
+	case 40:
+		return "13B"
+	case 48:
+		return "34B"
+	case 60:
+		return "30B"
+	case 80:
+		return "65B"
+	default:
+		return "unknown"
+	}
+}
+
+func (llm *llamaModel) ModelType() string {
+	return llamaModelType(llm.hyperparameters.NumLayer)
+}
+
+func (llm *llamaModel) FileType() string {
+	return fileType(llm.hyperparameters.FileType)
+}
+
+func (llm *llamaModel) NumLayers() int64 {
+	return int64(llm.hyperparameters.NumLayer)
+}
+
+type llamaHyperparameters struct {
+	// NumVocab is the size of the model's vocabulary.
+	NumVocab uint32
+
+	// NumEmbd is the size of the model's embedding layer.
+	NumEmbd uint32
+	NumMult uint32
+	NumHead uint32
+
+	// NumLayer is the number of layers in the model.
+	NumLayer uint32
+	NumRot   uint32
+
+	// FileType describes the quantization level of the model, e.g. Q4_0, Q5_K, etc.
+	FileType uint32
+}
+
+type Running struct {
+	Port          int
+	Cmd           *exec.Cmd
+	Cancel        context.CancelFunc
+	exitOnce      sync.Once
+	exitCh        chan error // channel to receive the exit status of the subprocess
+	*StatusWriter            // captures error messages from the llama runner process
+}
+
+type llama struct {
+	api.Options
+	Running
+}
+
+var errNoGPU = errors.New("nvidia-smi command failed")
+
+// CheckVRAM returns the free VRAM in bytes on Linux machines with NVIDIA GPUs
+func CheckVRAM() (int64, error) {
+	cmd := exec.Command("nvidia-smi", "--query-gpu=memory.free", "--format=csv,noheader,nounits")
+	var stdout bytes.Buffer
+	cmd.Stdout = &stdout
+	err := cmd.Run()
+	if err != nil {
+		return 0, errNoGPU
+	}
+
+	var freeMiB int64
+	scanner := bufio.NewScanner(&stdout)
+	for scanner.Scan() {
+		line := scanner.Text()
+		vram, err := strconv.ParseInt(strings.TrimSpace(line), 10, 64)
+		if err != nil {
+			return 0, fmt.Errorf("failed to parse available VRAM: %v", err)
+		}
+
+		freeMiB += vram
+	}
+
+	freeBytes := freeMiB * 1024 * 1024
+	if freeBytes < 2*format.GigaByte {
+		log.Printf("less than 2 GB VRAM available, falling back to CPU only")
+		freeMiB = 0
+	}
+
+	return freeBytes, nil
+}
+
+func NumGPU(numLayer, fileSizeBytes int64, opts api.Options) int {
+	if opts.NumGPU != -1 {
+		return opts.NumGPU
+	}
+	if runtime.GOOS == "linux" {
+		freeBytes, err := CheckVRAM()
+		if err != nil {
+			if err.Error() != "nvidia-smi command failed" {
+				log.Print(err.Error())
+			}
+			// nvidia driver not installed or no nvidia GPU found
+			return 0
+		}
+
+		// Calculate bytes per layer
+		// TODO: this is a rough heuristic, better would be to calculate this based on number of layers and context size
+		bytesPerLayer := fileSizeBytes / numLayer
+
+		// max number of layers we can fit in VRAM, subtract 8% to prevent consuming all available VRAM and running out of memory
+		layers := int(freeBytes/bytesPerLayer) * 92 / 100
+		log.Printf("%d MB VRAM available, loading up to %d GPU layers", freeBytes/(1024*1024), layers)
+
+		return layers
+	}
+	// default to enable metal on macOS
+	return 1
+}
+
+// StatusWriter is a writer that captures error messages from the llama runner process
+type StatusWriter struct {
+	ErrCh      chan error
+	LastErrMsg string
+}
+
+func NewStatusWriter() *StatusWriter {
+	return &StatusWriter{
+		ErrCh: make(chan error, 1),
+	}
+}
+
+func (w *StatusWriter) Write(b []byte) (int, error) {
+	var errMsg string
+	if _, after, ok := bytes.Cut(b, []byte("error:")); ok {
+		errMsg = string(bytes.TrimSpace(after))
+	} else if _, after, ok := bytes.Cut(b, []byte("CUDA error")); ok {
+		errMsg = string(bytes.TrimSpace(after))
+	}
+
+	if errMsg != "" {
+		w.LastErrMsg = errMsg
+		w.ErrCh <- fmt.Errorf("llama runner: %s", errMsg)
+	}
+
+	return os.Stderr.Write(b)
+}
+
+func newLlama(model string, adapters []string, runners []ModelRunner, numLayers int64, opts api.Options) (*llama, error) {
+	fileInfo, err := os.Stat(model)
+	if err != nil {
+		return nil, err
+	}
+
+	if len(adapters) > 1 {
+		return nil, errors.New("ollama supports only one lora adapter, but multiple were provided")
+	}
+
+	numGPU := NumGPU(numLayers, fileInfo.Size(), opts)
+	params := []string{
+		"--model", model,
+		"--ctx-size", fmt.Sprintf("%d", opts.NumCtx),
+		"--rope-freq-base", fmt.Sprintf("%f", opts.RopeFrequencyBase),
+		"--rope-freq-scale", fmt.Sprintf("%f", opts.RopeFrequencyScale),
+		"--batch-size", fmt.Sprintf("%d", opts.NumBatch),
+		"--n-gpu-layers", fmt.Sprintf("%d", numGPU),
+		"--embedding",
+	}
+
+	if opts.NumGQA > 0 {
+		params = append(params, "--gqa", fmt.Sprintf("%d", opts.NumGQA))
+	}
+
+	if len(adapters) > 0 {
+		// TODO: applying multiple adapters is not supported by the llama.cpp server yet
+		params = append(params, "--lora", adapters[0])
+	}
+
+	if opts.NumThread > 0 {
+		params = append(params, "--threads", fmt.Sprintf("%d", opts.NumThread))
+	}
+
+	if !opts.F16KV {
+		params = append(params, "--memory-f32")
+	}
+	if opts.UseMLock {
+		params = append(params, "--mlock")
+	}
+	if !opts.UseMMap {
+		params = append(params, "--no-mmap")
+	}
+	if opts.UseNUMA {
+		params = append(params, "--numa")
+	}
+
+	var runnerErr error
+
+	// start the llama.cpp server with a retry in case the port is already in use
+	for _, runner := range runners {
+		if runner.Accelerated && numGPU == 0 {
+			log.Printf("skipping accelerated runner because num_gpu=0")
+			continue
+		}
+
+		if _, err := os.Stat(runner.Path); err != nil {
+			log.Printf("llama runner not found: %v", err)
+			continue
+		}
+
+		port := rand.Intn(65535-49152) + 49152 // get a random port in the ephemeral range
+		ctx, cancel := context.WithCancel(context.Background())
+		cmd := exec.CommandContext(
+			ctx,
+			runner.Path,
+			append(params, "--port", strconv.Itoa(port))...,
+		)
+		cmd.Env = append(os.Environ(), fmt.Sprintf("LD_LIBRARY_PATH=%s", filepath.Dir(runner.Path)))
+		cmd.Stdout = os.Stderr
+		statusWriter := NewStatusWriter()
+		cmd.Stderr = statusWriter
+
+		llm := &llama{Options: opts, Running: Running{Port: port, Cmd: cmd, Cancel: cancel, exitCh: make(chan error)}}
+
+		log.Print("starting llama runner")
+		if err := llm.Cmd.Start(); err != nil {
+			log.Printf("error starting the external llama runner: %v", err)
+			continue
+		}
+
+		// monitor the llama runner process and signal when it exits
+		go func() {
+			err := llm.Cmd.Wait()
+			// default to printing the exit message of the command process, it will probably just say 'exit staus 1'
+			errMsg := err.Error()
+			// try to set a better error message if llama runner logs captured an error
+			if statusWriter.LastErrMsg != "" {
+				errMsg = statusWriter.LastErrMsg
+			}
+			log.Println(errMsg)
+			// llm.Cmd.Wait() can only be called once, use this exit channel to signal that the process has exited
+			llm.exitOnce.Do(func() {
+				close(llm.exitCh)
+			})
+		}()
+
+		if err := waitForServer(llm); err != nil {
+			log.Printf("error starting llama runner: %v", err)
+			llm.Close()
+
+			// default the runnerErr to the error returned by the most recent llama runner process
+			runnerErr = err
+
+			// capture the error directly from the runner process, if any
+			select {
+			case runnerErr = <-statusWriter.ErrCh:
+			default:
+				// the runner process probably timed out
+			}
+
+			// try again
+			continue
+		}
+
+		// server started successfully
+		return llm, nil
+	}
+
+	if runnerErr != nil {
+		// this is the error returned from the llama runner process that failed most recently
+		return nil, runnerErr
+	}
+
+	return nil, fmt.Errorf("failed to start a llama runner")
+}
+
+func waitForServer(llm *llama) error {
+	start := time.Now()
+	expiresAt := time.Now().Add(3 * time.Minute) // be generous with timeout, large models can take a while to load
+	ticker := time.NewTicker(200 * time.Millisecond)
+	defer ticker.Stop()
+
+	log.Print("waiting for llama runner to start responding")
+	for {
+		select {
+		case <-llm.exitCh:
+			// failed to start subprocess
+			return fmt.Errorf("llama runner process has terminated")
+		case <-ticker.C:
+			if time.Now().After(expiresAt) {
+				// timeout
+				return fmt.Errorf("timed out waiting for llama runner to start")
+			}
+
+			if err := llm.Ping(context.Background()); err == nil {
+				// success
+				log.Printf("llama runner started in %f seconds", time.Since(start).Seconds())
+				return nil
+			}
+		}
+	}
+}
+
+func (llm *llama) Close() {
+	// signal the sub-process to terminate
+	llm.Cancel()
+
+	// wait for the command to exit to prevent race conditions with the next run
+	<-llm.exitCh
+
+	if llm.StatusWriter != nil && llm.StatusWriter.LastErrMsg != "" {
+		log.Printf("llama runner stopped with error: %v", llm.StatusWriter.LastErrMsg)
+	} else {
+		log.Print("llama runner stopped successfully")
+	}
+}
+
+func (llm *llama) SetOptions(opts api.Options) {
+	llm.Options = opts
+}
+
+type prediction struct {
+	Content string `json:"content"`
+	Model   string `json:"model"`
+	Prompt  string `json:"prompt"`
+	Stop    bool   `json:"stop"`
+
+	Timings struct {
+		PredictedN  int     `json:"predicted_n"`
+		PredictedMS float64 `json:"predicted_ms"`
+		PromptN     int     `json:"prompt_n"`
+		PromptMS    float64 `json:"prompt_ms"`
+	}
+}
+
+const maxBufferSize = 512 * format.KiloByte
+
+func (llm *llama) Predict(ctx context.Context, prevContext []int, prompt string, fn func(api.GenerateResponse)) error {
+	prevConvo, err := llm.Decode(ctx, prevContext)
+	if err != nil {
+		return err
+	}
+
+	// Remove leading spaces from prevConvo if present
+	prevConvo = strings.TrimPrefix(prevConvo, " ")
+
+	var nextContext strings.Builder
+	nextContext.WriteString(prevConvo)
+	nextContext.WriteString(prompt)
+
+	request := map[string]any{
+		"prompt":            nextContext.String(),
+		"stream":            true,
+		"n_predict":         llm.NumPredict,
+		"n_keep":            llm.NumKeep,
+		"temperature":       llm.Temperature,
+		"top_k":             llm.TopK,
+		"top_p":             llm.TopP,
+		"tfs_z":             llm.TFSZ,
+		"typical_p":         llm.TypicalP,
+		"repeat_last_n":     llm.RepeatLastN,
+		"repeat_penalty":    llm.RepeatPenalty,
+		"presence_penalty":  llm.PresencePenalty,
+		"frequency_penalty": llm.FrequencyPenalty,
+		"mirostat":          llm.Mirostat,
+		"mirostat_tau":      llm.MirostatTau,
+		"mirostat_eta":      llm.MirostatEta,
+		"penalize_nl":       llm.PenalizeNewline,
+		"seed":              llm.Seed,
+		"stop":              llm.Stop,
+	}
+
+	// Handling JSON marshaling with special characters unescaped.
+	buffer := &bytes.Buffer{}
+	enc := json.NewEncoder(buffer)
+	enc.SetEscapeHTML(false)
+
+	if err := enc.Encode(request); err != nil {
+		return fmt.Errorf("failed to marshal data: %v", err)
+	}
+
+	endpoint := fmt.Sprintf("http://127.0.0.1:%d/completion", llm.Port)
+	req, err := http.NewRequestWithContext(ctx, http.MethodPost, endpoint, buffer)
+	if err != nil {
+		return fmt.Errorf("error creating POST request: %v", err)
+	}
+	req.Header.Set("Content-Type", "application/json")
+
+	resp, err := http.DefaultClient.Do(req)
+	if err != nil {
+		return fmt.Errorf("POST predict: %v", err)
+	}
+	defer resp.Body.Close()
+
+	if resp.StatusCode >= 400 {
+		bodyBytes, err := io.ReadAll(resp.Body)
+		if err != nil {
+			return fmt.Errorf("failed reading llm error response: %w", err)
+		}
+		log.Printf("llm predict error: %s", bodyBytes)
+		return fmt.Errorf("%s", bodyBytes)
+	}
+
+	scanner := bufio.NewScanner(resp.Body)
+	// increase the buffer size to avoid running out of space
+	buf := make([]byte, 0, maxBufferSize)
+	scanner.Buffer(buf, maxBufferSize)
+	for scanner.Scan() {
+		select {
+		case <-ctx.Done():
+			// This handles the request cancellation
+			return ctx.Err()
+		default:
+			line := scanner.Bytes()
+			if len(line) == 0 {
+				continue
+			}
+
+			if evt, ok := bytes.CutPrefix(line, []byte("data: ")); ok {
+				var p prediction
+				if err := json.Unmarshal(evt, &p); err != nil {
+					return fmt.Errorf("error unmarshaling llm prediction response: %v", err)
+				}
+
+				if p.Content != "" {
+					fn(api.GenerateResponse{Response: p.Content})
+					nextContext.WriteString(p.Content)
+				}
+
+				if p.Stop {
+					embd, err := llm.Encode(ctx, nextContext.String())
+					if err != nil {
+						return fmt.Errorf("encoding context: %v", err)
+					}
+
+					fn(api.GenerateResponse{
+						Done:               true,
+						Context:            embd,
+						PromptEvalCount:    p.Timings.PromptN,
+						PromptEvalDuration: parseDurationMs(p.Timings.PromptMS),
+						EvalCount:          p.Timings.PredictedN,
+						EvalDuration:       parseDurationMs(p.Timings.PredictedMS),
+					})
+
+					return nil
+				}
+			}
+		}
+	}
+
+	if err := scanner.Err(); err != nil {
+		if strings.Contains(err.Error(), "unexpected EOF") {
+			// this means the llama runner subprocess crashed
+			llm.Close()
+			if llm.StatusWriter != nil && llm.StatusWriter.LastErrMsg != "" {
+				return fmt.Errorf("llama runner exited: %v", llm.StatusWriter.LastErrMsg)
+			}
+			return fmt.Errorf("llama runner exited, you may not have enough available memory to run this model")
+		}
+		return fmt.Errorf("error reading llm response: %v", err)
+	}
+
+	return nil
+}
+
+type TokenizeRequest struct {
+	Content string `json:"content"`
+}
+
+type TokenizeResponse struct {
+	Tokens []int `json:"tokens"`
+}
+
+func (llm *llama) Encode(ctx context.Context, prompt string) ([]int, error) {
+	endpoint := fmt.Sprintf("http://127.0.0.1:%d/tokenize", llm.Port)
+	data, err := json.Marshal(TokenizeRequest{Content: prompt})
+	if err != nil {
+		return nil, fmt.Errorf("marshaling encode data: %w", err)
+	}
+
+	req, err := http.NewRequestWithContext(ctx, http.MethodPost, endpoint, bytes.NewBuffer(data))
+	if err != nil {
+		return nil, fmt.Errorf("encode request: %w", err)
+	}
+	req.Header.Set("Content-Type", "application/json")
+
+	resp, err := http.DefaultClient.Do(req)
+	if err != nil {
+		return nil, fmt.Errorf("do encode request: %w", err)
+	}
+	defer resp.Body.Close()
+
+	body, err := io.ReadAll(resp.Body)
+	if err != nil {
+		return nil, fmt.Errorf("read encode request: %w", err)
+	}
+
+	if resp.StatusCode >= 400 {
+		log.Printf("llm encode error: %s", body)
+		return nil, fmt.Errorf("%s", body)
+	}
+
+	var encoded TokenizeResponse
+	if err := json.Unmarshal(body, &encoded); err != nil {
+		return nil, fmt.Errorf("unmarshal encode response: %w", err)
+	}
+
+	return encoded.Tokens, nil
+}
+
+type DetokenizeRequest struct {
+	Tokens []int `json:"tokens"`
+}
+
+type DetokenizeResponse struct {
+	Content string `json:"content"`
+}
+
+func (llm *llama) Decode(ctx context.Context, tokens []int) (string, error) {
+	if len(tokens) == 0 {
+		return "", nil
+	}
+	endpoint := fmt.Sprintf("http://127.0.0.1:%d/detokenize", llm.Port)
+	data, err := json.Marshal(DetokenizeRequest{Tokens: tokens})
+	if err != nil {
+		return "", fmt.Errorf("marshaling decode data: %w", err)
+	}
+
+	req, err := http.NewRequestWithContext(ctx, http.MethodPost, endpoint, bytes.NewBuffer(data))
+	if err != nil {
+		return "", fmt.Errorf("decode request: %w", err)
+	}
+	req.Header.Set("Content-Type", "application/json")
+
+	resp, err := http.DefaultClient.Do(req)
+	if err != nil {
+		return "", fmt.Errorf("do decode request: %w", err)
+	}
+	defer resp.Body.Close()
+
+	body, err := io.ReadAll(resp.Body)
+	if err != nil {
+		return "", fmt.Errorf("read decode request: %w", err)
+	}
+
+	if resp.StatusCode >= 400 {
+		log.Printf("llm decode error: %s", body)
+		return "", fmt.Errorf("%s", body)
+	}
+
+	var decoded DetokenizeResponse
+	if err := json.Unmarshal(body, &decoded); err != nil {
+		return "", fmt.Errorf("unmarshal encode response: %w", err)
+	}
+
+	return decoded.Content, nil
+}
+
+type EmbeddingRequest struct {
+	Content string `json:"content"`
+}
+
+type EmbeddingResponse struct {
+	Embedding []float64 `json:"embedding"`
+}
+
+func (llm *llama) Embedding(ctx context.Context, input string) ([]float64, error) {
+	endpoint := fmt.Sprintf("http://127.0.0.1:%d/embedding", llm.Port)
+	data, err := json.Marshal(TokenizeRequest{Content: input})
+	if err != nil {
+		return nil, fmt.Errorf("error marshaling embed data: %w", err)
+	}
+
+	req, err := http.NewRequestWithContext(ctx, http.MethodPost, endpoint, bytes.NewBuffer(data))
+	if err != nil {
+		return nil, fmt.Errorf("error creating embed request: %w", err)
+	}
+	req.Header.Set("Content-Type", "application/json")
+
+	resp, err := http.DefaultClient.Do(req)
+	if err != nil {
+		return nil, fmt.Errorf("POST embedding: %w", err)
+	}
+	defer resp.Body.Close()
+
+	body, err := io.ReadAll(resp.Body)
+	if err != nil {
+		return nil, fmt.Errorf("error reading embed response: %w", err)
+	}
+
+	if resp.StatusCode >= 400 {
+		log.Printf("llm encode error: %s", body)
+		return nil, fmt.Errorf("%s", body)
+	}
+
+	var embedding EmbeddingResponse
+	if err := json.Unmarshal(body, &embedding); err != nil {
+		return nil, fmt.Errorf("unmarshal tokenize response: %w", err)
+	}
+
+	return embedding.Embedding, nil
+}
+
+// Ping checks that the server subprocess is still running and responding to requests
+func (llm *llama) Ping(ctx context.Context) error {
+	resp, err := http.Head(fmt.Sprintf("http://127.0.0.1:%d", llm.Port))
+	if err != nil {
+		return fmt.Errorf("ping resp: %w", err)
+	}
+	if resp.StatusCode != http.StatusOK {
+		return fmt.Errorf("unexpected ping status: %s", resp.Status)
+	}
+	return nil
+}

llm/llm.go

@@ -5,10 +5,12 @@ import (
 	"fmt"
 	"log"
 	"os"
+	"runtime"
 
 	"github.com/pbnjay/memory"
 
 	"github.com/jmorganca/ollama/api"
+	"github.com/jmorganca/ollama/format"
 )
 
 type LLM interface {
@@ -21,7 +23,7 @@ type LLM interface {
 	Ping(context.Context) error
 }
 
-func New(model string, adapters []string, opts api.Options) (LLM, error) {
+func New(workDir, model string, adapters []string, opts api.Options) (LLM, error) {
 	if _, err := os.Stat(model); err != nil {
 		return nil, err
 	}
@@ -32,52 +34,61 @@ func New(model string, adapters []string, opts api.Options) (LLM, error) {
 	}
 	defer f.Close()
 
-	ggml, err := DecodeGGML(f, ModelFamilyLlama)
+	ggml, err := DecodeGGML(f)
 	if err != nil {
 		return nil, err
 	}
 
-	switch ggml.FileType().String() {
-	case "F32", "Q5_0", "Q5_1", "Q8_0":
-		if opts.NumGPU != 0 {
-			// F32, F16, Q5_0, Q5_1, and Q8_0 do not support Metal API and will
-			// cause the runner to segmentation fault so disable GPU
-			log.Printf("WARNING: GPU disabled for F32, Q5_0, Q5_1, and Q8_0")
-			opts.NumGPU = 0
+	if runtime.GOOS == "darwin" {
+		switch ggml.FileType() {
+		case "Q8_0":
+			if ggml.Name() != "gguf" && opts.NumGPU != 0 {
+				// GGML Q8_0 do not support Metal API and will
+				// cause the runner to segmentation fault so disable GPU
+				log.Printf("WARNING: GPU disabled for F32, Q5_0, Q5_1, and Q8_0")
+				opts.NumGPU = 0
+			}
+		case "F32", "Q5_0", "Q5_1":
+			if opts.NumGPU != 0 {
+				// F32, Q5_0, Q5_1, and Q8_0 do not support Metal API and will
+				// cause the runner to segmentation fault so disable GPU
+				log.Printf("WARNING: GPU disabled for F32, Q5_0, Q5_1, and Q8_0")
+				opts.NumGPU = 0
+			}
+		}
+
+		var requiredMemory int64
+		var f16Multiplier int64 = 2
+
+		switch ggml.ModelType() {
+		case "3B", "7B":
+			requiredMemory = 8 * format.GigaByte
+		case "13B":
+			requiredMemory = 16 * format.GigaByte
+		case "30B", "34B", "40B":
+			requiredMemory = 32 * format.GigaByte
+		case "65B", "70B":
+			requiredMemory = 64 * format.GigaByte
+		case "180B":
+			requiredMemory = 128 * format.GigaByte
+			f16Multiplier = 4
+		}
+
+		systemMemory := int64(memory.TotalMemory())
+
+		if ggml.FileType() == "F16" && requiredMemory*f16Multiplier > systemMemory {
+			return nil, fmt.Errorf("F16 model requires at least %s of total memory", format.HumanBytes(requiredMemory))
+		} else if requiredMemory > systemMemory {
+			return nil, fmt.Errorf("model requires at least %s of total memory", format.HumanBytes(requiredMemory))
 		}
 	}
 
-	totalResidentMemory := memory.TotalMemory()
-	switch ggml.ModelType() {
-	case ModelType3B, ModelType7B:
-		if ggml.FileType().String() == "F16" && totalResidentMemory < 16*1024*1024 {
-			return nil, fmt.Errorf("F16 model requires at least 16GB of memory")
-		} else if totalResidentMemory < 8*1024*1024 {
-			return nil, fmt.Errorf("model requires at least 8GB of memory")
-		}
-	case ModelType13B:
-		if ggml.FileType().String() == "F16" && totalResidentMemory < 32*1024*1024 {
-			return nil, fmt.Errorf("F16 model requires at least 32GB of memory")
-		} else if totalResidentMemory < 16*1024*1024 {
-			return nil, fmt.Errorf("model requires at least 16GB of memory")
-		}
-	case ModelType30B, ModelType34B:
-		if ggml.FileType().String() == "F16" && totalResidentMemory < 64*1024*1024 {
-			return nil, fmt.Errorf("F16 model requires at least 64GB of memory")
-		} else if totalResidentMemory < 32*1024*1024 {
-			return nil, fmt.Errorf("model requires at least 32GB of memory")
-		}
-	case ModelType65B:
-		if ggml.FileType().String() == "F16" && totalResidentMemory < 128*1024*1024 {
-			return nil, fmt.Errorf("F16 model requires at least 128GB of memory")
-		} else if totalResidentMemory < 64*1024*1024 {
-			return nil, fmt.Errorf("model requires at least 64GB of memory")
-		}
-	}
-
-	switch ggml.ModelFamily() {
-	case ModelFamilyLlama:
-		return newLlama(model, adapters, ggmlRunner(), opts)
+	switch ggml.Name() {
+	case "gguf":
+		opts.NumGQA = 0 // TODO: remove this when llama.cpp runners differ enough to need separate newLlama functions
+		return newLlama(model, adapters, chooseRunners(workDir, "gguf"), ggml.NumLayers(), opts)
+	case "ggml", "ggmf", "ggjt", "ggla":
+		return newLlama(model, adapters, chooseRunners(workDir, "ggml"), ggml.NumLayers(), opts)
 	default:
 		return nil, fmt.Errorf("unknown ggml type: %s", ggml.ModelFamily())
 	}